Strategic Pursuit of Goals in the U.S. Courts of Appeals

Like judges more generally, federal circuit judges pursue a variety of goals. These goals can include making policy in accordance with their ideological preferences, preserving or enhancing professional reputation, and efficiently disposing of cases (Reference BaumBaum 2009; Reference Epstein, Landes and PosnerEpstein, Landes, and Posner 2013; Reference KleinKlein 2002). In addition to pursuing such goals directly, without reference to other actors, judges may act strategically to maximize such goals in light of the anticipated behavior of other relevant actors in the judicial hierarchy. For example, circuit judges may strategically seek all three goals listed above by pursuing a fourth goal of avoiding reversal on appeal (Reference CrossCross 2005; Reference Klein and HumeKlein and Hume 2003; Reference SmithSmith 2006). Avoiding reversal may or may not have independent significance to a judge, but it does serve to maximize the other three goals. Being overturned on appeal negates policy influence, potentially damages reputation, and creates additional work. While it may be possible for a judge to maximize all goals simultaneously, most of the time a judge must strike a balance amongst multiple goals. Exploring when and why circuit judges balance the anticipated actions of others against their own goals is important to understanding judicial decisionmaking in the federal circuits. This study focuses on examining how circuit judges balance the twin goals of following their own ideological preferences and avoiding reversal by the full circuit.

Although the federal circuit courts are subordinate to only one court, circuit judges face potential reversal from two sources. The vast majority of appeals filed in the federal circuit courts are resolved by randomly assigned panels of three judges (Reference Collins and MartinekCollins and Martinek 2011; Reference KimKim 2009). Such panel decisions may be reviewed either by the U.S. Supreme Court or by the full circuit in an en banc proceeding. The empirical evidence of circuit judges acting strategically with respect both sources of potential reversal is mixed (Reference AtkinsAtkins 1972; Reference Blackstone and CollinsBlackstone and Collins 2014; Reference Bowie and SongerBowie and Songer 2009; Reference Hettinger, Lindquist and MartinekHettinger, Lindquist, and Martinek 2004). Following the example of existing work, I focus on only one of the two potential sources of review. Specifically, this article examines panel behavior in light of the possibility en banc review. Reference Van WinkleVan Winkle (1997) finds that judges who are in the ideological minority in a circuit are less likely to vote according to their ideology when they have fewer allies in their circuit. In a similar vein, scholars have found that panel minority judges who are empowered through an ideological alliance with the majority of their circuit are more likely to dissent (Reference Blackstone and CollinsBlackstone and Collins 2014; Reference Cross and TillerCross and Tiller 1998; Reference KimKim 2009). However, Reference Hettinger, Lindquist and MartinekHettinger, Lindquist, and Martinek (2004) fail to find such evidence of strategic dissent.

Changing one's vote is a high-cost compromise that might rarely (if ever) be justified in light of the low probability of en banc review (Reference Bowie and SongerBowie and Songer 2009). Crafting an opinion offers innumerable opportunities for making much smaller concessions that may, nevertheless, help obviate review. The few studies that look beyond votes provide some evidence that lower court judges craft the text of their opinions to reduce the likelihood of reversal. There is evidence from administrative law cases that circuit judges strategically select legal grounds that are more difficult to overturn on appeal when facing potential review from an ideological opponent (Reference Smith and TillerSmith and Tiller 2002). Another study demonstrates that federal district judges utilize more hedging language, which makes conclusions more difficult to falsify, when they are more ideologically distant from the reviewing circuit (Reference HinkleHinkle et al. 2012). Federal circuit judges use Supreme Court precedent with an eye to how closely the current Supreme Court is aligned with a particular precedent (Reference WesterlandWesterland et al. 2010). However, Spriggs and Boyd (2009) fail to find a connection between the ideology of the reviewing circuit and how federal district judges use U.S. Supreme Court precedents. Like these studies, I tap into the rich informational content available in the text of judicial opinions. Specifically, I look for strategic balancing of panel ideology and reversal avoidance in the way precedents are used in panel opinions.

