The Tyrant s Death: Supreme Court Retirements and the Staying Power of Judicial Decisions Stuart Minor Benjamin and Georg Vanberg Introduction When a Supreme Court Justice is replaced, commentators and academics generally focus on the ways in which the Court s future opinions are likely to be affected by the Court s altered membership (e.g. Baum 1992, Segal 1985). A replacement of a Justice by someone who appears to have a very different ideological orientation seems significant, particularly so if the change moves the Court s median Justice. 1 The replacement of a Justice by a new Justice who is expected largely to agree with her predecessor, meanwhile, is of modest interest, precisely because it is not expected to significantly change future case outcomes (e.g., Ruckman 1993, Shipan and Shannon 2003). In contrast to this perspective, which emphasizes the impact of retirements on future Supreme Court decisionmaking, we investigate the manner in which turnover on the Court affects the impact and staying power of past Supreme Court decisions. In the case-based American legal tradition, the influence and importance of particular judicial decisions derives in large part not from the immediate resolution of a dispute, but rather from the way in which an opinion influences and shapes decisions in related future cases. The puzzle with which we are concerned is whether the retirement of Supreme Court Justices matters for the staying power of majority opinions these justices joined. Put succinctly, controlling for ideological shifts on the Supreme Court, do lower court judges react to judicial retirements by taking a more critical view of past decisions when Justices who had been in the majority retire? Benjamin is the Douglas B. Maggs Professor of Law, Duke Law School (benjamin@law.duke.edu); Vanberg is a Professor, Department of Political Science, Duke University (georg.vanberg@duke.edu). Tim Calloway, Sean Chen, and Doug VanDerwerken provided tremendous assistance in creating the dataset. Jane Bahnson, Brittany Edwards-Franklin, Andrew Jennings, Kelly Leong, Cassidy Nolan, and Benjamin Oster provided valuable research assistance. 1 One important question in the literature concerns whether appointments will influence future decisions only when they move the median (Moraski and Shipan 1990) or whether appointments can have a significant impact even if they do not affect the Court s median (Carrubba et al. 2012).
As a matter of legal doctrine, the precedential value of a Supreme Court majority opinion is unaffected by changes in the Court s membership. The precedent is still the precedent, and the jurisprudence remains the same. Indeed, a central element of precedent for continuing bodies like the Supreme Court is that a past decision remains authoritative even if the specific Justices who created the precedent leave the Court. Thus, the Supreme Court has stated forcefully that lower courts should not reject applicable Supreme Court decisions based on their assessment of a potential change in the Supreme Court s direction. As Justice Kennedy stated for the Court in Hohn v. United States (524 U.S. 236), (o)ur decisions remain binding precedent until we see fit to reconsider them, regardless of whether subsequent cases have raised doubts about their continuing vitality (253). 2 Notwithstanding pronouncements like that in Hohn, there is a longstanding normative debate among legal academics about whether appeals courts should predict how the contemporary Supreme Court will respond to one of its precedents. Some reject the Supreme Court s position that lower courts should not do so, and others offer support for the Supreme Court s position (Dorf 1995, Caminker 1994, Posner 1990, Bhaghwat 2000, Bradford 1990). Moreover, there are, of course, strong theoretical reasons to suspect that lower court judges will, in fact, pay attention to the preferences of the current Supreme Court in their decisions (Cameron, Segal, and Songer 2000, Kastellec 2007). In this paper, we hope to advance this debate by providing a positive understanding of whether, and to what extent, circuit judges change the manner in which they treat Supreme Court decisions in response to one specific and important kind of preference change on the Supreme Court: The retirement of justices whose votes were critical to these decisions. 2 See also Rodriguez de Quijas v. Shearson/American Express, Inc., 490 U.S. 477, 484 (1989) (The Supreme Court noted that one of its opinion was inconsistent with an earlier opinion, but stated flatly: We do not suggest that the Court of Appeals on its own authority should have taken the step of renouncing [the earlier case]. If a precedent of this Court has direct application in a case, yet appears to rest on reasons rejected in some other line of decisions, the Court of Appeals should follow the case which directly controls, leaving to this Court the prerogative of overruling its own decisions. ); State Oil v. Kahn, 522 US 3, 20 (1997) ( Despite what Chief Judge Posner aptly described as Albrecht v. Herald Co. s infirmities, [and] its increasingly wobbly, moth-eaten foundations, there remains the question whether Albrecht deserves continuing respect under the doctrine of stare decisis. The Court of Appeals was correct in applying that principle despite disagreement with Albrecht, for it is this Court's prerogative alone to overrule one of its precedents. ) (Internal citation omitted); Agostini v. Felton, 521 U.S. 203, 237 (1997) (requiring the Courts of Appeals to adhere to the Supreme Court s directly controlling precedents, even those that rest on reasons rejected in other decisions); Hutto v. Davis, 454 U.S. 370, 375 (1982) ( [U]nless we wish anarchy to prevail within the federal judicial system, a precedent of this Court must be followed by the lower federal courts no matter how misguided the judges of those courts may think it to be. ). 2
