Median-equivalent nominees versus utility-equivalent nominees

Crucial to understanding the outcomes of MTM games is the relationship between three quantities: first, the ideal point of the exiting justice (e); second, the ideology of the nominee (n); and third, the resulting ideal point of the new median justice (j ¹ ₅), conditional on confirmation. Importantly, the location of the new median justice j ¹ ₅ can only be j ⁰ ₄, j ⁰ ₅ (the old median justice), j ⁰ ₆, or n itself, with n bounded within [j ⁰ ₄, j ₆ ⁰]. The nominee can become the median justice only when the opening and the nominee lie on opposite sides of the old median justice and n lies between j ⁰ ₄ and j ⁰ ₆.

Because the new median justice is restricted to just a few values, many different appointees can have the same impact on the Court’s median. For example, if the opening is between j ⁰ ₁ and j ⁰ ₄ then all nominees n ⩽ j ⁰ ₅ induce no change in the median. Thus, these nominees are median equivalent. A critical question then is: should senators and the president view median-equivalent nominees as utility equivalent? Or, should they distinguish among otherwise median-equivalent nominees? To put it another way, do senators and the president care at least somewhat about the nominee’s ideology per se, irrespective of her immediate impact on the Court’s median?

The answer to this question is surely yes, for several reasons. First, nominee ideology may have direct political import. For example, a conservative senator may find it distasteful or politically inexpedient to vote for a liberal nominee even if the nominee would not move the Court’s median. Similarly, the president may gratify ideological allies by selecting the most proximate nominee from among a large group of median-equivalent ones (Nemacheck Reference Nemacheck2008; Yalof Reference Yalof2001). Second, a nominee who may not be the median today may become the median in the future. Hence, future-oriented actors may see more-proximate nominees as more attractive. Finally, the Court may not be a fully median-oriented body; rather, nonmedian justices may have some impact on policy (Carrubba et al. Reference Carrubba, Friedman, Martin and Vanberg2012; Lauderdale and Clark Reference Lauderdale and Clark2012). If so, presidents and senators may prefer more proximate nominees even if they are median equivalent. Indeed, with respect to the Senate, the literature on Supreme Court nominations has demonstrated a strong and persistent relationship between the likelihood of a vote for confirmation and the ideological distance between a senator and the nominee (Cameron, Cover, and Segal Reference Cameron, Cover and Segal1990; Epstein et al. Reference Epstein, Lindstadt, Segal and Westerland2006).

To capture the tradeoffs between the nominee’s ideology versus the median justice, we assume that the president and senators’ evaluation of the impact of a nominee (if confirmed) reflects a weighted sum of two quantities. The first is the ideological distance between each actor’s ideal point and the location of the new median justice. The second is the distance between each actor’s ideal point and the confirmed nominee’s ideal point. Formally, let λ _p and λ _s respectively denote this weight for the president and senators, with 0 ⩽ λ _p ⩽ 1 and 0 ⩽ λ _s ⩽ 1. For simplicity, we assume that all senators share the same value of λ. While this assumption is surely false, and relaxing it would be a worthy endeavor for future work, for our purposes its costs are not great since we can observe neither λ _p or λ _s . (We do, however, conduct tests for senator voting that are robust to any value of λ _s for a given senator.)

What are the substantive implications of differing values of λ _p and λ _s ? If λ _p = 1, the president is purely median oriented (that is, oriented around the outcome of the Court’s collective decision making). If λ _p = 0, the president is purely nominee oriented—note, however, that he compares his utility with the appointment against his utility without the appointment. The same holds true for a senator; when λ _s < 1 she is also interested in the nominee’s ideology per se, perhaps because of position taking or an orientation toward the future. Alternatively, one may see λ _s < 1 as reflecting a belief that, with some probability, the nominee will prove pivotal on some issues.

Thus, if the nominee is confirmed, the president receives − λ _p |p − j ¹ ₅| − (1 − λ _p )|p − n| in utility. If the nominee is rejected, he receives − |p − q| − ε, where ε > 0 is a turn-down cost (this may reflect public evaluation of the president). For senators, we adopt the standard convention that voting over two one-shot alternatives is sincere, so each senator evaluates her vote as if she were pivotal. If a senator votes to confirm, she receives − λ _s |s_i − j ¹ ₅| − (1 − λ _s )|s_i − n|. If she votes no, she receives − |s_i − q|.

Varieties of move-the-median models

The values of the parameters λ _p and λ _s create different variants of MTM models. We display the four key model variants in Table 1:

TABLE 1.