Data and Research Design

Studying judges’ decisions regarding which precedents to cite requires identification of not only a set of opinions to study, but also a relevant choice set of potentially applicable precedents (Reference NiblettNiblett 2010). I focus on a subset of cases in one particular issue area: Fourth Amendment search and seizure law. This topic is well-suited for this study because it incorporates a discrete set of legal issues that are routinely raised in litigation, and relevant cases can be identified by the simple expedient of finding cases that cite the Fourth Amendment of the United States Constitution.Footnote ³ Using Lexis, I collected every such published circuit case from 1953 to 2010. After excluding all opinions that do not address the merits and all opinions that do not contain the word “search” or “seizure” (or derivations thereof) at least once, the resulting dataset contains 13,334 cases. I analyze citation patterns in the panel opinions from 1990 to 2010 (n = 7,299).Footnote ⁴ The cases in the dataset from 1953 to 1989 and all en banc cases are used only as precedents that may be cited.

There are many types of legal authority a judge may cite including case law from a variety of sources. The legal doctrine of stare decisis assigns different levels of significance to different types of authority. To hold the effect of legal doctrine as constant as possible, I focus on the use of one particular type of binding precedent, published opinions from within the same circuit. For each opinion, this choice set potentially includes every precedent in the dataset from its own circuit from 1953 up to the day before the opinion was issued.Footnote ⁵ However, even within one issue area many precedents will remain uncited simply because they raise different legal questions. To solve the fundamental problem of identifying those precedents (among thousands of cases) that are so dissimilar that citation is exceedingly unlikely, I turn to the field of machine learning.

When you type a query into Google, the results are produced based on an algorithm used to automatically rank how similar the text of a query is to each website. Using a basic version of such technology, I calculate how similar each opinion is to every binding precedent preceding it in the dataset.Footnote ⁶ These similarity scores for each opinion are then sorted to create a ranking of all possible precedents from least similar to most similar. For the sake of computational efficiency, the scoring does not take the order of words into account. Similarity is assessed based on the number and importance of words that occur in both opinions. Words that appear in fewer cases within the entire dataset are given a higher weight as they carry more information. For example, the appearance of the word “curtilage” in two cases would increase the similarity score more than the appearance of the word “defendant” in both. Two search and seizure cases that discuss curtilage are more likely to be similar cases than two search and seizure cases that mention a defendant.

This method provides a feasible way to objectively assess a large body of caselaw involving millions of pairwise combinations and narrow down the choice set for each opinion in a principled way. Moreover, the process is analogous to the actual procedure judges and their clerks use to conduct legal research. The human task of legal research usually begins with typing relevant terms into Westlaw or Lexis and searching for precedents. Such a search will produce a list of cases sorted by anticipated relevance, possibly restricted by time or court, that the researcher will review to form her own conclusions about relevance. The sophisticated algorithms powering Westlaw and Lexis searches (together with the researcher's skill in constructing useful query terms) largely perform the initial task of filtering out precedents that are utterly irrelevant. The primary difference for my purposes is that rather than constructing a query, the text of the opinion is used to evaluate the relative relevance of existing precedents. One limitation of this approach is that there may be some endogeneity in this measure. While it would be preferable to use text generated prior to the citation/treatment decision (such as the district court ruling being appealed or legal briefs submitted by the parties) this type of information is not readily available for most cases in the dataset. However, I am able to avoid the most obvious cause of endogeneity by excluding all quoted text from each opinion before calculating similarity scores. When an opinion quotes text from a precedent, the same words in both documents will increase the similarity score. Excluding quoted text eliminates this potential source of endogeneity.Footnote ⁷

