IV. StateLevel Factors Related to Serious, Reversible Error in Capital Trials and Verdicts: Aggressive Use of the Death Penalty, Ineffective Law Enforcement, Politics and RaceA. Differences in How States Use the Death Penalty and How Often Their Verdicts Are Reversed As we point out in discussing Figures 1121, pp. 11832 above, our data reveal wide variation among states on many aspects of death sentencing, including how often their death verdicts are reversed due to serious error. Statistical analysis can sometimes identify factors that explain such differences across place and time. Those factors might include the other differences among states noted above—e.g., in how often they impose death verdicts and how often verdicts get bogged down in the courts—or a variety of other conditions. We turn now to the results of the regression analyses we designed to identify factors that are significantly related to capital reversal rates. B. Results of Main Analysis 1 and Analysis 2 of Reversal Rates at All Review Stages Combined: Aggressive Death Sentencing, Ineffective Law Enforcement, Politics and Race
Main Analysis 1 and our first followup analysis (Analysis 2) consider differences in reversal rates as a proportion of death verdicts imposed by each of the 34 states that used the death penalty in the 23year period between 1973 and 1995 in each of the years when the state imposed at least one death verdict. Each year in which a particular state imposed one or more death verdicts is an observation—because we can observe the reversal rate, over the remainder of the study period, for death verdicts imposed by the state in that year. If a state imposed death verdicts in 18 of the 23 study years, it provides 18 observations, or observed rates of reversible error. All told, the 34 states contributed 519 observed error rates to Analyses 1 and 2. That rate, in these analyses, is the proportion of the death verdicts imposed in the relevant state and year that, during the rest of the 23year study period, were reversed at one or another of the three stages of review—direct appeal in the state supreme court, state postconviction review in a state trial or appellate court or federal habeas review. As we note above, Main Analysis 1 and Analysis 2 differ in their assumptions about patterns of capital reversal rates.^{302} Analysis 1 uses binomial logistic regression to explain differences among rates that range fairly evenly from 0% to 100%. Analysis 2 uses Poisson regression to explain differences among rates that clump towards the low end of the spectrum. As we will see, the results are nearly the same no matter which assumption is made. In our first test of the usefulness of these analyses,^{303} we conduct a baseline inquiry to see how much of the variation in reversal rates among states and years is explained by three general traits of each of the 519 observed reversal rates—(1) the state in which the death verdicts subject to review were imposed, (2) the year in which they were imposed and (3) the general trend of reversal rates over the 23year period. The second trait identifies changes in reversal rates that are associated with particular years; the third trait identifies any linear trend in those rates over all years.^{304} State, year and time trend almost certainly appear to explain differences in reversal rates that in fact are related to more specific conditions that happen to vary from state to state, year to year or over time. Even so, we proceed with confidence to search for those specific explanations only if state, year and trend are weak proxies for more specific explanations for error rates, leaving significant differences to be accounted for. The baseline inquiries for Analyses 1 and 2^{305} reveal that even after state, year and time trend are considered, a significant amount of variance in reversal rates remains among states (as indicated by statistically significant random intercepts, or variation in the state coefficients^{306}), and among years (as indicated by statistically significant slopes, or variation in the year coefficients^{307}). 2. Main Analysis 1 and Analysis 2 identify sets of factors that explain differences in reversal rates and fit the data better than the baseline analyses and other sets of factors. Having found that differences in reversal rates across state and time are significant and are not fully explained by the state and year in which the death verdicts were imposed or the trend of reversals over time, we put Analyses 1 and 2 to four more tests: Whether, based on what is known about the capital system, as well as experience and common sense, we can identify specific explanatory factors that, individually, (1) are significantly related to reversal rates; and that, as a group, (2) do a better job than the baseline analysis of matching the actual reversal rates to those the explanatory factors predict; (3) explain more of the differences across state and time than the baseline analysis; and (4) provide a better explanation than other sets of explanatory factors (again, given their fit with the data and ability to explain differences across state and time).
Analyses 1 and 2—the former using a binomial regression analysis, the latter using a Poisson regression—identify similar sets of explanatory factors that are significantly related to reversal rates. In conducting each analysis, we identified two overlapping but not identical sets of factors that are significantly related to reversal rates—referred to here as Analyses 1A, 1B, 2A and 2B. By significantly related, we mean that:
Before discussing the significant explanations for reversal rates identified by Analyses 1 and 2, we ask whether, beyond each individual condition's statistically significant relationship to reversal rates, there are reasons for confidence in the overall set of explanatory factors the analyses identify.
