Leveraging Death - Journal of Criminal Law and Criminology - Volume 103 | Issue 2 Article 3 - Spring 2013 - Sherod Thaxton

Follow this and additional works at: https://scholarlycommons.law.northwestern.edu/jclc

Part of the Criminal Law Commons

This Criminal Law is brought to you for free and open access by Northwestern University School of Law Scholarly Commons. It has been accepted for inclusion in Journal of Criminal Law and Criminology by an authorized editor of Northwestern University School of Law Scholarly Commons.

Recommended Citation

Sherod Thaxton, Leveraging Death, 103 J. Crim. L. & Criminology 475 (2013).




Copy right © 2013 by Northwestern University School of Law Printed in U.S.A.



Empirical research addressing the use of the death penalty as leverage

in plea negotiations is virtually nonexistent. This is particularly surprising

given the fact that both plea bargaining and capital punishment have been

the focus of much scholarly attention. The U.S. Supreme Court has

explicitly approved guilty pleas induced out of fear of the death penalty, yet

the impact of the threat of the death penalty on the likelihood of parties

reaching a plea agreement is far from obvious. On the one hand,

prosecutors, defense attorneys, and defendants may have especially strong

incentives to plea bargain in death-eligible cases. On the other hand, many

of these advantages may be offset by forces pushing against compromise on

both sides of the aisle precisely because the death penalty is an option, so

the role the death penalty is playing in plea negotiations in the aggregate

remains ambiguous. To date, the only empirical study to explore this issue

concluded that the threat of capital punishment does not impact the

likelihood of reaching a plea agreement. Unfortunately the study suffers

from several limitations that may have ultimately masked any true effect

that the death penalty has on plea-bargaining rates. This Article

reexamines this question using an originally constructed data set of recent

capital charging-and-sentencing decisions in Georgia (1993–2000) that is

able to avoid many of the shortcomings of the sparse prior research. The

results provide strong evidence that the threat of the death penalty has a

robust causal effect on the likelihood of a plea agreement—the threat of the

death penalty increases the probability of a plea agreement by

approximately 20 to 25 percentage points across various model

specifications. Not only is this finding important in its own right by

illuminating capital defendants’ behavioral response to the death penalty, it

also has meaningful implications for other purported benefits of plea

bargaining in the capital context. The paper briefly considers one of the

most commonly identified benefits of plea bargaining—cost reduction—and

concludes that the death penalty fails to deter sufficient numbers of murder

defendants from opting for trial to offset the significant expense of a capital

case and subsequent appeals.

* Assistant Professor of Law, UCLA School of Law. Former Staff Attorney, Capital

Habeas Unit, Office of the Federal Defender for the Eastern District of California. I would

like to thank Daniel Abebe, Anya Bernstein, Alexander Boni-Saenz, Bernard Harcourt,

William Hubbard, Aziz Huq, Richard McAdams, Tracey Meares, Thomas Miles, Jennifer

Nou, Andres Sawicki, Julia Simon-Kerr, and participants of The University of Chicago Law

School Faculty Workshop for useful comments on drafts of this Article. Naturally, all

remaining errors are my own.


INTRODUCTION................................................................................... 477

I. PLEA BARGAINING IN THE SHADOW OF DEATH .................................. 484

A. Bargaining Incentives ................................................................. 484

B. Bargaining Disincentives ............................................................ 487

II. PRIOR RESEARCH ........................................................................... 492

III. GEORGIAS DEATH PENALTY ......................................................... 498

A. History and Description of the Modern Statute ............................. 498

B. Life Without the Possibility of Parole .......................................... 502

C. Capital Case Progression ............................................................ 503

IV. DATA ........................................................................................... 505

V. EMPIRICAL STRATEGY .................................................................... 507

A. Designation of Treatment and Control Groups ............................. 507

B. Statistical Model ........................................................................ 513

C. Missing Data.............................................................................. 519

VI. RESULTS....................................................................................... 521

A. Fixed-Effects logit specifications ................................................ 521

B. Sensitivity Analyses ................................................................... 524

VII. FINANCIAL IMPLICATIONS AND DISCUSSION .................................. 540

VIII. CONCLUSION .............................................................................. 549


Plea bargaining1 is a crucial feature of our criminal justice system, as

approximately 95% of convictions that occur within a year of arrest are

obtained by a guilty plea.2 Despite its current centrality, however, “plea

bargaining did not occur with any frequency until well into the nineteenth

century,”3 and the Supreme Court did not specifically address its

constitutionality until long after it was common practice in the criminal

justice system.4 In Brady v. United States,5 the Court reasoned that plea

bargaining benefited both sides of the adversarial system and was “inherent

in the criminal law and its administration.”6 The following year, in

Santobello v. New York,7 the Court defended the practice of plea

bargaining, calling it “an essential component of the administration of

justice”8 that was to be encouraged as long as it was properly administered

(i.e., as long as pleas were intelligent and voluntary). Several years later in

Bordenkircher v. Hayes,9 the Court endorsed prosecutorial threats of stiffer

penalties when defendants refuse to accept a plea offer.10 In fact, since its

1 Plea bargaining is “[a] negotiated agreement between a prosecutor and a criminal

defendant whereby the defendant pleads guilty to a lesser offense or to one of multiple

charges in exchange for some concession by the prosecutor, usu. a more lenient sentence or a

dismissal of the other charges.” BLACK’S LAW DICTIONARY 1270 (9th ed. 2009). Plea

bargaining may involve three areas of negotiation: charge bargaining (negotiation to plead

guilty or no contest in exchange for the dismissal of some counts or reduction of the charge);

sentence bargaining (negotiation to plead guilty or no contest in exchange for the

prosecutor’s recommendation to the court for a lighter sentence); and fact bargaining

(negotiation to stipulate to certain facts in exchange for an agreement not to introduce other

facts into evidence). Id.



also Padilla v. Kentucky, 130 S. Ct. 1473, 1485 n.13 (2010) (recognizing that only 5% of

federal and state felony prosecutions are resolved by trial).

3 Albert W. Alschuler, Plea Bargaining and Its History, 79 COLUM. L. REV. 1, 5 (1979);

see also id. at 19 (“It was only after the Civil War that cases of plea bargaining began to

appear in American appellate court reports.”).

4 See John H. Langbein, Torture and Plea Bargaining, 46 U. CHI. L. REV. 3, 9 (1978) (“If

you turn to the American Constitution in search of authority for plea bargaining, you will

look in vain. Instead, you will find—in no less hallowed a place than the Bill of Rights—an

opposite guarantee, a guarantee of trial.”).

5 397 U.S. 742 (1970).

6 Id. at 751.

7 404 U.S. 257 (1971).

8 Id. at 260.

9 434 U.S. 357 (1978).

10 Id. at 363. Eight years before Bordenkircher, the Court ruled that the mere fact that a

defendant pleads guilty to murder to avoid the death penalty does not make the plea

involuntary, especially when the defendant is represented by competent counsel who can

assess the weight of the evidence against the defendant. North Carolina v. Alford, 400 U.S.

formal endorsement of plea bargaining, the Court has been reluctant to

regulate plea bargains, recently noting that “[h]indsight and second guesses

are inappropriate . . . where a plea has been entered without a full trial . . . .”11

Most judges support the system of plea bargaining because it allows

them to alleviate the need to schedule and hold a trial on what are typically

already overcrowded dockets.12 Prosecutors desire both the reduced

caseload and assurance of a conviction from plea bargaining. Plea

negotiations also allow prosecutors to strengthen their cases against

codefendants by offering certain defendants a plea arrangement in exchange

for testimony against one or more codefendants. This practice assures

prosecutors at least one conviction while also enhancing the chances of a

25, 31 (1970).

11 Premo v. Moore, 131 S. Ct. 733, 745 (2011). Notable exceptions include the Court’s

recent decisions in Missouri v. Frye, 132 S. Ct. 1399 (2012), Lafler v. Cooper, 132 S. Ct.

1376 (2012), and Padilla v. Kentucky, 130 S. Ct. 1473 (2010). In Padilla, the Court held

that defense counsel’s incorrect advice to a client regarding the deportation risk associated

with pleading guilty may violate the Sixth Amendment guarantee to effective assistance of

counsel if it prejudices the client’s decision. 130 S. Ct. at 1478. The Court also rejected the

argument that its ruling would ultimately open the “floodgates” of litigation by providing

defendants with new causes of action. Id. at 1484–85. But see Derek Wikstrom, Note, “No

Logical Stopping-Point”: The Consequences of Padilla v. Kentucky’s Inevitable Expansion,

106 NW. U. L. REV. 351, 374 (2012) (“Even if Padilla has an eventual logical stoppingpoint,

that point comes after extension to many other collateral consequences that are as

serious as deportation.”).

In early 2012, the Court decided Cooper and Frye, extending Padilla’s logic to

nondeportation cases and holding that a defendant’s Sixth Amendment right to effective

assistance of counsel may be violated when (1) defense counsel fails to properly inform the

defendant of a beneficial plea agreement offered by the prosecution (Frye), or (2) defense

counsel incorrectly advises the defendant on the state of the law, leading the defendant to

reject a beneficial plea agreement (Cooper). Frye, 132 S. Ct. 1399; Cooper, 132 S. Ct. 1376.

The Court recognized that defense counsel’s duty to inform a client of formal plea offers

may be subject to exceptions. Frye, 132. S. Ct. at 1408. It also pointed out that claims

raised under “ineffective advice” must satisfy three conditions: (1) the ineffective advice was

the “but for” cause of the plea not being accepted by the defendant, (2) the trial court would

have ultimately accepted the terms of the bargain, and (3) the conviction or sentence would

have been lesser than those imposed after trial. Cooper, 132 S. Ct. at 1385; see also Justin F.

Marceau, Embracing a New Era of Ineffective Assistance of Counsel, 14 U. PA. J. CONST. L.

1161, 1191–92 (2012) (analyzing the Court’s recent plea-bargaining decisions). The impact

of these decisions on plea bargaining is debatable, however. As one federal trial judge has

explained, most of the problems that occur in the plea-bargaining process do not result from

ineffective assistance of counsel, but rather prosecutorial overconfidence in the face of

questionable evidence and sources. Jed S. Rakoff, Frye and Lafler: Bearers of Mixed

Messages, 122 YALE L.J. ONLINE 25, 26 (2012).

12 Blackledge v. Allison, 431 U.S. 63, 71 (1977) (noting that plea bargaining permits

“[j]udges and prosecutors [to] conserve vital and scarce resources”). But see Jo Dixon, The

Organizational Context of Criminal Sentencing, 100 AM. J. SOC. 1157, 1177 (1995)

(suggesting that the level of plea bargaining is high irrespective of caseload pressure).

subsequent conviction. Defendants are allowed to avoid a more serious

charge or sentence and, if represented by private counsel, avoid the cost of a

trial.13 So, on balance, the practice of plea bargaining is generally believed

to be superior to trials due to reduced costs, improvements in the speed and

efficiency of case processing, and increases in the certainty of

convictions.14 The practice is not without its detractors, however, as legal

academics and practitioners continue to debate its fairness and desirability.

Critics of the plea-bargaining system emphasize that it encourages

prosecutors to “overcharge” at the start of the case in an effort to coerce

defendants into accepting a plea, allows prosecutors to “cure” defects in

their cases by avoiding trial, and encourages defendants to plead guilty to

crimes that they did not commit.15 A defendant who agrees to a plea

bargain may also be required to relinquish certain constitutional and

statutory rights in exchange for a negotiated plea.16 Opponents of plea

bargaining also suggest that the practice allows defendants to avoid the

13 See, e.g., Frank H. Easterbrook, Criminal Procedure as a Market System, 12 J. LEGAL

STUD. 289, 297–98 (1983) (defending plea bargains as an element of a well-functioning

market system and cautioning that efforts to restrict actors’ discretion in criminal procedure

will have negative consequences); Frank H. Easterbrook, Plea Bargaining as Compromise,

101 YALE L.J. 1969 (1992); William M. Landes, An Economic Analysis of the Courts, 14

J.L. & ECON. 61, 66–69 (1971) (describing factors influencing the likelihood of parties

reaching plea agreements); Robert E. Scott & William J. Stuntz, Plea Bargaining as

Contract, 101 YALE L.J. 1909, 1918–34, 1967–68 (1992) (explaining the risks and

entitlements parties trade in plea bargaining).


BARGAINING IN AMERICA 200–01 (2003) (discussing the growth of plea bargaining in

response to the operational goals of judges and prosecutors); Donald J. Newman, Pleading

Guilty for Considerations: A Study of Bargain Justice, 46 J. CRIM. L. CRIMINOLOGY &

POLICE SCI. 780, 790 (1956) (describing plea bargaining “as a natural, expedient outgrowth

of deficiencies in the administration of our ‘trial-by-combat’ theory of justice”); Edward A.

Ruttenburg, Plea Bargaining Analytically—The Nash Solution to the Landes Model, 7 AM. J.

CRIM. L. 323, 353 (1979) (“Plea bargaining should be accepted openly as a system which

can accomplish the goals of justice as completely as can a pure trial system, while at the

same time releasing resources to serve society in other areas of life.”).

15 See Langbein, supra note 4, at 3 (discussing the parallels between the coerciveness of

current plea-bargaining practices and torture in medieval European courts); see also Oren

Bar-Gill & Omri Ben-Shahar, The Prisoners’ (Plea Bargain) Dilemma, 1 J. LEGAL ANALYSIS

737, 769 (2009) (positing that plea bargaining may benefit individual defendants, but due to

coordination problems, may not benefit defendants as a group); Oren Gazal-Ayal, Partial

Ban on Plea Bargains, 27 CARDOZO L. REV. 2295, 2306 (2006) (arguing that prosecutors use

negotiated pleas in cases with weak—possibly inadmissible—evidence and when specific

defenses might be established in court); Stephen J. Schulhofer, Plea Bargaining as Disaster,

101 YALE L.J. 1979, 1980–91 (1992) (describing structural flaws—e.g., innocence and

conflicts of interest—in the plea-bargaining system).

16 Stephanos Bibas, Regulating the Plea-Bargaining Market: From Caveat Emptor to

Consumer Protection, 99 CALIF. L. REV. 1117, 1123 (2011) (discussing the rights defendants

generally forfeit that could otherwise provide a cause of action for an appeal).

appropriate punishment for their crimes (as established by state legislative

bodies) and that the practice heavily favors defendants with savvy lawyers,

irrespective of the defendants’ actual culpability.17 Also, claims of plea

bargaining as a “necessary” and “inevitable” component of our criminal

justice system have been challenged, as critics of plea bargaining point to

jurisdictions that have experimented with partial or complete bans on the practice.18

The debate over the promises and pitfalls of plea bargaining is perhaps

most contentious in the context of the death penalty,19 yet scholars have



(explaining that many prosecutors disfavor litigating pretrial motions, and defense attorneys

who engage in intensive pretrial motion work are likely to obtain favorable plea bargains for

their clients); Albert W. Alschuler, Plea Bargaining and the Death Penalty, 58 DEPAUL L.

REV. 671, 674 (2009) (suggesting that plea bargaining for a reduced punishment undermines

both the symbolic and instrumental purposes of punishment); Stephanos Bibas, Plea

Bargaining Outside the Shadow of Trial, 117 HARV. L. REV. 2463, 2493 n.116 (2004)

(arguing that plea bargaining tends to favor repeat offenders).

18 See Michael L. Rubinstein & Teresa J. White, Alaska’s Ban on Plea Bargaining, 13

LAW & SOC’Y REV. 367, 367 (1979) (describing Alaska’s attempt to ban plea bargaining for

the vast majority of cases); Robert A. Weninger, The Abolition of Plea Bargaining: A Case

Study of El Paso County, Texas, 35 UCLA L. REV. 265, 311–13 (1987) (explaining that

efforts to abolish plea bargaining have achieved mixed results). But several scholars have

noted that the incentives to plea bargain are so strong for all the primary actors in the

criminal justice system that they simply find alternative ways to achieve the same end even

when the practice has been expressly prohibited. See, e.g., Joseph L. Hoffmann et al., Plea

Bargaining in the Shadow of Death, 69 FORDHAM L. REV. 2313, 2390 (2001).

19 See infra Part I. Compare, e.g., ROBERT M. BOHM, DEATHQUEST II: AN INTRODUCTION


(2003) (commenting that the availability and use of plea bargaining can contribute to

arbitrariness and discrimination in the implementation of the death penalty), WHITE, supra

note 17, at 54 (conducting interviews with capital defense attorneys and concluding that “the

likelihood of a plea bargain in a capital case will be dramatically affected by factors that

have nothing to do with the nature of the crime or the strength of the evidence against the

defendant”), Alschuler, supra note 17, at 674 (suggesting that plea bargaining “undermines

the [death penalty’s] most common rationale . . . some crimes are so horrible that they

simply require it”), and Daniel Givelber, The New Law of Murder, 69 IND. L.J. 375, 410–11

(1994) (explaining that the existence of statutorily defined special circumstances for death

eligibility strengthens the prosecutor’s bargaining position and facilitates arbitrariness), with


LITIGATING IN THE SHADOW OF DEATH] (explaining that plea bargaining is an important tool

for defense counsel to save their clients’ lives, but also recognizing how the practice leads to

increased arbitrariness and potentially wrongful convictions), and Russell Stetler,

Commentary on Counsel’s Duty to Seek and Negotiate a Disposition in Capital Cases (ABA

Guideline 10.9.1), 31 HOFSTRA L. REV. 1157, 1157–58 (2003) (noting that, in 2003, the

American Bar Association Guidelines recognized that attorneys representing a capital

defendant have an obligation to seek negotiated pleas, and suggesting that the current

number of death row inmates could have been drastically reduced if their defense attorneys

conducted very little research on the relationship between the death penalty

and plea negotiations.20 Instead, they have focused the bulk of their

attention on the examination of the possible deterrent effect of the death

penalty on potential murderers.21 This Article offers an empirical

examination of the causal impact of the threat of the death penalty on the

likelihood of parties reaching a plea agreement. This type of inquiry is

particularly relevant because the effect of the death penalty on plea

bargaining is theoretically ambiguous. The threat of the death penalty may

induce defendants who may not have otherwise accepted a plea agreement

to plead to avoid the risk of possible execution, so the overall number of

cases proceeding to trial is reduced. An opposite effect is plausible as well:

armed with the threat of the “ultimate penalty,” prosecutors may be less

willing to offer capital defendants desirable plea bargains, if any bargain at

all, so the number of cases going to trial may increase.22 It is also possible

worked more diligently to obtain a negotiated plea).

20 See Ilyana Kuziemko, Does the Threat of the Death Penalty Affect Plea Bargaining in

Murder Cases? Evidence from New York’s 1995 Reinstatement of Capital Punishment, 8

AM. L. & ECON. REV. 116, 141 (2006) (conducting the first study of the effect of the death

penalty on plea bargaining); see also Susan Ehrhard, Plea Bargaining and the Death

Penalty: An Exploratory Study, 29 JUST. SYS. J. 313, 315 (2008) (noting that there has been

only one systematic examination of the impact of the death penalty on plea bargaining); Kent

S. Scheidegger, The Death Penalty and Plea Bargaining to Life Sentences 1–2 (Criminal

Justice Legal Found., Working Paper No. 09-01, 2009) (same); cf. ANDREW WELSHHUGGINS,


COUNTRY’S BUSIEST DEATH PENALTY STATES 96–100 (2009) (discussing the high prevalence

of plea bargaining in capital cases in California, New York, Ohio, and at the federal level).

21 To be sure, social scientists—particularly those working in fields outside of

economics—have conducted empirical research on various aspects of the death penalty for

more than 75 years, and over the past 30 years, socio-legal research has figured prominently

into the debate over the appropriateness and effectiveness of capital punishment in the

United States. See David C. Baldus, Keynote Address: The Death Penalty Dialogue Between

Law and Social Science, 70 IND. L.J. 1033, 1041 (1995); Michael L. Radelet & Marian J.

Borg, The Changing Nature of Death Penalty Debates, 26 ANN. REV. SOC. 43, 43–44 (2000).

But no other issue related to the death penalty has received more systematic attention from

legal and academic communities. William C. Bailey & Ruth D. Peterson, Murder, Capital

Punishment, and Deterrence: A Review of the Literature, in THE DEATH PENALTY IN

AMERICA: CURRENT CONTROVERSIES 135, 135 (Hugo Adam Bedau ed., 1997).

22 See Kuziemko, supra note 20, at 117 (asserting that the death penalty may make

prosecutors more aggressive or recalcitrant, so they may be less willing to strike deals with

defendants); see also WHITE, LITIGATING IN THE SHADOW OF DEATH, supra note 19, at 157–

58 (presenting the problem that some defendants prefer execution rather than life

imprisonment); Ehrhard, supra note 20, at 318 (highlighting that many prosecutors only

make “take it or leave it” offers of life without the possibility of parole in death-eligible

cases); James S. Liebman, Opting for Real Death Penalty Reform, 63 OHIO ST. L.J. 315,

322–24 (2002) (showing that the availability of the death penalty can trump the usual

pressure for prosecutors and defendants to reasonably compromise).

Both chief prosecutors and rank-and-file prosecutors may anticipate huge career

that the death penalty will have no impact on the likelihood that a defendant

accepts a plea bargain and will only impact the terms of the bargain.23 As

two scholars have recently noted, “opposing hypotheses about the effect of

the death penalty on prosecutorial discretion have never been rigorously tested.”24

Examining the impact of capital punishment on plea bargaining is

important for several reasons. First, it helps inform our understanding of

how sentencing law influences plea bargaining. Given the centrality of

pleas for the disposition of criminal cases, studying how sentencing

structure impacts the incentives of prosecutors and defendants in plea

negotiations is important to our understanding of the criminal justice


Second, the use of the death penalty as leverage in plea negotiations

raises important legal and ethical issues. Defendants are typically required

to waive important constitutional rights as a condition of the plea

agreement.25 Many of these rights are considered crucial to the accurate

determination of guilt and punishment at trial, so the absence of these

protections may undermine our confidence in that determination. This may

be of particular concern in the capital context because although defendants

who plead guilty avoid the risk of execution, they still receive very lengthy

payoffs from pursuing a capital case even when the defendant ultimately receives a sentence

less than death or the capital sentence is reversed on appeal. See, e.g., WHITE, DEATH

PENALTY, supra note 17, at 55 (arguing that capital defense lawyers routinely remark that

upcoming prosecutorial elections are the most important factor in predicting the likelihood of

a plea bargain in a capital case, and savvy defense attorneys attempt to delay trial until after

an election in order to increase the likelihood that a prosecutor will offer a plea); Jonathan E.

Gradess & Andrew L. B. Davies, The Cost of the Death Penalty in America: Directions for


GENERATION OF CAPITAL PUNISHMENT RESEARCH 397, 409 (Charles S. Lanier et al. eds.,

2009) (“Prosecutors are often hawkish about the death penalty during election campaigns,

increasing the probability that they will press for it in office and use it when available.”);

Liebman, supra note 22, at 324–25 (arguing that trial error is virtually costless to

prosecutors). But see Ronald F. Wright, How Prosecutor Elections Fail Us, 6 OHIO ST. J.

CRIM. L. 581, 606 (2009) (suggesting that campaigning prosecutors invoke the death penalty

infrequently). Alaska, Connecticut, Delaware, New Jersey, and Rhode Island are the only

states that do not popularly elect their district attorneys. John A. Horowitz, Note,

Prosecutorial Discretion and the Death Penalty: Creating a Committee to Decide Whether

to Seek the Death Penalty, 65 FORDHAM L. REV. 2571, 2575 n.24 (1997).

23 Kuziemko, supra note 20, at 140. See generally George L. Priest & Benjamin Klein,

The Selection of Disputes for Litigation, 13 J. LEGAL STUD. 1 (1984) (explaining why parties

fail to reach settlement agreements when such agreements are apparently mutually


24 Gradess & Davies, supra note 22, at 409.

25 Hoffmann et al., supra note 18, at 2317–30 (discussing the U.S. Supreme Court’s

heightened concern over statutory schemes that “needlessly burden” the exercise of

constitutional rights).


sentences—typically life imprisonment. The threat of the death penalty has

also induced innocent defendants to plead guilty (and even falsely implicate

others) to avoid execution.26

Finally, the study of the plea-negotiation process in the capital context

permits us to gain a better understanding of the financial and administrative

costs of capital punishment to states and the federal government. Capital

trials are extremely expensive and they rarely reduce prison costs because

of the infrequency of executions and the added expense associated with

housing inmates on death row.27 The threat of capital punishment may

result in substantial savings, however, if the threat of execution deters

sufficient numbers of individuals from pursuing trial.