Although choosing a cutoff point is necessarily somewhat arbitrary, the potential precedents ranked in the bottom half in terms of similarity to an opinion are not viable candidates for citation. Therefore, I utilize only the precedents in the top fifty percent of the similarity ranking as the choice set for the first stage analysis of which cases judges choose to cite (and which they ignore). This leaves a fairly large choice set in play, which will still include many cases that are not very similar to the opinion. However, an examination of the relative similarity between opinion-precedent pairs for precedents that are actually cited confirms that this cutoff point does include most cited precedents in the choice set. Figure 1 illustrates that the vast majority of cited precedents (94.2 percent) are in the top half of an opinion's similarity scores. Raising the cutoff point to examine only the top 20 percent most similar precedents would exclude nearly one-fourth of cited precedents. Dropping the cutoff point to look at the top 80 percent most similar precedents would increase the size of the choice set well beyond what a judge could realistically be expected to consult. Moreover, Table 4 in Appendix B demonstrates that these alternative thresholds lead to similar empirical results.

Figure 1. Distribution of Similarity Percentile for Opinion-Cited Precedent Pairs: The Solid Vertical Line Represents the Threshold for Establishing the Choice Set for Each Opinion. The Dotted Lines Show Alternative Thresholds for Which the Relevant Empirical Analyses are Provided in Appendix B. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]

The unit of analysis is an opinion-precedent pair. There is an observation for each pair between an opinion and every precedent in its choice set. The first stage decision of whether to cite a precedent in the choice set is modeled using a probit model. For the opinion-precedent pairs where the precedent was cited, I use a multinomial probit model to examine whether the precedent was treated negatively, neutrally, or positively. A two-stage decision-making process such as this sometimes raises the problem of correlated residuals between the two stages biasing the estimates at the second stage (Reference HeckmanHeckman 1979). However, two separate Heckman models examining negative and positive treatment as separate binary outcomes fail to reveal any statistically significant correlation in the residuals.

Although the treatment categories are ordinal, some explanatory variables may make both positive and negative treatment more likely than neutral treatment. For example, more similar precedents may be expected to receive more in-depth treatment overall, both positively and negatively, than less similar precedents. As a consequence, an ordered probit (or logit) is not appropriate in this context because the parallel slope assumption is likely violated (Reference BorooahBorooah 2002). For both the citation and treatment models each opinion appears in the dataset every time it is paired with a precedent. Therefore, I estimate robust standard errors clustered on the opinion in both models.

Shepard's Citations provides data on both the citation and treatment of precedents. As some treatment categories in Shepard's can be both ambiguous and heterogeneous, I follow the advice of Reference Spriggs and HansfordSpriggs and Hansford (2000) and only utilize treatment categories that clearly indicate either positive or negative treatment. I employ their classification of which treatments are positive and negative. “Followed’ is the only Shepard's treatment classified as positive while negative treatments are the following: ‘Distinguished’, ‘Criticized’, ‘Limited’, ‘Questioned’, ‘Overruled’, and ‘Disapproved’. The (rare) Shepard's treatment, ‘Superseded’, while negative in character, actually signals a precedent has been rendered irrelevant due to a subsequent statute. Consequently, I exclude such superseded precedents from the choice set of all subsequent opinions.”Footnote ⁸

The key explanatory variables of Panel Alignment and Circuit Alignment are constructed using Judicial Common Space (“JCS”) scores. JCS scores are based on the ideology of the political elites who appointed a judge and are located on a scale from −1 (liberal) to 1 (conservative) (Reference EpsteinEpstein et al. 2007; Reference Giles, Hettinger and PeppersGiles, Hettinger, and Peppers 2001; Reference PoolePoole 1998). First, the outcome of each precedent is coded as conservative or liberal.Footnote ⁹ As JCS scores denote the level of conservativeness, when a precedent is conservative Panel Alignment is simply the JCS score of the panel median. When a precedent is liberal Panel Alignment is the panel median JCS score multiplied by −1. The variable Circuit Alignment is constructed the same way using the circuit median JCS score instead of the panel's. For example, if the circuit median has a fairly conservative JCS score of 0.35, when a precedent is conservative Circuit Alignment will equal 0.35 but when a precedent is liberal Circuit Alignment will equal −0.35. Increasingly conservative circuits will be both more aligned with conservative precedents and less aligned with liberal precedents than a more moderately conservative circuit. The interaction between Panel Alignment and Circuit Alignment rounds out the key explanatory variables.

