Replication Lab · Study 4
George & Epstein (1992): On the Nature of Supreme Court Decision Making
Law or politics? A legal model and a political model of the Court's death-penalty decisions, tested head to head. We re-coded the cases from the opinions, re-estimated both models on the paper's years, repeated its 1989 check, and carried the models forward to OT2025.
The study
Question. Does the law (the doctrine the Court has announced) or politics (who appointed the Justices, which party governs, who the litigants are) better explain how the Court decides death-penalty cases?
Design. Two probit models of every death-penalty decision from the 1971 to the 1988 terms, each judged by how many outcomes it classifies correctly, then combined; with a look at the 1989 term, which neither model had seen.
Data
- 64 decisions involving the imposition of the death penalty, 1971–1988 terms, from the U.S. Reports. The outcome: the Court upheld the death sentence (45%) or not.
- Legal variables, coded from the parties' briefs and the lower-court opinions: whether the crime was an intentional murder, and whether the defense raised four doctrinal claims (a jury death-qualified by excluding opponents of capital punishment; a failure to consider the defendant's particular, mitigating circumstances; vague or overbroad aggravating factors; a state psychiatric examination).
- Extralegal variables: Court change (one point per appointment from Justice Stevens on), the political environment (one point each for a Republican President and Senate), expert defense counsel (the NAACP Legal Defense Fund, the ACLU, or appointed by the Court), a repeat-player state (Texas, Georgia, Florida), a Solicitor General amicus brief, and whether the state appealed.
Method
- Probit regression of the outcome on each set of variables, then on both together.
- Each model judged by the share of cases it classifies correctly, against 55% for guessing that every case comes out for the defendant.
- Errors examined by era and by ideological direction.
What they found
| Result | Notes | |
|---|---|---|
| Legal model | 75% correct | Every claim variable in the expected direction; the strongest, whether the crime was an intentional murder. |
| Extralegal model | 81% correct | Strongest effects: the political environment, who appealed and the Solicitor General's support for the state. |
| Integrated model | 88% correct | Extralegal factors carry about twice the weight of legal ones overall; legal factors dominated only at the start (1971–74). |
| The 1989 term (out of sample) | Legal model: 10 of 11 for the defendant; extralegal: all 11 for the state | The state won 10. Each model drifts in its own direction: the legal one toward the defendant, the political one toward the state. |
The takeaway. Neither law nor politics alone explains the death-penalty cases; the two models make opposite mistakes, and together they classify nearly nine cases in ten.
- Tracey E. George & Lee Epstein, On the Nature of Supreme Court Decision Making, American Political Science Review 86(2): 323–337 (1992). DOI →
- Full text on Lee Epstein's site: epstein.wustl.edu/onnature. Full text →
How we rebuilt it
The paper's data set was available only on request, so we rebuilt it from the opinions.
| Paper | Our rebuild | |
|---|---|---|
| Cases | 64 decisions involving the imposition of the death penalty, 1971–1988 terms, from the U.S. Reports | Found by searching CourtListener's Supreme Court opinions for capital-case language (294 capital decisions through OT2025), kept when the sentence itself is at issue: the Supreme Court Database codes it as a death-penalty case, or the defense raised one of the paper's sentencing claims. 70 decisions in 1971–1988 |
| Legal claims | Coded from the parties’ briefs and the lower-court opinions | Coded by Claude from the Supreme Court opinion: a claim counts if the opinion reports the defense raised it. Opinions report claims selectively, which the authors warned about |
| Counsel and Solicitor General | From the U.S. Reports | From the counsel list CourtListener reproduces from the U.S. Reports (LDF, ACLU and the Fund’s known capital litigators count as expert; so does counsel appointed by the Court); from the opinion when the list is missing |
| Court change | One point per appointment from Stevens on (0–4 by 1988) | The number of sitting Justices appointed by Republican presidents from Ford on: the same index through 1988, extended past it (it falls when such a Justice is replaced by a Democratic appointee) |
| Political environment | 0 Democratic President and Congress, 1 split, 2 Republican President and Senate | One point each for a Republican President and a Republican Senate on the decision date |
| Repeat-player states | Texas, Georgia, Florida | The three states with the most capital cases in the training window (Florida, Georgia, Texas for 1971–88) |
| Model | Probit | Probit by maximum likelihood |
Check: the paper's years, 1971–1988
70 decisions (the paper had 64); the death sentence was upheld in 45.7% (paper: 45%). Share of cases classified correctly:
| Model | Paper | Our rebuild |
|---|---|---|
| Legal | 75% | 67.1% |
| Extralegal | 81% | 72.9% |
| Integrated | 88% | 81.4% |
| Guess the more common outcome | 55% | 54.3% |
Counting every capital case instead (87 decisions, including challenges to the conviction only): legal 65.5%, extralegal 70.1%, integrated 72.4%.