Part I of the Article discusses many of the unique aspects of plea

bargaining in the capital context and how these factors cut both in favor and

against successful plea negotiations. Part II describes the limited empirical

research on the impact of the threat of the death penalty on plea bargaining

and how certain weaknesses of that particular work undermine our ability to

draw firm conclusions about the relationship between capital punishment

and plea bargaining. This section also explains how the limitations of this

prior research are avoided in the study conducted in this Article using an

originally constructed data set on capital charging-and-sentencing decisions

in Georgia. Part III briefly outlines Georgia’s modern capital punishment

and life without the possibility of parole (LWOP) statutes, Part IV describes

the data collected on Georgia’s capital charging-and-sentencing system that

are used to analyze the impact of the death penalty on plea negotiations, and

Part V discusses the empirical strategy employed to analyze the data. Part

VI reports the results from various model specifications. According to

these findings, my conservative estimate is that the threat of the death

penalty increases the likelihood of reaching a plea agreement by

approximately 20 percentage points. In practical terms, the death penalty

increases the plea-bargaining rate from approximately 40% to 60%. In

other words, the threat of capital punishment deters roughly two out of

every ten death-noticed defendants from pursuing a trial. Part VII discusses

the implications of the findings for the administrative and financial costs of

26 See, e.g., Samuel R. Gross et al., Exonerations in the United States 1989 Through

2003, 95 J. CRIM. L. & CRIMINOLOGY 523, 544–46 (2005) (describing cases in which

innocent defendants pleaded guilty to murder, and even falsely implicated others, in order to

avoid the death penalty); Paul Hammel, ‘Beatrice 6’ Cleared; ‘100 Percent Innocent,

OMAHA WORLD-HERALD, Jan. 27, 2009, at B1 (discussing five exonerated convicted

murderers who falsely pleaded guilty after being threatened with the death penalty).

27 John K. Roman et al., Reassessing the Cost of the Death Penalty Using Quasi-

Experimental Methods: Evidence from Maryland, 11 AM. L. & ECON. REV. 530, 551–53, 571

(2009); see also infra Part VII (detailing financial and administrative costs of capital cases).

484 SHEROD THAXTON [Vol. 103

the death penalty. Based on the high costs associated with litigating a

single capital trial and the rather modest ability of the death penalty to deter

defendants from pursuing trial, capital punishment does not appear to be a

cost-justified bargaining chip.



In addition to the perceived increase in plea-bargaining leverage

resulting from the severity of the punishment, prosecutors typically enjoy

huge advantages by merely seeking the death sentence. First, it enables the

government to empanel a “death-qualified” jury. The primary purpose of

“death qualification” during voir dire is to remove jurors who

unequivocally oppose the death penalty or, conversely, who believe that the

death sentence is required in every homicide case.28 Although the Supreme

Court has upheld the constitutionality of death-qualified juries in the face of

the claim that these juries placed defendants at an unfair risk of

conviction,29 available evidence strongly suggests that death-qualified juries

are much more likely to convict than non-death-qualified juries.30 The

death-qualification process also permits prosecutors to identify jurors with

mildly skeptical views of the death penalty (or the criminal justice system

in general) who may survive exclusion for cause, and subsequently exercise

their peremptory challenges to remove them from the jury.31

28 See, e.g., Wainwright v. Witt, 469 U.S. 412, 424 (1985) (explaining that the standard

for determining when the Court may exclude a prospective juror because of his views on

capital punishment is whether “the juror’s views would ‘prevent or substantially impair the

performance of his duties as a juror in accordance with his instructions and his oath’”

(footnote omitted)).

29 Witherspoon v. Illinois, 391 U.S. 510, 517–18 (1968) (“We simply cannot

conclude . . . that the exclusion of jurors opposed to capital punishment results in an

unrepresentative jury on the issue of guilt or substantially increases the risk of conviction.”).


PSYCHOLOGICAL SYSTEM 118–21 (2005) (describing the conviction proneness of deathqualified

juries); William J. Bowers, The Capital Jury: Is It Tilted Toward Death?, 79

JUDICATURE 220, 222–23 (1996) (same); Robert Fitzgerald & Phoebe C. Ellsworth, Due

Process vs. Crime Control: Death Qualification and Jury Attitudes, 8 LAW & HUM. BEHAV.

31, 48 (1984) (same); Walter E. Oberer, Does Disqualification of Jurors for Scruples

Against Capital Punishment Constitute Denial of Fair Trial on Issue of Guilt?, 39 TEX. L.

REV. 545 (1961) (same); Robert L. Young, Guilty Until Proven Innocent: Conviction

Orientation, Racial Attitudes, and Support for Capital Punishment, 25 DEVIANT BEHAV. 151,

155 (2004) (same).

31 James S. Liebman, The Overproduction of Death, 100 COLUM. L. REV. 2030, 2097

(2000) (explaining that voir dire in capital cases allows prosecutors to “jettison[] the segment

of the jury pool that is most likely to be skeptical of informer, police, and forensic testimony

and to take seriously the beyond a reasonable doubt standard”); F. Thomas Schornhorst,


Second, seeking the death penalty substantially increases the defense’s

burden with very little increase, if any, in the government’s burden by

vastly expanding the defense attorney’s role and the requisite skill set and

financial resources.32 The overwhelming majority of criminal defense

attorneys, both in private practice and employed by the government, are

routine negotiators of deals that permit their clients to avoid trial.

Representation of a capitally charged client now possibly requires defense

counsel to become serious investigators of a horrific crime (or at least

coordinators of complex investigations), competent consumers (and

communicators) of mental health and forensic science, and experts on

complicated (and constantly evolving) specialized constitutional and

statutory law.33 Prosecutors are well aware that many highly skilled trial

lawyers will simply refuse to represent a capital defendant, and evidence

suggests that capital defendants represented by court-appointed counsel are

more likely to receive the death sentence and more likely to have their

appeals denied.34 According to one scholar:

Preliminary Screening of Prosecutorial Access to Death Qualified Juries: A Missing

Constitutional Link, 62 IND. L.J. 295, 325 (1987) (arguing for prescreening of capital cases

for probable cause of the existence of “death-eligible” special circumstances prior to death

qualification of a jury); Young, supra note 30, at 151 (explaining that death-qualified jurors

are more likely to prefer convicting the innocent over acquitting the guilty).

32 Liebman, supra note 31, at 2097–98; Liebman, supra note 22, at 322; see also Stephen

B. Bright, Counsel for the Poor: The Death Sentence Not for the Worst Crime but for the

Worst Lawyer, 103 YALE L.J. 1835, 1844 (1994) (discussing interrelated reasons for the poor

quality of representation in capital cases).

33 Liebman, supra note 22, at 322–23; see also Bibas, supra note 16, at 1141 (“Good

defense lawyers must know, for example, whether a defendant’s small children, ill health,

apology, cooperation, or restitution can lower his sentence.”); Richard G. Dudley, Jr. &

Pamela Blume Leonard, Getting It Right: Life History Investigation as the Foundation for a

Reliable Mental Health Assessment, 36 HOFSTRA L. REV. 963, 988 (2008) (emphasizing the

importance and accompanying complexity of thorough mental health investigation for

mitigation in capital cases). See generally Stephen B. Bright, Death by Lottery—Procedural

Bar of Constitutional Claims in Capital Cases Due to Inadequate Representation of Indigent

Defendants, 92 W. VA. L. REV. 679 (1990); Stephen B. Bright, In Defense of Life: Enforcing

the Bill of Rights on Behalf of Poor, Minority and Disadvantaged Persons Facing the Death

Penalty, 57 MO. L. REV. 849, 857–62 (1992) (providing examples of gross ineffective

assistance of defense counsel in capital cases).

34 See James C. Beck & Robert Shumsky, A Comparison of Retained and Appointed

Counsel in Cases of Capital Murder, 21 LAW & HUM. BEHAV. 525, 538 (1997); see also


CAPITAL PUNISHMENT 237 (2008) (reporting that court-appointed lawyers representing

capitally charged clients in the most active death penalty jurisdictions were significantly

more likely to have been professionally disciplined prior to the appointment); Jules Epstein,

Death-Worthiness and Prosecutorial Discretion in Capital Case Charging, 19 TEMP. POL. &

CIV. RTS. L. REV. 389, 400–01 (2010) (suggesting a direct consequence of overly broad

capital statutes, and the resulting sky-rocketing capital docket, is the difficulty of finding

competent counsel for representation); Liebman, supra note 22, at 322 (noting that, when

486 SHEROD THAXTON [Vol. 103

“There are many small communities that do not have surgeons. But that does not

mean that we allow chiropractors to do brain surgery in those communities.” We do,

however, let “chiropractors” with law degrees perform the equivalent of brain surgery

in capital cases and, predictably, the “patient” often dies.35

Third, prosecutors generally understand that defense attorneys will

adopt minimalist, risk-averse pretrial and litigation strategies and practices,

focusing primarily on penalty-phase investigation and preparation and on

preserving “credibility” at the penalty phase. Consequently, the risk of an

acquittal even in cases with genuine evidentiary problems regarding guilt is

particularly low.36 In fact, a national study of jurors who served on capital

trials revealed that jurors were more likely to vote for the death sentence

when defense counsel’s guilt-phase and penalty-phase presentations were

logically inconsistent.37 The Supreme Court has also recognized that

“[a]ttorneys representing capital defendants face daunting challenges in

developing trial strategies [for guilt and penalty phases]”38 and “must strive

at the guilt phase to avoid a counterproductive course.”39 Finally,

prosecutors recognize that judges at both the trial and appellate levels—

especially elected judges—are less likely to require capital trials to strictly

follow reliable procedure out of fear that prosecutors will publicly blame

judges for losses based on “legal technicalities.”40

taking into consideration the vast responsibilities associated with defending a capital

defendant, the number of qualified attorneys in any jurisdiction dwindles to a handful or

completely disappears); Ronald J. Tabak & J. Mark Lane, The Execution of Injustice: A Cost

and Lack-of-Benefit Analysis of the Death Penalty, 23 LOY. L.A. L. REV. 59, 70 (1989)

(reporting that 90% of inmates on death row could not afford private counsel).

35 Vivian Berger, The Chiropractor as Brain Surgeon: Defense Lawyering in Capital

Cases, 18 N.Y.U. REV. L. & SOC. CHANGE 245, 254 (1990–1991) (citations omitted).

36 Ehrhard, supra note 20, at 318 (describing capital defense attorneys’ concerns about

compromised credibility at the penalty phase after aggressively pursuing an innocence


37 Scott E. Sundby, The Capital Jury and Absolution: The Intersection of Trial Strategy,

Remorse, and the Death Penalty, 83 CORNELL L. REV. 1557, 1589–91 (1998).

38 Florida v. Nixon, 543 U.S. 175, 191 (2004) (discussing the potential tensions between

guilt and penalty phase strategies).

39 Id.

40 See Stephen B. Bright & Patrick J. Keenan, Judges and the Politics of Death:

Deciding Between the Bill of Rights and the Next Election in Capital Cases, 75 B.U. L. REV.

759, 834–35 (1995); Richard R. W. Brooks & Steven Raphael, Life Terms or Death

Sentences: The Uneasy Relationship Between Judicial Elections and Capital Punishment, 92

J. CRIM. L. & CRIMINOLOGY 609, 638–39 (2002); Andrew Gelman et al., A Broken System:

The Persistent Patterns of Reversals of Death Sentences in the United States, 1 J. EMPIRICAL

LEGAL STUD. 209, 260 (2004); Liebman, supra note 31, at 2111–14; Ashley Rupp, Death

Penalty Prosecutorial Charging Decisions and County Budgetary Restrictions: Is the Death

Penalty Arbitrarily Applied Based on County Funding?, 71 FORDHAM L. REV. 2735, 2736–

37 (2003) (discussing the tremendous political and community pressure on judges in capital




The aforementioned advantages that prosecutors enjoy in seeking the

death penalty would appear to increase the likelihood of a plea bargain;

however, these advantages may be offset by forces pushing against

compromise on both sides of the aisle. Both systematic and anecdotal

evidence suggests that plea negotiations are most likely to occur when (1)

both parties perceive a similar expected probability of conviction, (2) the

expected trial penalty is relatively small (i.e., low-severity offenses), (3) the

settlement discount offered by the government is substantial, (4) the

defendant’s risk aversion is high, (5) pretrial detention and court delays

substantially increase opportunity costs for defendants, and (6) the parties

(particularly the defendant) will directly incur the full costs of litigation.41

But many of these features may be missing or substantially weakened in the

majority of capital cases.

The central point of contention in a capital trial is not the expected

probability of conviction—approximately 90% of all murder trials result in

conviction;42 rather, it is the likelihood of a death sentence (or a sentence

greater than the statutory minimum) at the penalty phase,43 and perhaps

even the expected likelihood of ultimately being executed.44 Juries wield

enormous discretion in deciding whether to impose the death sentence,

which is inherently a subjective enterprise.45 Rigorous examinations of jury

41 See, e.g., Gazal-Ayal, supra note 15, at 2299 (explaining that prosecutors obtain guilty

pleas when they can offer substantial concessions); Landes, supra note 13, at 99 (presenting

evidence that the likelihood of a plea agreement depends, inter alia, on the severity of the

crime, differences in the expected probability of conviction, resources available to the

parties, and parties’ risk aversion).

42 COHEN & KYCKELHAHN, supra note 2, at 11; THOMAS H. COHEN & BRIAN A. REAVES,





55 (2003) (describing the penalty phase of a capital trial as a “status competition” between

the defendant and the victim’s family); Jon Kyl et al., On the Wings of Their Angels: The

Scott Campbell, Stephanie Roper, Wendy Preston, Louarna Gillis, and Nila Lynn Crime

Victims’ Rights Act, 9 LEWIS & CLARK L. REV. 581, 607–08 (2005) (describing the

proliferation of victims’ rights legislation and victim-impact statements during the penalty

phase of capital trials).


USES, ABUSES, AND THE FUTURE OF THE GREAT WRIT 147 (2011) (explaining that capital

cases are thirty-five times more likely to be granted relief upon federal habeas review than

noncapital cases); James S. Liebman et al., Capital Attrition: Error Rates in Capital Cases,

1973–1995, 78 TEX. L. REV. 1839, 1850, 1852 (2000) (noting that 68% of death sentences

were overturned on appeal from 1973 to 1995 and 82% of those defendants ultimately

avoided the death penalty at resentencing).

45 Zant v. Stephens, 462 U.S. 862, 891 (1983) (holding that aggravating circumstances

488 SHEROD THAXTON [Vol. 103

sentencing patterns are unable to identify any meaningful (i.e., legally

legitimate) defendant or crime characteristics that consistently distinguish

cases that receive the death penalty from cases that do not.46 This strongly

suggests that prosecutors and defendants in capital cases may significantly

differ in their assessments of the expected trial sentence. As a result,

prosecutors will only have imperfect information about a defendant’s

reservation price (i.e., the maximum plea sentence that the defendant would

accept to avoid a trial), and might inadvertently make a plea offer exceeding

that reservation price.47 While prosecutors may generally know the

probability that a case would result in a plea bargain, they cannot accurately

predict the actual result in specific cases.48 Nationally, only one-third of

capital trials result in a death sentence,49 so a capital defendant may require

a larger discount than what a prosecutor is willing to offer. This is

are only required to narrow death eligibility and not to channel jurors’ discretion at the

penalty phase, and nonstatutory aggravating circumstances may be considered by the jury as

well); accord Lee v. State, 365 S.E.2d 99, 105 (Ga. 1988), cert. denied, 488 U.S. 879 (1988);

see also Nance v. State, 623 S.E.2d 470, 473 (Ga. 2005) (stating that there is no requirement

under the Georgia capital statute that nonstatutory aggravating circumstances be proven

beyond a reasonable doubt), cert. denied, 549 U.S. 868 (2006); Chelsea Creo Sharon, The

“Most Deserving” of Death: The Narrowing Requirement and the Proliferation of

Aggravating Factors in Capital Sentencing Statutes, 46 HARV. C.R.-C.L. L. REV. 223, 245

(2011) (describing the U.S. Supreme Court’s refusal to place greater restrictions on the

factors juries may consider when deciding whether to impose the death sentence).

46 See Richard Berk et al., Statistical Difficulties in Determining the Role of Race in

Capital Cases: A Re-Analysis of Data from the State of Maryland, 21 J. QUANTITATIVE

CRIMINOLOGY 365, 387 (2005) [hereinafter Berk et al., Statistical Difficulties] (“It is difficult

to imagine that a few covariates exist that if included as predictors would lead to clear and

justified distinctions between defendants who are charged with a capital crime and

defendants who are not; likewise for death sentences. . . . [I]f idiosyncrasies associated with

the case, the defendant, or the adjudication process seem to determine a substantial part of

the outcome, the adjudication process is suspect whether race is important or not.”); see also


151 (1987) (noting that legally legitimate characteristics in capital cases only explain a small

portion of the variation in charging-and-sentencing decisions); Richard A. Berk et al.,

Chance and the Death Penalty, 27 LAW & SOC’Y REV. 89, 107–09 (1993) [hereinafter Berk

et al., Chance and the Death Penalty] (same); William J. Bowers & Wanda D. Foglia, Still

Singularly Agonizing: Law’s Failure to Purge Arbitrariness from Capital Sentencing, 39

CRIM. L. BULL. 51, 84 (2003) (demonstrating that constitutionally mandated requirements to

guide jury discretion and eliminate arbitrariness in sentencing are not working); Deon Brock

et al., Arbitrariness in the Imposition of Death Sentences in Texas: An Analysis of Four

Counties by Offense Seriousness, Race of Victim, and Race of Offender, 28 AM. J. CRIM. L.

43, 70 (2000) (reporting evidence of significant within-jurisdiction arbitrariness in the

imposition of the death penalty after taking into account offense seriousness).

47 Bar-Gill & Ben-Shahar, supra note 15, at 757.

48 Bibas, supra note 17, at 2467; Gazal-Ayal, supra note 15, at 2321.




important because prosecutors are concerned with more than just avoiding

trials. They also care about fairness and reputation, and this may lead

prosecutors to refuse to adjust plea offers in a manner that would avoid

trial, even when prosecutors accurately assess the defendant’s reservation

price.50 Pressure from victims’ family and the community may also factor

into the minimum plea a prosecutor is willing to offer.51

Statutory minima in capital cases—typically life with or without the

possibility of parole—also preclude prosecutors from offering substantial

concessions (at least in the minds of many defendants).52 Prosecutors are

also less willing to drop or reduce homicide charges because such actions

are politically costly, so they often feel bound to dedicate resources to

trying homicide cases even with evidentiary problems.53 Even plea offers

of a life sentence with the possibility of parole may not be considered a

meaningful concession relative to life without the possibility of parole in the

current climate in which determinate-sentencing and truth-in-sentencing

laws typically require inmates sentenced to life imprisonment to serve at

least twenty-five years, and parole boards are increasingly reluctant to

release inmates serving life sentences.54 As one scholar has explained, plea


prosecutor who becomes known as a pushover will be taken advantage of, not once but many

times . . . once [a] threat [is] made, it ha[s] to be carried out.”); Josh Bowers, Punishing the

Innocent, 156 U. PA. L. REV. 1117, 1152–53 (2008).

51 See, e.g., Kyl et al., supra note 43, at 621 (advocating increased victim and victim

family involvement in plea negotiations, including the ability to reopen a plea or sentence

when the accused has pleaded to a reduced offense).

52 See, e.g., WHITE, LITIGATING IN THE SHADOW OF DEATH, supra note 19, at 158

(describing the “free me or fry me” attitude of many capital defendants); Robert L. Misner,

Recasting Prosecutorial Discretion, 86 J. CRIM. L. & CRIMINOLOGY 717, 742 (1996)

(explaining how the prevalence of statutory minima has dramatically increased the

importance of charge bargaining relative to sentence bargaining for serious offenses); see

also Lucian E. Dervan, The Surprising Lessons from Plea Bargaining in the Shadow of

Terror, 27 GA. ST. U. L. REV. 239, 245 (2011) (suggesting that the ability to offer substantial

sentence reductions is key to prosecutors’ success in the use of plea bargaining).

53 Gazal-Ayal, supra note 15, at 2306; Daniel C. Richman & William J. Stuntz, Al

Capone’s Revenge: An Essay on the Political Economy of Pretextual Prosecution, 105

COLUM. L. REV. 583, 600–05 (2005) (arguing that state prosecutors are more often politically

obliged to prosecute a defendant for the offense she is suspected of committing, and thus

cannot drop the case or offer a charge bargain).


PRISON SENTENCES IN CONTEXT 12 (2004) (discussing the steady increase in time actually

served for life with the possibility of parole sentences across the nation); Bibas, supra note

16, at 1141 (noting trend in truth-in-sentencing laws and parole board practices); Ehrhard,

supra note 20, at 316 (explaining that the distinction between life with parole and without

parole sentences is disappearing in current political climate because parole boards are

resistant to granting early release); Press Release, Ga. State Bd. Pardons & Paroles, More

Violent-Crime Lifers Die in Prison Than Are Parole [sic] (June 1, 1998), available at

490 SHEROD THAXTON [Vol. 103

bargaining draws “sustenance . . . from the principles of the indeterminate

sentence,”55 so statutory minima and de facto determinate sentences (via

extremely inactive parole boards) severely restrict the magnitude of a plea


Many capital defendants also suffer from substantial cognitive and

emotional deficits and biases, and consequently are more likely to be riskseeking

(or risk-neutral at best) and less likely to accept favorable plea

bargains.56 As a result, defense counsel in capital cases must be particularly

skillful with potential mental health issues affecting their clients.57 In fact,

a recent study of federal habeas corpus actions discovered that client mental

health issues continue to provide a strong basis for relief.58 Capital

defendants are also more likely to be highly skeptical of their defense

attorneys during plea negotiations, especially in situations when the

defendant is a racial or ethnic minority and defense counsel is not.59

Pretrial detention, court delays, and litigation costs are much less of a

concern for capital defendants as well. Defendants charged with murder are

the least likely to be released pending trial (19%), have the highest bail

amounts (i.e., represent the largest percentage of defendants with bail

exceeding $50,000), and are decreasingly (over the past twenty years) likely

to be granted pretrial release.60 In many jurisdictions, defendants charged

with murder are “non-bondable” either through statute or practice (e.g.,


news_0005.html (refuting popular misconception that “straight lifers” are released and

reporting that twenty-one “straight lifers” died in prison while serving their sentences in the

previous twelve months).

55 FISHER, supra note 14, at 127.

56 See, e.g., WHITE, LITIGATING IN THE SHADOW OF DEATH, supra note 19, at 162

(explaining that capital defendants with mental health problems pose a significant obstacle in

plea negotiations and often express a “free me or fry me” position to their defense counsel);

Alschuler, supra note 17, at 674 (explaining that defendants are executed “for the crime of

being [] optimist[s] . . . [the] inability to think 100 yards in front of [themselves]”); Bibas,

supra note 17, at 2467 (highlighting that the combination of poor lawyering and irrational

behavior can lead some defendants to reject bargains they should otherwise accept).

57 Liebman, supra note 22, at 322.

58 KING & HOFFMANN, supra note 44, at 151 (examining habeas corpus petitions after the

enactment of the Anti-Terrorism and Effective Death Penalty Act (AEDPA)).