Control variables for both the citation and treatment models include a number of characteristics of the opinion, the precedent, and their relationship to each other. Similarity Percentile is the percentile of a precedent's similarity to the opinion compared with all other possible binding precedents in the choice set. Same Author equals one when the same judge wrote both the opinion and the precedent and zero otherwise. When the dataset contains a previous case in which the opinion author cited the precedent, Previous Cite by Author equals one, and it equals zero otherwise. Author Tenure is the number of years the author of the opinion has served on the circuit court. I also control for precedents with a dissenting opinion, precedents decided per curiam, those decided en banc, and those written by a particularly prominent jurist. Most of these factors are straightforward to code with the exception of the prominence of the author. I rely on the prestige scores calculated by Reference Klein and MorrisroeKlein and Morrisroe (1999). Unfortunately, these useful scores are not available for all circuit judges (let alone available for all 20 years). However, their list of the top-25 circuit judges calculated based on data from 1989-1991 does constitute a list of elite circuit court jurists who enjoyed such an established reputation throughout the 20 years studied here. Therefore, Elite Author equals one if a precedent is penned by a judge with a top-25 prestige score and zero otherwise.

Dynamic controls include the number of times a precedent has been both cited and treated (either positively or negatively) by a case from the same circuit as well as its vitality. Vitality is the number of positive treatments minus the number of negative treatments (Reference Hansford and SpriggsHansford and Spriggs 2006). For each opinion-precedent pair these measures only include use of the precedent that occurred prior to the opinion. Both very new and very old cases are less likely to be cited (Reference Black and SpriggsBlack and Spriggs 2008), so I control for both the age of a precedent and age squared. Measures of the logged word counts of both the opinion and precedent are included. The final feature of a precedent is the proportion of the text that is quoted.Footnote ¹⁰ Binary variables indicate whether the opinion was accompanied by a dissent or concurrence. Circuit Size indicates the number of active judges on the circuit in the relevant year. The opinion year is included to control for time. In addition, in the citation model (but not the treatment model) I control for the logged number of available precedents in the choice set and the citing circuit's caseload.Footnote ¹¹ A summary of the dataset is provided in Appendix B.

Results

The first place I look for evidence of strategic anticipation of en banc review is the decision regarding which precedents to cite and which to ignore. Figure 2 presents the results of the citation model.Footnote ¹² The coefficients for Circuit Alignment, Panel Alignment, and their interaction are all in the expected positive direction, but neither constituent variable is statistically significant independent of the other.Footnote ¹³ Whether the panel's ideology has a significant effect depends on the full circuit's ideology (and vice versa). Marginal effects are presented in Figure 3 along with their 95 percent confidence intervals to illustrate the results. Figure 3 shows the marginal effect of each variable over the range of the other (with all other variables set at their median). Panel (a) demonstrates that the marginal effect of Circuit Alignment is positive and statistically significant when the panel is at least somewhat aligned with the precedent (Panel Alignment >0.18). Panel (b) shows that the marginal effect of Panel Alignment is positive and statistically significant when the circuit median is at least moderately aligned with the precedent (Circuit Alignment >0.2).

Figure 2. Citation Model: Probit Regression Estimates of the Effect of Circuit Alignment, Panel Alignment, Their Interaction, and a Range of Control Variables on the Decision of Whether to Cite a Precedent. Dots and Diamonds Indicate Point Estimates. Bars Indicate 95 percent Confidence Intervals. Diamonds (Instead of Dots) and Bars in Black (Instead of Gray) Denote that a Coefficient has a p Value Less Than 0.05. The Full Regression Table is Presented in Table 2 in Appendix B.

Figure 3. Marginal Effects of Circuit Alignment and Panel Alignment on Citation: Panel (a) Shows the Marginal Effect of Circuit Alignment Over the Range of Panel Alignment While Holding all Other Variables at Their Median. Panel (b) Shows the Marginal Effect of Panel Alignment Over the Range of Circuit Alignment While Holding all Other Variables at Their Median. The Shaded Regions Around Each Line Delineate the 95 percent Confidence Intervals.