Our first basis for confidence in the overall explanation Analyses 1 and 2 supply is that each set of factors they generate—Analysis 1A, 1B, 2A and 2BCfits the data significantly better than the explanation provided by the two analyses' baseline inquiries. Fit is measured by the combined distance between each of the 519 actual reversal rates being studied and the reversal rates the set of explanatory factors predicts for each state and year. The higher the measurement, the poorer the fit.309 Fit helps us compare different sets of explanatory factors within the same analysis—each using the same type of regression to explain the same real world condition (here, the 519 particular reversal rates)Cto see which provides the closest match to the known reversal rates. We first compare the baseline analysis to each set of specific factors to see if the explanation for reversal rates the set provides fits the 519 reversal rates by state and year better than the baseline analysis. Then we compare each set of factors to all others to see which matches the actual outcomes the best. Fit measurements change from one type of regression analysis and one topic of study to another and are sensitive to the number of data points being explained. They thus are not assessed on an absolute scale and cannot be directly compared across analyses. A 3000 fit score in one analysis, using one type of regression analysis to study one set of reversal rates, is not necessarily worse than a 2000 score in another analysis, using another type of regression analysis or analyzing a different set of error rates. But if the degree and significance of any improvement in fit as one moves from a baseline analysis to a "best" analysis of multiple factors is greater than is true of another analysis, that may suggest that the former analysis is a better method of explaining the differences being studied (here, variation in capital reversal rates) than the latter. We first ask whether the groups of specific explanatory factors identified by Analyses 1 and 2 fit the data better than the baseline inquiry,^{310} in that they predict reversal rates closer to the actual ones than do analyses based solely on the state and year in which the death verdicts being studied were imposed and the trend of reversals over time. For Analyses 1A, 1B, 2A and 2B, the improvement in fit over the Analysis 1 and 2 baseline inquiries is highly statistically significant.311
As we note above, the Analysis 1 and 2 baseline analyses of state, year and time trend leave significant unexplained variation in reversal rates from state to state and year to year.312 We next consider whether the sets of specific explanatory factors identified by Analyses 1A, 1B, 2A and 2B leave less unexplained variation in reversal rates than both the Analysis 1 and 2 baseline inquiries and other sets of potentially explanatory factors. Statistical analysis is not expected to eliminate all significant unexplained variance. The goal is to eliminate as much as possible. Table 5 compares the coefficients indicating the amount of variance left unexplained by the baseline inquiries for Analyses 1 and 2 to those indicating the amounts of unexplained variance left by Analyses 1A2B. Smaller coefficients are better, indicating that less variance remains unexplained. Significance levels are in parentheses.313
As Table 5 reveals, Analyses 1A, 1B, 2A and 2B achieve the desired drop in the amount of unexplained variance that had remained after accounting for only state, year and time trend:
d. Better combination of significance, fit and explained variance than other groups of conditions. Tables 2 and 3, pp. 13640 above, list the categories of potentially explanatory factors we considered in studying state and county reversal rates, including in Analyses 1 and 2. Appendix E more fully defines the specific factor we discuss in the findings below. The groups of explanatory factors in Analyses 1A, 1B, 2A and 2B improved fit and diminished unexplained variance more successfully than other groups of factors we considered. The next section discusses the explanations for capital reversal rates identified by Analyses 1A2B.
Analyses 1 and 2 compare the rates at which death verdicts imposed in given states and years between1973 and1995 were reversed at one of the three phases of state and federal court review during the period. As we note above,^{314} there are two reasons why reversals as a proportion of imposed death verdicts—the reversal rates studied in Analyses 1 and 2Cmight be lower than otherwise:
These two distinct reasons for lower or higher reversal rates require that we include factors in our analysis that control for the effect of:
Controlling for these nonerrorrelated influences on reversal rates lets us link other significant factors to the presence or absence of serious error. As we discuss above,^{316} we use the year death verdicts were imposed to control for the nonerrorrelated effect on reversal rates of unfinished appeals. Here, we use three factors to control for the similar effect of court delay: the number of death verdicts awaiting review^{317} and two measures of court cases of all kinds awaiting decision. As predicted, these delayrelated factors are negatively associated with rates of imposed verdicts that were reversed by the end of the study period. In other words, they effectively measure the downward effect of unfinished and delayed review on reversal rates calculated as a proportion of imposed verdicts. We discuss delay first, then unfinished appeals. i. Court delay in reviewing capital verdicts. To control for the effect of delay in the review process, we first used a count of death verdicts imposed in each year by each state that had not been finally reviewed in the threestage appeals process by the end of the study period. As predicted, this nonerrorrelated explanatory factor was highly significant and negatively related to reversal rates in all four analyses (1A, 1B, 2A and 2B).^{318} States and years with more death verdicts backlogged in the courts awaiting review had fewer verdicts reversed as a proportion of imposed verdicts. This does not suggest delaying review of death verdicts in order to avoid reversals. States and years with many backlogged, and thus unreviewed, death verdicts had fewer verdicts affirmed as well as fewer reversed, because fewer were reviewed. Recall that Analyses 1 and 2 measure reversals as a proportion of imposed death verdicts. This creates three possible outcomes: the verdict was reversed; the verdict was affirmed; or the verdict was still under review and so was neither reversed nor affirmed. Because of this last option, a drop in reversal rates does not necessarily indicate a rise in affirmance rates. If the number of delayed verdicts rises, affirmance as well as reversal rates will both fall. If the goal is to carry out legally appropriate death verdicts while overturning unreliable ones, delaying review of all verdicts is the worst outcome. It is also probably true, however, that less serious, reversible error in death verdicts almost certainly will mean fewer delays on appeal. This increases the importance, in conducting our analyses, of controlling for the nonerrorrelated effects of court delay and time, so we can be relatively confident that other conditions related to capital reversals are also related to capital error. ii. Delays in all court cases. Although delays in capital appeals certainly keep death verdicts from being reviewed, artificially depressing reversal rates as a proportion of imposed verdicts, it also may be that delays in court cases of all kinds independently impede the progress of death verdicts through their appeals, with the same artificial effect on reversals. To test that possibility, we used two alternative measures of court congestion. In Analyses 1A, 1B, and 2A, we combined four individual indicators of court congestion—civil, criminal, felony and all cases filed and awaiting decision by state courts as a proportion of the state's population—into a single aggregate measure of court congestion.^{319} In Analysis 2B, we used one of the four individual indicators—the total number of filed cases awaiting decision by state courts as a proportion of the state's population.^{320} As predicted, states with high rates of court congestion tend to have lower rates of death verdicts making it through the courts and being reversed, even after controlling for the effect of delays in capital appeals themselves.^{321} In Analyses 1A, 1B and 2A, the combined (fourpart) measure of each state's rate of general court congestion was negatively related to reversals and significant at the .05 level, or at just slightly above that level.