Variants of Move-the-Median Games

• • Court-outcome based. In this variant, considered in Rohde and Shepsle (Reference Rohde and Shepsle2007), the president and senators care only about the impact of the nominee on the ideological position of the new median justice (both λ _p and λ _s = 1); i.e., presidents and senators only care about the outcome of the Court’s policy. Given the median equivalence of many nominees noted above, presidents are often indifferent over a wide range of possible nominees.

• • Nearly court-outcome based. This variant, considered in Moraski and Shipan (Reference Moraski and Shipan1999), is almost identical to the court-outcome based model, but allows the president to put at least some weight on nominee ideology per se (λ _s = 1, but λ _p < 1). Even a small such weight, however, has significant consequences on the president’s nominating strategy, as it prescribes a specific nominee for the president rather than a range of nominees.

• • Position-taking senators. In this variant, considered in Krehbiel (Reference Krehbiel2007), senators (and possibly the president) care only about the nominee’s ideology, and not her impact on the median justice (λ _s = 0). Thus, we characterize the senators as being purely interested in position taking with respect to the confirmation of the nominee himself, and not on the outcome of the Court’s policy following a successful nomination. However, the players continue to use the reversion policy q in their evaluation of the nominee. The strategies in the game are isomorphic to the standard one-shot take-it-or-leave-it Romer-Rosenthal (Reference Romer and Rosenthal1978) game.

• • Mixed-motivations model. In this variant, which is original to this article, senators and the president put some weight on both nominee ideology and nominee impact on the median justice (0 < λ _p < 1, 0 < λ _s < 1).Footnote ⁶

While our focus is squarely on the context of the Supreme Court, the theoretical step of allowing λ to vary in [0, 1] is quite general—it can encompass a wide variety of theories in several literatures that allow for tradeoffs between purely policy-outcome-oriented behavior (λ = 1) and purely position-taking behavior (λ = 0). Such theories include voter selection of candidate in multiparty elections (see, e.g., Austen-Smith Reference Austen-Smith1992) and theories of representation and elections in which members benefit from both policy information conveyed through party labels and position taking in individual roll call votes (Snyder and Ting Reference Snyder and Ting2003).

Court-outcome based and nearly court-outcome based models

In the court-outcome based and nearly court-outcome based models, senators compare the ideology of the new median justice on the Court induced by the appointment of the nominee with the ideological position of the old median justice. Thus, under these models a senator should vote for the nominee if and only if |s_i − j ¹ ₅| ⩽ |s_i − j ⁰ ₅|—that is, if the new median justice’s ideal point is as close or closer to the senator’s ideal point than is the ideal point of the old median justice. To conduct a direct test of this prediction, we calculate the cutpoint $\frac{j_{5}^{0}+j_{5}^{1}}{2}$ . All senators with ideal points at or on the new median justice’s side of this cutpoint are predicted to vote “yea;” all senators with ideal points on the old median justice’s side of this cutpoint are predicted to vote “nay.”

Position-taking senators model

In the position-taking senators model, senators compare the ideology of the nominee with the reversion policy (the old median justice) and vote for nominee if and only if |s_i − n| ⩽ |s_i − j ⁰ ₅|; that is, if the nominee’s ideal point is closer to the senator’s ideal point than that of the old median justice. For conducting a direct test of the position-taking senators model, the relevant cutpoint is the midpoint between the old median justice and nominee $\frac{n+j_{5}^{0}}{2}$ . Under the position-taking senators model, the Senate’s acceptance region will always be (weakly) smaller compared to the court-outcome based model, as the former model predicts rejection even in some instances where the median justice either does not move or is in the Senate’s acceptance region. If, for example, j ⁰ ₅ < s_m , under the position-taking senators mode the Senate should reject any nominee who is more conservative than 2s_m − j ⁰ ₅, even if such a nominee does not move the median.

Mixed-motivations model

In the mixed-motivations model, senators compare a weighted average of the distances to the nominee and the new median justice, with the distance to the old median justice. They vote for the nominee if and only if $\lambda _s |s_{i}-j_{5}^{1}|+(1-\lambda _s )|s_{i}-n|\text{ }\le |s_{i}-j_{5}^{0}|$ . That is, if the weighted average of the two distances (to the nominee and the new median justice) is less than the distance to the old median justice.