These results provide varying levels of support for the hypotheses. First, there is some evidence in support of the Ideological Hypothesis. The panel's own ideology does influence the citation decision for a substantial portion of the dataset. While Panel Alignment is only significant when Circuit Alignment is greater than 0.2, this includes 54 percent of the observations in this dataset. To the extent that the Ideological Hypothesis is not supported over the entire range of Circuit Alignment, this presents evidence in support of Conditional Strategic Hypothesis 1, which anticipates panels having greater flexibility to follow their own ideology when a circuit is more strongly aligned with a precedent. When the circuit is not aligned with a precedent, the effect of the panel's ideology is not statistically significant. Circuit Alignment only has a statistically significant positive marginal effect when Panel Alignment is greater than 0.18. This provides support for the Direct Strategic Hypothesis for the 46 percent of observations that meet this criteria.

The microlevel of analysis makes evaluating the substantive impact of statistically significant effects a difficult task. A judicial opinion is made up of many decisions about individual precedents. Also, the choice set contains hundreds of possibilities so the probability of citing each case is quite small. The marginal effects in Figure 3 appear to be minute because the baseline probability of citing a precedent when all variables are set at their median is only 0.0012.Footnote ¹⁴ As so many cases are not cited,Footnote ¹⁵ it is more relevant to assess the impact of variables of interest while holding the other explanatory variables at their median value among the cited precedents. The baseline probability of citing such a precedent is 0.056. When both Circuit Alignment and Panel Alignment are at their maximum, the predicted probability of citation increases to 0.061. Figure 3 show that both Circuit Alignment and Panel Alignment have the greatest impact when the other is at its maximum value. Keeping Panel Alignment at its highest value in the data and moving Circuit Alignment to its minimum value reduces the predicted probability of citation to 0.049, a roughly 20 percent reduction. Changing Panel Alignment to its minimum while holding Circuit Alignment at its maximum produces a similar result with a predicted probability of 0.051.

One way to evaluate the substantive meaning of the strategic and ideological effect sizes is to compare them to the effect of the variable expected to have the largest impact, Similarity Percentile. As I anticipate, given the importance of legal doctrine, the similarity of two cases plays a statistically and substantively significant role in the citation decision. The median cited precedent has a Similarity Percentile of 95, and the predicted probability of citation is 0.056. A similar precedent that is in the 85th percentile in terms of similarity to the opinion has a predicted probability of only 0.026. This 10-point reduction in Similarity Percentile more than halves the likelihood of citation. In short, strategy (i.e., circuit preferences) can move the predicted probability a maximum of 0.01, while moving the legal doctrine variable by less than one standard deviation changes the predicted probability by 0.03.

The next model explores the decision regarding whether to treat a cited precedent negatively, neutrally, or positively. These results are presented in Figure 4 with neutral treatment as the baseline category. There is no evidence that either the panel's ideology or strategic anticipation of en banc review significantly impacts the decision to positively treat a cited precedent. However, the hypotheses regarding treatment are supported with respect to negative treatment. The statistically significant negative impact of Panel Alignment supports the Ideological Hypothesis. The more closely aligned a panel is with a cited precedent, the less likely it is to treat it negatively. The Direct Strategic Hypothesis anticipates the same pattern with respect to Circuit Alignment. As expected, Circuit Alignment is also a significant negative predictor of negative treatment. There is no evidence of an interactive effect (either in the regression table or on more in-depth examination).

Figure 4. Treatment Model: Multinomial Probit Regression Estimates of the Effect of Circuit Alignment, Panel Alignment, Their Interaction, and a Range of Control Variables on the Decision to Treat a Cited Precedent Negatively, Neutrally, or Positively. Neutral Treatment is the Baseline Category. Dots and Diamonds Indicate Point Estimates. Bars Indicate 95 percent Confidence Intervals. Diamonds (instead of dots) and Bars in Black (Instead of Gray) Denote That a Coefficient has a p Value Less Than 0.05. The Full Regression Table is Presented in Table 3 in Appendix B.