^{322} In Analysis 2B, the single measure of rates of court congestion (total cases awaiting decision per the state's population) was also negatively and significantly related to reversal rates.^{323} iii. Year of death verdict. It takes time for courts to discover flaws in death verdicts: By the end of the study period, the average federal habeas reversal was occurring 13 years after the flawed death verdict was imposed.^{324} As a result, the later a death verdict was imposed, the less likely it is that the verdict was fully reviewed—and thus that flaws in it had been discovered—as of the study's cutoff date. More recent verdicts thus will automatically have lower rates of reversal within the study period than earlierimposed verdicts, not because later verdicts are less errorprone, but because later verdicts had less time to be screened for error.^{325} Additionally, reversals take longer on average than affirmances at the federal habeas stage. As a result, a disproportionate number of verdicts yet to be reviewed on federal habeas as of the study cutoff date are flawed. The effect is to artificially depress the reversal rate among all verdicts, but especially among more recent verdicts, because a higher proportion of them had yet to be reviewed on federal habeas as of the cutoff date than is true of earlier verdicts.^{326} In both the baseline inquiries and Analyses 1A2B of specific factors, we included the year death verdicts were imposed.^{327} Doing so controls for the nonerrorrelated effect of unfinished appeals, which are more probable in proportion to how recently the verdict was imposed and which, as we have just seen, depress review rates and thus reversal rates as a proportion of imposed verdicts. As predicted, the later a verdict was imposed—meaning the less time the verdict had to be reviewed for error by the study end date—the less likely it was that the verdict was reversed during the period. Comparing the effect of time in the Analysis 1 and 2 baseline inquiries and in Analyses 1A, 1B, 2A and 2B of specific factors tends to confirm our judgment that the effect of time in the latter analyses is mainly the dampening effect of delay on reversal rates, not the effect of the passage of time in lowering error rates. Recall that we included the passage of time in both our baseline analyses and our analyses of more specific factors. Recall also that in the baseline analyses, the passage of time probably serves as a standin for more specific conditions that tend to increase or decrease in intensity over time. One thus expects the effect of the passage of time (the linear time trend) to diminish as specific factors are added to the analysis—ideally leaving only the nonerrorrelated effect of the study cutoff date to be captured by the time factor. As predicted, the size of the effect of the time trend dropped substantially from the baseline analysis to each of our four analyses of specific factors.^{328} This tends to suggest (as do the results of Analyses 3, 4 and 10 below^{329}) that the effect of the passage of time is mainly the nonerrorrelated effect of the study cutoff date Cdecreased levels of review over time—not the effect of higher quality death verdicts over time. * * * * * The above efforts to isolate the nonerrorrelated effects on reversal rates of court delay and unfinished appeals prepare us to explore the effect of conditions, discussed below, that logic and experience suggest are primarily associated with the amount of serious error. All those factors were significantly associated with reversal rates, improved the fit between predicted and actual outcomes, and reduced unexplained variance.
The effect of race is commonly studied in American criminological research. In the death penalty context, studies suggest that the race of the defendant and especially that of the victim influences the decision whether a death sentence will be imposed for any given homicide—and particularly for homicides in the medium range of aggravation, where neither a life sentence nor a death sentence is the obviously appropriate punishment under law.^{330} What has not been studied is whether the racial makeup of the population generally and of the class of all homicide victims (as opposed to the race of particular defendants and victims) is associated with rates of capital error (as opposed to rates of capitalsentencing). Analyses 1 and 2 fill that void, finding that the racial makeup of both the state's population and its pool of homicide victims—and also the combination of the two—is associated with high rates of capital error. To be clear, we are not attempting to interpret our regression results on demographics causally. Rather, we are attempting to maximize the amount of information we can derive from a study of reversal rates in the 34 states that actively used the death penalty during the study period. With that finite number of states to study, it is useful to compare them in a number of ways that are strongly suggested by prior research, to get an idea of the factors associated with high capital reversal rates. i. Higher proportion of AfricanAmericans in the state population. Controlling for other factors, states with higher proportions of AfricanAmericans in their population tend to have higher rates of capital error.331 The relationship is highly significant in binomial Analyses 1A and 1B, and Poisson Analyses 2A and 2B.332 This result deserves attention. A threshold question is, The race of whom? Is it the proportion of blacks in the population generally that is related to higher capital error rates, as Analyses 1 and 2 suggest? Or, is the crucial factor the proportion of AfricanAmericans among the people the state sentences to die? In other words, do states that sentence AfricanAmericans to die more often than other states have more serious error? Or, alternatively, is the crucial factor the race of the actual defendant in each case? Each of these three relationships might masquerade as the other in a statistical study, and it is important to sort out their effects, because each has different implications. To test for the middle possibility—that there is a relationship between rates of reversible error and the racial makeup of the people states sentence to die or the race of their victims—we conducted subsidiary Analysis 1R. Analysis 1R uses Analysis 1A as a starting point and tests a number of additional factors relating to the race of individuals sentenced to die during the study period, and the race of the victims of the offenses for which the penalty was imposed. Although we tested a number of such factors (listed in the Appendix E), none improved the fit of the predicted and actual results achieved by Analysis 1A^{333} nor appreciably diminished the amount of unexplained variance. Nor was the percentage of AfricanAmericans among states' death row populations significantly related to error rates. Nor did adding factors based on the racial makeup of death row populations undermine the statistical significance of the percentage of AfricanAmericans in the state's general population. In other words, including the racial characteristics of people sentenced to die did not improve the explanatory power of Analysis 1A itself.^{334} We next consider whether the race of the capital defendant in a particular case—as opposed to the racial makeup of either the state's general population or its death row—is related to the probability of error. Table 6, p.159 below, compares reversal rates—at the state direct appeal, federal habeas and both stages combined^{335}—for death verdicts imposed on AfricanAmerican defendants to reversal rates for verdicts imposed on white defendants. (The table also compares reversal rates for death verdicts involving black versus white victims, and for death verdicts involving various combinations of defendants and victims by race. For now, though, we are concerned only with the race of the defendant.) As the table reveals, there is little or no, much less a statistically significant disparity in reversal rates for death verdicts imposed on black and white defendants: At both stages and the two combined, reversal rates for death verdicts mposed on black and white defendants are identical. As our caselevel study below finds, moreover, the race of the defendant has no significant relationship to reversal rates at the federal habeas stage, even when other factors are considered.^{336} Table 6:
Percent of Verdicts Reversed at State Direct Appeal
* In multiplevictim cases, if at least one victim is white, then the victim is classified as white. Sources: DADB, HCDB.