We cannot observe the weight (λ _s ) in each senator’s evaluation of the new median justice and the nominee, which complicates the creation of direct tests. However, because λ _s is bounded by 0 and 1, some votes are necessarily incorrect for some ranges of senators’ ideal points. Consider Figure 1, which considers the case when j ⁰ ₅ ⩽ j ₅ ¹ < n (there is a similar mirror case, j ⁰ ₅ ⩾ j ₅ ¹ ⩾ n). Senators with ideal points between the cutpoints $\frac{ j_{5}^{0}+j_{5}^{1}}{2}$ and $\frac{n+j_{5}^{0}}{2}$ could vote either yea or nay, depending on their value of λ _s . But all senators with ideal points less than $\frac{j_{5}^{0}+j_{5}^{1}}{2}$ must vote “nay” while all those with ideal points greater than $\frac{n+j_{5}^{0}}{2}$ must vote “yea,” irrespective of the size of λ _s . These unambiguous predictions allow a direct evaluation of the mixed-motivations model, focusing on senators in those two ranges.

FIGURE 1.

Predicted Votes in the Mixed-Motivations Model. See text for details

Robust predictions

There are two robust predictions for senators’ voting. First, recall that under the court-outcome based model, the senator should vote to reject whenever the new median justice is farther away from the senator than the old median justice. In fact, this prediction is robust. Why? By construction, this condition can only hold if the nominee is farther away from the senator than the old median, since the new median is bounded by j ⁰ ₄ and j ⁰ ₆. Thus, the court-outcome based model’s prediction about when to reject a nominee is robust: any time a senator should vote no under the court-outcome based model, he should also do so under any model. We call this robust prediction the too much movement prediction—the median justice moves too much for the Senate.

Second, recall that the position-taking senators model predicts a yes vote by a senator whenever the nominee is closer to the senator than the old median justice. This prediction is also robust, because in all models senators are (weakly) better off when this condition holds, and should vote yes. We call this robust prediction the attractive nominee prediction.

Choice of nominee in the court-outcome based model

We begin with the president’s selection strategy in the court-outcome based model, which is presented in Figure 2(A). A proximal vacancy creates what we call a “restoring” nomination. Because the president cares only about the median justice in this model, and all nominees n ⩾ j ⁰ ₅ result in an unchanged median justice, the president is indifferent among all such nominees. Hence, the court-outcome based model produces a range of possible nominees given such a nominee, and not a point prediction (see Rohde and Shepsle Reference Rohde and Shepsle2007).

Next, consider “distal” vacancies under the court-outcome based model. First, if the Senate median is on the other side of the old median justice, relative to the president, the result is what we call a “gridlock” nomination. Here the best the president can do is choose n = j ⁰ ₅, since the Senate will reject any nominee the president prefers more. Since the president and the Senate lie on opposite sides of the old medians, movement in the median is gridlocked.

On the other hand, if a distal vacancy occurs and the Senate median is on the same side of the old median justice as the president, he can move the median. The extent of this movement, however, depends on the relative locations of the Senate median and the president. If the Senate median is closer to the old median justice (type C), then the president offers what we call a “smaller shift” nominee that is the minimum of the president’s ideal point (p) and the indifference point of the Senate median around the old median (2s_m − j ⁰ ₅). If the Senate median is farther from the old median justice (type D), the president can make what we call a “maximum shift” nomination that moves the median justice as far as possible. Finally, if p > j ⁰ ₆, the court-outcome based model also predicts a range of possible nominees—all of which move the median justice to j ⁰ ₆, and thus similarly induce a maximum shift in the median justice.

Choice of nominee in the nearly court-outcome based model

Figure 2(B) indicates the president’s equilibrium choice of nominee in the nearly court-outcome based model. As discussed above, in this model the voting strategy of senators is exactly the same as in the court-outcome based model. But because the president is no longer indifferent over nominees who yield the same median justice, the ranges in the restoring and maximum shift nomination collapse to point predictions—in each the president nominates someone who mirrors his own ideology. Whether the president has a choice among (median-equivalent) nominees or is constrained to a single point has implications for work that evaluates how the president chooses among the “short list” of potential nominees—nominees who may look similar ideologically but differ on other important characteristics that the president may value (see, e.g., Nemacheck Reference Nemacheck2008).

Choice of nominee in the position-taking senators model

The nomination strategy for the position-taking senators model is shown in Figure 2(C). For ease of comparison with the rest of the panels, Figure 2(C) arrays nominating strategies for the same types of Senate medians. However, because senators do not care about the location of the new median justice and the president cares at least somewhat about the nominee’s ideology, whether a nomination is distal or proximal is irrelevant for determining the location of the nominee.Footnote ⁸ Rather, the president nominates a confirmable individual as close to his own ideal point as possible. When the median senator is opposed to the president, we again see a gridlock nomination. When the Senate median is on the same side as the president, the president can move the median justice. Again, he accomplishes a “smaller shift” in the median justice by appointing a nominee n = p or by choosing a nominee at 2s_m − j ⁰ ₅, depending on the relative locations of the Senate median and the president.