To do so, we begin by explaining why retirements of Supreme Court justices are likely to affect how appeals courts treat past Supreme Court decisions. Then we turn to a unique dataset that traces the treatment of all Supreme Court majority opinions by courts of appeals on a yearly basis from 1953 to 2012. These data allow us to investigate whether appeals court judges respond systematically to the retirement of Supreme Court justices in their treatment of Supreme Court opinions. We find strong evidence that, controlling for a wide variety of factors known to affect the treatment of Supreme Court precedent, including the general ideological shift of the Supreme Court, the retirements of justices whose votes were critical to a majority significantly shape the manner in which these decisions are treated. These findings are significant because they suggest that the effects of turnover on the Supreme Court go well beyond affecting how the Court decides future cases. Retirements also shape the manner in which appeals courts implement existing precedents. These effects are particularly important in light of the fact that the Supreme Court itself hears only a very small number of cases far too few to have a significant direct effect on the resolution of legal disputes. Instead, the Court s impact comes primarily through its control of the judicial hierarchy, that is, its ability to direct how lower courts resolve the vast bulk of cases that never reach the Court. If the manner in which appeals courts apply and use existing precedent changes simply as a function of retirements on the Supreme Court, the impact of judicial turnover on our legal system is potentially far greater than has traditionally been recognized. The title of this article borrows from Soren Kierkegaard s observation that [t]he tyrant dies and his rule is over; the martyr dies and his rule begins. 3 Insofar as appellate judges treat Supreme Court opinions more critically once the justices who joined a precedent become a minority on the Court, Supreme Court opinions are less like continuing precedents, and more like a ruler s decisions that lose force as soon as the ruler loses power. Justices may think of themselves as creating precedents that will long outlive them, but their opinions may instead diminish in importance as soon as the majority that created them ceases to exist. 3 Soren Kierkegaard, The Soul of Kierkegaard: Selections from his Journals (edited by Alexander Dru). 3
Supreme Court Retirements and the Treatment of Opinions Broadly speaking, there are four ways in which lower court judges can engage Supreme Court precedents that are relevant to a case they are currently considering. The first and presumably expected way is to explicitly follow the precedent. This is the strongest positive signal that lower court judges can send with respect to a specific opinion. A second is to cite the opinion without indicating a negative or positive reaction to it (e.g., a bare citation noting the existence of the case without discussing it). Third, judges may simply ignore the opinion and make no mention of it at all. Finally, lower court judges can have an explicitly negative response to the precedent, by casting doubt on it or limiting its implications. Such a negative treatment is a strong, and potentially risky, signal: A lower court s negative reaction to a Supreme Court precedent risks antagonizing a Supreme Court protective of its opinions and prerogatives. 4 As we noted above, as a doctrinal matter, judicial retirements should be irrelevant to how lower courts treat Supreme Court opinions: As long as the Supreme Court has not explicitly overruled its own precedent, lower courts are bound to apply it faithfully. But the incentives and constraints that affect how lower court judges (particularly judges on the federal courts of appeals, which sit immediately below the Supreme Court) choose among the ways in which they can respond to relevant precedents change when the Supreme Court s composition changes. To see why, it is useful to distinguish between two analytically separate support coalitions for a given majority opinion. First, there are the Justices who joined the opinion when it was issued. We refer to these Justices as the decision coalition. Second, there are those Justices on the contemporary Court who are broadly in support of a past decision (or at least favor disciplining lower courts who fail to faithfully apply the precedent). We refer to these Justices as the ongoing support coalition for that opinion. Importantly, it is the contemporary Supreme Court that polices lower court compliance with existing precedent, and retirements potentially affect an opinion s ongoing support coalition. As Justices who were members of the decision coalition retire or die, and are replaced by new Justices, the bargaining environment on the Supreme Court is reshaped: The loss of Justices who were in the decision coalition may make it easier to construct a new 4 In our empirical analysis below, we focus on the two strong signals that circuit judges can send with respect to a relevant precedent: Explicitly following it, or explicitly criticizing it. We do so because mere string citations without explicit discussion send no clear positive or negative signal, and because failure to cite an opinion raises a methodological challenge: It is difficult comprehensively to identify the Supreme Court precedents that should have been cited for tens of thousands of lower court cases. 4
coalition of Justices who disagree with at least some aspects of a decision, and prefer moving the law in a different direction. Importantly, this may be true even if Justices are replaced by ideologically similar new Justices (i.e., if there is little general ideological change on the Supreme Court as captured, for example, by such metrics as the Martin-Quinn (2002) scores): If there are aspects of a Justice s position in a case that are idiosyncratic, and not captured by more general ideological considerations, a new Justice may see the case differently. The fact that judicial retirements may erode the ongoing support coalition for a decision has implications for how judges on the courts of appeals are likely to approach Supreme Court precedents that are relevant to a case they are currently considering. The appeals courts stand in a hierarchical relationship vis-à-vis the Supreme Court, and one important reason for appeals court judges to adhere to Supreme Court precedents is the potential for review and reversal (Miceli and Cosgel 1995, Posner 1993, Epstein and Knight 2013). 