Our finding above stands. Race and capital reversal rates are significantly related, with the key factor being the racial makeup of the state's population. Controlling for other factors, states with higher AfricanAmerican populations relative to the total population have significantly higher rates of serious error than states with smaller AfricanAmerican populations. Before interpreting this relationship, we consider two other significant racial effects based on the victims of homicide. ii. Relatively high risk of homicide to whites as compared to blacks. Research suggests that the race of the victims of homicides affects whether those crimes are punished with death.337 This has led observers to concluded that homicides of white victims put more pressure on prosecutors, judges and jurors to charge and sentence capitally than do homicides of black victims.^{338} These results prompted us to consider whether high rates of homicides against whites are also significantly related to high rates of reversible capital error. One measure of how often whites are victims of homicide is the number of white homicide victims per 100,000 white residents in each relevant state and year—the "white homicide victimization rate." But in states with large proportions of white residents, this rate is essentially the state's overall homicide rate—the number of homicides per 100,000 residents of the state. And with minor exceptions noted below, neither homicide rates generally, nor the white victimization rate by itself, are significantly associated with rates of reversible error in our analyses.^{339} A different measure is needed to determine the extent to which the threat of homicide is concentrated on—i.e., is experienced with particular force by—the white community. This led us to devise an explanatory factor that compares the number of white homicide victims per 100,000 white population to the number of AfricanAmerican homicide victims per 100,000 black population: rate of white homicide victimization ? rate of black homicide victimization.^{340} Because the homicide risk to members of the black community is usually higher than the homicide risk to members of the white community, this factor in effect compares states and years based on how closely the homicide risk to whites approaches the homicide risk to blacks.^{341} In Analyses 1and 2, the more heavily the risk of homicide is concentrated on a state's white community compared to its black community, the higher the state's rate of reversible capital error.^{342} As with the percent of each state's population that is AfricanAmerican,^{343} we examined the race of homicide victims to see whether the effect we found was from the racial makeup of the state's overall pool of homicide victims (as our results thus far would suggest), or from the race of the victims of people sentenced to death in the state (Are states that sentence more people to death for killing whites than for killing blacks more prone to capital error?), or from the race of the victims in individual capital cases (Are capital verdicts for homicides against white victims more likely to be reversed than verdicts for homicides against other victims?) Our results here are similar to those in the preceding section. Analysis 1R, compares states on:
These conditions are only intermittently significant; when significant, they are sometimes positively and sometimes negatively related to reversible rates; and they do not improve the capacity of the analysis to fit the data and decrease unexplained variance. And in each case, adding these factors increased the significance of our original raceofvictim factor which compares state rates of white and black homicide victimization. Analysis 1R thus gives no basis for confidence that the percentage of whitevictim homicides that result in death verdicts has any relationship to reversal rates, but it does support our finding that states where whites face a higher risk of homicide relative to the risk faced by blacks have significantly higher rates of serious capital error. Table 6, at p. 159 above, further supports the conclusion that it is a state's overall proportion of white versus black homicide victims, and not some other category of white and black homicide victims, that is associated with high rates of reversible capital error. The table shows that the fact that the victims of particular homicides are white does not appear to be related to reversal rates for death verdicts imposed for those homicides. At the state direct appeal and the federal habeas stage combined, and at the direct appeal stage itself, the reversal rates for whitevictim and nonwhitevictim homicides are identical, or nearly so. At the federal habeas stage alone, death verdicts imposed for homicides against white victims do have a higher reversal rate (42%) than capital verdicts imposed for homicides against nonwhites (34%). But our Analysis 19 below finds that this relationship is not statistically significant when other factors are considered.^{345} iii. The interaction of a larger AfricanAmerican population and a relatively high risk of homicide to whites as compared to blacks. Both the proportion of AfricanAmericans in a state's population and the extent of the homicide risk to the state's white as compared to its black population are racial factors. But is there reason to think these two significant racial explanations for error rates are related, or are part of a single explanation for error rates? To help answer this question, we asked whether the interaction of the two factors is significantly related to rates of serious capital error. Where both racial conditions are high, are error rates especially high, and where both are low, are error rates especially low? More specifically, if we multiply the values for each of the two conditions, is the resulting value significantly associated with capital error rates, above and beyond the effect of the two factors themselves? Analyses 1, 1R and 2 reveal that the interaction of the percent of a state's population that is black and the state's white homicide victimization rate relative to its black homicide victimization rate is significantly associated with rates of reversible capital error.^{346} States with both a high proportion of blacks in their population and a high concentration of homicides of whites relative to blacks tend to have especially high rates of reversible capital error—above and beyond the positive effect on error rates of each component of that interaction and other significant effects.^{347} Again, it is the racial characteristics of the state's overall population and pool of homicide victims that has this effect, not the racial characteristics of the state's death row population and their victims, nor the race of individual capital defendants and victims. No analogous effect was found in Analysis 1R for the interaction of the race of offenders who were sentenced to die and the race of the victims of their offenses—although controlling for such factors did slightly increase the significance of the relationship between capital error rates and the interaction between the percent of the state's population that is black and the state's rate of white, relative to black, homicide victimization. And, as Table 6, p.159 above, reveals, death verdicts imposed on black defendants for homicides against white victims are not noticeably more or less likely to be reversed than other death verdicts. iv. Preliminary interpretation: racebased pressures to punish homicides with death. A final interpretation of our results, including that capital error rates are significantly associated with the percent of the state's population that is black, the state's rate of white compared to black homicide victimization and the interaction of the two, must await a full presentation of all the results. A preliminary interpretation is set out here. Three possible explanations for the relationship between capital error rates and the size of a state's AfricanAmerican population must be rejected at the outset, because available evidence rules them out. The first explanation notes that homicide defendants and victims in states with large AfricanAmerican populations tend to be AfricanAmerican themselves, and hypothesizes that trials with black defendants or victims are more likely to be poorly run and unreliable given the defendant's or victim's low status or outright discrimination against them and their community. As we have seen, however, reversible error is not more probable when the defendant is black than when he or she is white, and the same goes for the race of the victim.^{348} A second hypothesis is that large black populations mean high rates of participation in death cases by black judges, prosecutors, defense lawyers and jurors, and that they commit more error than white trial participants. Here, the evidence contradicts both the premise and the inference from it. AfricanAmerican representation among states judges, prosecutors and defense lawyers is extremely low throughout the country, including in states with high percentages of AfricanAmerican residents.^{349} Black participation on juries that impose death sentences is also low, for several reasons:
Participation by AfricanAmerican judges, prosecutors, defense lawyers and jurors in imposing the death penalty—and thus in committing capital error—appears to be too uniformly small across states, therefore, to account for the large amounts of error in such cases and wide differences among states. In addition, there is no evidence to suggest that capital trial participants of one race are any more prone to error than members of another race. On the contrary:
Low black support for the death penalty also undermines a third hypothesis—that pressure from the black community to use the death penalty accounts for high rates of capital error. Although such pressure would theoretically be higher in states with larger black populations, and might have an effect even where most trial participants are white, there is no evidence that such pressure exists, and considerable evidence that such pressures are much more likely to come from the white community. Even if blacks did press for use of the death penalty, the available evidence suggests that law enforcement officials are relatively unresponsive to pressures from the black community.^{357} The evidence contradicting these hypotheses suggests a different explanation for the significant association between capital error rates and the three racial factors identified by Analyses 1 and 2. All three racial factors tend to generate crime fears among members of politically influential communities. And those fears may generate pressure on officials to extend their use of the death penalty to weaker, more marginal cases where the need to cut corners to obtain capital convictions and sentences is greater. Longstanding racial stereotypes and prejudices have created a stubborn association in the minds of some between blacks and violent crime, especially violent crime against whites.^{358} The larger a state's AfricanAmerican minority,^{359} therefore, the more fear of violent crime some members of the majority may feel, and the more pressure politically influential members of that group may generate to use the death penalty as a protective measure.^{360} Second, a fair gauge of the threat of homicide felt by politically influential communities—and the best gauge for which evidence is available in all states—is whether and to what extent homicides committed in the state actually pose as much of a threat to the white, as to the black, community. As the level of white homicide victimization approaches or surpasses the level of black homicide victimization, pressure to use the death penalty may increase as well. Finally, if both racial forces have the suggested effect, states with both forces operating at once—high black populations and high rates of white, relative to black, homicide victimization—would likely generate especially heavy pressures to use the death penalty as a protective measure. If people with political influence feel not only that they live in a highcrime environment and are often the victims of violent crime, but, in addition, that they might be targeted for crime by members of a different group, pressure to increase the use of the death penalty might be especially high. This explanation is provisional, for now. If it is credible, capital error should also be associated with one or more of the following conditions:
We consider below whether our study results provide any evidence of these processes by which crime fears lead to increased use of the death penalty which, in turn, lead to increased rates of capital error. c. More aggressive use of the death penalty and higher capitalerror rates. As Figure 11, pp. 121 above, demonstrates, states vary widely in how often homicides resulted in death verdicts in the study period. Colorado, Connecticut, Maryland, New Jersey, New Mexico and Washington had 7 or fewer death verdicts per 1000 homicides; and California and Louisiana had around 10 per 1000 homicides. By contrast, Alabama had around 37 death verdicts per 1000 homicides; Arizona, Delaware, Nevada and Oklahoma had around 45 per 1000 homicides; and Idaho had 60 per 1000 homicides. Above, we note that a number of states with above average deathsentencing rates per 1000 homicides carry out below average proportions of the death verdicts they impose—and, conversely, that states with low deathsentencing rates often have high execution rates.^{361} Together with Figure 9, pp. 79 above—revealing a correlation between high error rates and low execution rates—these results suggest a relationship between high capitalsentencing rates and high error rates. Evidence that high rates of death verdicts per 1000 homicides may be related to high rates of reversible capital error was first detected by Cornell Law Professors John Blume and Theodore Eisenberg. Using Bureau of Justice Statistics (BJS) data for 198595, they found that "[t]he rate at which states impose [death] sentences strongly correlates with the rate at which relief was obtained from those sentences."^{362} After we replicated this twofactor correlation with the BJS data that Blume and Eisenberg used, and with our own data covering the longer period from1973 to 1995,^{363} we decided to include Blume and Eisenberg's explanatory factor in our multiple regression analyses^{364} to see if the twofactor relationship between high deathsentencing and high error rates persists when other factors are simultaneously considered. Analyses 1A, 1B, 1RA, 1RB, 1RC, 2A and 2B all reveal a highly significant relationship between high deathsentencing rates and high rates of reversible capital error. Because the relationship is between two rates (death verdicts per 1000 homicides, and reversals per 100 death verdicts), it is not a foregone conclusion—as it would be if it instead compared raw numbers. Although states with larger numbers of death verdicts (e.g., states with larger populations) should have larger numbers of reversals, because more cases to review should mean more reversals, there is no reason to presume that all death verdicts from particular states (e.g., populous states) are more likely to contain reversible error than all verdicts from other states.^{365} Yet, that is what Analyses 1 and 2 find: Controlling for other factors, states that are more likely to impose death verdicts per 1000 homicides are more likely to have the verdicts they impose reversed due to serious error. And states that are less likely to impose death verdicts per 1000 homicides are less likely to have their verdicts reversed. States with a propensity to impose death sentences are prone to serious capital error. We already have suggested a reason for this relationship:^{366} The more homicides officials treat as capital, the more often they may sweep in marginal cases where it is necessary to cut corners and commit other kinds of errors to obtain death verdicts. We explore this hypothesis further below. d. Poorer record of arresting and punishing serious criminals and higher capitalerror rates.