Figure 5 graphs the predicted probability of negative treatment across the range of Circuit Alignment and Panel Alignment to illustrate the impact of circuit and panel ideology respectively. The predicted probability a panel that is maximally aligned with a precedent negatively treats that precedent is only 0.029, while the minimum level of Panel Alignment increases that rate to 0.043. Circuit Alignment has an even larger effect. At the maximum value of Circuit Alignment the predicted probability is 0.027 and it more than doubles to 0.060 at the minimum value of Circuit Alignment. Once again, I compare this to the impact of Similarity Percentile. More similar precedents are significantly more likely to be both negatively and positively treated. It makes sense that less similar precedents are more likely to simply be cited rather than discussed at length. For purposes of comparison, I focus on the impact Similarity Percentile has on the predicted probability of negative treatment. The median case has a Similarity Percentile of 95 and a 0.034 predicted probability of negative treatment. When similarity is in the 85th percentile this drops to 0.023.

Figure 5. Effect of Circuit Alignment and Panel Alignment on Negative Treatment: This Graph Provides the Predicted Probability of Negative Treatment (Conditional on Citation) and 95 percent Confidence Intervals Over the Range of Circuit Alignment, Panel (a), and Panel Alignment, Panel (b). All Other Variables are Held at Their Median.

The final hypothesis left to be evaluated is Conditional Strategic Hypothesis 2, which posits that the circuit's ideology will have a larger direct impact on treatment than citation decisions. The predicted probability of citation changes from 0.053 to 0.057 over the range of Circuit Alignment. The predicted probability of negative treatment changes from 0.060 to 0.027 over the same range. These predicted probabilities hold all other variables at the median for cited precedents. Even though the baseline probability of negatively treating a cited precedent is lower overall than the corresponding probability of citing the same type of precedent, the absolute effect of moving Circuit Alignment from its minimum to maximum value is considerably larger for negative treatment (0.033) than for citation (0.004). The fact that Circuit Alignment has a statistically significant impact on negative treatment regardless of the panel's ideology further supports the conclusion that Conditional Strategic Hypothesis 2 is supported, at least with respect to negative treatment. Together, these results indicate that panels give more weight to circuit ideology in the more visible decision about negative treatment than in the less visible decision of whether to cite a precedent.

Across the models the control variables tend to perform as expected. As discussed, when the precedent is very similar to the opinion it is more likely to be cited, positively treated, and negatively treated. Authors are more likely to cite both precedents they wrote and those they have cited previously. Previous citation makes both negative and positive treatment significantly less likely. Per curiam cases are less likely to be cited, but curiously more likely to be treated positively. Precedents with a dissenting opinion are more likely to generate negative treatment. Somewhat strangely, en banc precedents and those written by circuit judges with an elite reputation are less likely to be cited. Various aspects of how a precedent has been dealt with by previous cases within the same circuit also influence citation and treatment. Precedents with greater vitality are more likely to be positively treated and less likely to be negatively treated. A greater number of total cites to a precedent increases the probability of citation, but decreases both types of treatment while an increase in the total number of previous treatments increases the probability of citation and both kinds of treatments. An opinion that is accompanied by a dissent is less likely to cite any given case, but conditional on citation both negative and positive treatment are more likely than in opinions unaccompanied by a dissent. Panels in larger circuits tend to be more likely to both cite a precedent and to negatively treat cited precedents. Other factors such as the age, length, and composition of the respective opinions involved also play a role in these decisions. Finally, when a judge has a greater number of circuit precedents from which to choose, the probability of any given case being cited decreases.