5 As a result, if appeals court judges perceive that support on the Supreme Court for a particular decision may have weakened, two potential effects follow. First, circuit judges who disagree with an opinion may feel less constrained in expressing their disagreement with it or, less aggressively, in finding ways to avoid following the precedent without explicitly criticizing it. Second, even appeals court judges who have no qualms about a precedent may anticipate how the contemporary Supreme Court views the precedent. If they believe that the current Court is likely to be critical of its own precedent, appeals court judges may become convinced that (ironically) enthusiastic adherence (i.e., explicitly following the precedent) may increase the chances of review and reversal by the Supreme Court. Put differently, the anticipation that the Supreme Court may look upon its precedent with less sympathy may encourage circuit judges to engage in an explicit critique or, more modestly, to refrain from following the precedent fully (Dorf 1995). There are, of course, a number of signs that may lead appeals court judges to suspect that Supreme Court support for a decision has weakened, and to adjust their own treatment of an opinion in response. Most obviously, circuit judges may react to explicit negative treatments of an opinion by the Supreme Court itself (a factor we control for in our empirical analysis). The focus of our current argument, however, is the potential impact of judicial retirements. 5 Of course, there are also district courts below the courts of appeals. But district court judges opinions are appealed to circuit courts (each of which has its own circuit jurisprudence to which stare decisis applies within the circuit). As a result, for district judges, the immediate reviewer is the circuit court, and review by the Supreme Court is a more distant concern. We therefore focus on circuit judges, for whom the only reviewer is the Supreme Court. 5
Retirements of Supreme Court justices who were critical to the establishment of a precedent can serve as a simple heuristic for circuit judges who are trying to anticipate the possible loss of support for a decision on the Supreme Court. Consider a circuit judge confronting a precedent for which at least five Justices from the original decision coalition remain on the Court. It is likely that our judge will perceive that attacking or criticizing the opinion will antagonize a majority of the Supreme Court. Similarly, she may also perceive that explicitly following the precedent may help to insulate her decision against rebuke from the Supreme Court. But once enough Justices from the decision coalition have left the Court such that the remaining members constitute only a minority on the contemporary Supreme Court (i.e., when there are four or fewer Justices left on the Court who joined a given majority opinion), the situation has changed after all, a majority of the Justices on the current Supreme Court have not signed on to the precedent, and may see it differently. 6 Our circuit judge may perceive that a more critical approach to a precedent is less risky, and less likely to result in rebuke. Similarly, she may feel under less pressure to engage in explicit positive treatment of the decision; indeed, refraining from explicitly treating a precedent positively might be the safest course of action for some imperiled Supreme Court precedents. This argument leads to our first two hypotheses: Hypothesis 1 ( Retirement Effect on Negative Reactions ): Ceteris paribus, courts of appeals judges are more likely to treat a Supreme Court precedent negatively once the Justices in the opinion s decision coalition constitute only a minority on the Supreme Court. Hypothesis 2 ( Retirement Effect on Positive Reactions ): Ceteris paribus, courts of appeals judges are less likely to treat a Supreme Court precedent positively once the Justices in the opinion s decision coalition constitute only a minority on the Supreme Court. 6 Put simply, the clearest signal of a Justice s support for a given majority opinion is her joining of it. A Justice who arrives after the opinion is decided cannot give that signal. Note in this regard that in only four cases in its history has the Supreme Court overruled one of its cases without a change in membership (Gerhardt 2005: 952). 6
While we expect that circuit judges will use retirements as a simple proxy for the likely support for a decision, thus giving rise to the consequences summarized in hypotheses 1 and 2, not all opinions are subject to this retirement effect to the same extent. Consider the contrast between a unanimous decision and a 5-4 decision. We would expect that the retirement effect is stronger for the 5-4 decision than for the unanimous decision. The reason is straightforward. The fact that an opinion attracted unanimous support suggests that it has broad appeal to Justices across the ideological and jurisprudential spectrum. The same is not true of the 5-4 decision, which rests on a more narrow foundation. Ceteris paribus, we would therefore expect that it is more likely that newly appointed Justices will take a critical view of the 5-4 decision than of the unanimous opinion. As a result, we expect that as the decision coalition for an opinion becomes smaller, court of appeals judges are more likely to take the loss of the decision coalition s majority status on the Court as a proxy for a loss of support for an opinion. This leads to the expectation that appeals court judges will be more willing to criticize opinions with small decision coalitions than those with large decision coalitions once the decision coalition s majority status is lost. Similarly, we expect that the drop-off in positive treatments is greater for more narrow decisions than for decisions supported by a broader coalition: Hypothesis 3 (Interactive Effect): The retirement effect is more pronounced the smaller the original majority coalition. Data and Measures To measure the subsequent treatment of Supreme Court majority opinions by the courts of appeals, we rely on Shepard s Citations, an approach that has become standard in the literature (e.g., see Benjamin and Desmarais 2012, Corley 2009, Hansford and Spriggs 2008, Westerland et al. 2010). Shepard s is a widely used, commercial legal research company that employs attorneys who examine every published state and federal court opinion, and engage in a content analysis of every citation in those opinions. Citations that have a substantive treatment of an opinion (i.e., discuss it rather than just mentioning it) are classified into the following main categories: Overruled, Questioned, Limited, Criticized, Distinguished, Explained, Harmonized, or Followed. Shepard s characterizes Overruled, Questioned, Limited, Criticized, and Distinguished as negative treatments. Of these, the first four are viewed as strongly negative, while 7