Originally, we hypothesized that states with high rates of incarcerated individuals would have high rates of serious capital error. We accordingly designed a factor comparing the number of people incarcerated in each state's prisons each year (a number influenced by how many people the state arrests, convicts and imprisons) to the number of FBI Index Crimes committed in that state and year. Index Crimes are a composite of property and violent crimes (including homicide) the FBI uses to measure crime rates.^{367} In all seven analyses (1A, 1B, 1RA, 1RB, 1RC, 2A and 2B), this factor indeed has a highly significant relationship to capital error rates. But the relationship runs in the opposite direction from what we predicted: States with higher prison populations relative to the number of serious crimes — i.e., states that apprehend, convict and imprison more of their serious criminals—have lower rates of reversible capital error than states that arrest, convict and imprison fewer serious criminals. This suggests that states with relatively more effective noncapital responses to crime—i.e., arrest, conviction and imprisonment—may be under less pressure than states with weaker law enforcement records to use the death penalty. And that in turn may dampen the penalty's use in weak cases in which the temptation to use unreliable procedures is high. e. Heavier political pressure on state judges and higher capitalerror rates. Is there any evidence that political pressure on capital officials influences capital error rates? To answer this question, we had to identify differences among states in the political pressure capital officials experience, to see if the differences are associated with differences in states' capital error rates. Two actors in the capital system face political pressures formal enough to measure: district attorneys^{368} and state judges. Both are subject to legally defined selection procedures that make them vulnerable to political pressure from the officials who appoint them or the voters who elect them. District attorneys do not provide a good point of comparison, however, because most are selected the same way—partisan elections every four years.^{369} This leaves too little statetostate variation to compare to differences in capital reversal rates. State judges, by contrast, are selected in a variety of ways that create the needed variability. In a few states, judges are appointed by other officials and never face direct elections. In other states, judges are appointed to long terms and can be removed only by recall elections triggered by fairly onerous petition requirements. In other states, judges are appointed to shorter terms, after which they face periodic retention elections. And in other states, judges are directly elected from the beginning in contested elections, for either longer or shorter terms, in either nonpartisan or partisan elections. These possibilities create a continuum of selection methods ranging from those placing less to those placing more political pressure on judges to conform their rulings to the desires of politically influential groups. Using these criteria, we developed two measures of political pressure on judges in different states to see if either has a significant relationship to rates of serious capital error. In all analyses under discussion, the amount of political pressure on state judges is related to higher error rates, and the relationship is highly significant. The two indexes of political pressure had similar effects.^{370} Death verdicts imposed at trials run by state judges who are subject to relatively more direct political pressure are more likely to be seriously flawed than verdicts presided over by judges facing less political pressure. f. Other factors related to higher capitalerror rates. We also hypothesized that demographic factors, and factors gauging the quality of the criminal justice system such as expenditures and caseloads, might affect rates of reversible error. Two factors were significant: population structure and capital and other caseload burdens on the courts. i. Higher population size and density. In research on the spatial patterning and structural covariates of county homicide rates, Professor Steven F. Messner and colleagues at the University of Albany and the University of Illinois developed a combined measure of population structure that compares jurisdictions based on their overall population and population density.^{371} This factor, adapted for use at the state level, was positively associated with error rates and significant in all seven analyses.^{372} In these analyses of reversal rates at all three stages of judicial review, states with larger populations that are more highly concentrated in cities tend to have higher reversal rates than less heavily and densely populated states. ii. Courts more burdened by capital and noncapital caseloads. As is noted above, Analyses 1 and 2 find that large numbers of capital verdicts awaiting court review, and general court congestion, are each associated with low capital reversal rates—as one would expect, given that court delay decreases the amount of court review, which in turn lowers rates of reversals per imposed verdicts.^{373} We also studied the interaction of the two types of court congestion, expecting it to coincide with especially low reversal rates.^{374} The interaction is indeed related to reversal rates, and the relationship is highly significant in all analyses, but its direction is the opposite of the one we had predicted. States whose courts are congested with large numbers of capital verdicts being reviewed and high rates of court filings tend to have high capital reversal rates. The positive association between reversal rates and the interaction of capital and noncapital court congestion suggests that a combination of many pending capital and noncapital cases may overwhelm courts, increasing the number of serious mistakes made in trying capital cases. 4. The size of the effects associated with each significant explanatory factor. a. Effect size generally. With enough observations, even tiny relationships may be statistically significant. This leads us to ask whether the significant relationships between capital reversal rates and the explanatory factors described above are large and important enough to deserve attention in, for example, crafting reform proposals. Statistical analysis lets us do this by estimating effect size—the predicted rise or fall in capital reversal rates associated with a given change in the amount or intensity of each significant explanatory condition. We use two methods of displaying effect size—one numerical, the other in graphs. The numerical estimates are reported in our detailed results in Appendix G, coded as "newestimates." The graphs are displayed below, starting with Figure 22A, p. 175 below. The different interpretation of these effectsize estimates in binomial regression analyses (Analyses 1 and 1R) and Poisson analyses (Analysis 2) is discussed above.^{375} Recall that, when using a particular analysis to predict the change in capital reversal rates associated with a given change in the amount or intensity of the explanatory condition under discussion, the graphs hold all other explanatory factors in the analysis constant at their average. As we explain above, the reversal rates Analyses 1 and 2 predict are reversals as a proportion of all imposed verdicts, not as a proportion of only reviewed verdicts. This enlarged base number of verdicts (i.e., the enlarged denominator) enables our statistical analyses to make the best use of all our data.^{376} But by inflating the base number (denominator) to include unreviewed as well as reviewed verdicts, these analyses deflate predicted reversal rates (the number of reversals divided by the larger base of all imposed verdicts) below the actual reversal rates (the number of reversals divided by actually reviewed verdicts).