Choice of nominee in the mixed-motivations model

Finally, Figure 2(D) depicts the nomination strategy in the mixed-motivations model. The strategy here is similar to that seen in the position-taking senators model, except now there is a “maximum shift” region; here the president chooses a nominee either at his ideal point or a location (x, defined in the caption of Figure 2) that depends on λ _s , but which leaves the median senator indifferent between the nominee and the old median justice.

Robust predictions across models

Using Figure 2, we can discern four robust predictions for presidential choice that hold across all the models:

1. 1. Own goals. Looking at all the variants of presidential strategies in Figure 2, it is clear that regardless of the regime, the president should never choose a nominee on the opposite side of the old median justice from himself. The worst-case scenario for the president is a gridlock nomination; across all model variants, in the gridlock scenario the president should choose a nominee exactly at the old median justice. Thus, if a president chooses a nominee on the opposite side of the old median justice, in soccer parlance he would be committing an “own goal.”

2. 2. Aggressive mistakes. Recall that a robust prediction for the Senate is that it should never confirm a nominee who moves the median justice farther away from the Senate than the old median justice. Accordingly, the president should never choose such an nominee, since she would be rejected. Such a nominee would thus constitute an “aggressive mistake.”

3. 3. Median locked. From Figure 2, it is clear that the “lower left quadrant” of each panel predicts that the president should choose a nominee exactly at the location of the old median justice. In this region, the president and Senate are on opposite sides of the old median justice, and hence the Senate would reject any nominee that would move the median in the president’s direction. We thus say that the president is “median locked.”

4. 4. Smaller shift. Finally, it can be seen that the “smaller shift” nomination regions of the court-outcome based and nearly court-outcome based models also apply to the position-taking senators and mixed-motivations models. Whenever the Senate is on the president’s side but is not too “extreme,” and the vacancy is opposite the president, each variant predicts a nominee at the minimum of the president’s ideal point and 2s_m − j ⁰ ₅.

Ideal points of presidents, senators, and justices

For the NOMINATE-based measures, we place all relevant actors in the Senate DW-NOMINATE space (Poole and Rosenthal Reference Poole and Rosenthal1997). For senators and presidents, we employ their relevant DW-NOMINATE score at the time of a nomination. To place the justices on the same scale, we follow the lead of Epstein et al. (Reference Epstein, Martin, Segal and Westerland2007) and begin with the Martin-Quinn (Reference Martin and Quinn2002) scores of the justices, which are based on the justices’ voting records. We transform these scores into DW-NOMINATE by using the DW scores of the appointing presidents as a bridge. While the specifics of this procedure are given in Online Appendix A.2, it worth noting that to conduct this bridging, Epstein et al. (Reference Epstein, Martin, Segal and Westerland2007) only use presidents who were seemingly unconstrained in their choice of nominees, based on the results in Moraski and Shipan (Reference Moraski and Shipan1999). Because this choice assumes that MTM predicts presidential selection well, which is exactly what we evaluate, it does not make sense for us to use the same set of presidents. Instead, we use all presidents to estimate the transformation, which means that our choice of observations is not endogenous to MTM theory.Footnote ¹⁰ Recall that the Bailey scores include estimates of presidents, senators, and justices on the same scale. Thus, for both sets of measures, it is straightforward to identify the median of the existing court (that is, the status quo), at the time of any given confirmation. To do so, we simply take the median of the ideal points of the nine justices (in the most recent Supreme Court term prior to a given nomination).

Estimated ideal points of nominee

Our next step is to place the location of the nominee into the same space as the other actors. Here we follow prior research and use the Segal-Cover scores (Reference Segal and Cover1989) as a proxy for the ideology of each nominee (Epstein et al. Reference Epstein, Lindstadt, Segal and Westerland2006; Moraski and Shipan Reference Moraski and Shipan1999). These scores are based on contemporaneous assessments of nominees by newspaper editorials. While not flawless, this measure is exogenous to the subsequent voting behavior of the confirmed nominees and it is not based on the president’s measured ideal point, which are both virtues. To place these scores into the same space as NOMINATE or Bailey scores, we regress the respective first-year voting score of each confirmed nominee on their Segal-Cover score. We use the linear projection from this regression to map the Segal-Cover scores into the relevant space. Because every nominee has a Segal-Cover score, this procedure results in comparable scores even for unconfirmed nominees.Footnote ¹¹ With this measure in hand, we can calculate the location of the new median justice (assuming the nominee would be confirmed), as well as necessary distances between a senator and the nominee, and the senator and the new median justice.