Distinguished is classified as a mildly negative treatment. Followed the most common treatment is characterized as a positive treatment. 7 Finally, Explained and Harmonized are neutral. 8 Because our argument leads to the expectation that retirements of key Justices will lead circuit courts to be more ready to negatively treat a precedent and less ready to positively treat it, we focus on negative treatments and on Followed. In so doing, we measure the two most significant sets of signals an appeals court can send. To construct our data, we begin with all orally argued cases in which the Supreme Court issued a signed majority opinion in the 1953-2012 terms, as reported by the US Supreme Court Database (Spaeth et al. 2013). 9 We then use Shepard s to generate an annual count of the number of negative treatments for each of these Supreme Court opinions in courts of appeals opinions issued between 1953 and the end of 2012. In other words, the unit of observation in the dataset is the opinion-year, organized by Supreme Court opinion (beginning with the year in which the opinion was issued), a structure that makes sense given our interest in how opinions are treated over time. Finally, we combine these data with information on the retirement dates of the Justices who constituted the majority in each case, allowing us to construct the measures discussed in more detail below. The full dataset comprises 5,813 majority opinions and 188,424 opinionyears. As we noted above, a negative treatment of a Supreme Court opinion by a lower court is a costly signal and a risky move for a circuit judge. Instead of simply citing or ignoring an earlier case, the lower court states that it is casting doubt on (in the case of Criticized and Questioned) or avoiding the implications of (in the case of Distinguished and Limited) a relevant precedent. By explicitly minimizing the application of a Supreme Court opinion, the circuit court draws attention to itself and risks rebuke. Not surprisingly, such treatments are generally rare. As the aggregate data displayed in Table 1 make clear, most opinions receive fewer than three negative treatments, with only 26% of opinions receiving more than seven negative treatments. Strongly negative treatments are even rarer roughly 83% of opinions are never criticized in this explicit 7 In addition to Followed, Shepard s also identifies a treatment called Paralleled that is characterized as positive. However, this treatment is exceedingly rare so rare, in fact, that in the courts of appeals data we employ below there is not a single instance of a court of appeals treatment that is classified as Paralleled. 8 Spriggs and Hansford (2000) investigate the reliability of Shepard s Citations by independently coding citations for a stratified sample of Supreme Court cases citing earlier Supreme Court cases, finding high levels of agreement between their coding and Shepard s coding. 9 These are those opinions coded as decisiontype=1 in the Supreme Court Database. 8
fashion, and only 2% of opinions receive 4 or more strong negative treatments. By contrast, positively treating a precedent that is on point is the expected and obvious way for circuit courts to engage with Supreme Court opinions. As a result, it is not surprising that Followed is a far more common substantive treatment. Note, however, that even for Followed, more than 50% of opinions receive seven or fewer such treatments. Not surprisingly, the fact that in the aggregate, substantive treatments are relatively rare (for negative treatments) and small in number (for positive treatments) implies that once we break the data down into yearly treatments of Supreme Court opinions a step necessary to evaluate changes in treatments over time most opinions receive either no substantive treatments in each of these categories during a given year, or perhaps one or two. Rather than model the count of each kind of treatment during a given year, we therefore generate binary variables that indicate whether an opinion received at least one negative/strong negative/positive treatment during a given year. These indicators are the key dependent variables for our analysis. Table 1: Number of Negative and Positive Treatments of Supreme Court Majority Opinions in the Courts of Appeals, 1953-2012 # of Treatments All Negative Treatments Strong Negative Treatments Positive Treatments None 1,107 (19%) 4,795 (83%) 682 (12%) 1-3 1,959 (34%) 894 (15%) 1,367 (23%) 4-7 1,223 (21%) 71 (1%) 1,063 (17%) 8 or more 1,524 (26%) 53 (1%) 2,701 (48%) N 5,813 5,813 5,813 Key Explanatory Variables The central argument we wish to evaluate is that judicial retirements that suggest to appeals court judges that a majority of Supreme Court Justices may no longer support a given 9
precedent will induce appeals court judges to show less respect for these opinions, that is, to treat them more negatively and less positively. As explained above, we expect this retirement effect to have a pronounced impact after enough Justices have retired so that the remaining members of the original majority coalition constitute only a minority on the Court. To capture this effect, we employ a binary variable ( Majority Gone ) that indicates those years in which a sufficient number of Justices from the majority supporting a given opinion have retired so that only a minority of Justices in the original majority remains on the Court. For example, for a 5-4 decision, this variable indicates all years following the retirement of at least one of the Justices in the 5-member majority. For a 6-3 decision, it indicates years following the retirement of at least two majority Justices, and so on. Our expectation is that opinions are more likely to be treated negatively and less likely to be followed in the years following these key retirements. Moreover, as outlined