^{377} Particularly given that the rates presented in these graphs are systematically depressed by their enlarged denominators, the best use of the effect size graphs is comparative. Rather than predicting the actual reversal rates we know to have occurred— e.g., 68% nationally over the 23year study period—the graphs are best used to indicate the percent increase or decrease in reversal rates associated with specified increases or decreases in the amount or intensity of significant explanatory factors. In this way, we can determine whether statistically significant factors are worth attention based on whether changes in them are associated with appreciable changes in capital reversal rates, holding other factors constant. On each effectsize graph, the horizontal (x) axis plots the value of the explanatory factor, and the vertical (y) axis plots the predicted reversal rate associated with that value (for binomial analyses) or the predicted change in the reversal rate associated with that value (for Poisson analyses). The range of values displayed on the horizontal axis for each explanatory factor is the actual range of values for that factor among states and years in our study. Thus, the rise or fall in predicted reversal rates (holding other factors constant) that the graph displays is the rise or fall across the actual spectrum of 34 capitalsentencing states and 23 years in our study. (The minimum, average and maximum values for each explanatory factor in each analysis are listed in Appendix F1.) In most cases, the graphs divide the range of values on the horizontal axis into seven equal intervals. As a check on the extent to which different sets of factors within the same analyses, and different analyses (binomial vs. Poisson analyses), make a difference in effect sizes, we proceed factor by factor, examining the size of its effect in all four analyses, with four graphs displayed on the same page. The steeper the line on the graph, the larger the effect size. Lines that get higher (meaning predicted reversal rates increase) as they move from left to right (as the explanatory factor increases) indicate that the explanatory factor is associated with higher reversal rates. Lines that get lower as they move from left to right indicate that the explanatory factor is associated with lower reversal rates. We do not report effect sizes for interaction effects, which are more complicated to interpret.^{378} b. Factors unrelated to the quality of death verdicts. i. Court delays in reviewing capital verdicts. Figures 22AD, p. 175 below, display the size of the effect on reversal rates of large backlogs of unreviewed capital verdicts in, respectively, Analyses 1A, 1B, 2A and 2B. As is noted above, this factor serves mainly to measure the nonerrorrelated, downward effect on reversal rates (as proportions of imposed verdicts) that delay causes by depressing rates of review.379 Figures 22AD indicate that, as backlogs of capital verdicts awaiting review increase, the number reviewed—and thus the number available to be and that actually are reversed—decreases sharply. At some point—about where the backlog of unreviewed verdicts reaches 20Cthe system appears to shut down, with virtually no cases being reviewed or reversed.^{380} This suggests that as the number of death verdicts awaiting review increases, they so clog the appellate system that it ceases to function as a means of moving valid death verdicts forward to execution and for diverting flawed verdicts back for retrials. In that event, unclogging the system would require fewer death verdicts, fewer flaws demanding extended review, or both. Displaying the effect size graphs for all four analyses on the same page reveals that they are fairly similar, as is true for most factors found significant by Analyses 1 and 2. Effect sizes for our main, binomial regressions tend to be larger than those for the Poisson regressions (Analysis 1 vs. Analysis 2), but the differences are modest, as are differences among analyses with slightly different groupings of explanatory factors (Analysis 1A vs. 1B vs. 2A vs. 2B). ii. Delays in all court cases. Figures 23AC, p. 176 below, show the size of the statistically significant effect on capital reversal rates of the combined fourfactor measure of undecided court cases of all types, holding other factors constant. Figure 23D shows the same for the alternative measure of court caseloads that uses only one of those four measures of general court caseloads. In both cases, effect size is too small to warrant additional attention to these alternative measures of general court cases—illustrating the usefulness of effectsize inquiries in distinguishing among explanatory factors all of which are statistically significant.^{381} Figure 22 AD: Effect Size: Backlog of Capital Verdicts (Analyses 1A, 1B, 2A, 2B) Figure 23 AD: Effect Size: State Court Caseloads (Analyses 1A, 1B, 2A, 2B) iii. Year of death verdict. As is discussed above, court review of death verdicts takes several years at each of three review stages, so that almost no death verdicts imposed in the last years of the study period were reviewed at any, much less at all, review stages by the end of that period.^{382} More generally, the later a verdict was imposed, the less likely it is that the verdict was reviewed and reversed as of 1995 at any and especially at the later review stages.^{383} This holds true for seriously flawed as well as highquality verdicts. At the federal habeas stage, the effect is magnified for flawed verdicts, which take longer to review than valid verdicts.^{384} The year the death verdict was imposed serves as a control for this nonerrorrelated, downward effect on reversal rates of unfinished appeals.^{385} If the expected decline over time in reversal rates (as a proportion of imposed verdicts) occurs—as it does in Analyses 1A2BCthat decline indicates that the factor is serving its purpose as a control for the dampening effect on review and thus reversal rates of unfinished appeals. But the decline conveys little information about changes in the probability of reversible error over time.^{386} Figures 24AD, p. 178 below, display the size of this effect as predicted by Analyses 1A2B. Because that effect is fairly small —a decline in reversal rates of about 1 or 2 percentage points per year in Analyses 1A and 1BCand because that decline is less steep than the dropoff in completed appeals indicated by Table 4, p.142 above, these graphs suggest the possibility that increases in flawed death verdicts over time, after accounting for other factors, may be neutralizing some of the dampening effect on reversal rates of unfinished appeals. If that is so, we should find increases in reversal rates over time (controlling for other factors) in the regressions discussed below387 in which reversal rates are calculated as proportions of actually reviewed, not all, verdicts, because those regressions avoid the nonerrorrelated effect of unfinished appeals.388 Our direct appeal regressions indeed reveal just that increase in reversal rates over time, accounting for other factors.^{389}
c. Factors that appear to be related to the quality of death verdicts. i. Larger AfricanAmerican population. Figures 25AD below indicate that increases in the proportion of AfricanAmericans in a state's population are associated with considerable percentage increases in predicted capital error rates. During the study period, the AfricanAmerican proportion of states' population ranged from .25% ( Montana, 1978) to 36% (Mississippi, 1975), and averaged about 14% in all 34 capital states and 23 years. Binomial Analyses 1A and 1B (Figures 25A and 25B, p. 180) below, predict that, holding other factors constant at their average, reversal rates increase more than 4fold when the black proportion of the population rises from its lowest level among and states and years in the study to the average level for all states and years; doubles when the black proportion of the population rises from 5% to 35%; and increases about 8fold across the entire spectrum of AfricanAmerican populations in the study. Poisson Analyses 2A and 2B (Figures 25C and 25D) predict about 4fold increases across that spectrum. ii. Relatively high homicide risk to whites as compared to blacks. Figures 26AD, p.182 below, estimate the size of the effect on capital error rates of the rate of white as compared to black homicide victimization.