59 See, e.g., WHITE, LITIGATING IN THE SHADOW OF DEATH, supra note 19, at 156–57

(describing how defendants’ mistrust of defense counsel in capital cases is often

compounded by racial or cultural differences); cf. Albert W. Alschuler, The Defense

Attorney’s Role in Plea Bargaining, 84 YALE L.J. 1179, 1180–82 (1975) (claiming that plea

bargaining is destructive to attorney–client relationships).






defendants lack the requisite financial resources to post bail),61 defense

counsel has a strategic incentive to delay trial absent any strong claims of

actual innocence,62 and the costs associated with defending a capital trial

make it virtually impossible for defendants to finance their own defense,63

so the state must almost invariably declare them indigent and appoint

counsel at the government’s expense.64 Local prosecutors also avoid the

burden of incurring the full cost of capital trials—particularly the

substantial costs of error correction at the appellate level.65

Finally, there are a variety of additional reasons why prosecutors and

defendants may not reach plea agreements in capital cases. The highly

publicized nature of capital cases, relative to noncapital cases, can increase

the reluctance of defendants to admit their participation in the crime in open

court.66 Complicating matters is the fact that many prosecutors and judges

(and even some capital statutes) do not permit defendants to enter Alford or

nolo contendere pleas in capital murder cases, but do allow such pleas in

noncapital murder cases.67 Prosecutors may also seek the death penalty

61 COHEN & REAVES, supra note 60, at 3–6.


CAPITAL APPEALS: A MULTISTATE STUDY 30 (2007) (remarking on the extraordinary amount

of time some defense attorneys take to file a notice of appeal in capital cases); Michael E.

Tigar, Judges, Lawyers and the Penalty of Death, 23 LOY. L.A. L. REV. 147, 148 (1989)

(discussing judges’ condemnation of defense counsel’s conduct in capital cases, including

delay tactics).

63 See infra Part VII for a discussion of the specific costs associated with capital trials.



available at http://www.americanbar.org/content/dam/aba/migrated/moratorium/assessment

project/georgia/report.authcheckdam.pdf (noting that appointment of counsel is required for

a defendant indicted for a capital felony if she can establish that she is indigent). To be

declared indigent, a person must “lack[] sufficient income or other resources to employ a

qualified lawyer to defend him or her without undue hardship on the individual or his or her

dependants [sic].” Id. at 143 n.73 (citations omitted) (internal quotation marks omitted); cf.

Alschuler, supra note 17, at 677 (discussing how defense counsel in capital cases often

threaten or attempt to make the trial as expensive as possible in order to improve their

bargaining positions).

65 Arthur L. Alarcón & Paula M. Mitchell, Executing the Will of the Voters?: A Roadmap

to Mend or End the California Legislature’s Multi-Billion-Dollar Death Penalty Debacle, 44

LOY. L.A. L. REV. S41, S88 (2011) (reporting the significant costs associated with federal

habeas review of state death sentences in California); Liebman, supra note 22, at 325

(explaining that state and federal courts incur the financial burden of correcting trial

mistakes); Misner, supra note 52, at 719–20 (arguing that our current system is seriously

flawed because prosecutors are not required to take into account finite criminal justice

resources when making charging-and-sentencing decisions).

66 WHITE, LITIGATING IN THE SHADOW OF DEATH, supra note 19, at 157.

67 An Alford plea is similar to a plea of nolo contendere where a defendant pleads guilty

to a crime but does not admit guilt. Alford v. North Carolina, 400 U.S. 25, 25–26, 37 (1970)

492 SHEROD THAXTON [Vol. 103

against a particular defendant in an attempt to extract information that may

have otherwise been impossible to obtain—e.g., the names of possible

accomplices and additional victims—by inducing the defendant’s

cooperation in exchange for a sentence reduction. If the defendant refuses

to cooperate, the prosecutor may be less likely to offer a favorable plea.68

Such cooperation may not be forthcoming in capital cases both because of

the high-profile nature of the cases and the severity of the potential

penalties for possible codefendants implicated by the defendants.

Due to the influence of these competing incentives on plea-bargaining

dynamics in the capital context, coupled with the lower likelihood of pleas

in murder cases in general, it is far from obvious what role capital

punishment is playing in plea negotiations in the aggregate. The empirical

analysis of the influence of the threat of the death penalty on parties’

propensity to reach a plea agreement assists in informing this debate.69


To date, only one study has attempted to systematically examine the

impact of the death penalty on plea bargaining.70 Professor Ilyana

(explaining that “[a]n individual accused of crime may voluntarily, knowingly, and

understandingly consent to the imposition of a prison sentence even if he is unwilling or

unable to admit his participation in the acts constituting the crime” and there is no “material

difference between a plea that refuses to admit commission of the criminal act and a plea

containing a protestation of innocence when . . . a defendant intelligently concludes that his

interests require entry of a guilty plea and the record before the judge contains strong

evidence of actual guilt”); see also Henderson v. Morgan, 426 U.S. 637, 644–45 (1976)

(explaining the constitutional prerequisites for Alford pleas). Georgia’s capital statutes

prohibit Alford pleas in capital cases. GA. CODE ANN. § 17-7-95(a) (West 2003).

68 See, e.g., David Garland, “Symbolic” and “Instrumental” Aspects of Capital

Punishment, in THE FUTURE OF AMERICA’S DEATH PENALTY, supra note 22, at 421, 437.

69 The focus of the empirical analyses in this Article is whether a plea agreement was

obtained, not the specific dynamics of the negotiations involved in obtaining a plea. In cases

that were ultimately disposed by trial, the data used in these analyses cannot distinguish

whether plea negotiations were attempted and, if so, why the negotiation failed.

Nevertheless, it is reasonable to assume that in the vast majority of cases, defendants rejected

the plea offer from the prosecutor because the offer was deemed unacceptable. Although no

systematic research on this topic has been conducted, experts estimate between 50% and

75% of inmates on death row rejected plea offers that would have avoided the death

sentence. WHITE, LITIGATING IN THE SHADOW OF DEATH, supra note 19, at 145–46;

Alschuler, supra note 17, at 671; Stetler, supra note 19, at 1157.

According to a recent empirical study of federal postconviction review of capital

cases, 3% of defendants sentenced to death plead guilty at the conviction phase without

obtaining a sentence bargain from the prosecutor. KING & HOFFMANN, supra note 44, at 147.

Some defendants may perceive additional benefits from accepting a plea apart from what the

government is willing or able to offer, such as leniency from the judge or jury during

sentencing. See Dervan, supra note 52, at 259.

70 Kuziemko, supra note 20, at 141 (“This work is, to the best of my knowledge, the first


Kuziemko’s study of the impact of the death penalty on plea bargaining in

New York State is the first study to examine whether the threat of the death

penalty deters capital defendants from taking their cases to trial.71 After the

reinstatement of the death penalty in 1995, several district attorneys from

across New York’s sixty-two counties publicly announced their refusal to

seek the death penalty in death-eligible murder cases (i.e., first-degree

murder cases).72 The study used the reinstatement of the death penalty in

1995, coupled with these refusals to seek the death penalty, as a “natural

experiment to estimate the effect of the death penalty on plea bargaining.”73

The study identifies defendants who were arrested for first- or seconddegree

murder as the “treatment group”—that is, the group of defendants

potentially susceptible to the death penalty—and defendants who were

arrested for burglary, forcible rape, or armed robbery as the “control

group,” and compares these groups before and after the policy change.74

The study concludes that the threat of the death penalty in New York did

not increase defendants’ overall propensity to plead guilty, but the death

penalty did lead defendants to accept plea bargains with harsher terms.75

The approach taken in the study was sensible considering the available

data on New York’s capital charging-and-sentencing process, but several

deficiencies with the data likely undermine the substantive conclusions.

The first problem is the jurisdiction selected for the study: New York State.

While it is understandable that New York was chosen because of the

“natural experiment” conditions presented by the reinstatement of the death

to examine the effect of the death penalty on plea bargaining.”); see also Ehrhard, supra note

20, at 315; Scheidegger, supra note 20, at 2–3.

71 See generally Kuziemko, supra note 20 (studying the effect of this threat on behavior).

72 Id. at 118, 121; see also Al Baker, Effort to Reinstate Death Penalty Law Is Stalled in

Albany, N.Y. TIMES, Nov. 18, 2004, at A1.

73 Kuziemko, supra note 20, at 118.

74 Id. at 118, 122. The capital statute enacted in 1995 also expanded the definition of

first-degree murder in the state. Before the act, first-degree murder was limited to

individuals who were convicted of willfully killing law enforcement officers, and seconddegree

murder was a catchall category that included a wide range of homicide offenses. In

an attempt to take into account the expanded definition of first-degree murder, and therefore

keep the treatment group consistent across time, the study aggregated first- and seconddegree

homicides. Id. at 120.

75 Id. at 140–41. Recognizing the possibly limited generalizability of the New York

findings, the study examines a national cross section of murder defendants in 1988 drawn

from the thirty-three largest counties in the United States. Results from the national data

also corroborate these findings. Id. at 135–40. Unfortunately, the national data suffer from

important limitations that also limit generalizability (i.e., improper focus on large urban

counties and improper identification of treatment and control groups). Stephanie Hindson et

al., Race, Gender, Region and Death Sentencing in Colorado, 1980–1999, 77 U. COLO. L.

REV. 549, 570–77 (2006) (discussing county variation in use of the death penalty within a


494 SHEROD THAXTON [Vol. 103

penalty, New York has a rather inactive death penalty when compared to

many other jurisdictions with capital statutes. For example, from 1995

through 2000, district attorneys in New York issued only thirty-six death

notices, an average of six notices per year,76 although New York averaged

nearly 500 first- and second-degree murder arrests each year.77

Furthermore, by the year 2000, only six people sat on death row in New

York, and no executions have taken place since the death penalty was


The second problem is that the study does not directly examine the

impact the death penalty has on the bargaining process in cases that are

actually noticed for the death penalty. The study posits that the effect of the

death penalty may not be limited to defendants who are actually noticed for

the death penalty because “its specter may have encouraged some

defendants to secure plea bargains after the [district attorney] merely

mentioned a death sentence as a possibility but before he actually issued a

death notice.”79 This assumption is problematic because it misrepresents

actual capital charging dynamics. District attorneys do not deem all firstdegree

murders as worthy of the death penalty.80 While the fact that a

defendant is charged with first-degree murder under New York’s statute is

sufficient to permit the prosecutor to seek the death sentence, the

prosecutor’s mere mention of the possibility of seeking the death penalty

against a defendant is unlikely to be viewed as a credible threat because the

death penalty is used so infrequently.81 Even in cases that are technically

76 Kuziemko, supra note 20, at 121.

77 Id.

78 Id. (stating the last execution in New York occurred in 1963). Three additional factors

may limit the study’s generalizability: (1) the limited scope of New York’s death penalty

(i.e., felony murder is ineligible for the death penalty); (2) New York’s provision allowing

defendants the right to plead guilty and automatically receive life imprisonment; and (3) the

dramatic increase in the quality of defense counsel available to capital defendants resulting

from the creation of the state-initiated Capital Defender Office, which was part of the capital

punishment statute. Id. at 135 & n.14. Furthermore, New York’s highest court invalidated

the death penalty statute in 2004, just four years after the time frame the study examines.

See People v. LaValle, 817 N.E.2d 341 (N.Y. 2004).

79 Kuziemko, supra note 20, at 133.

80 See Raymond Paternoster et al., Justice by Geography and Race: The Administration

of the Death Penalty in Maryland, 1978–1999, 4 U. MD. L.J. RACE RELIGION GENDER &

CLASS 1, 17–18 (2004) (explaining that different prosecutors in different offices, or even

within the same office, may not evaluate a case as death eligible, even when the case

satisfies all of the statutory criteria for death eligibility); Glenn L. Pierce & Michael L.

Radelet, Race, Region, and Death Sentencing in Illinois, 1988–1997, 81 OR. L. REV. 39, 41,

46 (2002) (describing the same phenomenon).

81 This observation is underscored by the fact that prosecutors only formally sought the

death penalty thirty-six times over a six-year period. See, e.g., Bar-Gill & Ben-Shahar,

supra note 15, at 738 (underscoring the importance of “credible threats” by prosecutors in


death eligible, it is more plausible that the defendant would wait until the

prosecutor actually carried out her threat to seek the death penalty before

agreeing to a plea bargain. In fact, defense counsel, who is typically a

repeat player with the prosecutor, is likely to advise her client as to whether

the prosecutor’s threat is credible.82

The study also implausibly assumes, without evidence, that district

attorneys (or defense counsel) actually mention the possibility of the death

sentence in all cases where an indictment for first-degree murder is

obtained. This is especially unlikely because capital defendants are

permitted to accept plea bargains allowing them to avoid the death sentence

after their cases are noticed for the death penalty, but they do not enjoy a

constitutional right to withdraw guilty pleas in capital cases when the pleas

were made voluntarily and knowingly.83 Nearly all death penalty

jurisdictions (including New York) permit capital defendants to accept plea

bargains up until the penalty phase of their capital trials.84 As mentioned

supra, the majority of the thirty-six death notices in the study were

withdrawn after plea bargains were reached. Therefore, if the assumption

that the mere possibility of a death notice influences defendants’ plea

calculus is incorrect, then the study’s treatment group includes defendants

who are not impacted by the possible threat of the death penalty and the

treatment effect will be biased towards zero because of systematic

measurement error. Stated differently, the treatment group will actually

contain individuals who should be either in the control group or entirely

excluded from the analysis.

The New York State study defends this approach by characterizing its

estimates as “intent to treat” (ITT) measures,85 rather than the conventional

estimates of “average treatment effect” (ATE) or average treatment effect

the plea-bargaining context).

82 STUNTZ, supra note 50, at 258 (“Plea bargaining is what academics call a ‘repeat-play’

game; the same lawyers negotiate pleas again and again.”).

83 Compare Santobello v. New York, 404 U.S. 257, 260 (1971) (explaining that a

defendant does not have a constitutional right to withdraw a guilty plea in all circumstances),

with Fair v. State, 268 S.E.2d 316, 323–24 (Ga. 1980) (holding that a defendant has no

constitutional right to withdraw a permissible guilty plea in a capital case and that motions to

withdraw a plea must be assessed on a case-by-case basis).


EMPIRICAL ANALYSIS 23 (1990) (discussing bifurcated death penalty trials post-Furman);

Kuziemko, supra note 20, at 119–20 (explaining that capital trials are conducted in two

parts: in the first phase, the court is only concerned with the question of guilt; if the

defendant is convicted of capital murder in the first phase, the same jury proceeds to the

sentencing phase of the trial to determine the appropriate punishment); see also infra Part

III.A (discussing Georgia’s capital charging-and-sentencing process and noting that

defendants are allowed to accept pleas up until the penalty phase).

85 Kuziemko, supra note 20, at 133.

496 SHEROD THAXTON [Vol. 103

on the treated (ATT).86 “The ITT is one of a number of possible parameters

of interest and may not always be of greatest scientific or policy

relevance.”87 It measures the effect of treatment assignment, but not the

effect of the treatment itself.88 But it is questionable that the study’s

estimates can be accurately characterized as “intent to treat” estimates

because of the reasons previously stated: (1) prosecutors do not deem all

death-eligible cases as “death worthy,” (2) not all death-eligible defendants

(or their counsel) believe their cases are at risk for the death penalty, and (3)

prosecutors do not “threaten” the death penalty in all (or even the vast

majority) of death-eligible cases. Furthermore, the study does not alleviate

this concern because it does not distinguish cases in which a threat was ever

mentioned from the larger group of defendants indicted for first-degree

murder.89 Problems of systematic measurement error notwithstanding,90

scholars have repeatedly remarked that the ATT is both easier to identify

and likely to be more theoretically informative because it describes the

impact of the treatment only among the units who are actually exposed to

it.91 Formally, assuming some selection on observables, ATT = E[Y(1) Y(0)

| X, T = 1]; where Y(1) and Y(0) denote the two potential outcomes under

treatment and control conditions, respectively, X indicates observable

86 Michael E. Sobel, Causal Inference in Randomized and Non-Randomized Studies: The

Definition, Identification, and Estimation of Causal Parameters, in THE SAGE HANDBOOK

OF QUANTITATIVE METHODS IN PSYCHOLOGY 3, 7–8 (Roger E. Millsap & Alberto Maydeu-

Olivares eds., 2009) (defining ITT, ATE, and ATT causal effects).

87 Id. at 3, 7. The ITT, ATE, and ATT address different causal questions: (1) ITT

measures the average effect of the treatment per person offered the treatment, irrespective of

how many treatment group members actually received it; (2) ATE measures the average

effect of the treatment if all members in the treatment group actually received the treatment;

and (3) ATT measures the average treatment effect per person receiving the treatment. Lisa

A. Gennetian et al., Constructing Instrumental Variables from Experimental Data to Explore


EVOLVING ANALYTIC APPROACHES 75, 86–87 (Howard S. Bloom ed., 2006).

88 Sobel, supra note 87, at 7. It is also important to emphasize that the intent to treat

effect “is commonly featured in connection with randomized clinical trials,” in order to

justify the assumption that the treatment effect is identifiable. Id.

89 But see Els Goetghebeur & Tom Loeys, Beyond Intention to Treat, 24 EPIDEMIOLOGIC

REVS. 85, 85 (2002) (arguing that the “upside” of noncompliance is that it more closely

resembles the heterogeneous population of future treatment groups).

90 See, e.g., Goetghebeur & Loeys, supra note 89, at 89 (“The more we seek to tailor

possibly dynamic treatments to individual characteristics . . . the more imperative it becomes

to acknowledge treatment actually received as an important source of variation in treatment



INFERENCE 43 (2007) (“[T]he average treatment effect among the treated is a theoretically

important quantity . . . .”); Markus Gangl, Causal Inference in Sociological Research, 36

ANN. REV. SOC. 21, 24 (2010) (same).


covariates in the model, and T is an indicator variable for treatment


The New York State study should be applauded for highlighting this

gap in the literature and making an important empirical contribution to the

existing debate, although the limitations of the data partly undermine the

reliability and generalizability of the conclusions that were reached.93 A

significant improvement on that seminal study would entail several

modifications. First, it would examine a jurisdiction that more actively

pursues the death penalty with respect to both sentencing individuals to

death and carrying out executions.94 Second, the study would directly

examine actual treatment effects and not merely “intent to treat” effects. In

fact, prior research has appropriately defined the “treatment effect” of the

death penalty as the government’s “decision to file a death notice that

formally announces [the] state’s intention to seek a death sentence.”95

Finally, it would control for a wider range of case factors relevant to

charging and plea-bargaining decisions.96 The present study incorporates

all of these improvements by analyzing a rich data set from Georgia, which

is better suited to test this hypothesis.

92 MORGAN & WINSHIP, supra note 91, at 42; see also Donald B. Rubin, Estimating

Causal Effects of Treatments in Randomized and Nonrandomized Studies, 66 J. EDUC.

PSYCHOL. 688, 689–90 (1974).

93 See infra Part V.

94 It is important to make a distinction between those jurisdictions that actively pursue

the death penalty but infrequently execute individuals and those that actively pursue the

death penalty and execute defendants. California, for example, has nearly double the death

row population of any other state (721 death row inmates as of January 1, 2011), but rarely

executes inmates who have been sentenced to death. CAL. COMM’N ON THE FAIR ADMIN. OF



ROW U.S.A. 34, 39–45 (2011).

95 See Roman et al., supra note 27, at 533.

96 The New York study examines the following case-level factors: defendant’s sex,

race/ethnicity, age, number of prior convictions, county of arrest, original charge, and plea

charge. Kuziemko, supra note 20, at 129. The New York study fails to explore victim

characteristics, contemporaneous convictions, and specific case factors identified in the

statute that make a case a death-eligible crime. The supplemental national cross-section

analysis includes both offender and victim demographic information, but lacks legally

relevant, case-specific information outside of the number of prior convictions. Id. at 136.

Taking into account more case information also permits the analyst to determine whether the

assumptions of the “natural experiment” actually hold—that is, whether the treatment and

control groups are truly indistinguishable except for the intervention. Richard A. Berk,

Randomized Experiments as the Bronze Standard, 1 J. EXPERIMENTAL CRIMINOLOGY 417,

421, 428 (2005).

498 SHEROD THAXTON [Vol. 103


Prosecutors in Georgia have aggressively sought the death penalty

since the practice first began in the colonies in 1608—Georgia ranks fifth in

executions carried out since that time (1,002).97 Georgia also ranks seventh

in the nation with respect to total executions since the death penalty was

reinstated (52) and ninth in terms of its death row population (102).98 At

the time of this writing, the most recent execution carried out in Georgia

was that of Andrew Cook on February 21, 2013.99 Georgia has also been

the most influential state in shaping national death penalty policy in the

modern era of capital punishment.100 No less than seventeen cases

originating in Georgia have set legal precedent with respect to the

administration of capital punishment, including Furman v. Georgia101 and

Gregg v. Georgia,102 which, respectively, were responsible for placing and

lifting the moratorium on executions in the United States in the 1970s.103

The following section provides a brief history of Georgia’s current capital



On June 29, 1972, in Furman v. Georgia,104 the U.S. Supreme Court

invalidated Georgia’s death penalty statute, ruling that the lack of



DEATH PENALTY (2012), available at http://www.deathpenaltyinfo.org/FactSheet.pdf

(providing information about executions in Georgia from 2003 through the present).

98 DEATH PENALTY INFO. CTR., supra note 97, at 2–3.

99 Id.; Rhonda Cook, Executed Man Makes Apology, ATLANTA J.-CONST., Feb. 23, 2013,

at B1.

100 BALDUS ET AL., supra note 84, at 3.

101 408 U.S. 238, 239–40 (1972) (per curiam) (holding that unguided-discretion death

penalty statutes are unconstitutional).

102 428 U.S. 153, 207 (1976) (holding that guided-discretion capital statutes are

constitutionally permissible).

103 Rhonda Cook, Georgia Cases Have Set Legal Precedent, ATLANTA J.-CONST., June 2,

1996, at C4. Other significant cases include McCleskey v. Kemp, 481 U.S. 279, 319 (1987)

(finding statistical evidence of system-wide discrimination irrelevant; the defendant must

show discrimination in the case at hand); Francis v. Franklin, 471 U.S. 307, 324–25 (1985)

(holding that capital defendants enjoy the presumption of innocence and do not have the

burden of proof in capital cases); Zant v. Stephens, 462 U.S. 862, 891 (1983) (holding that

death penalty may be imposed as long as the jury finds at least one valid aggravating

circumstance, and nonstatutory aggravating circumstances may also be considered by the

jury when deciding whether to sentence a defendant to death); Eberheart v. Georgia, 433

U.S. 917, 917 (1977) (declaring death penalty judgment for nonhomicidal kidnapping with

bodily injury unconstitutional); and Coker v. Georgia, 433 U.S. 584, 592 (1977) (declaring

death penalty for nonhomicidal rape of an adult unconstitutional).

104 408 U.S. 238.


sentencing guidance for capital jurors was a violation of the Eighth

Amendment, and subsequently invalidated all capital statutes that lacked

such standards.105 The Georgia General Assembly quickly drafted new

death penalty legislation the following January and the bill was signed into

law by then-Governor Jimmy Carter on March 28, 1973.106 The legislation

was soon codified; it provided for eleven separate instances where the death

penalty could be imposed on someone convicted of a capital offense.107

Less than a year later, Troy Leon Gregg was convicted of murder and

armed robbery and sentenced to death under Georgia’s new death penalty

statute. The U.S. Supreme Court would eventually grant certiorari in

Gregg’s108 case (consolidated with two other death penalty cases from

Florida109 and Texas110) and ultimately rule that Georgia’s new death

penalty statute was constitutionally acceptable.111 The Court’s decision

officially ended the nation’s four-year moratorium on the death penalty.112

105 Id. During the previous year, the Supreme Court ruled by a six-to-three vote that

neither the absence of sentencing guidelines nor single-verdict procedures violated the

Fourteenth Amendment. McGautha v. California, 402 U.S. 183, 196–208, 210–13 (1971).

The Furman ruling, which was decided by a five-to-four vote, was particularly surprising

because the composition of the Court had not changed from the McGautha ruling. HERBERT


AMERICA, 1972–1994, at 37–39 (1996).

The immediate effect of Furman was that approximately 558 death row inmates had

their sentences commuted to life imprisonment. Although there was widespread speculation

by death penalty proponents that many of these inmates would kill again once released from

prison, subsequent research revealed that only one of the 239 Furman-commuted inmates

released from prison committed a second murder in the fifteen years following the Furman

decision. James W. Marquart & Jonathan R. Sorensen, A National Study of the Furman-

Commuted Inmates: Assessing the Threat to Society from Capital Offenders, 23 LOY. L.A. L.

REV. 5, 23–24 (1989).


at 14–41 (1999) (providing a detailed history of the modern death penalty in Georgia). On

December 8, 1972, less than six months after Furman, Florida became the first state to

officially restore capital punishment when Governor Reubin Askew signed new death

penalty legislation into law. HAINES, supra note 105, at 45.