above, we expect this retirement effect to increase in strength as the original opinion is supported by more narrow majorities because new appointments to the Court are all else being equal more likely to upset the support for 5-4 decisions than for decisions carried by larger majorities. This implies that the retirement effect is strongest for 5-4 decisions, and declines as there are more votes in the original majority. To test this expectation, we include an interactive effect between the Majority Gone indicator, and the number of additional votes beyond 5 in the original majority (the original majority size can vary from 5 to 9). The coefficient for the Majority Gone indicator therefore provides the estimate of the retirement effect for 5-4 decisions, and the effect for decisions with larger majorities is calculated by adding the coefficient on Majority Gone and the (appropriately multiplied) coefficient on the interaction effect. Given our hypotheses, the coefficient on the Majority gone indicator should be positive, coupled with a negative coefficient on the interaction term. Control Variables Of course, there are a number of other factors that are likely to affect the treatment of Supreme Court opinions by appeals courts, and we need to control for these in our analysis. We can usefully group these into characteristics of the cited Supreme Court opinion that may directly affect how a decision is received, and dynamic factors that capture changes in the legal environment (relative to the environment in which the original decision was issued) that may affect how a decision is treated. 10
As previous work has shown, the most important characteristics of Supreme Court opinions that affect their subsequent treatment are the number Justices who signed on to the opinion, the ideological range of the majority coalition, and the number of concurrences published alongside the opinion (Benjamin and Desmarais 2012, Westerland et al. 2010). The intuition for why these characteristics matter is immediate: One would expect that appeals courts are less likely to treat negatively and more likely to treat positively opinions that announce legal rules that are broadly acceptable and sensible to judges and lawyers across the ideological and jurisprudential spectrum. The three characteristics indicate opinions for which this is likely to be the case: The greater the number of Justices who sign on to an opinion, and the broader the ideological range of that coalition (which we measure by the range of the majority in terms of Martin-Quinn (2002) scores, the most widely-used, dynamic estimate of the Justices preferences), the more broadly acceptable the decision is likely to be. We thus expect these variables to reduce the likelihood of negative treatments and increase the likelihood of positive treatments during a given year. In contrast, a greater number of concurring opinions signals that there are alternative legal justifications for the outcome reached in a case, which signals that the opinion is less broadly acceptable. We thus expect the number of concurrences to increase the likelihood of a negative treatment and to decrease the number of positive ones. Prior work provides evidence consistent with these expectations. For example, Benjamin and Desmarais (2012) show that aggregate negative treatments of opinions increase with the number of published concurrences, and decrease for larger majority coalitions, and for broader coalitions. Similarly, Spriggs and Hansford (2001) find that the probability that an opinion will be overruled rises in the number of concurrences. 10 The second set of control variables aims to capture the nature of the current legal environment relative to the Supreme Court that issued the citing opinion. As we outlined in the theoretical section, one reason for appeals court judges to treat an opinion more negatively and less positively is the belief that the legal environment has changed in such a way that the current, contemporary Supreme Court is more critical of the cited opinion. Although the Supreme Court stresses that appeals court judges should not anticipate how the current Court might approach a 10 In contrast, in their random sample of 500 Supreme Court opinions, Westerland et al. (2010) find that the number of concurrences has a small positive impact on the probability of compliance. However, they do not discuss this finding. 11
case, appeals court judges have strong reasons to anticipate the preferences of the current Court. Accordingly, we control for the relative shift in the location of the median Justice on the Supreme Court that issued the cited opinion and the current Supreme Court, measured as the absolute distance between the Martin-Quinn scores of these Justices. We expect that as this distance increases, so will the likelihood of a negative treatment, whereas positive treatments would decrease. Indeed, previous work has found evidence consistent with this expectation. Benjamin and Desmarais (2012) show that the aggregate number of negative treatments increases with the average distance between Supreme Court medians, and Spriggs and Hansford (2001) have demonstrated that the Supreme Court is more likely to overrule a precedent the greater the ideological distance between the decision-median and the contemporary median (see also Hansford and Spriggs 2008). Similarly, Westerland et al. (2010) show that increasing distance between the median member of the decision coalition and the contemporary Supreme Court median significantly reduces compliance with precedents by circuit court panels. In addition to the ideological shift of the Supreme Court, we control for several factors that may affect the likelihood of a negative or positive treatment. Negative treatments are more likely, and positive treatments less likely, after the Supreme Court itself has treated an opinion negatively, thus sending a signal to appeals courts that an opinion is open to criticism (Westerland et al. 2010). This effect should get stronger the more often the Supreme Court has treated an opinion negatively. We thus include a count of the number of times an opinion has been treated negatively by the Supreme Court. Similarly, if an opinion is explicitly overruled by the Supreme Court, we would of course expect appeals courts to