Incorporating uncertainty

As with any ideal point measure, both the NOMINATE and Bailey scores are measured with error, and it is important to account for this when testing MTM theory. To do so, we use the relevant ideal points and their corresponding standard errors to generate 1,000 random draws of each actor’s ideal point. With these distributions in hand, we can simulate the location of the existing median justice on the Court 1,000 times, as well as the location of every senator and the Senate median. Thus, for every nominee, we can run empirical tests of nominee location and senatorial voting decision 1,000 times, and use variation within those simulations to make probabilistic estimates of “correct” decisions, depending on the theory’s predictions. (The actual implementation depends on a given test and quantities of interest.Footnote ¹² ) This allows us to generate uncertainty in all the measures and tests based on the location of the nominee. (Figure A-1 in Online Appendix A.2 depicts the estimates of the nominees’ ideal points, while Figure A-2 depicts the estimates of the extent to which each nominee moves the median justice, assuming they are confirmed. Both figures include estimates of uncertainty for these quantities.)

Voting by Individual Senators

We begin our empirical analysis with direct tests of the Senate’s roll call voting on nominees, comparing the predictions of each MTM variant with actual voting behavior.Footnote ¹³ (We exclude from these analyses the three withdrawn nominees—Homer Thornberry, Douglas Ginsburg, and Harriet Miers—whose nominations thus created no Senate voting record.) Recall that under the court-outcome based and nearly court-outcome based models, a senator should vote for the nominee if and only if |s_i − j ¹ ₅| ⩽ |s_i − j ⁰ ₅|, while under the position-taking senators model a senator should vote yes if and only if |s_i − n| ⩽ |s_i − j ⁰ ₅|. Finally, for the mixed-motivations model, as described in Figure 1, we identify observations where the predictions are unambiguous, and then compare those predictions to actual votes. For simplicity, we treat voice votes as votes to confirm.Footnote ¹⁴

The top part of Table 2 displays the results of this analysis, across both the NOMINATE and Bailey measures. Each “model-measure” pair depicts a two-by-two table of cell proportions, with 95% confidence intervals in brackets (based on the simulations). The results are very similar across the two different measures. For reference, the shaded portions of a given two-by-two table depict where the robust tests can be evaluated. We return to these below.

TABLE 2.

Predicted versus Actual Votes by Individual Senators (top), and the Senate as a Whole (bottom), in Different Versions of the MTM Theory.

Note: For each two-by-two table, cell proportions are displayed, along with 95% confidence intervals in brackets. The shaded regions indicate the tests of the robust predictions for senatorial voting.

Our direct tests are simple. Given the structure of the two-by-two tables, correct classifications occur on-the-main diagonal, while errors occur off-the-main diagonal. The table reveals that voting errors were very numerous in all three models, but particularly so in the position-taking senators and mixed-motivations models. For the position-taking senators model, in nearly half of all senator observations the model predicted a “no” vote when the senator actually voted yes. The court-outcome based model performs best, correctly predicting about 68% of votes correctly. However, this means that a third of votes were incorrect, according to this variant.

Where do the model’s predictions go wrong? A striking feature across Table 2 is the asymmetry in errors across predicted yes and no votes. Across all three models, if a senator’s vote was predicted to be a “yea,” most votes were in fact “yeas.” Indeed, in the position-taking senators and mixed-motivations models, the percentage of instances in which a senator votes no when he is predicted to vote yes is less than five percent. However, if a senator was predicted to vote no, for each model errors outnumber correct classification by a ratio of at least 3:1. The conclusion is inescapable: historically, senators have been much more accommodating of the president’s nominee than MTM theory would suggest.

We now evaluate the robust predictions for Senate voting. Recall that the court-outcome based model’s prediction of when to reject is robust (the “too much movement” prediction). Due to the asymmetry in errors, this prediction does not perform well. As seen in the shaded area of the court-outcome based model tests in Table 2, when the model predicts a no vote, meaning that the new median justice is farther away from the senator than the old median justice, the senator is still three times more likely to vote yes. Next, recall that the position-taking senators model’s prediction of when to confirm is robust (the “attractive nominee” prediction). As seen in the shaded regions of the position-taking senators model tests, this prediction is supported: when the nominee is closer to a senator than the old median justice, senators almost always vote yes.

Confirmation Decisions

How consequential are these errors for MTM theory in terms of which nominees actually make it to the Supreme Court? One benign possibility is that nonpivotal senators engage in position taking by voting to support nominees even when they are inclined to oppose them for ideological reasons—especially high quality nominees, or nominees with public support in their home states (Kastellec, Lax, and Phillips Reference Kastellec, Lax and Phillips2010; Overby et al. Reference Overby, Henschen, Walsh and Strauss1992). If this were true, MTM theory would fail across many individual votes, but the Senate as a whole might still conform to the theory’s predictions.