107 1973 Ga. Laws 163–65, § 3 (originally codified at GA. CODE ANN. § 27-2534.1

(1973); current version at GA. CODE ANN. § 17-10-30 (West 2003)).

108 Gregg v. Georgia, 428 U.S. 153 (1976).

109 Proffitt v. Florida, 428 U.S. 242 (1976).

110 Jurek v. Texas, 428 U.S. 262 (1976).

111 Gregg, 428 U.S. at 169, 195 (holding, seven-to-two, that the death penalty for murder

did not by itself violate the Eighth Amendment and all three of the capital statutes contained

sufficient procedural reforms to warrant them constitutional under Furman); see also

HAINES, supra note 105, at 52–54. Interestingly, the court had no evidence suggesting that

the new statutes eliminated arbitrariness and bias in capital sentencing; rather, the court

based its decision on whether the procedural reforms enacted in each statute were capable of

producing outcomes different from those produced under the pre-Furman statutes.

112 MEARS, supra note 106, at 65–69.

500 SHEROD THAXTON [Vol. 103

As a result of the decision, thirty-four states, the federal government, and

the U.S. Armed Forces currently permit the death penalty in their


As noted supra, Georgia’s new death penalty was originally enacted in

1973 and enumerated eleven separate instances where the death penalty

could be imposed on someone convicted of a capital offense:

(a) The death penalty may be imposed for the offenses of aircraft hijacking or treason,

in any case.

(b) In all cases of other offenses for which the death penalty may be authorized, the

judge shall consider, or he shall include in his instructions to the jury for it to

consider, any mitigating circumstances or aggravating circumstances otherwise

authorized by law and any of the following statutory aggravating circumstances which

may be supported by the evidence:

(1) The offense of murder, rape, armed robbery, or kidnapping was committed by a

person with a prior record of conviction for a capital felony, or the offense of

murder was committed by a person who has a substantial history of serious

assaultive criminal convictions.114

(2) The offense of murder, rape, armed robbery, or kidnapping was committed

while the offender was engaged in the commission of another capital felony, or

aggravated battery, or the offense of murder was committed while the offender

was engaged in the commission of burglary or arson in the first degree.

(3) The offender by his act of murder, armed robbery, or kidnapping knowingly

created a great risk of death to more than one person in a public place by means of

a weapon or device which would normally be hazardous to the lives of more than

one person.

(4) The offender committed the offense of murder for himself or another, for the

purpose of receiving money or any other thing of monetary value.

(5) The murder of a judicial officer, former judicial officer, district attorney or

solicitor or former district attorney or solicitor during or because of the exercise of

his official duty.

(6) The offender caused or directed another to commit murder or committed

murder as an agent or employee of another person.

(7) The offense of murder, rape, armed robbery, or kidnapping was outrageously

or wantonly vile, horrible or inhuman in that it involved torture, depravity of mind,

or an aggravated battery to the victim.

(8) The offense of murder was committed against any peace officer, corrections

employee or fireman while engaged in the performance of his official duties.

(9) The offense of murder was committed by a person in, or who has escaped

from, the lawful custody of a peace officer or place of lawful confinement.

113 DEATH PENALTY INFO. CTR., supra note 97.

114 The italicized portion of the death penalty statute was subsequently declared

unconstitutionally vague. Arnold v. State, 224 S.E.2d 386, 392 (Ga. 1976).


(10) The murder was committed for the purpose of avoiding, interfering with, or

preventing a lawful arrest or custody in a place of lawful confinement, of himself

or another.

(c) The statutory instructions as determined by the trial judge to be warranted by the

evidence shall be given in charge and in writing to the jury for its deliberation. The

jury, if its verdict be a recommendation of death, shall designate in writing, signed by

the foreman of the jury, the aggravating circumstance or circumstances which it found

beyond a reasonable doubt. In non-jury cases the judge shall make such designation.

Except in cases of treason or aircraft hijacking, unless at least one of the statutory

aggravating circumstances enumerated in Code section 27-2434.1(b) is so found, the

death penalty shall not be imposed.115

With very few changes, Georgia’s death penalty legislation has

remained in place since Governor Jimmy Carter first signed it into law;116

however, there were several changes mandated by subsequent U.S.

Supreme Court rulings. The year after the Court officially reinstated

Georgia’s death penalty in Gregg, it invalidated the death penalty for

defendants convicted of non-homicidal rape and kidnapping with bodily

injury in, respectively, Coker v. Georgia and Eberheart v. Georgia.117

Georgia’s juvenile death penalty was also invalidated following the Court’s

decision in Roper v. Simmons,118 which forbade the death penalty for

defendants who were under the age of eighteen during the commission of

their crime. The Georgia statute had permitted the death penalty for

defendants who were seventeen at the time of their crime.119

115 1973 Ga. Laws 163–65, § 3 (emphasis added) (originally codified at GA. CODE ANN.

§ 27-2534.1 (1973); current version at GA. CODE ANN. § 17-10-30 (West 2003)).

Throughout the remainder of this Article, the specific elements of the capital statute listed in

subsection (b) of Georgia’s capital statute will be referred to as B1, B2, B3, etc.

116 In 1996 and 1997, there were two unsuccessful proposals to lower the age of

eligibility for the death penalty to sixteen. There were also two attempts to add an additional

aggravating circumstance that would allow the death penalty in the event a person was

convicted of the rape of a child under the age of twelve; however, this legislation was also

unsuccessful. MEARS, supra note 106, at 46. In 2006, an additional element, B11, was

added: “The offense of murder, rape, or kidnapping was committed by a person previously

convicted of rape, aggravated sodomy, aggravated child molestation, or aggravated sexual

battery.” GA. CODE ANN. § 17-10-30(b)(11) (West Supp. 2012).

117 See Eberheart v. Georgia, 433 U.S. 917 (1977); Coker v. Georgia, 433 U.S. 584

(1977). The current Georgia statute permits the death penalty for murder (malice or felony),

aircraft hijacking, and treason. See Collins v. State, 236 S.E.2d 759, 762 (Ga. 1977) (Jordan,

J., concurring) (“Of course the crimes of treason and aircraft hijacking, along with murder,

remain capital felonies . . . .”).

118 543 U.S. 551 (2005).

119 DEATH PENALTY INFO. CTR., supra note 97. For a critical treatment of the U.S.

Supreme Court’s rulings in Coker and Roper, see generally Adam S. Goldstone, The Death

Penalty: How America’s Highest Court Is Narrowing Its Application, 4 CRIM. L. BRIEF 23

(2009) (arguing that the Coker and Roper decisions are examples of judicial activism and

inappropriately limit the application of the death penalty).

502 SHEROD THAXTON [Vol. 103


Georgia’s life sentence without the possibility of parole statute was

enacted in May 1993 and allowed juries to deny parole to defendants

convicted of certain high felonies, including murder.120 Presently, all thirtyfour

states that authorize the death penalty have enacted similar

legislation.121 Georgia’s LWOP legislation may have significantly altered

the administration of capital punishment in Georgia by restricting LWOP to

murder cases in which the prosecution has filed notices of intent to seek the

death penalty.122 As a result, prosecutors may seek the death sentence in

cases they do not believe are deserving of the death penalty, but they do

believe warrant LWOP.123 For example, in 2001, Devonia “Eddie” Inman

was convicted in Adel, Georgia, of the murder of Donna Brown and

sentenced to life without the possibility of parole.124 In commenting on the

verdict and sentence, Alapaha Judicial Circuit District Attorney Bob Ellis

remarked, “Had we not sought the death penalty, we could have not gotten

life without parole.”125 Ellis further explained that by seeking the death

penalty, he gave the jury the opportunity to deny parole to Inman.126

Even in jurisdictions that do not restrict the LWOP sentencing option

to death cases, it is unlikely that a defendant charged with a death-eligible

homicide offense would agree to plea to an LWOP sentence without the

threat of a death sentence at trial. Absent the risk of a death sentence at

trial, the defendant could do no worse at trial than the LWOP plea offered

by the prosecutor. The defendant would be better off taking her chances,

120 1993 Ga. Laws 1656–57, § 4 (codified at GA. CODE ANN. § 17-10-30.1 (repealed


121 DEATH PENALTY INFO. CTR., supra note 97.

122 In 2009, after two failed attempts, the Georgia General Assembly enacted legislation

permitting the imposition of life without the possibility of parole in murder cases,

independent of a death penalty prosecution. 2009 Ga. Laws 227, § 10; see also H.R. 142,

§ 17, 151st Gen. Assemb., Reg. Sess. (Ga. 2011) (codifying the 2009 law at GA. CODE ANN.

§ 17-10-16.1 (West Supp. 2012)).

123 See, e.g., Tony Perry, Drug Lord Avoids Death Penalty with Plea Deal: Arellano

Felix Pleads Guilty to Charges that Will Put Him in Prison for Life Without the Possibility of

Parole, L.A. TIMES, Sept. 18, 2007, at B1 (discussing defendant agreeing to plea to life

without the possibility of parole in exchange for withdrawal of death penalty); Gene

Johnson, Strategy Changing on Death Penalty, NEWS TRIB. (Tacoma, Wash.) (July 30, 2007,

1:00 AM), http://www.thenewstribune.com/2007/07/30/121534/strategy-changing-on-deathpenalty.

html (quoting a prosecutor explaining that the threat of the death penalty is the only

leverage available in some cases).

124 Peter Failor, Man Gets Life for 1998 Adel Murder, VALDOSTA DAILY TIMES, June 28,

2001, at A1.

125 Id.

126 Id.


however slim, at trial for the possibility of receiving a lesser sentence.127

This is particularly true in jurisdictions such as Georgia that permit judges

and juries to impose a life with the possibility of parole sentence for anyone

convicted of murder or felony murder.128 In Georgia, the defendant and

prosecutor may enter into a plea agreement at any time up until the jury

renders its sentence in the penalty phase. The following section briefly

outlines the major stages of a case from indictment through the penalty

phase that differentiate capital from noncapital cases.


Indictment Through Arraignment. Georgia is an indictment

jurisdiction, so a grand jury is required to decide formally whether there is

probable cause to believe that the accused has committed the specified

crime.130 Following the indictment, the accused may be eligible for the

appointment of counsel.131 If deemed eligible for appointed counsel and the

charge is a capital felony, two attorneys must be appointed before the

accused is called upon to plea to the charges, which generally occurs at the

arraignment. Prior to arraignment, a pretrial conference is held and the

prosecuting attorney must announce whether she intends to seek the death

penalty and then file a notice of intent with the clerk of the superior

127 Ehrhard, supra note 20, at 313 (summarizing statements from prosecutors explaining

that the death penalty is often the only leverage they have in plea negotiations in murder

cases). Indeed, the likelihood of receiving a straight life sentence at trial in a capital murder

case in Georgia does not appear to be particularly slim. From 1993 to 2000, 31% of capital

cases disposed by trial resulted in straight life sentences, whereas 36% received LWOP and

32% received the death sentence.

128 GA. CODE ANN. § 17-10-30 (West 2003). In economics parlance, there is no “price”

or “penalty” associated with a defendant invoking her constitutional right to trial under the

Georgia regime in place at the time of this study if the prosecutor only offers LWOP in a

noncapital case. To be sure, individuals who finance their own defenses will incur those

costs, but the vast majority of murder defendants are represented by court-appointed counsel.

See generally Beck & Shumsky, supra note 34, at 525; Tabak & Lane, supra note 34, at 59.

129 In the interest of space, only the most relevant stages of the “typical” progression of a

Georgia death penalty case through automatic appeal are described. The qualifier “typical”

is used because there are numerous factors that may cause a case to deviate from this

abbreviated description.

130 GA. CODE ANN. §§ 15-12-60 et seq.

131 Upon a showing of indigence, an individual indicted for a capital felony is eligible for

appointed counsel. Pursuant to the Georgia Indigent Defense Act of 2003 (GIDA), an

indigent is “[a] person charged with a . . . [crime] punishable by imprisonment who earns

less than 100 percent of the federal poverty guidelines unless there is evidence that the

person has other resources that might reasonably be used to employ a lawyer without undue

hardship on the person or his or her dependents.” GA. CODE ANN. § 17-12-2 (West 2003 &

Supp. 2012); see also 2003 Ga. Laws 192–217, § 1.

504 SHEROD THAXTON [Vol. 103

court.132 The superior court must then transmit the notice to the clerk of the

Supreme Court of Georgia.133 During the arraignment, the court must read

the indictment and ask the defendant to plead to the capital felony and any

lesser-included offenses charged. The defendant is allowed to plead guilty,

not guilty, or mentally incompetent to stand trial.134

Capital Trial. The court must empanel forty-two prospective jurors

from whom the state and defense must select a total of twelve jurors and

one or more alternative jurors, if deemed necessary by the judge.135 All

capital cases are heard before the Georgia Superior Court136 and conducted

in two phases: the conviction phase (also commonly referred to as the

guilt/innocence phase) and, if the defendant is found guilty of a capital

felony, the penalty phase. Immediately prior to the conviction phase, the

court must conduct a conference with the state, defense counsel, and the

defendant to resolve several matters, including, inter alia, any last-minute

motions, stipulations, and objections to defense counsel.137

In situations where the defendant is found guilty of capital murder at

the conclusion of the conviction phase, the case proceeds to the penalty

phase (i.e., a presentencing hearing) where both the prosecutor and defense

counsel may present witnesses and evidence regarding the statutory

aggravating circumstances, as well as nonstatutory aggravating and

mitigating circumstances.138 The jury may sentence the defendant to death

only if they find one or more statutory aggravating circumstances beyond a

reasonable doubt, but a death sentence is never required.139 Following a

conviction for a capital felony and a sentence of death, the defendant may

challenge her conviction or death sentence by: (1) filing a motion for a new

trial with the superior court, or (2) filing a direct appeal with the Georgia

132 UNIF. APP. R. IIC(1). The specific aggravating circumstances the government intends

to prove at trial need not be included in the indictment. See generally Lewis v. State, 620

S.E.2d 778 (Ga. 2005).

133 UNIF. APP. R. IIC(1).

134 A defendant indicted for a capital felony may not plead nolo contendere. GA. CODE

ANN. § 17-7-95(a).

135 Id. §§ 15-12-160, -168.

136 GA. CONST. art. 6, § 4.

137 UNIF. APP. R. IIIA(1).

138 GA. CODE ANN. § 17-10-2(c); see also Zant v. Stephens, 462 U.S. 862, 891 (1983)

(holding that nonstatutory aggravating circumstances may be considered by the jury during

the presentence hearing).

139 GA. CODE ANN. § 17-10-31.1(c). Prior to Ring v. Arizona, 536 U.S. 584 (2002),

which prohibited judge-imposed death sentences, judges in Georgia were permitted to

sentence a capital defendant to death if the defendant waived her right to a jury at the penalty

phase. This situation, however, never occurred in the years under investigation for this



Supreme Court.140 If the defendant does not initiate any sort of review, the

case will automatically be appealed to the Georgia Supreme Court within

ten days of the filing of the trial transcript by the court reporter of the

Georgia Superior Court.141 This automatic review will occur even if the

defendant does not wish to appeal her conviction or sentence.142

It is worth emphasizing that it is very common for multiple years to

elapse between the initial arrest and the trial in death penalty cases, so both

parties have ample time to negotiate a plea agreement. The data analyzed

for this study143 reveal that the average time between arrest and sentencing

for defendants noticed for the death penalty and opting for trial was 31.9

months. The average time for defendants noticed for the death penalty but

ultimately pleading at some point before the penalty phase was

approximately 24.6 months. For defendants noticed for the death penalty,

irrespective of whether they pleaded or took their cases to trial, the average

amount of time between arrest and sentencing was 27.9 months. By

comparison, the average time between arrest and sentencing for deatheligible

defendants not facing the death penalty but ultimately convicted of

murder was 17.4 months for those opting for trial and 13.5 months for those

who pleaded (and 15.9 months irrespective of method of disposition). The

specific data collected in Georgia and analyzed in this Article are discussed

in the next section.


The current study analyzes eight years of death penalty charging-andsentencing

data from Georgia (1993–2000) and addresses each of the

aforementioned modifications in an effort to assess more accurately the

impact of the death penalty on the plea-bargaining process.144 The data

used for these analyses were collected from the Georgia Bureau of

Investigation (GBI), the Georgia Department of Corrections (GDC), the

Office of the Georgia Capital Defender (GCD), the Clerk’s Office of the

140 GA. CODE ANN. § 17-10-35; UNIF. APP. R. IVA(1)–(2).

141 GA. CODE ANN. § 17-10-35.

142 Id.

143 See infra Part IV.

144 This particular time frame was selected for two important reasons. First, Georgia’s

life without the possibility of parole legislation was enacted in 1993. The legislation was

specifically designed as a sentencing alternative in capital murder trials, therefore potentially

having a significant impact on prosecutorial discretion. GA. CODE ANN. § 17-10-30.1

(repealed 2009); 1993 Ga. Laws 1656–57, § 4. Second, in October 1992, the Georgia

General Assembly, along with the Georgia Supreme Court, established a statewide agency to

actively monitor all death penalty cases in Georgia’s 159 counties. MEARS, supra note 106,

at 4. For a discussion of additional advantages of examining Georgia’s capital charging-andsentencing

process, see BALDUS ET AL., supra note 84, at 3.

506 SHEROD THAXTON [Vol. 103

Georgia Supreme Court (CO), the Atlanta Journal-Constitution (AJC),145

and the U.S. Census Bureau.146 These data contain detailed information on

each homicide case in Georgia with respect to the defendant,

codefendant(s), victim(s), judge, prosecutor, defense counsel, and the crime.

As mentioned supra,147 the Georgia death penalty statute lists eleven

elements making a crime eligible for the death penalty.148 These data allow

for the determination of which defendants are actually eligible for the death

penalty in Georgia. Information obtained concerning prosecutors’ actual

decisions to seek the death penalty in each case permit the examination of a

genuine treatment effect.149 The more nuanced data also permit the

inclusion of a much richer set of statistical controls than were considered in

prior research.

During the period under investigation (1993–2000), prosecutors filed a

notice of intent to seek the death penalty in 400 cases and fifty-four

defendants ultimately received the death penalty.150 Of the 395 capitally

charged cases in which the method of disposition is known, 59% (234) were

resolved by plea and 41% (161) were resolved by trial. With respect to

cases that were technically death eligible under the Georgia statute but in

which the prosecutor declined to seek the death penalty, 39% (350) were

disposed by plea and 61% (551) disposed by trial. Of the 724 cases

resulting in a murder conviction that were ineligible for the death penalty,

30% (212) were disposed of by plea and 70% (505) by trial.151 The plea

rate for cases noticed for the death penalty ranged from 38.7% (in 1998) to

145 Bill Rankin et al., A Matter of Life or Death: An AJC Special Report: High Court

Botched Death Reviews, ATLANTA J.-CONST., Sept. 26, 2007, at A1 (discussing a collection

of data on 2,328 murder convictions in Georgia between 1995 and 2004); Raymond

Paternoster, The Death Penalty in Georgia, 1995–2004 (Sept. 17, 2007) (unpublished report)

(on file with the Journal of Criminal Law and Criminology).

146 See infra Appendix A.

147 Supra Part III.A.

148 The original statute permitted the death penalty for the crimes of murder, rape, armed

robbery, or kidnapping with bodily injury, but rape, armed robbery, and kidnapping with

bodily injury were removed as death-eligible offenses following the Supreme Court’s

decisions in Coker v. Georgia, 433 U.S. 584, 584 (1977) (declaring the death penalty for

rape unconstitutional), and Eberheart v. Georgia, 433 U.S. 917, 917 (1977) (holding that the

death penalty for armed robbery was unconstitutional). See supra note 103.

149 See Gangl, supra note 91, at 24.

150 These 400 death notices and the fifty-four death sentences are with respect to incident

dates, not disposition dates. Seven of these sentences (12.9%) were imposed after 2000.

151 Forty-five percent (584) of all death-eligible cases were disposed by plea.

Information concerning the method of disposition was missing for seventeen cases in the

sample (0.8%). A total of 1,628 non-capitally-charged murder convictions were obtained in

the period under study—34% were disposed by plea. In these data, there are twenty-eight

cases in which the defendant was initially noticed for the death penalty but was ultimately

acquitted, had charges dismissed, or was convicted of a lesser offense.


75% (in 1999). The plea rate for death-eligible cases that were disposed by

plea in which the defendant did not face the death penalty ranged from 31%

(in 1998) to 53% (in 1993).



The “treatment group” for the purposes of the study consists of

defendants who were noticed for the death penalty. The “control group”

was comprised of defendants eligible for the death penalty, but against

whom the prosecutor chose not to seek the death penalty.152 The difficulty

in defining the control group is specifying what qualifies as a “deatheligible”

case. Some analysts suggest that any homicide committed in

Georgia could be death eligible resulting from the state’s felony-murder

statute and the B7 statutory aggravating circumstance.153 Recall that B7

reads: “The offense of murder, rape, armed robbery, or kidnapping was

outrageously or wantonly vile, horrible, or inhuman in that it involved

torture, depravity of mind, or an aggravated battery to the victim.”154 Due

to this ambiguity, two different approaches were adopted to identify deatheligible


The first approach categorizes defendants as eligible for the death

penalty based on the presence of at least one of the special aggravating

circumstances listed in Georgia’s capital statute. The presence of these

specific factors in each case was assessed in two ways. First, data from a

study conducted by the Atlanta Journal-Constitution on 2,328 murder

convictions obtained between 1995 and 2004 were used to determine the

number of aggravating circumstances present in the 967 murder convictions

with incident dates from 1995 through 2000 in that study.155

152 Roman et al., supra note 27, at 533 (defining treatment and control groups in death

penalty studies in a similar fashion).

153 Kuziemko, supra note 20, at 137 n.15; Kathryn W. Riley, The Death Penalty in

Georgia: An Aggravating Circumstance, 30 AM. U. L. REV. 835, 853–54 (1981) (explaining

that the vagueness and overbreadth of the B7 circumstance is in conflict with the narrowing

requirement articulated in Furman and Gregg); Richard A. Rosen, The “Especially

Heinous” Aggravating Circumstance in Capital Cases—The Standardless Standard, 64 N.C.

L. REV. 941, 945 (1986) (arguing that the “especially heinous” aggravating factor is

overinclusive, has been applied inconsistently, and fails to guide prosecutorial discretion).

154 GA. CODE ANN. § 17-10-30(b)(7) (West 2003); see also supra Part III.A for a full

description of Georgia’s capital statute. In Godfrey v. Georgia, the Court ruled the B7

special circumstance was not unconstitutionally vague on its face. 446 U.S. 420, 420 (1980).

155 See Rankin et al., supra note 145; Paternoster, supra note 145.

508 SHEROD THAXTON [Vol. 103

Table 1

Death Notices in Georgia by County (1993–2000)


Table 2

Death Notices in Georgia by Judicial Circuit (1993–2000)

Judical Circuit Death Notices

% of Total

Death Notices



















Blue Ridge



















































Lookout Mountain







































South Georgia









Stone Mountain












Western 18 4.5

Total Death Notices: 400

Percent of all judicial circuits filin g a death notice: 94%

510 SHEROD THAXTON [Vol. 103

The second manner in which the presence of statutorily defined

elements was determined was based on the presence of B1, B2, or B4

special circumstances from inmate records from the Georgia Department of

Corrections and the Georgia Bureau of Investigation. Recall that under B1,

a defendant is eligible for the death penalty when “[t]he offense of murder,

rape, armed robbery, or kidnapping was committed by a person with a prior

record of conviction for a capital felony.”156 Under Georgia’s statute,

capital felonies are defined as murder, rape, armed robbery, or

kidnapping.157 A capital offense refers to statutorily defined capital

offenses, not necessarily death-eligible offenses.158 The predicate offenses

for which the death penalty can be imposed are murder (malice or felony),

aircraft hijacking, and treason.159 According to the B2 statutory aggravating

circumstance, a defendant is eligible for the death penalty when “[t]he

offense of murder, rape, armed robbery, or kidnapping was committed

while the offender was engaged in the commission of another capital felony

or aggravated battery, or the offense of murder was committed while the

offender was engaged in the commission of burglary or arson in the first

degree.”160 Defendants who were convicted of contemporaneously

committing these crimes—or had one of these crimes initially listed in their

arrest reports—were categorized as death eligible.161 Defendants convicted

156 GA. CODE ANN. § 17-10-30(b)(1); see also supra Part III.A.

157 GA. CODE ANN. § 17-10-30(b)(1).

158 Merrow v. State, 601 S.E.2d 428, 431 (Ga. Ct. App. 2004) (“A capital offense . . .

refers to offenses defined by statute as capital offenses, not necessarily offenses for which

the state could or actually does seek the death penalty.” (quoting White v. State, 414 S.E.2d

296, 297 (Ga. Ct. App. 1991))); accord Peek v. State, 238 S.E.2d 12, 20 (Ga. 1977)

(explaining that nonhomicide “capital offenses” listed in the death penalty statute qualify as

“capital felonies” for purposes of applying the aggravating circumstance provision of the

capital statute).