treat this opinion more negatively and less positively (Benesh and Reddick 2002). We control for this possibility by including a binary variable that indicates opinion-years after an opinion has been overruled. Because opinions may become less relevant over time, we include the age of the opinion. Finally, we must take account of the fact that opinions do not have equal opportunities to be treated negatively or positively. Opinions vary in the extent to which they are pertinent to the issues currently before the courts of appeals. Some decisions are relevant in many cases, and thus are cited frequently. Others are more obscure and cited infrequently. Naturally, an opinion that is relevant in many cases, and is cited often, has a higher likelihood of being treated negatively or positively than an opinion that is hardly ever relevant, and therefore has few opportunities to be treated negatively or positively. To capture this, we include the (logged) count of the number of 12
times an opinion is Followed in the courts of appeals. For our analysis of negative treatments, we use the number of Follows during the current year; for the analysis of positive treatments, we use a one-year lag (since we obviously cannot include the contemporaneous value). 11 Estimation and Results Because our dependent variable is a binary indicator of the presence of at least one negative or positive treatment of a Supreme Court opinion in a given year in an opinion issued by a court of appeals, we employ a logit model to test our hypotheses. In doing so, we must confront an important complication. Although the covariates introduced above are intended to control for a number of factors that in addition to judicial retirements are likely to affect the likelihood that an appeals court will treat an opinion negatively or positively, it is likely that we confront unobserved heterogeneity in the data: Some opinions are simply more likely to be treated negatively or positively for reasons that we have not (and perhaps cannot) control for explicitly. For example, the particular constellation of facts in a specific case may make it more likely that subsequent cases will be distinguished from the opinion. We can imagine other idiosyncratic features of an opinion that might affect the reception of the decision by circuit courts. There are two widely-used options for confronting the challenge of unobserved heterogeneity: estimating either a random or fixed effects model (grouped, in our case, by Supreme Court opinion). The advantage of the fixed effects approach is that it can deal with unobserved heterogeneity that is correlated with covariates that are included in the model. For example, we might expect that particular constellations of case facts may affect subsequent treatment of an opinion and the nature of the majority coalition that emerges in a case. A fixed effects specification avoids by design any omitted variable bias linked to unobserved factors that make particular opinions more susceptible to negative or positive treatments. Of course, estimating a fixed effects model comes at a price. Because the model must estimate a far larger number of parameters, and can only make use of within-opinion variation rather than also exploiting across-opinion variation to estimate the effect of the covariates in the model, a fixedeffects model is less efficient than a model without fixed effects. Perhaps more importantly, 11 To be precise, we use ln(# of contemporary follows+1) in order to deal with the fact that some opinions are not treated at all during some years. The logic of employing a one-year lag in the analysis of positive treatments is that an opinion that is highly relevant to cases this year is also likely to have been highly relevant to many cases last year. 13
because it only uses within-opinion variation, the fixed effects model cannot estimate the effects of time-invariant covariates, nor can it make use of opinions for which the dependent variable does not vary across time. This implies that the fixed effects model can only be estimated on a subsample of our data (those opinions for which the dependent variable varies over time), and that fixed opinion characteristics (the number of majority votes, the range of the majority coalition, and the number of concurrences) cannot be included in the model. 12 In contrast to the fixed effects approach, a random effects model is able to incorporate information on variation across opinions. As a result, the model can be estimated on a sample that includes opinions for which the dependent variable does not vary across time, and it is possible to derive estimates for the effects of time-invariant covariates. The disadvantage of the random effects approach is that (unlike the fixed effects model) the model assumes that any omitted variables are independent of the covariates included in the model, an assumption that may be violated in our data if any omitted variables affect not only the likelihood of negative or positive treatment, but are also correlated with other covariates included in the model (e.g., the nature of the majority coalition). Faced with this trade-off, we estimate both a random and a fixed effects model. As we discuss below, the results are highly similar under both modeling approaches. Results: Negative Treatments We begin our analysis by investigating the likelihood of a generic negative treatment during a given year, i.e., the likelihood that an opinion receives any kind of treatment classified as negative by Shepard s. In table 2, we report results from a random effects (column 1) and fixed effects (column 2) logit model, predicting the presence of at least one negative treatment of a given Supreme Court opinion during a given year in a courts of appeals decision. Before focusing on the effects of the key explanatory variables, note that the control variables are in line with the existing literature for both models. Greater distance (as measured in Martin-Quinn scores) between the Supreme Court median at the time of the decision and the contemporary 12 For the analyses we perform below, the fact that a fixed effects model can only be estimated on the subsample of opinions for which the dependent variable varies implies that instead of the full complement of 5,813 opinions (all of which we can use for the random effects models), we can estimate the fixed-effects models for only 4,706 opinions when we analyze all negative treatments, 1,018 opinions when we analyze strongly negative treatments, and 4,790 opinions when we analyze positive treatments. 14