^{390} In the state and year with the highest whitetoblack victimization rate (New Mexico, 1987), the number of white homicide victims per 100,000 whites in the population was just slightly greater than the number of black homicide victims per 100,000 blacks in the population. In the state and year with the lowest rate (Utah, 1974), the number of white homicide victims per 100,000 whites was about 5% that of the black victimization rate. Again, the effectsize graphs enable us to gauge the predicted rise in reversal rates across this spectrum of comparative homicide rates for whites and blacks, holding other factors constant at their averages:^{391} Other things being equal, where whites and blacks face the same risk of being killed by homicide, Analyses 1A and 1B predict that the rate of serious capital error is about twice as high as when the homicide risk to white citizens is only 5% of the homicide risk to black citizens. The predicted increase in reversal rates is about 67% when the homicide risk faced by whites increases from 10% of the homicide risk faced by AfricanAmericans to the same risk. Poisson Analyses 2A and 2B (Figures 26C and 26D) predict somewhat smaller increases. Figure 26 AD: Effect Size: Risk of Homicide to Whites Versus Blacks (Analyses 1A, 1B, 2A, 2B) iii. More aggressive use of the death penalty. Analyses 1 and 2 find that higher deathsentencing rates are linked to higher capital error rates.^{392} Across states and years in the study, death verdicts per 1000 homicides varied from about 1 (e.g., Georgia, 1995; Pennsylvania, 1979) to 208 (Idaho, 1982), and averaged about 23.5 for all states and years. As shown in Figures 27AD, p. 184 below, all four analyses predict large increases in capital reversal rates as deathsentencing rates rise, holding other factors constant. Binomial Analysis 1A (Figure 27A) predicts:
Poisson Analysis 2A (Figure 27C) predicts a quadrupling of reversal rates over that same spectrum of deathsentencing rates. As noted, the predicted reversal rates shown here are not true error rates— reversals as a proportion of reviewed death verdicts—but, instead, rates of reversals as a proportion of imposed verdicts (whether or not the verdicts finished being reviewed for error), which are systematically lower than the true error rate.^{393} That, even so, predicted reversal rates rise from 13% to over 75% across the spectrum of deathsentencing rates in the study indicates the sizeable impact of deathsentencing rates on reversal rates. Together with the clogging effect of large numbers of capital verdicts awaiting review,^{394} the sharply higher reversal rates linked to higher deathsentencing rates begin to suggest that less is more. Singling out fewer—only the most egregious—homicides for capital charges and sentences may mean higher quality death verdicts, fewer reversals, less frequent and shorter delays on appeal and a more efficient system overall. Figure 27 AD Effect Size: State Death Sentencing Rate (Analyses 1A, 1B, 2A, 2B) iv. Weaker record of apprehending and imprisoning serious criminals. Death verdicts from states with fewer prisoners per 100 serious crimes are significantly more likely to be overturned than verdicts from states with higher rates of prisoners per serious crime.^{395} States and years range from just about 1, to about 13, prisoners per 100 FBI Index Crimes, averaging about 5.^{396} Figures 28AD, p. 186, display the size of this effect, which again is large. Everything else equal, binomial Analysis 1A (Figure 28A) predicts capital error rates (per imposed verdicts) of about 75% where there is only 1 prisoner per 100 Index Crimes—but only about 36% where there are 4 prisoners per 100 Index Crimes, and only about 13% for the highest number of prisoners per 100 Index Crimes among states and years in the study. Poisson Analysis 28C predicts a fivefold drop in predicted error rates across the same spectrum. These analyses predict that states with better records of arrest, conviction and incarceration—i.e., with more effective alternatives to capital punishment that may relieve pressure to sentence capitally in close cases—do a better job of avoiding flawed capital verdicts than states with weaker law enforcement records. v. More political pressure on state judges. Starting in the 1950s with Supreme Court Justice Felix Frankfurter, observers have noted that death penalty cases put officials under intense pressure to use the penalty.^{397} If that is so, and if the pressures are not always congruent with the quality of the evidence favoring a death verdict—as may occur when the crime is horrible, but the evidence implicating a suspect is weak—capital error may result. As is noted above, we tested this hypothesis by asking whether error rates are related to two alternative gauges of political pressure placed on judges by states' distinct methods of selecting judges. Each gauge rates state's judicial selection methods based on a number of traits that may increase the amount of political pressure judges face to conform rulings to the outcome most voters or other politically influential individuals would prefer.^{398} The first index measures 9 such traits. Study states had anywhere from 2 to all 9 traits, averaging 6.5. The second index measures 8 traits, with states varying between 2 and all 8, and averaging 5.7. Supporting the above hypothesis, one or the other political pressure index was significantly and positively related to error rates in all of our Analysis 1 and 2 analyses.^{399} Figures 29AD, p.188 below, show that increases in the amount of political pressure on state judges are linked to considerable increases in the probability of reversible capital error. In binomial Analyses 1A and 1B (Figures 29A and 29B), states with selection methods putting the least political pressure on judges have expected reversal rates (as a proportion of imposed verdicts) of about 16%. States with selection methods putting the most political pressure on judges have 3 times that expected rate of error. The comparable increase in Poisson Analyses 2A and 2B (Figures 29C and 29D) is between 2 and 3 times. Effect size is similar for the first (9criteria) and second (8criteria) political pressure indexes. These results may provide a reason to use methods to select judges that emphasize professionalism and limit political pressure, or a caution about a penalty that is susceptible to political pressure and resulting mistakes. Figure 29 AD: Effect Size: Method od Selecting State Judges (Analyses 1A, 1B, 2A, 2B) vi. Higher population size and density. The remaining significant explanatory factor for which effect sizes may be calculated is the size and density of the states' population in the relevant years. The combined scale of population and density is built around a national average score of 0, with highly and densely populated states having scores ranging up to 1.75, with sparsely populated states having scores ranging down to 2.40, and with states averaging slightly above 0 (.175) during the 23year study period. More heavily and densely populated states have significantly higher reversal rates in these analyses of all three review stages than do more sparsely populated states. As shown by Figures 30AB, p. 190 below, binomial Analyses 1A and 1B, predict 4 to 5fold increases in reversal rates when this factor is varied from its lowest to its highest value represented by states in our study, holding other factors constant. The corresponding increases in Poisson Analyses 2A and 2B (Figures 30C, 30D) are 2 to 3fold. Figure 30 AD: Effect Size: Population Structure (Analyses 1A, 1B, 2A, 2B) d. Summary: evidence of the role of race, politics and zeal for the death penalty. Except for general court caseloads, all factors that are significantly related to higher rates of serious capital error and for which effect sizes can be calculated have appreciable effects. This is true of:
The listed factors may all warrant attention from policy makers seeking solutions to high capital reversal rates. Before adopting this conclusion, we consider whether our other analyses support them.