This is not the case, however. The bottom part of Table 2 examines predicted versus actual confirmation decisions, using the predicted votes of the Senate median and comparing it to whether the Senate actually confirmed a nominee. (We omit the mixed-motivations model from this analysis because for some nominations the predicted vote of the Senate median is ambiguous.) The results for confirmation decisions are generally very similar to the individual voting analysis. For both measures, the court-outcome based model classifies only about 60% of confirmations correctly. The performance of the position-taking senators model is even more dismal. The former classifies only about 40% of confirmation decisions correctly. Again, when all model variants predict rejection, confirmation is the much more likely outcome.

Because the court-outcome based model’s prediction of when to reject is robust, this means that in nearly one of out every three nominations, the Senate is approving nominees that all variants of MTM theory predict should be rejected. If presidents are selecting nominees to further their own ideological interests on the Court, the Senate’s behavior means the president has much more leeway than MTM theory would suggest.

Own goals

The first two robust predictions are independent of the model-specific regions seen in Figure 2 and hence are straightforward to test. Recall that the president should never commit an “own goal” by choosing a nominee on the “opposite” side of the old median justice, since the worst the president can do is to select a nominee exactly at the location of the old median justice. Figure 3 depicts the distance between the old median justice and the nominee, scaled in the direction of the president, for both the NOMINATE and Bailey measures. The points show the median estimate across simulations for each nominee, along with 95% confidence intervals. Thus, positive values mean that the nominee is on the “correct” side of the president, while negative values (those in the shaded region) indicate an own goal. For nominees in the latter category, the numbers depict the probability that the estimate is statistically less than zero.

FIGURE 3.

Evaluation of “Own Goals” by Presidents. See text for details

Figure 3 reveals that, in general, presidents have avoided scoring “own goals.” In fact, according to the Bailey measures, zero nominees display a statistically significant probability that the nominee was on the wrong side of the old median justice. For the NOMINATE measure, however, for eight nominees the probability that the nominee was on the wrong side of the old median justice is highly statistically significant. This means that in more than 15% of nominations from 1937 to 2010, presidents did make self-induced errors. Moreover, of these nominees, five potentially had the effect of moving-the-median justice in the opposite direction. Thus, in these instances presidents failed to clear the easiest hurdle of MTM theory: do not move the median away from you.

Notably, all such nominations were made by Presidents Roosevelt, Truman, and Eisenhower—including perhaps the most famous own goals, Eisenhower’s nominations of Earl Warren and William Brennan. This means that the last own goals occurred more than six decades ago. This fact accords with the conventional wisdom that presidents have shifted over time towards a policy-making focus in their Supreme Court appointments (Yalof Reference Yalof2001), and means that the modern threat to MTM theory is presidents selecting nominees that move-the-median too far in the direction of the president, rather than away.

Aggressive mistakes

The second robust prediction is that the president should never make “aggressive mistakes”—selecting a nominee who moves the median father away from the Senate median than the old median justice. Before evaluating this prediction, we first examine the incidence of the necessary condition for such a mistake to occur: that the nominee himself is farther from the Senate median than the old median justice. Recall that under the position-taking senators model, the president should not select such a nominee. Thus, for the robust prediction to fail, the position-taking senators prediction must first fail, such that the nominee has the potential to move the median justice too far (relative to the Senate median).

Figures 4(A) and 4(B) depict estimates of the absolute value of the distance between the nominee and the Senate median, minus the absolute value of the distance between the old median justice and the Senate median, along with 95% confidence intervals. Positive values thus indicate that the nominee is father away from the Senate median than the old median justice, while negative values (the shaded regions) indicate that the nominee is closer. Solid dots indicate confirmed nominees, while open dots indicate failed nominees. The plots show that, using the NOMINATE measure, 33 out of 46 (72%) of nominees were “too extreme” relative to the Senate median (i.e., have positive values). For the Bailey measure, some 23 out of 33 (70%) of nominees were too extreme. Moreover, these conclusions generally hold even when accounting for uncertainty. Using the NOMINATE measure, 22 of 33 nominees with positive values have at least a 95% probability of being too extreme (i.e., their confidence interval does not include zero). Under the Bailey measure, 17 of 23 nominees with positive values have at least a 95% probability of being too extreme. Thus, the prediction of the position-taking senators model frequently fails, as the president nominated someone more extreme than the model would predict.

FIGURE 4.