159 Collins v. State, 236 S.E.2d 759, 762 (Ga. 1977) (Jordan, J., concurring) (“Of course

the crimes of treason and aircraft hijacking, along with murder, remain capital

felonies . . . .”).

160 GA. CODE ANN. § 17-10-30(b)(2).

161 The major limitation of this measure is that the Georgia Department of Corrections

does not technically distinguish between offenses committed during the actual commission

of the murder and offenses the defendant was simply convicted of during the same trial as

the murder. As a check, several cases were randomly selected from the Department of

Corrections website that lists the separate offenses (if a multiple-offense case) by the

incident date. In the vast majority of these cases, the murder and other felony were

committed on the same day. This suggests that, for most cases, the B2 measure is valid for

determining death eligibility. I also cross-referenced the conviction data from the Georgia

Department of Corrections with the arrest data from the Georgia Bureau of Investigation in

order to determine whether the contemporary felony present at time of arrest was consistent

with the conviction data for lesser felonies. This additional check supports the assertion that

the murder and the other felony conviction were truly contemporaneous.


of murdering multiple victims were also categorized as death eligible

because multiple victim homicides satisfy the requirements of the B2

statutory aggravating circumstance.162

Death eligibility based on the presence of the B4 circumstance—

“offender committed the offense of murder for himself or another, for the

purpose of receiving money or any other thing of monetary value”—was

determined by a contemporaneous conviction for an economically

motivated crime (i.e., robbery and theft). Admittedly, the three factors

employed to determine death-eligible defendants are narrow and ignore a

host of other factors listed in the statute, but it is worth noting that prior

research strongly suggests that the presence of multiple victims and a

contemporaneous felony are “the most commonly used factors in death

sentence cases, and thus account for a high proportion of death eligible

cases.”163 Consequently, these two types of aggravating circumstances are

also the most common factors used by judges and jurors to justify death

sentences, as well as the strongest predictors of a death sentence, even after

holding other legally relevant factors constant.164 Perhaps more

162 The presence of multiple victims or a contemporaneous felony is commonly

employed by researchers to identify death-eligible cases. See generally SAMUEL R. GROSS &


(1989); Pierce & Radelet, supra note 80, at 72–91 (2002); Glenn L. Pierce & Michael L.

Radelet, The Impact of Legally Inappropriate Factors on Death Sentencing for California

Homicides, 1990–1999, 46 SANTA CLARA L. REV. 1, 21–25 (2005). Not only are these two

types of aggravating circumstances the most common set of aggravating circumstances used

by prosecutors, jurors, and judges to justify death sentences, but the number of victims is

consistently one of the strongest predictors of a death sentence, even after holding other

legally relevant factors constant. Pierce & Radelet, supra note 80, at 72–91; Steven F. Shatz

& Nina Rivkind, The California Death Penalty Scheme: Requiem for Furman?, 72 N.Y.U. L.

REV. 1283, 1328–32 (1997).

To be sure, the multiple victim measure is imperfect because it is possible that a

defendant murdered multiple victims on the same day but in unrelated situations. The vast

majority of cases, however, are single victim (87.3%), so this measure, in and of itself, does

not influence the categorization of most defendants. Perhaps more importantly, the B1 and

B2 criteria are not mutually exclusive—in fact, they share considerable overlap. It is

possible, even likely, then, that a defendant’s categorization as death eligible will be valid on

one or more of these measures.

163 Compare Pierce & Radelet, supra note 80, at 66, with Shatz & Rivkind, supra note

162, at 1329 (“[T]he felony murder special circumstances play the predominant role in

defining death-eligibility.”).

164 Pierce & Radelet, supra note 80, at 61 (noting that juries were most likely to impose

the death sentence in cases involving multiple victims); Shatz & Rivkind, supra note 162, at

1329–30 (explaining three-quarters of death-sentenced cases involved a felony-murder

circumstance); see also Gregg v. Georgia, 428 U.S. 153, 225 (1976) (White, J., concurring)

(“[T]he standards by which [prosecutors] decide whether to charge a capital felony will be

the same as those by which the jury will decide the questions of guilt and sentence.”);

William J. Bowers & Glenn L. Pierce, Arbitrariness and Discrimination Under Post-Furman

512 SHEROD THAXTON [Vol. 103

importantly, data on death eligibility from the more detailed study

conducted by the Atlanta Journal-Constitution overlap with nearly 80% of

the cases in the current sample, so a more “fine-grained” assessment of

statutorily defined elements is available for the vast majority of cases.165

Death eligibility was also limited to situations in which the defendant

was ultimately convicted of murder. Obviously prosecutors seek the death

penalty against defendants prior to obtaining a guilty plea or guilty verdict

at trial, but limiting the pool of death-eligible defendants to those who are

actually convicted of murder serves as a proxy for strength of evidence.

Other scholars have employed this limiting strategy when examining capital

charging-and-sentencing processes.166 Perhaps of equal significance is the

fact that Georgia’s murder statute does not include “degrees” of murder like

many other states. Instead, the statute specifies that individuals may be

charged with malice murder (intentional) or felony murder (unintentional,

but during the commission of any other felony),167 and only these two types

of murder may be death eligible, depending on the presence of at least one

statutory aggravating circumstance. “Lesser” degrees of homicide are

categorized as voluntary and involuntary manslaughter.168 It is very

unlikely that prosecutors will offer a charge bargain from murder to

manslaughter when the available evidence permits a conviction for murder,

which requires a mandatory minimum life sentence.169 Similarly,

Capital Statutes, 26 CRIME & DELINQ. 563, 627 (1980) (discovering that the vast majority of

death sentences imposed in Georgia and Florida involved cases with a felony circumstance).

165 The “crude” measure of death eligibility based on the presence of B1, B2, or B4

special circumstances failed to classify 36% of cases identified by the AJC study (and 16%

of cases noticed for the death penalty). Thirty-nine percent of the cases misclassified were

technically eligible for death based upon the presence of the B7 circumstance.

166 See, e.g., BALDUS ET AL., supra note 84, at 40–42, 477 n.72 (discussing the use of

murder conviction as a proxy for strength of evidence); accord David C. Baldus et al., Racial

Discrimination and the Death Penalty in the Post-Furman Era: An Empirical and Legal

Overview, with Recent Findings from Philadelphia, 83 CORNELL L. REV. 1638, 1668–70

(1998); Paternoster et al., supra note 80.

167 GA. CODE ANN. § 16-5-1(a) (West 2003) (malice murder); id. § 16-5-1(c) (felony


168 Voluntary manslaughter carries a maximum sentence of twenty years and involuntary

manslaughter carries a maximum sentence of ten years. GA. CODE ANN. § 16-5-2 (voluntary

manslaughter); id. § 16-5-3(a) (involuntary manslaughter).

169 COHEN & KYCKELHAHN, supra note 2, at 10–11 (noting that, regardless of method of

adjudication, the vast majority of defendants initially charged with murder are ultimately

convicted of murder). See William J. Stuntz, Plea Bargaining and Criminal Law’s

Disappearing Shadow, 117 HARV. L. REV. 2548, 2563 (2004) (commenting that prosecutors

generally pursue every murder case they can, which is why the acquittal rate in murder cases

is higher than for other violent felonies); see also Bowers, supra note 50, at 1153 (same);

Gazal-Ayal, supra note 15, at 2320 (explaining that dropping or reducing murder charges

can be politically costly for prosecutors and this is part of the reason they dedicate resources


prosecutors are not likely to seek the death penalty in cases in which they

believe a conviction for manslaughter is warranted, even in the presence of

aggravating circumstances that would make the case eligible for the death

penalty if a murder conviction were obtained.170

Focusing exclusively on defendants subsequently convicted of murder

suffers from the drawback of excluding death-noticed defendants who are

not ultimately convicted of murder. In these data, twenty-eight individuals

(7% of death-noticed defendants) were initially noticed for the death

penalty, but later were acquitted, had their charges dismissed, or were

convicted of an offense other than murder. Nevertheless, a closer

inspection of the data supports the previously stated intuition that

prosecutors do not typically offer charge bargains in exchange for guilty

pleas. Of the twelve cases noticed for the death penalty that resulted in a

plea bargain for a charge other than murder, only two were for

manslaughter.171 Thus, the more cautious approach adopted in these

analyses (i.e., underinclusion) in an attempt to most accurately compare

“apples to apples” would appear to outweigh its disadvantages.

The second approach utilized to identify death-eligible defendants was

to define all cases that ultimately resulted in a murder conviction as death

eligible. Because some scholars claim that capital statutes like Georgia’s

permit any homicide to be deemed death eligible,172 this very broad

definition of death eligibility was employed in order to determine whether

the results are robust to the specific criteria used to identify the control



There are, essentially, two reasons why death-noticing and pleabargaining

decisions might be related. First, the decisions may be causally

related.173 Second, death noticing and plea bargaining may be related

to trying even weak murder cases).

170 See Liebman, supra note 31, at 2097–98 (charging a case capitally increases the

chances of winning, but it also increases the embarrassment and publicity of losing); William

J. Stuntz, The Pathological Politics of Criminal Law, 100 MICH. L. REV. 505, 570 (2001)

(explaining that defeats at trial for prosecutors are so vivid because they are rare, so

prosecutors are less likely to pursue cases that are unwinnable).

171 The remaining plea bargains to nonhomicide charges were for aggravated assault,

armed robbery, burglary, concealing the death of another, and kidnapping.

172 See, e.g., John Blume & Theodore Eisenberg, Judicial Politics, Death Penalty

Appeals, and Case Selection: An Empirical Study, 72 S. CAL. L. REV. 465 (1999); Rosen,

supra note 153.

173 Death noticing typically precedes plea negotiations. In fact, it is common for

prosecutors to announce the intent to seek the death penalty before having obtained an

indictment or meeting with the defendant. It remains plausible, however, that some

514 SHEROD THAXTON [Vol. 103

because they are caused by other shared factors (i.e., “third-variables”) that

may or may not be observed, and if these factors are taken into account, the

relationship between death noticing and plea bargaining disappears; that is,

the relationship between death noticing and plea bargaining is not

independent of those other factors.174

In an effort to determine the impact of the death penalty on the

decision to go to trial, a conditional fixed-effects logit model is estimated

according to the following equation:

Pr(Trial𝑖𝑐𝑡 = 1) = exp (𝛼𝑐+𝛾𝑘𝑋𝑖𝑐𝑡+𝛽×DPNotice𝑖𝑐𝑡 + 𝜁𝑖𝑐𝑡)

1+exp (𝛼𝑐+𝛾𝑘𝑋𝑖𝑐𝑡+𝛽×DPNotice𝑖𝑐𝑡 + 𝜁𝑖𝑐𝑡)

, [1]

where i indexes the defendant, c indexes the judicial circuit, and t indexes

the year.175 In all of the analyses employed, Xict is a matrix of case

characteristics, including, inter alia, t – 1 incident-year dummies, and

DPNotice is a dummy (binary) variable indicating whether the prosecutor

sought the death penalty against the defendant. The model makes the

following assumptions: (a) conditional on Xict and DPNotice, Trialict is an

independent Bernoulli random variable176 with probability given by

Equation [1]; (b) Pr(Trialict = 1) depends on Xict and DPNotice through the

logistic function; (c) Pr(Trialict = 1) is governed by parameters γ and β,177 a

prosecutors elect to file a death notice following an initial failed attempt to obtain a plea

bargain, but ultimately withdraw the death notice after a defendant agrees to a negotiated

plea. This potential complication, known as simultaneous causation, is addressed more fully

in Part VI.B.

174 For a detailed discussion of the key requirements of causal inference, see RICHARD A.


emphasize that researchers need not control for every conceivable variable possibly

influencing plea bargaining. Candidate variables must meet three conditions: (1) correlated

with the key causal variable (i.e., death-noticing decision); (2) affects the outcome variable

(i.e., plea-bargaining decision); and (3) causally prior to the key causal variable. Lee Epstein

& Gary King, The Rules of Inference, 69 U. CHI. L. REV. 1, 78 (2002). If one of these three

conditions is absent, then controlling for the rival variable is not only unnecessary when

examining the causal impact of the key variable of interest, but it may also lead to incorrect

inferences if the variable is a consequence of the key causal variable (i.e., “post-treatment”


INFERENCE IN QUALITATIVE RESEARCH 78 (1994) (controlling for a consequence of the cause

produces the incorrect causal effect).


176 A random variable, Y, that can only take on two values, 0 and 1, with Pr(Y=1) = p is a

Bernoulli random variable with parameter p. This variable has a mean of p and a variance of


(4th ed. 2012).

177 More precisely, the γ and β coefficients represent the expected change in the

probability of trial corresponding to changes in each predictor in the model (i.e., X and


unit-specific (e.g., jurisdiction-specific) parameter αc, and (d) ζ is a vector

of residual error terms with a mean of zero and variance of π2/3.178

The fixed-effects specification is particularly advisable with these data

because the models control for unobserved heterogeneity across judicial

circuits and years.179 Failing to account for these fixed effects can result in

omitted variable bias and lead to inconsistent estimates of a hypothesized

causal effect.180 Georgia’s Administrative Office of the Courts (AOC)

DPNotice, respectively). Gelman et al., supra note 40, at 239. There are a total of k + 1

parameters estimated in the model.

178 GREENE, supra note 175, at 70. I employ the conditional fixed-effects estimator, id. at

839, due to the bias introduced by using the unconditional fixed effects (indicator set) when

cluster sizes are relatively small. See Tom Coupé, Bias in Conditional and Unconditional

Fixed Effects Logit Estimation: A Correction, 13 POL. ANALYSIS 292, 295 (2005); Ethan

Katz, Bias in Conditional and Unconditional Fixed Effects Logit Estimation, 9 POL.

ANALYSIS 379, 384 (2001). An acceptable probit alternative specification does not exist

because there is no sufficient statistic that allows the fixed effects to be conditioned out of

the likelihood function. William Greene, The Behaviour of the Maximum Likelihood

Estimator of Limited Dependent Variable Models in the Presence of Fixed Effects, 7

ECONOMETRICS J. 98, 102–03 (2004). I also reanalyzed the data with the unconditional

estimator and obtained similar results.

Alternative specifications were analyzed employing a random-effects estimator (i.e.,



MODELS 49–50 (2004). Random effects models rely on the strong assumption that the

unobserved cluster-specific influences are uncorrelated with individual-level case

characteristics, so fixed-effects models are preferred when that assumption is unrealistic.

GREENE, supra note 175, at 576–77. It is possible, however, to allow the random effect (i.e.,

intercepts) to be correlated with the individual-level variables by creating an aggregated

measure of the individual variables for each group and including that measure as a predictor

in the model. Any covariance between an individual-level predictor and a group-level

random effect must operate through the covariance between the group-level average of the

individual-level predictor and that random effect; therefore, inclusion of the group-level

average of the predictor as a covariate in the model will eliminate any confounding between

the individual-level predictor and omitted variables at the group level. STEPHEN W.




(2007). Results (not reported) from these models were very similar to those obtained from

the fixed-effects specifications.

179 The variance of the jurisdiction-level effects tells us the extent to which there is

variability among jurisdictions in the data beyond that explained by the other regression

predictors. Similarly, the year effects represent unexplained variation among years. Gelman

et al., supra note 40, at 238 n.84; see also Thomas R. Ten Have et al., Deviations from the

Population-Averaged Versus Cluster-Specific Relationship for Clustered Binary Data, 13

STAT. METHODS MED. RES. 3, 9 (2004) (“In the binary response case, conditional likelihood

estimation is the only approach that is less susceptible [to confounding of treatment effect

due to cluster-level unobserved heterogeneity], because it conditions out all cluster-level

information that may confound within-cluster effects.”).

180 GREENE, supra note 175, at 839–40.

516 SHEROD THAXTON [Vol. 103

organizes the state’s 159 counties into forty-nine superior court judicial

circuits.181 As a result, county-level data are nested in the judicial circuits

for the analyses conducted in this paper.182 The distribution of death

notices by county and judicial circuit are presented in, respectively, Tables

1 and 2. The specific variables used, coding conventions adopted, and

summary statistics can be found in Appendix A and Tables 3, 4, and 5.

Although the model controls for average differences across

jurisdictions and average differences across years, it does not take into

account omitted covariates that are case specific.183 Prosecutors do not

randomly select cases for the death penalty, so the estimation of the true

impact of capital punishment on trials may be biased and inconsistent if the

death-noticing decision is correlated with some other unobserved

variable(s) that also has a causal impact on the decision to go to trial (i.e.,

endogeneity bias). Relatedly, if a death-noticing decision is also influenced

by a prior failed plea negotiation (i.e., simultaneous causality), death

noticing is also endogenous because it will be correlated with an

unobserved variable influencing both decisions: preliminary plea

negotiation. The fixed-effects logit model discussed earlier implicitly

assumes that all factors simultaneously influencing the death noticing and

trial decisions are included in the model via the covariates and circuit and

year fixed effects. This may be a questionable assumption, so several

alternative models that explicitly account for confounding omitted variables

at the case level are examined in Part VI.B.


182 This is necessary for two important reasons. First, in Georgia, there is one district

attorney per judicial circuit. While large counties comprise a single judicial circuit, many

smaller counties are grouped together to form a single judicial circuit. As a result, a single

prosecutor may be responsible for charging and plea-bargain decisions for several counties

in her judicial circuit. Also, if a judicial circuit consists of multiple counties, trial judges

rotate throughout these counties in the circuit. Treating counties that share a single judicial

circuit as if they were independent ignores the similarities they share in the administration of

capital punishment resulting from shared decisionmakers. Second, death penalty cases are

extremely rare events, so aggregating county-level data to the judicial-circuit level allows

one to observe more cases per contextual unit and better statistically estimate relationships

occurring at both the case- and contextual-level without altering the dependence structure of

the cases due to their clustering. RAUDENBUSH & BRYK, supra note 178, at 45 (noting that

aggregating data at the highest level of nesting preserves the dependence structure of the

units); Gary King & Langche Zeng, Logistic Regression in Rare Events Data, 9 POL.

ANALYSIS 137, 163 (2001) (discussing the difficulties associated with analyzing rare events

in binary data).

183 When the death-noticing decision is correlated with the case-specific error term rather

than just the judicial-circuit or year effect, a simultaneous equation approach is necessary.

John Antonakis et al., On Making Causal Claims: A Review and Recommendations, 21

LEADERSHIP Q. 1086, 1092 (2010).


Table 3

Summary Statistics (DE Subsample)

Table 4

Summary Statistics (MC Subsample)

518 SHEROD THAXTON [Vol. 103

Table 5


Summary Statistics (Death-Noticed Subsample)


An additional complication with analyzing official homicide records is

incomplete information (i.e., missing data).184 The vast majority of

statistical analyses must be performed on a full data matrix; therefore, the

common practice among social scientists is to perform casewise deletion by

eliminating observations that have missing data on one or more variables.185

Casewise deletion is problematic because it (1) potentially forces

researchers to discard much useful information about the relationships

between variables, (2) results in inefficient parameter estimates due to a

184 See generally Wendy C. Regoeczi & Marc Riedel, The Application of Missing Data

Estimation Models to the Problem of Unknown Victim/Offender Relationships in Homicide

Cases, 19 J. QUANTITATIVE CRIMINOLOGY 155 (2003) (suggesting ways to deal with missing

data problems when analyzing official homicide data).

185 Roderick J. A. Little, Regression with Missing X’s: A Review, 87 J. AM. STAT. ASS’N

1227, 1228 (1992).

520 SHEROD THAXTON [Vol. 103

reduction of sample size, and (3) may bias parameter estimates if the data

are not missing completely at random (i.e., if the missing data are not a

random subset of the overall population).186 There is missing data on at

least one variable in approximately one-third of the observations in the

Georgia data; however, when it is possible to predict the probability that a

variable is missing information for an observation (using information from

other covariates in the data), the most appropriate strategy is to attempt to

predict those missing values.187 This is particularly true when examining

the capital punishment process because death penalty cases in Georgia (and

elsewhere) are very rare occurrences, so it is crucial to retain as much

information as possible.

Over the past two decades, quantitative methodologists have

developed several approaches to “guess” the values of missing data by

using information about the association of the variable of interest with other

variables in the data. A regression-based multiple imputation approach is

employed in these analyses, which provides a significant improvement over

simple imputation methods and traditional single imputation strategies.188

In brief, Stef van Buuren and colleagues’ “fully conditional specification”

(FCS) approach is used because it offers the greatest flexibility in creating

multivariate imputation models by allowing for specialized methods that are

impractical under the other approaches.189 The FCS approach imputes the

data on a variable-by-variable basis by specifying an imputation model for

each variable, thereby allowing the analyst to preserve unique features of

the data such as bounds, skip patterns, interactions, and bracketed

responses, and to incorporate appropriate constraints between different

variables in order to avoid logical inconsistencies in the imputed data.190

A somewhat simplified description of the algorithm is that observed

data are used to impute missing values and incorporate estimation

uncertainty (resulting from analyzing a finite number of observations) and

186 Id. at 1227.

187 Id. at 1230, 1233.

188 Id. at 1234–35.

189 Stef van Buuren, Multiple Imputation of Discrete and Continuous Data by Fully

Conditional Specification, 16 STAT. METHODS MED. RES. 219, 219 (2007).

190 Id. at 219, 222. The statistical properties of FCS are not fully understood, but

simulation studies suggest that FCS performs well in a variety of applications. S. van

Buuren et al., Fully Conditional Specification in Multivariate Imputation, 76 J. STAT.

COMPUTATION & SIMULATION 1049, 1061 (2006); Trivellore E. Raghunathan et al., A

Multivariate Technique for Multiply Imputing Missing Values Using A Sequence of

Regression Models, 27 SURV. METHODOLOGY 85, 92–93 (2001) More importantly, when

there are missing variables that follow a mixture of distributions (e.g., continuous, ordinal,

categorical), FCS is the only sensible parametric approach. Van Buuren et al., supra, at



fundamental uncertainty (resulting from unmodeled variation in the

dependent variable and represented by the stochastic component of the

model) in their prediction of plausible values. For the present study, this

process was repeated five times to create five complete data sets, with each

data set containing different plausible values for missing variables to

account for the uncertainty surrounding the imputations. After these data

sets were created, a complete-case analysis was repeated on each data set.

The overall point estimate of each parameter was obtained by averaging

across the five separate point estimates for that particular parameter.191 The

variance of the point estimate was computed by averaging across the five

estimated variances from within each completed data set, plus the sample

variance in the point estimate across the data sets (multiplied by a factor

that corrects for bias because the number of imputed data sets is finite).192

The next section presents results from models using both casewise

deletion and multiple imputation strategies.



Table 6 displays results from four different specifications.193 Across

all four models, defendants noticed for the death penalty were significantly

less likely to opt for trial (i.e., significantly more likely to accept a plea).194

Models 1 and 3 analyze the impact of the death penalty on the probability

of going to trial for defendants classified as death eligible according to the

first criteria discussed: eligibility based on the presence of statutorily

defined elements (hereinafter, “DE”). Models 2 and 4 analyze the impact of

the death penalty on defendants classified as death eligible based on the

191 Little, supra note 185, at 1235.

192 Id.

193 Binary regression models in this Article report Cragg and Uhler’s pseudo-R2 statistic,

defined as:

1 exp(2×[𝐿𝐿null 𝐿𝐿full]/𝑁)

1 exp(2×[𝐿𝐿null 𝐿𝐿max]/𝑁)


where LL0 is the log-likelihood for the null model (i.e., constant-only model), LL1 is the full

regression model, LLmax is the maximum possible likelihood (i.e., perfect fit), and N is the

sample size. See John G. Cragg & Russell S. Uhler, The Demand for Automobiles, 3 CAN. J.