Legal model coefficients
| Paper | Ours (s.e.) | |
|---|---|---|
| Constant | -5.35 | -0.27 (1.17) |
| Intentional murder | 1.49 | 1.29 (0.45) |
| No death-qualified-jury claim | 0.94 | -0.25 (0.50) |
| No particularized-circumstances claim | 1.46 | 0.17 (0.38) |
| No aggravating-factors claim | 1.01 | -0.83 (0.43) |
| No state-psychiatric-exam claim | 1.38 | -0.14 (0.58) |
Extralegal model coefficients
| Paper | Ours (s.e.) | |
|---|---|---|
| Constant | -5.87 | -3.29 (0.80) |
| Court change (Republican appointees) | 0.70 | 0.19 (0.16) |
| Defense counsel not expert | 1.20 | 1.04 (0.41) |
| Repeat-player state | 1.29 | 1.13 (0.39) |
| Solicitor General amicus | 2.67 | 1.60 (0.79) |
| State appealed | 1.74 | 0.75 (0.46) |
| Political environment | 1.88 | 0.99 (0.30) |
Positive coefficients favor upholding the death sentence. The paper's table prints the Solicitor General coefficient with a minus sign, but its text says a Solicitor General brief raised the state's chances; we show it as positive.
Forward to 2025: the 1971–88 models on later cases
The three models estimated on every 1971–88 capital decision, unchanged, applied to every later one. (For the comparison above we kept the cases where the sentence itself is at issue, closest to the paper's 64; the database files most modern capital cases, which come through federal habeas, under other issues, so the forward test uses every capital case: 87 in 1971–88.) In parentheses: the share of cases each model predicted the sentence would be upheld, to show the drift the paper looked for.
The paper's own check, the 1989 term. 13 capital decisions; the death sentence was upheld in 10. The legal model got 53.8% right (predicting "upheld" in 46.2%), the extralegal 38.5% (15.4%). The paper: the state won 10 of 11; the legal model called 10 for the defendant, the extralegal all 11 for the state.
| Terms | Cases | Upheld | Legal | Extralegal | Integrated | More common outcome |
|---|---|---|---|---|---|---|
| OT1989–OT1993 | 52 | 55.8% | 46.2% (67.3%) | 42.3% (25.0%) | 42.3% (28.8%) | 55.8% |
| OT1994–OT2004 | 62 | 53.2% | 53.2% (74.2%) | 50.0% (54.8%) | 51.6% (50.0%) | 53.2% |
| OT2005–OT2015 | 62 | 45.2% | 45.2% (64.5%) | 66.1% (40.3%) | 62.9% (37.1%) | 54.8% |
| OT2016–OT2025 | 31 | 38.7% | 41.9% (77.4%) | 38.7% (61.3%) | 45.2% (54.8%) | 61.3% |
OT2022–OT2025
Each term predicted by the three models re-estimated on every earlier capital decision since 1971 (the repeat-player states re-chosen each time). The share is the model's probability that the sentence is upheld.