Evaluation of “Aggressive Mistakes” by Presidents. Top: In terms of the nominee. Bottom: In terms of the new median justice. See text for more details

Having established this result, we now evaluate the robust prediction of no aggressive mistakes. Figures 4(C) and 4(D) depict the absolute value of the distance between the new median justice and the Senate median minus the absolute value of the distance between the old median justice and the Senate median, for both measures. Because many nominations do not provide presidents with an opportunity to move the median, the number of nominations in which nominees actually move the median too far is smaller than the number of nominees who themselves are too extreme. But, using the NOMINATE measure, 20 of 46 nominees (43%) moved the median too far, relative to the Senate median. Notably, and consistent with the Senate voting results above, fully 16 of these nominees were confirmed by the Senate, rather than rejected. Out of these 20 nominees, for 11 there exists at least a 95% probability that they moved the median too far (i.e., the point estimate is significantly greater than zero). The results are similar under the Bailey measure: 17 out of 33 nominees moved the median too far; however, only five of these nominees have statistically significant positive values (due in large part to the greater uncertainty in the Bailey measures).

Is the president ever median locked?

The prevalence of aggressive mistakes shows that presidents often select nominees who are too extreme under all variants of MTM theory. But it does necessarily mean that the Senate cannot act as as a greater constraint across different types of nomination. Specifically, recall the third robust prediction, which we denoted “median locked”: for all gridlocked nominations, meaning the vacancy falls on the opposite side of the presidency, the president must select a nominee at the location of the old median justice. Conversely, in other regions, he is free to move the nominee either to his ideal point, or least closer to it, depending on the model variant. A complication arises in evaluating regime-specific predictions given the uncertainty in the data. For some nominations, the predicted location of a nominee (for a given model variant) will not vary significantly across simulations. For other nominations, there exists much greater variance. For instance, in the vacancy that led to Stephen Breyer’s nomination, 100% of simulations result in “restoring” nominations in the court-outcome based and nearly court-outcome based models. Conversely, for Hugo Black, 35% of his simulations place him as a “smaller shift” nomination, 43% as a gridlock nomination, and 21% as a maximum shift nomination. For such nominees, the data are simply too noisy for us to make firm point predictions.

Accordingly, to test the median locked prediction, we select nominees where we are at least 50% confident that the nomination falls into the median locked category—that is, nominees where a majority of simulations place them in this region. (Below we conduct a more systematic regression analysis in which we both use all nominees and distinguish among the different predicted locations across different nomination regimes.) For each of these nominees, we then estimate the difference between the nominee’s estimated ideal point and the old median justice, and well as 95% confidence intervals around that distance. The robust prediction is that the confidence intervals for median locked nominees should include zero (meaning the nominee is located at the old median justice, accounting for uncertainty).

Figure 5(A) depicts the results of this analysis using the NOMINATE-based measures, Figure 5(B) using the Bailey-based measures. (Note that the set of nominees across the two measures differ based on whether the data place them in the gridlock region.) The point estimates show the median difference between the nominee and the old median justice (the confirmed and unconfirmed nominees have, respectively, solid and open circles). Thus, positive (negative) values indicate that the nominee was more conservative (liberal) than the old median justice. We order the nominees by party—Democratic appointees appear in the shaded regions—and then by decreasing differences.

FIGURE 5.

Evaluation of the “Median Locked” Prediction. See text for details

Two strong patterns emerge from Figure 5. First, the robust median-locked prediction fails much more often than not: only rarely do the confidence intervals around the difference between the nominee and the old median justice include zero. For the NOMINATE measures, this occurs in only four out of 21 nominees; for the Bailey measures it occurs in five out of 14. Second, the errors are not random: presidents tend to choose nominees on “their side” away from the old median justice. This is particularly noticeable among Republican appointees, who across both measures are almost always significantly more conservative than the old median justice (the exceptions are Eisenhower’s appointments of Warren, Harlan, and Brennan, using the Bailey scores). To be sure, many of the nominees were ultimately rejected by the Senate. But many aggressive mistakes by Republican presidents nevertheless resulted in confirmation.

For Democratic appointees, the picture is less clear cut. The Bailey measures place only three nominees in the median-locked region—the confidence interval for each includes zero. Under NOMINATE, four Democratic appointees are significantly more liberal than the old median justice, while three Roosevelt appointees are more conservative (Burton, Stone, and Byrnes). Interestingly, the last time a Democratic president was clearly median locked was in 1967, when Lyndon Johnson nominated Thurgood Marshall. This means that the asymmetric polarization among nominees that Cameron, Kastellec, and Park (Reference Cameron, Kastellec and Park2013) document, where Republican nominees have become increasingly conservative over time, has come even as Republican presidents have tended to face greater theoretical constraint from the Senate, in terms of MTM theory.