ECON. 386, 400 n.20 (1970). The Cragg and Uhler pseudo-R2 statistic is most analogous to

the traditional R2 statistic used in ordinary least squares (OLS) regressions because, unlike

most other pseudo-R2 statistics, it is “normed” so the upward bound approaches unity. See J.


106 (1997).

194 Recall that the models estimate the probability of a defendant taking her case to trial,

so a negative sign on the DPNotice coefficient indicates that defendants noticed for the death

penalty are less likely to have their cases resolved by trial, which is equivalent to being more

likely to have their cases resolved by plea agreement.

522 SHEROD THAXTON [Vol. 103

second criteria: murder conviction (hereinafter, “MC”). All specifications

include judicial-circuit and year fixed effects, as well as controls for the

number of codefendants; defendant’s race/ethnicity, gender, and age;

defendant’s employment status at time of arrest; defendant’s marital status;

number of statutory aggravating factors; contemporaneous felony

conviction; prior felony conviction; whether defendant graduated from high

school; number of children defendant has; number of victims; victim

race/ethnicity, gender, and age; relationship between the victim and

offender; and whether the homicide was interracial.195

The natural coefficients from the logistic regression model, “logit

coefficients or log odds,” lack an intuitive interpretation, so marginal effects

are presented. The marginal effect represents the change in the probability

of a case being disposed by trial, holding all other variables constant.196

Model 1 (DE) and Model 2 (MC) reveal that being noticed for the death

penalty reduces the probability of a defendant taking her case to trial by,

respectively, .23 and .22. For Model 1, this means the probability of deathnoticed

defendants going to trial is .37, whereas the probability of

defendants not noticed for death going to trial is .60, all else equal (see

Figure 1). For Model 2, the probabilities are .41 and .63, respectively.

Figure 1

Method of Disposition

195 See infra Appendix A for a detailed description of these variables and Tables 3, 4,

and 5 for summary statistics. Appendix B presents the marginal effects for all of the

covariates in the fixed-effects logit model. Model 1 lists the point estimates for covariates

predicting the probability of a case being disposed by trial. Model 2 displays the effects of

the same covariates on the probability a case is noticed for the death penalty.

196 LONG, supra note 193, at 71–74. The conditional fixed-effects model does not

provide estimates of the judicial-circuit fixed effects, αc, which are needed to compute

marginal effects. Coupé, supra note 178, at 292. Thus, marginal effects were obtained using

the unconditional fixed-effects logit model. The conditional and unconditional fixed-effects

estimates are essentially the same when cluster sizes average at least eight, and the average

cluster sizes are twenty and twenty-seven in, respectively, the DE and MC models. Id. The

marginal effects obtained from the unconditional fixed-effects linear model were nearly

identical, suggesting that the results are quite robust to various model specifications. See

infra Part VI.B.

Population-average effects (sometimes referred to as marginal effects) were also

obtained, and these estimates were similar to the conditional and unconditional fixed-effects

estimates. The unconditional fixed-effects estimates represent the difference in the

probability of trial, depending on being noticed for the death penalty, for the same defendant.

The population-average effect, on the other hand, represents the difference in probability of

trial of the average defendant noticed for the death penalty versus the average defendant not

noticed for the death penalty—that is, the estimates do not control unobserved circuit-level

effects. RAUDENBUSH & BRYK, supra note 178, at 304, 334.


As noted, supra, approximately one-third of the cases in the data have

missing information on at least one variable. Specifically, the DE models

lose 31.2% of the cases and the MC models lose 35.9% of the cases. Table

6 presents results from the multiple imputation models. As with the

casewise deletion results previously reported, the coefficient for DPNotice

is statistically significant across all specifications, although coefficients are

slightly smaller with respect to their absolute value. In the Model 3 (DE),

being noticed for the death penalty decreases the probability of trial by .18,

and by .17 in Model 3 (MC). The baseline probabilities for the DE (Model

3) and MC (Model 4) specifications are very similar to the casewise

deletion models (.60 and .64, respectively).

Table 6

Marginal Effect of Death Penalty on Trial (Fixed-Effects Logit Models)









DP Case Non-DP Case


Plea Trial

524 SHEROD THAXTON [Vol. 103

Model 1 Model 2 Model 3 Model 4

DP Notice

-0.234*** -0.218***

-0.184*** -0.168***

(0.038) (0.038)

(0.032) (0.031)


0.603 0.628

0.603 0.635


847 1238

1236 1932




Year Dummies



R-Squared 0.32 0.28

*p<.05; **p<.01; ***p<.001

Standard errors adjusted for clustering on judicial circuit.

Models 1 & 3: Death Eligible Subsample.

Models 2 & 4: Murder Conviction Subsample.

Models 3 & 4: Multiple Imputation Estimates.

Pr(Trial) = Probability of trial without a death noticed filed.

All specifications include controls for the number of codefendants, victims,

statutory aggravating factors, contemporaneous felony convictions, prior

felony convictions, and children of defendant; the race/ethnicity, gender, and

age of defendant and victim(s); defendant’s employment status at time of

arrest, marital status, high school graduation status; offender/victim

relationship; and whether the homicide was interracial.


Table 7 presents the linear (unconditional) fixed-effects estimates for

the DE and MC models using casewise deletion (Models 1 and 2) and

multiple imputation (Models 3 and 4). The classic fixed-effects linear

model takes the form:

𝐸(Trial) = Pr(Trial = 1) = α𝑐 +𝛾𝑋 +𝛽 ×DPNotice+ζ [2]

where α, 𝛾, β, ζ, X, and DPNotice are defined in Equation [1], but ζ are now

identically and independently normally distributed: ζ~N(0,σ2).197 Linear

regression models applied to binary dependent variables are commonly

referred to as linear probability models.198 These models are generally

deemed inappropriate for binary data because of heteroscedasticity,

197 See GREENE, supra note 175, at 560. Some subscripts have been omitted for


198 LONG, supra note 193, at 35.


nonnormality, nonsensical predictions, and functional form

misspecification.199 These results are presented simply as a robustness

check for the conditional fixed-effects logit estimates presented in Table 6.

Unconditional fixed-effects models are most appropriate in the linear

regression context, so the purpose of showing these estimates is to check

that the DPNotice coefficients are similar in terms of direction, magnitude,

and statistical significance.200 The interpretation of coefficients from the

linear probability is similar to the linear regression model, so according to

Model 1 (DE and casewise deletion), for example, being noticed for the

death penalty decreases the probability of going to trial by .22 from a

baseline probability of .62, holding all other variables constant.

Table 7

Marginal Effects of Death Penalty on Trial (Fixed-Effects LPMs)

Model 1 Model 2 Model 3 Model 4

199 Id. at 38–40.

200 GREENE, supra note 175, and accompanying text; accord David S. Abrams & Albert

H. Yoon, The Luck of the Draw: Using Random Case Assignment to Investigate Attorney

Ability, 74 U. CHI. L. REV. 1145, 1168 (2007) (employing a linear model for a binary

outcome, rather than a logit or probit model, in order to estimate unconditional fixed effects).

526 SHEROD THAXTON [Vol. 103

DP Notice -0.217*** -0.212*** -0.186*** -0.173***

(0.048) (0.044) (0.036) (0.034)

Pr(Trial) 0.618 0.645 0.622 0.658

N 852 1238 1238 1932

Fixed-Effects Y Y Y Y

Year Dummies Y Y Y Y


0.18 0.16

*p<.05; **p<.01; ***p<.001

Standard errors adjusted for clustering on judicial circuit.

Models 1 & 3: Death Eligible Subsample.

Models 2 & 4: Murder Conviction Subsample.

Models 3 & 4: Multiple Imputation Estimates.

Pr(Trial) = Probability of trial without a death noticed filed.

All specifications include controls for the number of codefendants, victims,

statutory aggravating factors, contemporaneous felony convictions, prior

felony convictions, and children of defendant; the race/ethnicity, gender, and

age of defendant and victim(s); defendant’s employment status at time of

arrest, marital status, high school graduation status; offender/victim

relationship; and whether the homicide was interracial.

As noted supra, another concern with the models estimated in this

study is possible endogeneity bias201 resulting from either case-level

201 Endogeneity occurs when “the values our explanatory variables take on are

sometimes a consequence, rather than a cause, of our dependent variable.” KING ET AL.,

supra note 174, at 185.

The level of concern over endogeneity bias in observational studies varies across

disciplines—e.g., econometricians tend to be much more concerned about endogeneity than,

for example, epidemiologists, biostatisticians, psychologists, statisticians, and other social

scientists. See Antonakis et al., supra note 183, at 1100 (remarking that attention to

remedying possible endogeneity bias “has not had a big impact on other social science

disciplines including psychology and management research”); Robert Gibbons, What Is

Economic Sociology and Should Any Economists Care?, 19 J. ECON. PERSP. 3, 6 (2005);

James J. Heckman, The Scientific Model of Causality, 35 SOC. METHODOLOGY 1, 5 (2005)

(noting that epidemiological and statistical models often fail to take into account

simultaneous causality and other sources of randomness generating unobservables in their

models); S. Rabe-Hesketh & A. Skrondal, Parameterization of Multivariate Random Effects

Models for Categorical Data, 57 BIOMETRICS 1256, 1256 (2001) (explaining that

econometricians have given greater attention to identification problems than biostatisticians).


omitted variable bias or possible reverse causality.202 Due to the fact that in

nonexperimental research, predictor and outcome variables may covary

because of factors outside the control (and knowledge) of the researcher,203

standard regression techniques will result in biased and inconsistent

estimators when unobserved factors affecting the response are correlated

with unobserved factors affecting the causal variable of interest.204 While

the problems of omitted variable bias and reverse causality may be

theoretically distinct, they result in the same source of bias—correlation

between the causal variable and the unobserved factors affecting the

response variable205—so similar corrective approaches can be used to

address both forms of this potential bias. Three alternative approaches were

employed to examine the robustness of the previously reported results.

First, a nonrecursive simultaneous-equation model in which a

dependent variable indicating selection into a treatment group (i.e.,

DPNotice) appears as an explanatory variable in a substantive equation

predicting the outcome (i.e., Trial).206 “[M]ultiple equation models have

202 KING ET AL., supra note 174, at 185. Measurement error is a third source of

endogeneity bias, but is not of particular concern in these analyses.

203 Id. at 186.

204 James J. Heckman, Dummy Endogenous Variables in a Simultaneous Equation

System, 46 ECONOMETRICA 931, 931 (1978).

205 KING ET AL., supra note 174, at 185.

206 Simultaneous equation models can be divided into two major types: recursive and

nonrecursive. A nonrecursive model occurs when there are reciprocal relationships (i.e.,

feedback loops) between the outcome variables of two or more equations in the system or at

least some of the disturbances are correlated. PAMELA M. PAXTON ET AL., NONRECURSIVE



16–17 (2000) (noting that nonrecursive models have non-zero off-diagonal elements in the

residual variance–covariance matrix); Rex B. Kline, Reverse Arrow Dynamics: Formative

Measurement and Feedback Loops, in STRUCTURAL EQUATION MODELING: A SECOND

COURSE 43, 56 (Gregory R. Hancock & Ralph O. Mueller eds., 2006) (“Nonrecursive models

have feedback loops or disturbance covariances for endogenous variables with direct effects

between them.”). Most econometricians, however, refer to models with correlated

disturbances as recursive models if there are no feedback loops present. E.g., GREENE, supra

note 174, at 659 (explaining that a model is recursive when the matrix of coefficients of the

endogenous variables is triangular); accord Joachim Wilde, Identification of Multiple

Equation Probit Models with Endogenous Dummy Regressors, 69 ECON. LETTERS 309, 310

(2000). For the purposes of these analyses, the simultaneous models are labeled

nonrecursive to maintain consistency with the larger structural equation modeling literature.

Labeling the simultaneous models nonrecursive also underscores the fact that the models

take into account possible endogeneity bias resulting from reciprocal causation.

Bivariate logistic models were also estimated and produced nearly identical results.

Bivariate probit models are generally preferred in the literature because the various extant

multivariate logistic distributions have properties such as restrictions on possible values of

correlation coefficients and asymmetric nonelliptical distributions that make such a direct

528 SHEROD THAXTON [Vol. 103

been the key tools for many researchers [to] study complicated cause-andeffect

relationships. . . . The regression equations are explicitly meant to

represent the mechanisms by which causes have their effects.”207 This

model is typically referred to as an endogenous bivariate probit or

endogenous switching model due to the fact that DPNotice is a binary

variable and the observational units (i.e., defendants) are allocated to a

specific regime (i.e., death noticed/non-death noticed) depending on the

value of this decision variable.208 The model is estimated from the

following equation:

𝜋𝑗𝑘 = Φ2[𝑑2(γ2𝑋2 + β2 × DPNotice + λξ +δ2) + 𝑑1(𝛾1𝑋1 + ξ + δ1)], [3]

where πjk can represent four different joint probabilities, depending on the

values of the Trial and DPNotice variables (j indexes the binary outcome

for Trial and k indexes the binary outcome for DPNotice).209 So, for

example, π11 = P r(Trial = 1, D PNotice = 1), π10 = Pr(Trial = 1, DPNotice =

0), etc., and d1 and d2 are signs variables, being equal to 1 or –1 depending

on whether the observed binary outcome equals 1 or 0.210 The parameters γ

and β, as well as the variables X and DPNotice, are defined the same as in

Equation [1],211 delta (δ) is the error term that is unique to each equation,

approach less practical, and convergence problems are common. Barry C. Arnold,

Multivariate Logistic Distributions, in HANDBOOK OF THE LOGISTIC DISTRIBUTION 237, 244–

45 (N. Balakrishnan ed., 1992); Murray D. Smith & Peter G. Moffatt, Fisher’s Information

on the Correlation Coefficient in Bivariate Logistic Models, 41 AUSTL. & N.Z. J. STAT. 315,

317–19 (1999).

207 BERK, supra note 174, at 190; accord ERIC A. HANUSHEK & JOHN E. JACKSON,

STATISTICAL METHODS FOR SOCIAL SCIENTISTS 227 (1977) (“[S]tructural [i.e., multiple]

equations represent the theoretical model hypothesized to underlie the observed data; this is

the causal structure assumed to generate the data.”).

208 SKRONDAL & RABE-HESKETH, supra note 178, at 437. The endogenous switching

model with a binary outcome is also known as the “multivariate probit model with structural

shift.” Heckman, supra note 204, at 932.

209 Note that some subscripts have been omitted to simplify the presentation. The

standard representation of simultaneous equation models lists β as the effect of endogenous

variables and γ as the effect of exogenous variables. See PAXTON ET AL., supra note 206,

at 4.

210 Alfonso Miranda & Sophia Rabe-Hesketh, Maximum Likelihood Estimation of

Endogenous Switching and Sample Selection Models for Binary, Ordinal, and Count

Variables, 6 STATA J. 285, 288 (2006); accord Lorenzo Cappellari & Stephen P. Jenkins,

Calculation of Multivariate Normal Probabilities by Simulation, with Applications to

Maximum Simulated Likelihood Estimation, 6 STATA J. 156, 166 (2006). Technically, dm =

2ym – 1, where m indexes the particular equation (m = 1, 2). So when y = 1, dm = 1 and when

y = 0, dm = –1. Id.

211 GREENE, supra note 175, at 852–56 (noting that the endogenous nature of an

explanatory variable can be ignored in formulating the log-likelihood in the bivariate probit

model). Including observed endogenous variables in a system of probit equations yields


and Φ2 is the bivariate normal cumulative distribution function (CDF).212

Xi (ξ) is a second-order latent variable—i.e., a latent variable whose

indicators are themselves latent variables213—and “merely represents the

combined effect of all unobserved covariates.”214 The inclusion of latent

variables “in statistical models is a common way of taking unobserved

heterogeneity into account.”215 Lambda (λ) is a factor loading, representing

the covariance between the disturbances (i.e., the covariance between the

omitted variables after the influence of the included factors are accounted

for). Due to the standardized parameterization of the model—i.e., the

variables ξ and δ are normally distributed with mean equal to zero and

variance equal to one—the covariance (λ) is equal to the correlation, rho

(ρ).216 The bivariate probit model assumes that unobserved factors

influencing the treatment variable (i.e., death penalty notice) and the

outcome variable (i.e., case disposed by trial) manifest themselves in the

correlation of the error terms of the two equations.217

likelihoods whose maximization generates consistent parameter estimates. G.S. MADDALA,


212 GREENE, supra note 175, at 849–52, 854.

213 The latent indicators are the “first-order” factors and “may be found to satisfy a factor


VARIABLES 313 (1989) (“Less widely appreciated is that more general and abstract latent

variables may determine the ‘first-order’ latent variables.”); SKRONDAL & RABE-HESKETH,

supra note 178, at 18 (“[L]atent variables pervade modern statistics and . . . are used to

represent widely different phenomena such as true variables measured with error,

hypothetical constructs, unobserved heterogeneity, missing data, counterfactuals and latent

responses underlying categorical variables.”); David W. Gerbing & James C. Anderson, On

the Meaning of Within-Factor Correlated Measurement Errors, 11 J. CONSUMER RES. 572,

574 (1984); Anders Skrondal & Sophia Rabe-Hesketh, Latent Variable Modelling: A Survey,

34 SCANDINAVIAN J. STAT. 712, 712 (2007) (“[L]atent variables are referred to by different

names in different parts of statistics, examples including ‘random effects’, ‘common factors’,

‘latent classes’, ‘underlying variables’ and ‘frailties’.”).

214 SKRONDAL & RABE-HESKETH, supra note 178, at 9; Heckman, supra note 204, at 935

(“[T]he error term in each equation consists of the sum of continuous and discrete random

variables that are correlated.”).

Second-order factor models have at least three distinct advantages: (1) permit the

testing of whether the hypothesized factor accounts for the relationships between the firstorder

factors (i.e., the residuals); (2) impose a structure on the pattern of the covariance

between the first-order factors; and (3) separate the variance due to specific factors from

measurement error. Fang Fang Chen et al., Testing Measurement Invariance of Second-

Order Factor Models, 12 STRUCTURAL EQUATION MODELING 471, 473 (2005).

215 SKRONDAL & RABE-HESKETH, supra note 178, at 9; see also Francesca Francavilla et

al., Mothers’ Employment and Their Children’s Schooling: A Joint Multilevel Analysis for

India, 41 WORLD DEV. 183, 186 (2012) (“Systems of random effects [i.e., latent variables]

equations have been used to deal with endogenous covariates . . . . In such cases the

outcome of an equation appears as a covariate in another equation.”).

216 Miranda & Rabe-Hesketh, supra note 210, at 287–88.

217 Robert C. Luskin, Estimating and Interpreting Correlations Between Disturbances

530 SHEROD THAXTON [Vol. 103

In contrast to Equation [1], the model now represents a system of

equations, so the numbered subscripts refer to the different equations (e.g.,

X1 and X2 index the explanatory variables for, respectively, the deathnoticing

and trial-decision equations), where the main response (i.e., Trial)

and the switching dummy (i.e., DPNotice) are nested (i.e., clustered) within

cases.218 This is easily shown by writing the two equations separately:

Pr(DPNotice = 1) = Φ(γ1𝑋1 +ζ1 ) [4a]

Pr(Trial = 1) = Φ(γ2𝑋2+β2 ×DPNotice+ζ2). [4b]

The relationship between the errors across the two equations can be

observed with the following equations:

ζ1 = ξ +δ1 [5a]

ζ2 = λξ + δ2, [5b]

where the error in each equation consists of a part that is unique to that

equation, δ, and a second part, ξ, that is common to both.219 Each error term

(ζ) now depends, in part, on the value of ξ, which in turn means that ζ1 and

ζ2 will be related to one another.220

and Residual Path Coefficients in Nonrecursive (and Recursive) Causal Models, 22 AM. J.

POL. SCI. 444, 450 (1978) (“Stated simply, the correlation between the disturbances of the []

structural equations expresses the extent to which those equations fail to recognize major

causes of their dependent variables that are either the same or correlated.”). These

disturbances represent effects of random influences or omitted covariates that are casespecific

and immediate, whereas the fixed-effects models represent the effects of omitted

influences that remain constant within a particular jurisdiction or year. GREENE, supra note

175, at 852–56 (explaining that the key advantage of the bivariate probit model is its ability

to explicitly control for unobservable confounding factors); SKRONDAL & RABE-HESKETH,

supra note 178, at 87. This approach shares similarities with propensity score adjustment,

which has been used in prior research to estimate the causal impact of filing a notice to seek

the death penalty on associated costs. Roman et al., supra note 27, at 556 (noting that

propensity score models are a viable solution to modeling selection bias in models with

binary treatments); see also Paul R. Rosenbaum & Donald B. Rubin, The Central Role of the

Propensity Score in Observational Studies for Causal Effects, 70 BIOMETRIKA 41, 55 (1983).

In fact, sensitivity analyses reveal that the estimates from propensity score models were very

similar. See infra Part VI.B.

218 Miranda & Rabe-Hesketh, supra note 210, at 288.

219 BOLLEN, supra note 213, at 314; SKRONDAL & RABE-HESKETH, supra note 178, at 91;

Gerbing & Anderson, supra note 213, at 574–76.

220 “We can induce dependence between responses by including factor structures [i.e.,

latent variables] in the linear predictor.” SKRONDAL & RABE-HESKETH, supra note 178, at



Substituting [5a] into [4a] and [5b] into [4b] gives the following


Pr(DPNotice = 1) = Φ(γ1𝑋1 +ξ + δ1) [6a]

Pr(Trial = 1) = Φ(γ2𝑋2+β2 × DPNotice+λξ + δ2), [6b]

and combining [6a] and [6b] results in Equation [3].221 The relationship

between the two decision processes may be more easily observed in the

(causal) path diagram of the model depicted in Figure 2. The bivariate

probit model takes into account any unobserved causes influencing both

decision processes through ξ, including potential simultaneous/reverse

causality,222 so β2 can be considered the unbiased causal effect of the threat

of the death penalty (via the filing of a notice of intent to seek the death

penalty) on the probability of a case being disposed by trial.223

221 Technically, the combination of Equations [6a] and [6b] results in a specific instance

of Equation [3]: when πjk = π11, therefore, Equation [3] represents a more general

formulation because it can estimate all four of the joint probabilities.

222 The model depicted in Figure 2 does take into account both omitted variables and

potential reverse causality, even though it does not estimate a causal relationship from the

plea-decision variable to the death-noticing variable. Although the statistical literature is

replete with incorrect examples of feedback arrows between two endogenous variables in a

system of simultaneous equations, as well as correlated disturbances, such a formulation has

been shown to be logically inconsistent in the bivariate probit context. The correlated

disturbances properly adjust for reciprocal causation. See Peter Schmidt, Constraints on the

Parameters in Simultaneous Tobit and Probit Models, in STRUCTURAL ANALYSIS OF


McFadden eds., 1981); see also MADDALA, supra note 211, at 117–18 (same).

223 BOLLEN, supra note 213, at 314 (explaining that second-order factors can eliminate

bias resulting from correlated measurement errors). The model is fitted via maximum

likelihood and the unobserved heterogeneity term, ξ, is integrated out using either ordinary

Gauss–Hermite quadrature or adaptive quadrature. Miranda & Rabe-Hesketh, supra note

210, at 288.

532 SHEROD THAXTON [Vol. 103

Figure 2

Path Diagram of Model Predicting Trial

The two equations share identical explanatory variables except for

DPNotice, which only appears in the trial equation. No exclusion

restrictions are required to identify the model—the multi-equation probit

model is identified as long as each equation contains one varying

exogenous variable.224 In fact, it is unlikely that a valid “instrument”

exists—that is, a variable that induces substantial variation in the

endogenous covariate is independent of unmeasured confounders and has

no direct effect on the outcome.225 Researchers have recognized that

224 SKRONDAL & RABE-HESKETH, supra note 178, at 438 (“Although beneficial for

identification, the [exclusion] restrictions are not necessary for identification.”); Heckman,

supra note 204, at 957 (remarking that “the restriction to exclusion restrictions is overly

stringent” and that “[i]dentification through use of covariance restrictions is also permitted”);

Giampiero Marra & Rosalba Radice, Estimation of a Semiparametric Recursive Bivariate

Probit Model in the Presence of Endogeneity, 39 CAN. J. STAT. 259, 263 (2011) (noting that

theoretical identification in the recursive bivariate probit context does not require the

availability of any instrumental variables because the linear combination of the two

equations does not contain the same variables as the original); Wilde, supra note 206, at 312

(exclusion restrictions are only necessary if there is no variation in exogenous regressors,

and this is a rather weak assumption in economic applications).