| Term | Case | Legal | Extralegal | Integrated | Outcome |
|---|---|---|---|---|---|
| OT2022 | Cruz v. Arizona Capital defendant sentenced to death for murdering a Tucson police officer sought Simmons relief via successive state postconviction petition after Lynch v. Arizona; Court held Arizona Supreme Court's ruling that Lynch was not a 'significant change in the law' was not an adequate state ground, vacated and remanded. |
Upheld 50% ✗ | Not upheld 35% ✓ | Not upheld 37% ✓ | Not upheld |
| OT2022 | Reed v. Goertz Death-row inmate Rodney Reed's § 1983 procedural due process challenge to Texas's post-conviction DNA testing law was held timely (limitations period runs from end of state litigation, including rehearing denial); Court reversed the Fifth Circuit's dismissal. |
Upheld 50% ✗ | Not upheld 34% ✓ | Not upheld 33% ✓ | Not upheld |
| OT2023 | Thornell v. Jones Arizona sought review of the Ninth Circuit's grant of habeas relief on an ineffective-assistance-at-sentencing claim; the Court reversed, holding Jones showed no Strickland prejudice given weak new mitigation and weighty aggravators, leaving the death sentence in place. |
Upheld 51% ✓ | Upheld 77% ✓ | Upheld 79% ✓ | Upheld |
| OT2024 | Hamm v. Smith Alabama sought review of the Eleventh Circuit's affirmance of vacatur of Smith's death sentence on Atkins intellectual-disability grounds; the Court granted cert, vacated the judgment, and remanded for clarification on how multiple IQ scores are evaluated. |
Not upheld 50% ✓ | Upheld 77% ✗ | Upheld 81% ✗ | Not upheld |
| OT2024 | Andrew v. White Brenda Andrew, sentenced to death for murdering her husband, argued that irrelevant, prejudicial sex-life evidence violated due process; the Court held the Tenth Circuit wrongly found no clearly established law under Payne v. Tennessee, reversed, and remanded for prejudice analysis under AEDPA. |
Not upheld 50% ✓ | Not upheld 33% ✓ | Not upheld 34% ✓ | Not upheld |
| OT2024 | Glossip v. Oklahoma Glossip, sentenced to death for paying Sneed to murder Van Treese, challenged his conviction under Napue for prosecution's failure to correct Sneed's false testimony about lithium/psychiatrist; Court reversed OCCA and remanded for a new trial. |
Not upheld 32% ✓ | Not upheld 33% ✓ | Not upheld 13% ✓ | Not upheld |
| OT2024 | Gutierrez v. Saenz Death-sentenced Gutierrez sued under § 1983 over Texas's DNA-testing procedures; the Court held he has standing (redressable under Reed v. Goertz) and reversed the Fifth Circuit, remanding, which leaves the death sentence not upheld. |
Not upheld 32% ✓ | Not upheld 31% ✓ | Not upheld 11% ✓ | Not upheld |
| OT2025 | Pitchford v. Cain Capital habeas challenge to Mississippi death sentence based on Batson race-discriminatory peremptory strikes; Court reversed the Fifth Circuit, holding the state supreme court unreasonably found waiver of the pretext argument and restoring the District Court's habeas relief (remanded). |
Not upheld 31% ✓ | Not upheld 31% ✓ | Not upheld 12% ✓ | Not upheld |
Correct calls: legal 6 of 8, extralegal 7, integrated 7. In the earlier cases the sentence stood slightly less than half the time, so the simplest forecast, "the sentence falls", gets 7 of 8; the Court upheld the sentence in 1 of these 8. Eight cases are too few to tell the models apart.
What we learned
- The political model travels better than the legal one, within the paper's years. Re-coded from the opinions, the extralegal variables keep the paper's signs and land near its size for defense counsel and repeat-player states; the legal claims lose almost all their pull. Opinions are a poor stand-in for the briefs the authors coded: the Court reports the claims it decides, not every claim raised.
- Out of sample, the models rarely beat a simple guess. Carried forward unchanged, the combined 1971–88 model classifies 42.3% (1989–1993), 51.6% (1994–2004), 62.9% (2005–2015), 45.2% (2016–2025) of later capital cases, against 55.8%, 53.2%, 54.8%, 61.3% for guessing the more common outcome. The one era where it does better is OT2005–OT2015, where the political variables alone reach 66.1%.
- Half of the paper's warning holds. The paper found its political model drifting toward the state and its legal model toward the defendant. Our political model does drift toward the state, predicting "upheld" in 25.0%, 54.8%, 40.3%, 61.3% of cases by era while the Court upheld 55.8%, 53.2%, 45.2%, 38.7%. Our legal model drifts the same way rather than the opposite (67.3%, 74.2%, 64.5%, 77.4%).
- The Court's capital docket has changed. Most recent capital cases reach the Court through federal habeas or civil-rights suits rather than direct review of the sentence, and in OT2022–OT2025 the Court ruled against the state in 7 of 8. A model built on the doctrinal fights of the 1970s and 1980s has little to say about them.
Data and code
Every number comes from /data/replication/george-epstein.json, including each case's model probabilities. The coded cases (with Claude's one-line note on each) are in the repository as data/curated/capital-cases.json. Opinions and counsel lists: CourtListener (Free Law Project). Case data: Supreme Court Database, 2025 Release 01.
Updated 2026-10-05. Questions or corrections: contact us.