Discussion

In light of the mixed results the literature contains regarding strategic voting by panel judges, the findings, here, illustrate the utility of delving into the content of judicial opinions. Focusing on the use of precedent reveals patterns of strategic choices consistent with what some scholars have found in voting behavior (Reference Blackstone and CollinsBlackstone and Collins 2014; Reference Cross and TillerCross and Tiller 1998; Reference KimKim 2009; Reference Van WinkleVan Winkle 1997). The evidence of strategic behavior presented, here, is quite broad. There is some support for each of the strategic hypotheses. First, the Direct Strategic Hypothesis is supported by the findings that the ideological preferences of the entire circuit directly influence both citation (when Panel Alignment >0.18) and negative treatment. Second, Conditional Strategic Hypothesis 1 is supported by the finding that a panel is less likely to rely on its own ideology when the full circuit is more ideologically divergent from the precedent. In fact, Panel Alignment only significantly predicts citation when the circuit is at least moderately aligned with a precedent (Circuit Alignment >0.2). Finally, Conditional Strategic Hypothesis 2 is supported by the finding that Circuit Alignment has a larger (and more consistent) impact on negative treatment than on citation. These patterns are consistent with the intuition that panels give greater weight to circuit preferences when making decisions that are easier for the circuit to review.

For the most part, the preferences of the full circuit fail to have a statistically significant impact in precisely the circumstances predicted by strategic anticipation of en banc review. However, the one major exception is the positive treatment model. There is no evidence that the circuit's ideology plays a role, either direct or indirect, in the panel's decision about expanding the scope of a precedent. These results are consistent with the research that has failed to find any evidence of strategic behavior (Reference Hettinger, Lindquist and MartinekHettinger, Lindquist, and Martinek 2004). However, there is also no evidence that the panel's own ideology influences positive treatment. Most likely this lack of explanatory power is due to similarity between a simple citation (i.e., neutral treatment) and a positive treatment. A neutral treatment can be viewed as a soft positive treatment, different in degree rather than type. A mere citation to a precedent tacitly indicates acknowledgment and applicability. The time pressures under which panel opinions are drafted may lead to judges putting less effort into the relatively fine distinction between mere citation and positive treatment. The fact that many of the control variables also fail to reach statistical significance for positive treatment further supports this conclusion.

The institutional structure that makes review from the full circuit possible is the most plausible explanation for the patterns that emerge in this study. The effect of legal doctrine is largely controlled for by looking at the use of only one type of precedent, so cases have the same general status under stare decisis. To the extent that more relevant cases carry greater weight according to legal doctrine, controlling for the similarity between an opinion and precedent accounts for this variation as well. In fact, the statistical and substantive significance of Similarity Percentile in both the citation and treatment decisions reinforces the importance of legal considerations. The model accounts for the impact a panel's own ideology has on the selection and use of precedent as well. Even after controlling for law and ideology (and several control variables), the composition of the full circuit emerges as a factor in how opinions are drafted at the panel level. Moreover, circuit impact is conditional in precisely the manner theorized based on the asymmetric visibility of noncitation. These patterns provide evidence that under some circumstances panels seek to avoid en banc review entirely or avoid reversal en banc by tailoring their opinions to be persuasive to the full circuit.

Reference Bowie and SongerBowie and Songer (2009) argue that rational judges should not be expected to anticipate review because of the rarity of that procedure and uncertainty over when review will be exercised. Yet, there is evidence presented here that panels strategically use precedent. Two features of these findings do reflect the underlying reality that en banc review is quite rare. The first is the microlevel decision of the use of a particular precedent. Compromising on this level imposes only a small cost. Crafting an opinion while taking the circuit's preferences into account may be viewed by panels as a cheap way to protect against even the unlikely possibility of en banc review. Second, in the circumstances where circuit preferences have a statistically significant impact, the size of that impact is, understandably, quite small. While strategy evidently plays some role in how panels use precedent, the small baseline probability of en banc review makes it implausible that strategic anticipation of such review would consistently outweigh other considerations.