225 This underlying identifying assumption of the instrumental variable approach is both

very strong and unverifiable. MORGAN & WINSHIP, supra note 91, at 196–200 (detailing the

shortcomings of traditional instrumental variable estimation and explaining how analysts are



# Stat Aggs


Def. Race

Def. Sex

Def. Age

Crim. History

HS Grad


Marital Status



Vic. Race

Vic. Sex

Vic. Age

Def./Vic. Rel.












situations frequently arise in practice where identical explanatory variables

influence selection and a subsequent outcome of interest, and the analysts

may be required to base identification on distribution assumptions about the

residuals alone.226 The assumption of joint normality of the residual terms

in the bivariate probit model may be reasonable under certain weaker

assumptions: (1) the selection equation and the equation of interest

represent closely related decisions or goals; (2) the decisions have the same

causes; and (3) the decisions occur within a short time frame or are close to

each other geographically.227 These conditions would appear to hold rather

well for the current analysis. Prior simulation studies also report that the

bivariate probit model outperforms instrumental variable models in many

applications and is rather robust to nonnormality of error terms, especially

estimated covariate effects and variance of the random effects.228

Moreover, analysts have established that identification is less of a concern

when causal effects, rather than structural parameters, are of primary

interest.229 In fact, calculation of the treatment effect in the bivariate probit

model directly lends itself to the counterfactual/potential outcomes

mistaken when believing the assumption is empirically testable).

226 See Anne E. Sartori, An Estimator for Some Binary-Outcome Selection Models

Without Exclusion Restrictions, 11 POL. ANALYSIS 111–12 (2003); cf. GREENE, supra note

175, at 616 (“The case of identical regressors [across a system of equations] is quite

common . . . .”).

227 Sartori, supra note 226, at 112; cf. James J. Heckman et al., Matching as an

Econometric Evaluation Estimator, 65 REV. ECON. STUD. 261, 264–65 (1998) (discussing

bias in estimation of treatment effects resulting from geographic mismatch).

228 See, e.g., Jay Bhattacharya et al., Estimating Probit Models with Self-Selected

Treatments, 25 STAT. MED. 389, 399–402 (2006); Marra & Radice, supra note 224, at 260

(“[I]t is well known, from both a theoretical and empirical point of view, that simultaneous

likelihood estimation methods are superior to conventional two-stage instrumental variable

procedures.”); Charles E. McCulloch & John M. Neuhaus, Misspecifying the Shape of a

Random Effects Distribution: Why Getting It Wrong May Not Matter, 26 STAT. SCI. 388, 400

(2011) (“Theory and simulation studies indicate that most aspects of statistical inference are

highly robust to this assumption [of normality for random effects] . . . including estimation

of covariate effects, [and] estimation of the random effects variance . . . .”); Gary Young et

al., Multivariate Probit Models for Conditional Claim-Types, 44 INS.: MATHEMATICS &

ECON. 214, 222 (2009).

229 Joshua D. Angrist, Estimation of Limited Dependent Variable Models with Dummy

Endogenous Regressors: Simple Strategies for Empirical Practice, 19 J. BUS. & ECON. STAT.

2, 3–5 (2001) (endorsing the use of the potential outcomes/counterfactual framework instead

of structural parameters, particularly in models examining limited dependent variables

(LDV) with dummy endogenous variables and also remarking that identifying assumptions

for structural parameters are largely unnecessary for causal inference in LDV models in light

of the potential outcomes/counterfactual framework); Guido W. Imbens & Joshua D.

Angrist, Identification and Estimation of Local Average Treatment Effects, 62

ECONOMETRICA 467, 475 (1994) (same).

534 SHEROD THAXTON [Vol. 103

framework.230 A case “is only allocated to one of the regimes and never

both, the responses in the regimes thus represent potential outcomes.”231 As

Nobel Prize-winning econometrician James Heckman explains, the

“Neyman-Fisher-Cox-Rubin model of potential outcomes . . . is also the

switching regression model of Quandt.”232

The average marginal effect of DPNotice on the likelihood of trial is

the difference between two conditional probabilities: Pr(Trial = 1|DPNotice

= 1, X1, X2) – Pr(Trial = 1|DPNotice = 0, X1, X2). In words, the marginal

effect is the probability that a case results in trial, given that a death penalty

notice has been filed in that case, minus the probability that a case results in

a trial, given that a death penalty notice has not been filed in that case.233

The results from the bivariate probit model support the earlier analyses (see

Table 8). Being noticed for the death penalty decreases the probability of a

case going to trial from .61 to .34 in the DE model. The correlation,

rho (ρ), of the residuals across the two equations after controlling for the

covariates is statistically significant (rho = -.559, p < .05). In the MC

model, being noticed for the death penalty decreases the probability of a

case going to trial from .61 to .36. The correlation between the

disturbances was statistically insignificant (rho = -.518, p < .05).234 These

effect sizes are somewhat larger than those obtained from the previous

models, but it is important to emphasize that the bivariate probit models do

not control for unobserved judicial-circuit-level factors, and this is likely to

impact the causal estimates. The proportion of the variance in the residuals

that is attributable to shared omitted variables is equal to the square of the

correlation coefficient, rho, across the models.

230 GREENE, supra note 175, at 853; Bhattacharya et al., supra note 228, at 400.

231 SKRONDAL & RABE-HESKETH, supra note 178, at 437.

232 James J. Heckman & Edward J. Vytlacil, Local Instrumental Variables and Latent

Variable Models for Identifying and Bounding Treatment Effects, 96 PROC. NAT’L ACAD.

SCI. 4730, 4730 (1999) (citing Richard E. Quandt, A New Approach to Estimating Switching

Regressions, 67 J. AM. STAT. ASS’N 306 (1972)).

233 GREENE, supra note 175, at 853. The conditional probabilities are obtained by:

Pr(Trial=1|DPNotice=1, X1, X2)=Pr(Trial=1, DPNotice=1|X1, X2)⁄(Pr(DPNotice=1|X1)) and

Pr(Trial=1|DPNotice=0, X1, X2)=Pr(Trial=1, DPNotice=0|X1, X2)⁄(Pr(DPNotice=0|X1)).

234 Even if the error terms from the two equations are not correlated in the overall

population (as indicated by significant tests), they are correlated in the selected sample and


QUASI-EXPERIMENTS 73–81 (1986); Steven Goodman, A Dirty Dozen: Twelve P-Value

Misconceptions, 45 SEMINARS HEMATOLOGY 135, 136 (2008) (stating that the “effect best

supported by the data from a given experiment is always the observed effect, regardless of its

significance”). Nonetheless, the close correspondence between the single-equation results

and the bivariate probit results strongly suggest that any bias in the estimates is minimal.

Moreover, additional sensitivity checks reveal that the parameter estimates were not unduly

influenced by endogeneity bias.


Second, a semi-nonparametric version of the bivariate probit model

(SNP) was analyzed to check the robustness of the previously estimated

parametric bivariate model.235 The fully parametric bivariate probit model

in Equation [3] assumes joint normality of residuals, and although

simulation studies suggest the model is rather robust to nonnormality,236

bias in the causal estimates resulting from the violation of that assumption

remains a possibility. The SNP model makes less restrictive assumptions

about the distribution of unobservables, and therefore can handle a broader

class of error distributions.237 The intuition behind the SNP approach is to

approximate the unknown distributions of the residuals by Hermite

polynomial expansions (i.e., the product of a squared polynomial and a

normal density) and use the approximations to derive a pseudo-maximum

likelihood estimator for the model parameters.238 The polynomial

expansion can handle distributions with arbitrary skewness and kurtosis.239

The SNP models reveal that the threat of capital punishment reduces

probability of trial by .249 (DE) and .258 (MC).240 These estimates are

similar to those obtained via the classic bivariate probit model. The

estimates are also somewhat larger than those obtained from the fixedeffects

logit and linear probability models, however the SNP models also do

not account for judicial-circuit-level fixed effects.

Finally, a propensity score-matching model is used to adjust for

nonrandom selection into capital prosecution.241 The intuition behind the

model is that bias in estimates of treatment effects is reduced when the

comparison of outcomes is performed using “treated” and “control”

subjects who are as similar as possible along a large number of relevant

dimensions.242 Exact matching,243 or even coarsened exact matching,244 is

typically infeasible when the number of relevant variables is large, so

235 Giuseppe De Luca, SNP and SML Estimation of Univariate and Bivariate Binary-

Choice Models, 8 STATA J. 190, 192 (2008); Marra & Radice, supra note 224, at 259–60.

236 See Bhattacharya et al., supra note 228; Young et al., supra note 228.

237 Siegfried Gabler et al., Seminonparametric Estimation of Binary-Choice Models with

an Application to Labor-Force Participation, 11 J. BUS. & ECON. STAT. 61, 63 (1993).

238 Gabler et al., supra note 237, at 63.

239 De Luca, supra note 235, at 194; Mark B. Stewart, Semi-Nonparametric Estimation of

Extended Ordered Probit Models, 4 STATA J. 27, 30 (2004).

240 DE model: rho = -.324, p < .10; MC model: rho = -.123, p > .10.

241 See, e.g., Roman et al., supra note 27, at 556–58 (employing propensity score models

to account for nonrandom selection into prosecution).

242 Rosenbaum & Rubin, supra note 217, at 55.

243 Daniel E. Ho et al., Matching as Nonparametric Preprocessing for Reducing Model

Dependence in Parametric Causal Inference, 15 POL. ANALYSIS 199, 217 (2007).

244 Stefano M. Iacus et al., Causal Inference Without Balance Checking: Coarsened

Exact Matching, 20 POL. ANALYSIS 1 (2012).

536 SHEROD THAXTON [Vol. 103

propensity scores summarize pretreatment characteristics into a singleindex

variable.245 By definition, capital and noncapital cases with the same

value of the propensity score have the same distribution of the full vector of

regressors, so it is sufficient to match cases on their propensity score to

obtain the same probability distribution of X for treated and nontreated

match samples.246 Propensity score models rest on the strong, yet

unverifiable, assumption that differences between cases are captured by

their observable attributes (“conditional independence assumption”).247

This assumption is unlikely to hold perfectly with these data; nonetheless,

propensity score models have been shown to reduce, but not eliminate, the

bias generated by unobservable confounding factors. The extent to which

this potential bias is reduced depends on the richness and quality of control

variables used to compute the propensity scores. Thus, the models are

estimated for the simple purpose of comparing their results to the

previously estimated models in order to determine how sensitive the

estimates are to model specifications. The propensity score (PS) is the

conditional probability of a death-notice filing and can be expressed as:

Pr(DPNotice= 1) = exp(𝛼𝑐+𝛾𝑋+𝜁)


, [7]

where α, γ, and 𝜁 are defined in Equation [1].248 As is customary in the

literature, the estimation sample is restricted to the region of common

support: 20 cases were dropped from the DE model (N = 832) and 175

cases were dropped from the MC model (N = 1063).249 In practice, no two

variables will share the exact same propensity score because it is a

continuous variable, so a kernel-matching estimator is used.250 Following

245 Rosenbaum & Rubin, supra note 217.

246 The propensity scores satisfy the balance condition when observations with the same

propensity score have the same distribution of observable (and hopefully unobservable)

characteristics independent of treatment status—i.e., for a given propensity score, exposure

to the treatment is random so treatment and control units should be, on average,

observationally identical. Id.; see also Sascha O. Becker & Andrea Ichino, Estimation of

Average Treatment Effects Based on Propensity Scores, 2 STATA J. 358, 360 (2002)

(describing the standard balancing algorithm).

247 Rosenbaum & Rubin, supra note 217, at 43.

248 Some subscripts have been omitted for simplicity.

249 The region of common support is the overlapping distribution of propensity scores for

the treatment and the control groups. Roman et al., supra note 27, at 557 (explaining that

“the average treatment effect is defined only in the region of common support”); see Gary

King & Langche Zeng, The Dangers of Extreme Counterfactuals, 14 POL. ANALYSIS 131,

146–151 (2006) (noting that using data outside the region of common support induces some

degree of model dependence and increases the risk of bias for almost any model chosen); see

also Iacus et al., supra note 244, at 11 (same).

250 James J. Heckman et al., Matching as an Econometric Evaluation Estimator, 65 REV.


the matching of cases based on propensity scores, the average treatment

effect can be estimated by simply taking the difference in the potential

outcomes in the two counterfactual situations.251 Again, the results from

the propensity score models should be viewed with caution because of the

likelihood of bias resulting from unobservable confounding factors. With

this caveat in mind, it is worth noting that propensity score models

corroborate results from the previously estimated models: a death notice

decreased the probability of trial by .25 and .24 in, respectively, the DE and

MC models.252

In summary, the current study was able to examine directly the causal

impact of the death penalty on defendants’ propensity to go to trial.

Estimates from the various specifications suggest that the death penalty

decreases the probability of a trial anywhere from .17 to .27, from a

baseline probability of approximately .60, although estimates in the range

of .17 to .23 are likely to be most accurate due to the consideration of

unobserved heterogeneity across judicial circuits and years. As explained

earlier this roughly equates to the death penalty deterring two out of every

ten death-noticed defendants from pursuing a trial.

ECON. STUD. 261, 271 (1998) (describing kernel-regression-based matching estimators).

Matching estimators based on nearest neighbors, local linear regression, and Mahalanobis

distances provided similar results. See also MORGAN & WINSHIP, supra note 91, at 107–16

(describing various matching estimators).

251 Iacus et al., supra note 244, at 1; see also supra note 92 and accompanying text for

discussion of the calculation of treatment effects under the counterfactual framework.

252 Results not reported, but available from author upon request.

538 SHEROD THAXTON [Vol. 103

Table 8

Marginal Effect of Death Penalty on Trial (Multivariate Models)

Model 1

Model 2

Model 3

Model 4

DP Notice -0.268*** -0.265*** -0.246*** -0.258***

(0.048) (0.041) (0.047) (0.034)

Pr(Trial) 0.607 0.606 0.684 0.622

N 852 1238 852 1238

Fixed-Effects N N N N

Year Dummies Y Y Y Y

R-Squared N/A N/A N/A N/A

*p<.05; **p<.01; ***p<.001

Standard errors adjusted for clustering on judicial circuit.

Models 1 & 3: Death Eligible Subsample.

Models 2 & 4: Murder Conviction Subsample.

Models 3 & 4: Semi-Parametric Model.

Pr(Trial) = Probability of trial without a death noticed filed.

All specifications include controls for the number of codefendants, victims,

statutory aggravating factors, contemporaneous felony convictions, prior felony

convictions, and children of defendant; the race/ethnicity, gender, and age of

defendant and victim(s); defendant’s employment status at time of arrest,

marital status, high school graduation status; offender/victim relationship; and

whether the homicide was interracial.

Defendants generally prefer charge bargains to sentencing bargains

because a less serious charge is accompanied by a lower penalty (and,

perhaps, less severe collateral consequences).253 Unfortunately, it was

impossible to examine directly the impact of the threat of the death penalty

on charge bargaining because the data used in this study do not contain

information on whether the prosecutor modified her initial charge. As

discussed in Part V.A, however, it is debatable whether much charge

bargaining occurs when a defendant is initially charged with murder (and a

prima facie case exists for the charge).254 Recall that Georgia’s murder

statute does not include “degrees” of murder like many other states. A

253 Kuziemko, supra note 20, at 126.

254 COHEN & KYCKELHAHN, supra note 2, at 10–11 (explaining that the vast majority of

defendants initially charged with murder are ultimately convicted of murder, irrespective of

method of adjudication).


defendant is either charged with murder (mandatory minimum life

sentence), voluntary manslaughter (twenty-year maximum), or involuntary

manslaughter (ten-year maximum).255 The statutory minimum for both

manslaughter offenses is one year, and inmates convicted of manslaughter

are generally eligible for parole after serving 65% of their sentences.256 The

dramatic reduction in potential punishment can make it politically costly for

a prosecutor to offer a charge bargain from murder to manslaughter simply

to avoid trial—especially when he is faced with pressure from victims’

families and the electorate.257

Although prosecutors may lack the flexibility to offer charge or

sentence bargains in the noncapital context,258 they still may present

defendants with other incentives to plea. Such incentives might include: (1)

dropping less serious offenses against the defendant, (2) dropping charges

against codefendants,259 (3) keeping potentially embarrassing facts about

the crime or defendant confidential, (4) asking the judge to impose multiple

sentences concurrently rather than consecutively, (5) assisting with

detention facility placement, or (6) agreeing to assist with parole board

hearings. According to the Georgia data, approximately 40% of noncapital

murder convictions were obtained via a plea bargain. This strongly

suggests that noncapital murder defendants are willing to accept plea

bargains for consideration other than a charge or sentence reduction.

Charge bargaining in the capital context seems equally unlikely.260 That is,

it is doubtful that prosecutors would seek the death penalty against

defendants absent a prima facie case for murder solely to obtain a plea for

manslaughter.261 The Georgia data, in fact, support this intuition: only two

255 GA. CODE ANN. § 16-5-2(b) (West 2003) (voluntary manslaughter); § 16-5-3(a)

(involuntary manslaughter).

256 Id. §§ 16-5-2(b), 16-5-3(a).

257 See supra notes 51–53 and accompanying text. Both scholars and victims’ rights

groups have advocated expanding victims’ involvement in plea negotiations, including the

ability to reopen a plea or sentence when the accused has pleaded to a reduced offense. Kyl

et al., supra note 43, at 621. Others have noted that state prosecutors are more restricted

than federal prosecutors with respect to offering charge bargains due to actual and perceived

political constraints and consequences. Gazal-Ayal, supra note 15, at 2306; Richman &

Stuntz, supra note 53, at 600–05; see also WILLIAM S. MCFEELY, PROXIMITY TO DEATH 19,

80–82 (2000) (describing the intense pressure politicians and prosecutors face from the

electorate to aggressively pursue the death penalty).

258 See, e.g., Ehrhard, supra note 20, at 322 (describing interviews conducted with

prosecutors and defense attorneys who have litigated capital cases and reporting that both the

prosecution and defense believe that LWOP is often necessary to get defendants to take an

LS plea).

259 Prosecutors may also agree not to file charges against potential codefendants.

260 See supra Part V.A.

261 See supra Part V.A.

540 SHEROD THAXTON [Vol. 103

cases that were noticed for the death penalty during the period of the study

resulted in a plea bargain for manslaughter.262


The U.S. Supreme Court has reasoned that plea bargaining is “an

essential component of the administration of justice” when properly

conducted.263 Plea bargaining is purported to provide benefits in the form

of reduced costs, increased efficiency, and certainty to defendants,

prosecutors, defense counsel, judges, and victims.264 When plea bargaining

in the “shadow of death,” incentives for parties to plea bargain may be

significantly magnified, although some have suggested that disincentives to

bargain may be overriding.265 Not only has the Court approved the use of

the death penalty by prosecutors to secure guilty pleas,266 but prosecutors

have openly stated that they (or other prosecutors with whom they are

familiar) routinely use the death penalty as leverage in plea negotiations.267

Due to the absence of methodologically rigorous research on the subject,

however, the actual impact of the threat of capital punishment on pleabargaining

dynamics has been largely speculative. This Article has

demonstrated that, at least in Georgia, the death penalty does reduce the

total number of cases proceeding to trial. Based on the magnitude of that

effect, however, it does not appear that the threat of the death penalty deters

enough murder trials to be cost-effective. This is especially relevant

because the high price tag associated with pursuing the death penalty—

coupled with the infrequency of death sentences and executions268—has

262 See supra Part V.A.

263 Santobello v. New York, 404 U.S. 257, 260 (1971).

264 See supra note 13 and accompanying text; see also Stephanos Bibas, Harmonizing

Substantive-Criminal-Law Values and Criminal Procedure: The Case of Alford and Nolo

Contendere Pleas, 88 CORNELL L. REV. 1361, 1367 (2003) (summarizing procedurally based

arguments in favor of plea bargaining).

265 See supra Part I. See generally Ehrhard, supra note 20, at 313; Hoffmann et al.,

supra note 18, at 2390.

266 Alford v. North Carolina, 400 U.S. 25, 31 (1970); Brady v. United States, 397 U.S.

742 (1970).

267 See supra note 123 and accompanying text (describing instances of prosecutors

admitting using the death penalty as leverage in plea negotiations); see also Ehrhard, supra

note 20, at 319 (describing interviews with prosecutors who admitted that the death penalty

is often used as a bargaining chip).

268 Clark Calhoun, Note, Reviewing the Georgia Supreme Court’s Efforts at

Proportionality Review, 39 GA. L. REV. 631, 632 (2005) (underscoring that less than 2% of

homicide cases occurring in the modern era of the death penalty have resulted in a death

sentence); Liebman et al., supra note 44, at 1844 (showing that 68% of death sentences

handed down between 1973 and 1995 were reversed on appeal and less than 2% of death

row inmates are executed in any given year).


caused many state and local officials to rethink seriously the feasibility of

maintaining the death penalty.269 These concerns have only been

exacerbated in recent years due to the current financial crisis.

In 2009, at least eleven state legislatures considered bills to abolish the

death penalty, citing associated costs as one of their primary concerns.270

Abolition bills were reintroduced in at least five of those states in 2011.271

Georgia lawmakers introduced their own abolition bill in early 2012, also

citing the financial burden of administering the death penalty.272 Over the

past five years, New Jersey, New Mexico, New York, Illinois, and

Connecticut have abolished the death penalty,273 and the high cost of

maintaining a properly functioning death penalty system figured

prominently in all of those debates.274 A recent national study suggests that

each additional capital trial causes an increase in county spending of more

than $2 million and these costs are borne primarily by increasing taxes.275

269 Widespread cuts have been made (or proposed) to courthouse staff, attorney general

offices, district attorney offices, and public defender offices. See, e.g., Greg Bluestein, State

Budget Cuts Clog Criminal Justice System, NBC NEWS (Oct. 26, 2011),

http://www.nbcnews.com/id/45049812/ns/us_news-crime_and_courts/. The American Bar

Association reported that most states have decreased court funding by 10% to 15%,

including significantly scaling back indigent defense and collateral review. Id. Nationally,

twenty-six states have been unable to fill judgeships for budgetary reasons and 14 states have

reduced court staff. In San Francisco, for example, 40% (25 of 63) of the courtrooms have

been closed, resulting in huge backlogs in both the civil and criminal dockets. Id. State and

local governments are also forced to divert funding from hospitals and health care, police

and public safety, education, and roads and infrastructure to pay for capital trials. See Abby

Goodnough, States Turning to Last Resorts in Budget Crisis, N.Y. TIMES, June 22, 2009, at

A1 (stating that in states across the nation, governors and legislators have recommended

increasing taxes and fees, deepening spending cuts, and extending furloughs for government

workers in the face of a $121 billion budget gap).

270 These states included Colorado, Connecticut, Illinois, Kansas, Maryland, Montana,

Nebraska, New Hampshire, New Mexico, Texas, and Washington. DIETER, supra note 49, at

14; see also James S. Liebman, The New Death Penalty Debate: What’s DNA Got to Do with

It?, 33 COLUM. HUM. RTS. L. REV. 527, 528 (2002) (noting that legislation to either abolish

or moderate the use of the death penalty was considered in twenty-six death penalty

jurisdictions in 2000–2001 and passed a committee vote in at least twelve states).

271 Death Penalty Abolished in Illinois, PROJECT PRESS (Am. Bar Ass’n Death Penalty

Representation Project), Spring 2011, at 1.

272 S. 342, 151st Gen. Assemb., Reg. Sess. (Ga. 2012).

273 Ashby Jones & Steve Eder, Costs Test Backing for Death Penalty: Some Former

Supporters Say Capital Punishment Isn’t Worth Huge Sums Spent on Drawn-Out Cases;

Californians to Vote, WALL ST. J. (Oct. 5, 2012), http://online.wsj.com/article/SB10000

872396390443493304578036792267666794.html. Prior to 2007, when New Jersey and

New York abolished capital punishment, it had been nearly twenty-three years since a state

had officially abolished its death penalty. Massachusetts and Rhode Island both dismantled

their capital statutes in 1984.