Supreme Court increases the likelihood of a negative treatment. Opinions are more likely to be treated negatively by the appeals courts as the Supreme Court has treated them negatively, and once they are overruled. Cases that are relevant to many circuit court cases (i.e., followed often in a given year) are more likely to garner a negative treatment. The time-invariant covariates (which can only be estimated in the random effects model) are also largely in line with expectations: Opinions signed by larger majorities are less likely to be treated negatively, and opinions accompanied by concurrences are more likely to be treated negatively. Interestingly, opinions are more likely to be treated negatively the greater the range of the coalition that supported the opinion, a finding in tension with previous results (see Benjamin and Desmarais 2012). For the random effects model, the estimate of the parameter ρ indicates that roughly 26% of the variance in negative treatments can be explained by the random (opinion-level) component of the model. Turning to our main variables of interest the impact of the retirement effect captured by the Majority Gone variable and its interaction with the number of votes for the original opinion we find strong evidence consistent with our hypotheses in the random effects and the fixed effects models. The coefficient on the Majority Gone variable (which provides the estimated effect for a 5-4 decision) is positive, indicating that opinions are more likely to be treated negatively after the decision coalition loses its majority status (Hypothesis 1). The coefficient on the interaction term is negative, which indicates that the retirement effect diminishes for opinions signed by larger majorities (Hypothesis 3). Because this is an interactive effect, it is difficult to interpret the total effect and its statistical significance directly from these estimated coefficients. To help with interpretation, we therefore turn to a graphic approach that illustrates both the substantive impact, and the uncertainty surrounding our estimates. 15
Table 2: Negative Treatments of Supreme Court Majority Opinions in the Courts of Appeals, 1953-2012 Random-Effects Logit Majority Gone 0.105 (0.038) Majority Gone x # of Extra Votes -0.030 (0.013) Fixed Effects Logit 0.112 (0.040) ** -0.034 (0.014) ** Shift in SC Medians 0.230 (0.021) # of Majority Votes -0.119 (0.016) Range of Majority 0.031 (0.010) # of Concurrences 0.280 (0.021) Previous Negative Treatment 0.108 (0.016) Case Overruled 0.782 (0.094) # of Contemporary Follows (log) 1.037 (0.015) Opinion Age -0.042 (0.001) Constant -1.564 (0.107) σ u 1.067 (0.016) 0.269 (0.023) 0.068 (0.015) 0.522 (0.102) 0.828 (0.017) -0.042 (0.001) ρ 0.257 (0.006) N of opinion-years 188,424 154,148 N of opinions 5,813 4,706 BIC 120,695 93,362 AIC 120,843 93,292 Note: Dependent variable is presence of at least one negative treatment in a court of appeals decision in a given year. significant at 0<0.01, ** significant at p<0.05, * significant at p<0.10. 16
Figure 1: Average marginal effect of retirement on predicted probability of at least one negative treatment, conditional on size of original majority Figure 1 plots the average marginal effect of losing the decision coalition s majority status on the predicted probability of a negative treatment (holding the random intercept at 0), conditional on the size of the original majority supporting an opinion, along with 95% confidence intervals. 13 The results are clearly consistent with our hypotheses. For 5-4 decisions, the predicted probability of a negative treatment increases from roughly 0.12 to about 0.13 in the period after judicial retirements reduce the original majority to a minority (and increase of 8%), and this difference is statistically significant (as can be seen by the fact that each point estimate is not contained in the confidence interval of the other). For 6-3 and 7-2 decisions the retirement effect remains statistically significant, but diminishes in size, as predicted by Hypothesis 3. For 8-1 decisions, as well as for unanimous opinions, there is no statistically discernible impact of losing the majority on negative treatments. Of course, these differences in predicted probabilities are in an absolute sense small. At the same time, it is worth bearing in mind that negative 13 Given that the results are similar across the fixed effects and random effects specification, we use the results form the random effects model to generate these predictions since this model is estimated on the larger sample, and includes time-invariant covariates. We obtain these estimates from the margins command in STATA 13. 17
treatments represent a highly risky, costly strategy for circuit court judges, and that these are predicted probabilities of a negative treatment for a specific decision in a specific year. More importantly, perhaps, this analysis may understate the impact of the retirement effect on the willingness of appeals courts judges to criticize a Supreme Court opinion. So far, we have modeled the probability of any negative treatment of a given Supreme Court opinion in a given year, including instances in which an appeals court distinguishes its case from a relevant Supreme Court precedent. But we might be concerned that the decision by an appeals court to distinguish a case is qualitatively different from engaging in explicit criticism. Shepard s characterization of Distinguished as mildly negative (as opposed to strongly negative) reflects that distinguishing a precedent is a less hostile act, and often reflects the application of sound legal criteria to an attempt at stretching a precedent. Sometimes litigants invoke a Supreme Court holding that, in light of its facts, does not squarely apply to the case at hand, usually in the hopes that the lower court will decide to apply the holding expansively say, arguing that a given precedent s protections for suspects in custodial interrogations should be applied to interrogations beyond those the Court had previously