An important caveat to the conclusions drawn here is that they are based exclusively on one legal issue area. Search and seizure law is in important topic with broad implications across society. Nevertheless, the question remains whether the results are generalizable to other issue areas. The data do incorporate both civil and criminal cases, so there is little reason to expect that the findings are exclusive to one domain or the other. However, one way search and seizure law is distinct from at least some other legal topics is the sheer bulk of precedents available. Panel-circuit dynamics involving the use of precedent may vary when the universe of potential precedents is smaller. For example, panels may have less discretion when the applicable choice set is smaller and easier for the full circuit to monitor. Under such circumstances the buy-in cost to exercise review over noncitation decisions would be smaller. This suggests that looking for patterns of strategic behavior in an issue area like search and seizure with a vast array of precedents is a fairly difficult case. As circuit preferences exert an effect even under these difficult circumstances, there is reason to believe such patterns exist in other issue areas as well. Ultimately, though, this is an empirical question that can be explored in future research.

In addition to future work expanding this line of inquiry to apply across multiple topics, the results suggest that a number of other related lines if inquiry would prove interesting as well. Perhaps the most obvious move forward is to apply the theoretical expectations about strategic anticipation of en banc review to look for evidence of strategic anticipation of Supreme Court review. Due to the overlap in full circuit and Supreme Court preferences, disentangling these two types of strategy will require extensive additional data collection targeted at circuits that diverge from the Supreme Court ideologically. This study also reinforces the utility of looking for evidence of other types of strategic behavior manifesting in the use of precedent. For example, future work can examine whether an author's use of precedent strategically anticipates the preferences of their panel colleagues. If evidence can be found that circuit preferences impact these microlevel decisions, it suggests that other key actors might influence them as well. Exploring questions like these provides opportunities to supplement our knowledge of how judges vote with knowledge of how they craft legal opinions.

The evidence of strategic citation behavior presented here has important implications for how we understand legal development. Each opinion dynamically shapes the contours of the law. If circuit panels seek to use precedent in a manner that does not stray too far from the preferences of the full circuit, this will result in the law of the circuit gradually moving toward the circuit median. This pattern may already be evident in the data used for this study. Table 1 in Appendix B shows the summary statistics. Overall, the median value of Circuit Alignment is 0.23 while the median value of Panel Alignment is 0.14. This pattern is consistent with the logical implications strategic citation has for long-term legal development. Of course, these conclusions should not be overstated. The evidence provided here is derived from opinions dealing with a single legal issue. Nevertheless, there is reason to believe that published panel opinions are drafted in the shadow of the full circuit even though the probability of en banc review is ultimately quite low.

Exploring how panels relate to the full circuit is also key to understanding issues such as the interplay between stare decisis and ideological shifts brought about by personnel changes within circuits. When the location of a circuit's median shifts dramatically, the presence (or absence) of strategic use of precedent has implications for how quickly the ideological tenor of circuit law will change. If panels disregard the full circuit's preferences, then all panels still dominated by the old guard will continue to rely on old circuit law that is no longer in line with circuit preferences. This will slow down the overall shift in circuit law. However, if all panels strategically account for the current circuit median, even the panels farthest from the circuit median will at least partially participate in the (still gradual) process of moving circuit law toward the new circuit median.

Finally, the evidence of strategic action on the part of panels suggests the importance of examining the efficacy of that strategy. Studying the seemingly minuscule effects of microlevel decisions is important because these can aggregate up to help answer questions carrying more obvious substantive significance. For example, does strategic citation usage in panel opinions influence the granting of en banc review or the probability of reversal if such review is granted? The importance of such issues should not be underestimated as the vast majority of federal case law is established by the U.S. Courts of Appeals rather than the U.S. Supreme Court (Reference CrossCross 2007; Reference Hettinger, Lindquist and MartinekHettinger, Lindquist, and Martinek 2006; Reference KleinKlein 2002). Scholars have increasingly recognized that both law and ideology play important roles in judicial decisionmaking and the nuances of when, how, why, and under what conditions these factors matter present challenging questions. This study contributes to the work investigating these questions by offering insight into how strategic considerations shaped by institutional structure influence the manner in which the ideology of different actors influences how judges apply existing case law.