274 Id.

275 Katherine Baicker, The Budgetary Repercussions of Capital Convictions, 4

542 SHEROD THAXTON [Vol. 103

Such costs have forced many counties to seek help from state legislatures to

create programs to diffuse death penalty costs across counties (even those

that do not choose to use the death penalty).276 Total taxes and

expenditures for capital trials from 1983 to 1999 were more than $5.5


Some of the most thorough examinations of costs associated with the

death penalty over the past twenty years have expressly noted that the threat

of the death penalty may actually produce financial savings by increasing

capitally charged defendants’ propensity to accept a plea, thereby avoiding

trial costs.278 Unfortunately, none of these studies were able to offer any

definitive answers to this question because, based on available evidence, the

impact of capital punishment on plea bargaining was too speculative.279

Studies in California,280 Indiana,281 Kansas,282 and North Carolina283 all

ADVANCES ECON. ANALYSIS & POL’Y 1, 10 (2004) (explaining that death penalty convictions

have cost counties more than $5.5 billion between 1982 and 1999).

276 See Richard Willing & Gary Fields, Geography of the Death Penalty, USA TODAY,

Dec. 20, 1999, at 1A (reporting that “[f]ifteen counties account for nearly a third of all

prisoners sentenced to death [in the United States] but only one-ninth of the population of

the states with capital punishment”).

Prosecutors from rural and suburban counties in Georgia account for a

disproportionate number of death penalty cases. See supra note 75. District attorneys from

ninety-two counties in Georgia (57.8%) have filed all of the 400 death penalty notices for

homicides occurring between 1993 and 2000 (see Table 1). Three counties (1.9%)—Fulton

(Atlanta), Dekalb (Decatur), and Chatham (Savannah)—account for nearly one-half of the

reported homicides, but only 13.5% of death notices during this period. Particularly

interesting is that nearly one-third of Georgia homicides (and 22% of murder convictions)

occurred in Fulton County, although only 4.8% of death notices (and 1.9% of death

sentences) came from Fulton County between 1993 and 2000. The seven counties with the

highest death-noticing rates (i.e., percentage of murder convictions noticed for the death

penalty)—Oconee (88%), Morgan (82%), Putnam (73%), Ware (70%), Appling (63%),

Bartow (56%), and Lowndes (54%)—account for approximately 3% of the state’s population

and less than 4% of murder convictions (and less than 3% of total reported homicides), but

13% of death notices between 1993 and 2000. Collectively, these seven counties sought the

death penalty in 68% (52 of 77) of homicide cases that ultimately resulted in a murder


277 Baicker, supra note 275, at 1321 (also estimating that each capital case costs

approximately $2.5 million to prosecute); see also Public Policy Choices on Deterrence and

the Death Penalty: A Critical Review of New Evidence: Hearing on H.B. 3834 Before the J.

Comm. on the Judiciary, 2005 Leg. (Mass. 2005) (statement of Jeffrey Fagan, Professor,

Columbia University), available at http://www.deathpenaltyinfo.org/MassTestimony

Fagan.pdf (summarizing studies of the financial cost of the death penalty).

278 DIETER, supra note 49, at 16–17 (citing studies that have recognized potential savings

from capital statutes by increasing the likelihood of pleas).

279 Id.

280 CAL. COMM’N ON THE FAIR ADMIN. OF JUSTICE, supra note 94, at 80–81 (commenting

that, in some cases, the risk of the death penalty provides an incentive to plead to life without

parole so removing the death penalty might result in more trials, but explaining that even if


acknowledge that a truly comprehensive study of the cost of the death

penalty in their respective jurisdictions would require a close examination

of the causal impact of the death penalty on plea bargaining in deatheligible

cases. Capital trials are extremely expensive, so numerous plea

bargains in potentially capital trials may be required to offset the cost of a

single capital trial. Capital cases are more expensive and time-consuming

than comparable noncapital murder cases at every stage of the process:

more time for pretrial preparation, more pretrial motions, more experts,

more attorneys for the defense (and typically the government as well), more

time to conduct voir dire, longer trials, longer jury deliberations, and more

appeals that take longer.284 Georgia has yet to conduct a comprehensive

study of the costs of its capital punishment system, but trial costs for the

median capital case have been estimated as at least four times more

expensive than the median noncapital murder trial.285 The elapsed time

from arrest to murder conviction in capital cases in Georgia for the period

under investigation in this study was nearly 1.8 times longer than the

elapsed time from arrest to murder conviction in death-eligible noncapital

all murder cases eligible for life without parole went to trial, both trial and appellate costs

would be considerably less expensive than capital cases, in large part because of the absence

of a penalty phase and right to counsel for habeas petitions).



for the Office of the Governor and the Indiana General Assembly) (recognizing that plea

bargaining also influences the cost of resolving a capital case, but also acknowledging that

the impact of the death penalty on plea bargaining may not always be in both parties’ best




CORRECTIONS 19, 32 (2003) (stating a potential benefit of the death penalty savings from

avoiding trials, but acknowledging the instant study did not examine those potential savings).

283 Philip J. Cook & Donna B. Slawson, The Costs of Processing Murder Cases in North

Carolina 31 (May 1993) (unpublished manuscript) (on file with the Terry Sanford Inst. of

Pub. Policy, Duke Univ.) (noting that 21% of capitally charged defendants plead guilty to

murder and avoid trial, but concluding that “[s]ince we have no direct evidence on the effect

of the death penalty option on the likelihood of trial, and since there are plausible arguments

in both directions, we proceed on the assumption that there are neither more nor fewer trials

as a result of the death penalty option”).

284 See Roman et al., supra note 27, at 536 (explaining that case processing is more

expensive at every stage of a capital case relative to a noncapital murder case); Robert L.

Spangenberg & Elizabeth R. Walsh, Capital Punishment or Life Imprisonment? Some Cost

Considerations, 23 LOY. L.A. L. REV. 45, 46–47 (1989) (conducting the first cost analysis of

every stage of the capital charging-and-sentencing process and concluding that life

imprisonment is a more economical alternative).

285 Stephen Gurr, The High Cost of Death, GAINESVILLE TIMES, Aug. 10, 2008, at 1D

(reporting that defense expenses alone average about $150,000, and juror and bailiff costs

are seven times higher in capital cases).

544 SHEROD THAXTON [Vol. 103

murder cases (27.9 months versus 15.9 months). Interestingly, the average

time between arrest and conviction via trial in death-eligible noncapital

murder cases (17.4 months) was less than the time between arrest and plea

in capital cases in Georgia (24.6 months).286

A recent study in New Jersey revealed that capital cases resulted in

two to five times more pretrial motions filed, three to five times longer

pretrial defense investigation, sixty-six times longer voir dire, and thirty

more court days.287 Capital cases were also ten times as likely to proceed to

trial, had twice as many lawyers (by statute), and resulted in longer and

more complicated appeals.288 An examination of death penalty costs in

Kansas discovered that, relative to noncapital murder cases, capital cases

had investigation costs that were three times higher, trial costs that were

sixteen times greater, and appeal costs that were twenty-one times more.289

The added expense associated with capital cases at every stage of the

process has also been documented in the administration of the federal death

penalty: attorney costs, expert costs, transcript costs, and out-of-court costs

were significantly greater for capital cases.290 In general, capital cases were

four times more expensive than noncapital cases for death-eligible

defendants from 1989 to 1997, and 6.5 times more expensive from 1998 to


One of the few methodologically rigorous studies attempting to take

both state and federal costs into account when comparing cost differentials

between capital and noncapital death-eligible cases revealed that

prosecuting a single capital case in Maryland adds at minimum $1 million

in total costs even after taking into account differential imprisonment

costs.292 Trial costs were five times more expensive in capital cases

compared to noncapital cases ($823,000 versus $160,000), and appellate

286 See supra Part III.C.



288 Id.

289 LEGIS. DIV. OF POST AUDIT, supra note 282 (discovering that, relative to noncapital

murder cases, capital cases had investigation costs that were three times higher, trial costs

that were sixteen times greater, and appeal costs that were twenty-one times more).




4.6 times more hours worked by defense counsel and 15.7 times more spent in expert fees in

federal capital cases versus death-eligible noncapital cases).

291 Id. (reporting that mean capital case was $490,000 and the mean noncapital murder

case was $76,000; the median costs for capital and noncapital cases were, respectively,

$350,000 and $45,000).

292 Roman et al., supra note 27.


costs were almost four times more expensive ($340,000 versus $88,000).293

Another relevant study, coauthored by Senior Judge Arthur L. Alarcón of

the U.S. Court of Appeals for the Ninth Circuit, suggests that federal habeas

review of a state capital conviction adds between $635,000 and $1.58

million to each capital case.294 These federal costs are in addition to the

$1.2 million more California spends on each capital case at the state


It is important to underscore that because capital cases are more

expensive and time-consuming at every stage, those cases that are

ultimately resolved by plea bargain may be more expensive than noncapital

trials because of pretrial and pre-penalty-phase costs. According to data

from the federal death penalty system, capital cases eventually disposed by

plea were over four times more expensive than the median noncapital case,

regardless of method of disposition.296 Costs for noncapital cases were not

disaggregated according to trial and plea, but given the fact that capital

trials were 1.3 times more expensive than the median capital case,

irrespective of method of disposition ($465,602 versus $353,185), it may be

reasonable to assume that noncapital trial costs exceed noncapital plea costs

by a similar multiplier. Based on that assumption, the median capital case

resolved by plea is over 3.3 times more costly than the median noncapital

trial in the federal system ($200,933 versus $60,000).

Similar estimates have been discovered at the state level. For example,

in North Carolina (20022006) the median cost of a capital case ultimately

disposed by plea bargain was 1.6 times more expensive than the median

cost of a death-eligible noncapital case disposed by trial ($31,000 versus

$19,000).297 Pleas obtained in capital cases were also three times as costly

as pleas obtained in death-eligible noncapital cases ($31,000 versus

$10,000).298 Capital trials were nearly 4.5 times more expensive than

death-eligible noncapital trials ($82,000 versus $19,000), and 2.6 more

expensive than capital cases disposed by plea ($82,000 versus $31,000).299

Even dismissals in capital cases were four times more expensive than

293 Id.

294 See Alarcón & Mitchell, supra note 65, at S88.

295 CAL. COMM’N ON THE FAIR ADMIN. OF JUSTICE, supra note 94 (noting the difference

between the least expensive capital trial and most expensive noncapital trial was $1.1


296 GOULD & GREENMAN, supra note 290, at 27 (reporting a total cost of $200,993 for

capital cases disposed by plea compared to $44,809 for noncapital cases, irrespective of

method of disposition).



298 Id.

299 Id.

546 SHEROD THAXTON [Vol. 103

dismissals in death-eligible noncapital cases ($19,000 versus $4,500).300

The bulk of these cost differentials can be attributed to the fact that once a

case is noticed for the death penalty, most death penalty statutes (including

North Carolina’s) require the appointment of two attorneys for the

defendant, and attorney fees comprise the bulk of the cost of capital

cases.301 It should be emphasized that these cost estimates only include the

defense’s attorney fees and expert witness expenses; they do not take into

account expenses incurred by the government.302 We would expect

government expenditures to be significantly higher in capital cases, but

assuming arguendo that the cost to the government is similar for capital and

noncapital cases, the relative gap in spending would remain the same

(although the multiplier would change).

Now consider a few examples to fix ideas. Using the rather

conservative estimates obtained from a recent Maryland study, the

government can save, at most, $250,000 in a death-eligible noncapital case

by avoiding trial and subsequent appeals.303 Of course, there are still costs

associated with obtaining a plea. The Maryland study did not report

estimates of costs in cases disposed by plea prior to the start of trial, but

based on estimates from the federal government and North Carolina, it is

reasonable to assume that significant costs are associated with pleas in both

capital and noncapital cases. Discounting potential trial savings by plea

costs by a conservative estimate of one-third suggests that the government

can save approximately $195,000 by avoiding trial (and subsequent

appeals) in a noncapital case (plea costs = $160,000 trial costs × .33 =

$53,280). Also recall that a capital case in Maryland costs at least an

additional $1 million to process fully, even after considering prison costs.

This suggests that the threat of the death penalty would need to deter more

than five murder defendants from pursuing trial to offset the price of one

capital trial. If the cost of obtaining a plea is higher—e.g., 50% as in North

Carolina—then the death penalty would have to deter more than six murder

defendants from opting for trial. Based on the analysis presented in this

Article, the death penalty only deters between 1.5 and 2 murder defendants

300 Id.

301 Id. North Carolina compensates capital defense attorneys at a rate of $85 per hour,

whereas the federal government paid attorneys $125 per hour during a similar time period.

Id. at 5; see also GOULD & GREENMAN, supra note 290, at 28 (explaining that attorney fees

constitute a large fraction of the cost of capital cases).

302 N.C. OFFICE OF INDIGENT DEF. SERVS., supra note 297, at 1, 10.

303 Roman et al., supra note 27, at 565. The estimates of costs associated with capital

trials obtained from the Maryland study are likely to be biased downward. For example, the

study estimated the median cost of federal habeas review at $25,000 (std. dev. $105,000),

whereas a recent California study estimated federal habeas costs between $635,000 and

$1.58 million.


from opting for trial for every one capital trial in Georgia.

Of course these calculations ignore the fact that obtaining a plea

bargain in a capital case may be more expensive than the total trial and

appellate costs for a noncapital death-eligible case. This would preclude the

possibility of the death penalty serving any cost-saving function in light of

the fact that it does not appear that there is a corresponding decrease in

prison costs.304 More cost estimates from death penalty jurisdictions based

on methodologically rigorous study designs are necessary to explore fully

this counterintuitive implication. But even if the costs associated with

obtaining pleas in capital cases merely constituted half of the total trial and

appellate savings, then the threat of capital punishment would need to

dissuade eight defendants from choosing trial. And more than twelve

defendants would need to be discouraged from going to trial if plea

expenses comprised two-thirds of total noncapital trial savings.

Due to the fact that most capitally prosecuted defendants are not

sentenced to death, perhaps it is more illuminating to explore cost estimates

that take account of cases that begin as capital cases, but do not accomplish

the stated purpose of a death penalty system: executions.305 These “cohort

costs” (i.e., per-execution costs) have been estimated at $3.2 million in

North Carolina,306 $3.5 million in Texas,307 $24 million in Florida,308 $37.2

million in Maryland,309 and $250 million in California.310 New York and

304 In fact, prison costs appear to be greater for death row inmates than non-death row

inmates, although death row inmates’ total incarceration time is less. Id.

305 Nationally, only one in every three capital trials results in a death sentence (because

of mercy from the judge or jury) and only one in every ten death sentences may result in an

execution, so the total cost to reach that one execution may be prohibitively high. DIETER,

supra note 49, at 14 (discussing different approaches to assessing the cost of the death

penalty); see also supra note 127 and accompanying text (noting that in Georgia, from 1993

to 2000, only 32% of capital cases disposed by trial received the death sentence).

306 This figure represents the cost per execution above life imprisonment. The $3.2

million price tag is the inflation-adjusted number from $2.16 million in 1993.

307 This is the inflation-adjusted figure. Costs of the Death Penalty and Related Issues:

Hearing on H.B. 1094 Before the H. Comm. on the Judiciary, 2007 Leg. (Colo. 2007)

(statement of Richard C. Dieter, Executive Director, Death Penalty Information Center)

(stating that Texas spends three times more on capital cases from trial through execution

than for life imprisonment for forty years in a maximum security single cell). Reliable data

on the cost of the death penalty are lacking, but recent reports suggest that the average

nonmurder trial in Texas costs about $3,000, whereas death penalty trials range from

$200,000 to $1 million. Logan Carver, Paying the Price: Death Penalty Cases More

Expensive than Lifetime Imprisonment, but Local CDA Says Cost Never a Consideration,

LUBBOCK AVALANCHE-J., Dec. 13, 2009, at A1.

308 S.V. Dáte, The High Price of Killing Killers: Death Penalty Prosecutions Cost

Taxpayers Millions Annually, PALM BEACH POST, Jan. 4, 2000, at 1A.

309 Jennifer McMenamin, Death Penalty Costs Md. More than Life Term, BALTIMORE

SUN, Mar. 6, 2008, at 2B.

548 SHEROD THAXTON [Vol. 103

New Jersey have spent, respectively, $170 million and $253 million on their

capital punishment systems, but have failed to execute a single condemned

inmate.311 Even so, conservative estimates based on continued annual

expenditures would place cost-per-execution in the $20–$40 million range

for these two jurisdictions.312 The number of plea bargains induced by the

threat of the death penalty required to offset per-execution costs, obviously,

would be significantly higher. Continuing with the Maryland example, and

assuming no cost to obtain a plea, nearly 150 death-eligible defendants

would need to be deterred from opting for trial to offset the cost of one

execution. Assuming that plea costs constitute one-third of trial costs, then

190 death-eligible defendants would need to be discouraged from going to


Based on these stark differences in costs between capital and

noncapital cases, it is quite possible that prosecutors are fully aware that the

threat of capital punishment cannot serve any cost-saving function, and they

merely leverage the death penalty to impose harsher bargaining terms—

most notably, life imprisonment. More than 70% of inmates serving life

sentences were convicted of murder, and more than one in four of all

inmates serving a life sentence have no possibility of parole.313 But this use

of the death penalty has important cost implications as well. The

mandatory minimum for most first-degree murder (and equivalent)

convictions is life with the possibility of parole,314 and several states only

allow life without parole, so it is not clear that the threat of the death

penalty (and the associated expenses) is necessary to obtain a sufficiently

harsh sentence (the average time served for an inmate serving life

imprisonment is thirty years across all offenses). In most death penalty

jurisdictions, governors and pardon and parole boards are extremely

reluctant to grant convicted murderers early release, so in practice, facially

indeterminate sentences have become de facto determinate life sentences.315

310 Rone Tempest, Death Row Often Means a Long Life, L.A. TIMES, Mar. 6, 2005, at


311 DIETER, supra note 49, at 14 (citing studies and compiling statistics).

312 Id. at 15. New York sentenced seven individuals to death from 1995 through 2004—

a cost of $24 million per death sentence. JOSEPH LENTOL ET AL., THE DEATH PENALTY IN

NEW YORK 7 (2005). Juries in New Jersey returned sixty death sentences ($4.2 million per

death sentence), but fifty-seven were overturned on appeal and only nine inmates were on

death row as of 2007 ($28 million per death sentence). See N.J. DEATH PENALTY STUDY

COMM’N, supra note 287, at 7.


AMERICA (2009).

314 See supra note 54 and accompanying text.

315 See supra Part III.B.


When the Supreme Court ruled in Furman that the death penalty was

unconstitutional as applied, the majority was particularly concerned that the

death penalty was not being reserved exclusively for the worst crimes and

worst criminals.316 Many commentators have remarked that the death

penalty is an incredibly powerful tool at the disposal of prosecutors,

especially when it is used as leverage to induce defendants to forfeit their

right to trial (and accompanying rights). The empirical findings in this

Article suggest that the threat of the death penalty has a substantial causal

effect on the likelihood that a defendant accepts a plea agreement.

Nevertheless, the magnitude of the effect is clearly insufficient to offset the

substantial administrative and financial costs arising from the occasional

capital defendant taking her chances at trial (or, in some instances, even the

capital case that incurs significant pretrial or pre-penalty-phase costs prior

to a plea agreement). The government’s use of the death penalty to obtain

convictions quickly and cheaply appears to fail on both of these

dimensions—and this may be particularly true in marginal cases because

the likelihood of trial, a sentence other than death, or a reversal on appeal is

particularly high. Prosecutors are likely to continue to use the threat of the

death penalty in this highly inefficient manner unless they are required to

internalize more of the costs of making poor screening decisions on the

front end. Given the tremendous human and financial costs associated with

the use of the “ultimate punishment,” prosecutors must be made to think

carefully about using the death penalty as a “bargaining chip” in situations

when such a severe sanction may be unwarranted.317

316 There was no controlling opinion in Furman. Each of the Justices comprising the 5–4

majority (Brennan, Marshall, Douglas, Stewart, and White) differed over the basis of the

decision; nonetheless, three recurring themes ran through their individual opinions. The

Furman court was primarily troubled by three glaring problems with the existing practice of

capital punishment: (1) the small number of death sentences handed out relative to

potentially capital crimes; (2) the lack of statutory restrictions upon the sentencing discretion

of judges and jurors; and (3) sentencing disparities based on social class and race. DAVID


225–30 (2010); see also Calhoun, supra note 268, at 632; Schornhorst, supra note 31, at 301.

317 “[T]he practice of charging the death penalty solely for the purpose of obtaining plea

bargains is an unethical and unconstitutional interference with a defendant’s Sixth

Amendment right to trial [and impermissibly] risks convicting innocent defendants who

plead guilty solely to avoid the possibility of a death sentence—which has occurred on

numerous occasions.” DIETER, supra note 49, at 17.

550 SHEROD THAXTON [Vol. 103

Appendix A





DP Notice

Death Penalty Notice Filed (Yes=1)



Plea/Verdict Case Disposed by Trial or Plea Bargain




Incident Year Date of Incident (YYYY) GDC; CO;


Offender Age Defendant’s Age at Time of Incident (in




Offender Sex Defendant’s sex/gender (Male=1) GDC; CO;


Offender Race Defendant’s Race (Asian, Black, Hispanic,

Other, White)



Offender HS


Defendant Graduated from High School






Defendant’s Marital Status (Married=1) GDC

# of Children Defendant’s Number of Children GDC



Defendant was convicted of committing a

contemporaneous felony (Yes=1)


Prior Felony

Defendant had prior felony conviction



# of Offenders Total Number of Codefendants CO; GCD



Defendant Convicted of Murder (Yes=1) GDC; CO;




Number of statutory aggravating

circumstances present in case









# of Victims Number of deceased victims in the case CO; GCD;


Victim Age Victim’s age at time of incident (in Years) CO; GCD;


Victim Race Victim’s race (Asian, Black, Hispanic, Other,




Victim Sex Victim’s sex/gender (0=No; 1=Yes) CO; GCD;




Victim(s) and defendant were strangers






The defendant differed in race from at least

one of the victims (Yes=1)



County County in which the trial took place GDC; CO;


Circuit Circuit in which the trial took place GDC; CO;


DP Eligible Defendant was eligible for the DP based on a

murder conviction and the presence of at least

one Statutory Aggravating Circumstance



LEGEND: U.S. Bureau of the Census (CENSUS); Ga. Department of Corrections

(GDC); Ga. Sup. Ct. Clerk’s Office (CO); Ga. Bureau of Investigation (GBI);

Office of the Georgia Capital Defender (GCD); Supplementary Homicide Reports

(SHR); Atlanta-Journal Constitution (AJC).

Appendix B

Fixed-Effects Logit Models (Marginal Effects)

Pr(Trial) Pr(Notice)

DP Notice -0.234*** (0.038) -- --

# of Stat Aggs -0.042* (0.019) 0.108*** (0.020)

# of Offenders -0.038* (0.017) -0.005 (0.013)

Offender Black -0.033 (0.048) -0.107* (0.044)

Offender Male -0.149 (0.076) -0.102 (0.065)

Offender Age 0.000 (0.002) -0.002 (0.002)

Contemp. Felony 0.043*** (0.013) 0.029* (0.013)

Prior Felony 0.040*** (0.010) 0.006 (0.010)

High School Grad 0.142*** (0.035) 0.032 (0.036)



0.073* (0.035) 0.072** (0.025)

Offender Married 0.085 (0.063) -0.029 (0.030)

# of Children -0.045 (0.027) 0.017 (0.033)

# of Victims 0.028 (0.037) 0.104*** (0.024)

Victim White -0.097* (0.044) 0.115*** (0.035)

Victim Female -0.085** (0.033) 0.117*** (0.027)

Victim Age -0.001 (0.001) 0.000 (0.001)

Victim Stranger 0.055 (0.036) 0.087** (0.031)



0.006 (0.046) -0.012 (0.038)

N 852 856

Year Dummies Y Y

Pseudo R-Squared 0.32 0.37

*p<.05, **p<.01, ***p<.001

Standard errors adjusted for clustering on judicial circuit.