Regression analysis of presidential selection

Despite the failure of these robust predictions, it could still be the case that presidents are more constrained when they do face gridlock nominations than when they do not. To evaluate this possibility, we conduct a more systematic (but weaker) test of presidential location: does the ideology of the nominee move in accordance with the predictions of MTM theory? Because the court-outcome based model predicts a range of possible nominees under certain conditions, and because the predicted location is sometimes unobservable in the mixed-motivations model, we can only conduct tests of the nearly court-outcome based and the position-taking senators models. We follow the switching regression approach of Moraski and Shipan (Reference Moraski and Shipan1999), in which the predicted location varies across a given region. From Figure 2, it can be seen that for both models, there are three possible predicted locations: the ideal point of the president, the Senate’s indifference point (or “flip” point) around the old median (2s_m − j ⁰ ₅), and the old median justice. The key difference across the models, of course, is that they will often place the same nominee in a different region, and thus create a different prediction under the same configuration of preferences across actors. Thus, let G denote a gridlock nomination, F a “flip” nomination (where the predicted location is 2s_m − j ⁰ ₅), and P denote a nomination where the president can appoint someone at his ideal point (which, recall, is denoted with a lowercase p). For each model, we can then estimate the following linear model, which we call the “main” regression:

Under MTM theory, the predicted coefficients for β₁, β₂, and β₃ is 1, while the predicted coefficient for the constant is 0. In addition, testing each model requires evaluating whether each respective quantity (j ⁰ ₅, p, and (2s_m − j ⁰ ₅)) does not predict nominee location in the regions where it is not supposed to. Let Not G, Not P, and Not F denote instances where a nominee is not in those respective regions. We then fit the following “placebo” regression:

The predicted coefficients for β₁, β₂, and β₃ is 0.

Table 3 presents eight models—the dependent variable in each is the nominee’s estimated location. Each regression accounts for the uncertainty in the independent variables; the brackets under each estimate depict 95% confidence intervals.Footnote ¹⁵ There are four regressions each for the nearly court-outcome based and position-taking senators models: the models alternate between the NOMINATE- and Bailey-based measures.

TABLE 3.

Linear Regression Models of Presidential Selection

Notes: In each model the dependent variable is the estimated location of the nominee. 95% confidence intervals in brackets, which are estimated via simulation. The R ² values presented are the mean R ² estimate across all simulations, for a given model.

Beginning with the nearly court-outcome based model, Models (1) and (2) present the main regressions. While the coefficients on the Gridlock × j ⁰ ₅ are in the predicted direction, the confidence interval for each includes 0 (though they both also include 1). In contrast, the coefficients on President predicted × p are both statistically larger than 0; however, they are statistically less than 1, meaning nominee location does not vary as strongly with presidential ideology as MTM theory would predict. Finally, the coefficients on Flip × 2s − j ⁰ ₅ are indistinguishable from both 1 and 0 (the confidence intervals are much larger due to the small number of observations that fall into the flip region). Thus, the main regressions show at best weak support for the nearly court-outcome based model.

The next key question is whether a given actor’s ideology does not predict nominee location in the regions where it is not supposed to. Models (3) and (4) test the placebo regression for the nearly court-outcome based model. The coefficients on Not gridlock × j ⁰ ₅ are statistically indistinguishable from zero. However, the coefficients on Not president predicted × p are positive and significantly different from zero, meaning that presidents choose nominees based on their own ideology even when they should not be able to. Moreover, the magnitude of the effect of the president’s ideal point is statistically indistinguishable when we compare the coefficient on President predicted × p in the main regressions to the coefficient on Not president predicted × p in their placebo counterparts.

Turning to the position-taking senators model, the results tell mostly a similar story. The main regressions in Models (5) and (6) show that Gridlock × j ⁰ ₅ is both positive and either statistically distinguishable from 0 or very close to it (the confidence interval in Model (6) only barely includes 0). The coefficients on President predicted × p are both positive, although under NOMINATE the confidence interval includes 0. (Recall that the president is much more constrained in the position-taking senators model, since the Senate evaluates the nominee against the old median justice; this means that there are many fewer observations in which the predicted location is at the ideal point of the nominee, thereby increasing the uncertainty of the estimate.) Both coefficients, however, are also statistically indistinguishable from 1, as the theory predicts. Finally, the coefficients on Flip × 2s − j ⁰ ₅ are indistinguishable from both 1 and 0.

As with the nearly court-outcome based model, these results provide weak support at best for the position-taking senators model. Moreover, when we turn to the placebo models in Models (7) and (8), we again see that the president’s ideal point predicts nominee location even under conditions when it should not. Thus, combining these results with our robust tests above, it is clear that the president has much more influence over the location of Supreme Court nominees than MTM theory would predict.