recognized as custodial, and that the Court had never confronted. Depending on how close the circumstances in the new case were to that in the Supreme Court precedent, a lower court s distinguishing of the precedent might not be best seen as a negative reaction indeed, it might reasonably be seen as the most faithful reading of the precedent. For this reason, we now turn to a second set of analyses, in which we focus on only those treatments that are labeled as strongly negative by Shepard s, i.e., treatments that clearly criticize a Supreme Court opinion. In Table 3, we present the results from our random and fixed effects models, where this time the dependent variable is the presence of at least one strongly negative treatment of a given opinion by an appeals court during a given year. Once again, the results are similar across the two estimation methods. The control variables continue to have the expected effects, although the shift in Supreme Court medians loses statistical significance. Turning to the main variables of interest, the loss of the decision coalition majority s status has a significant, positive effect on the predicted probability of a strongly negative treatment, as Hypothesis 1 predicts. Moreover, as the negative coefficient on the interaction indicates, this retirement effect decreases in size for opinions carried by larger majorities. 18
Table 3: Strongly Negative Treatments of Supreme Court Majority Opinions in the Courts of Appeals, 1953-2012 Random-Effects Logit Majority Gone 0.723 (0.125) Fixed Effects Logit 0.739 (0.135) Majority Gone x # of Extra Votes -0.058 (0.045) Shift in SC Medians 0.018 (0.063) -0.099 (0.049) 0.033 (0.075) ** # of Majority Votes -0.086 (0.044) ** Range of Majority 0.011 (0.020) # of Concurrences 0.150 (0.043) Previous Negative Treatment 0.135 (0.029) Case Overruled 3.345 (0.145) # of Contemporary Follows (log) 0.699 (0.037) Opinion Age -0.036 (0.003) Constant -5.361 (0.316) σ u 1.308 (0.043) 0.091 (0.029) 2.531 (0.170) 0.448 (0.047) -0.037 (0.003) ρ 0.342 (0.015) N of opinion-years 188,424 37,272 N of opinions 5,813 1,018 BIC 18,084 10,939 AIC 17,962 10,879 Note: Dependent variable is presence of at least one strongly negative treatment in a court of appeals decision in a given year. significant at 0<0.01, ** significant at p<0.05, * significant at p<0.10. 19
In line with our expectation that the inclusion of Distinguished in the previous analyses may have diluted the impact of the retirement effect, the substantive impact of the effect is significantly more pronounced for strongly negative treatments. This can be readily seen in Figure 2, which plots the average marginal effect of losing the decision coalition s majority status on the predicted probability of a strongly negative treatment during a given year. For 5-4 decisions, the predicted probability of a strongly negative treatment more than doubles. The effect diminishes as the original opinion garnered more votes, but remains statistically significant across the range of majority sizes. Even for unanimous opinions, the probability of a strongly negative treatment during a given year increases by 60% after retirements reduce the decision coalition to minority status. Figure 2: Average marginal effect of retirement on predicted probability of at least one strong negative treatment, conditional on size of original majority Results: Positive Treatments As in the case of our analysis of all negative treatments, the absolute size of the retirement effect on the probability of a strong negative treatment is small. Given how risky an 20
open critique of a relevant Supreme Court precedent is for a circuit judge, this is perhaps not surprising. Explicit negative treatments represent an extreme form of trying to limit the impact of a relevant precedent and are, in this sense, much like the tip of the proverbial iceberg, which is only an indicator of a much larger mass of ice that is visible only beneath the waves. Rather than engaging in overt resistance, circuit court judges who wish to avoid the implications of a relevant precedent, and feel freer to do so once the decision coalition s majority status has been lost, are more likely to engage in a less confrontational style of noncompliance refraining from treating a precedent positively (Hypothesis 2). We now turn to an analysis of this phenomenon by focusing on an analysis of Followed treatments by circuit judges. In Table 4, we present the results of our random and fixed effects models, predicting the presence of at least one Followed for a given Supreme Court Opinion during a given year. Before turning to the key variables of interest, note that the control variables largely continue to have the expected impact. Most importantly, increasing ideological distance between the enacting Supreme Court and the contemporary Supreme Court results in a lower likelihood of a positive treatment that is, circuit judges are less likely to explicitly follow a precedent if the Court that enacted the precedent is ideologically distant from the current Supreme Court. Similarly, opinions that have been explicitly overruled by the Supreme Court are less likely to be followed. Interestingly, opinions supported by broader coalitions both in terms of the number of majority votes, as well as the ideological range of the decision coalition are less likely to be followed, a finding that stands in contrast to negative treatments. Of course the main variables of interest are the indicator of lost majority status for the decision coalition, and the interaction between this variable and the number of additional majority votes. Once again, the results clearly support our expectations. Once the decision coalition no longer constitutes a majority on the Supreme Court, an opinion is less likely to be followed during a given year, as indicated by the negative coefficient on Majority Gone. Moreover, the retirement effect decreases for opinions supported by larger coalitions, as the positive coefficient on the interaction term shows. As before, we turn to a graphic illustration to demonstrate the substantive impact of this effect. In Figure 3, we plot the average marginal effect of losing the decision coalition s majority status on the predicted probability of at least one Followed treatment during a given year. 21