Robert M. Howard and Jeffrey Lazarus. The Sentencing Reform Act of 1984 created the United States Sentencing Commission, which had the responsibility to promulgate sentencing guidelines. These guidelines created a range of determinate sentences for all categories of federal offenses, and were binding upon all federal judges. In a curious aftermath of legislation designed to toughen the response of courts to crime, scholars have noticed that federal judges' conviction rates have fallen since the 1980's (Leipold 2004). This unexpected trend raises the possibility that the outcome of the Sentencing Act is opposite of what Congress intended. When faced with mandatory sentencing guidelines, judges sometimes try cases in which they believe the defendant to be guilty, but the legally mandated sentence is too harsh given the circumstances. In these cases, judges may choose to acquit. Since these defendants would have been convicted in the absence of guidelines, aggregate conviction rates should be lower when guidelines are in effect than otherwise. We test these hypotheses with data on federal bench trials from 1970 to 1995. We find that federal judges' conviction rates have fallen substantially and that the trend is attributable, at least in part, to the Sentencing Reform Act of 1984 and the resulting binding sentencing guidelines. We conclude that although the Act allowed Congress to appear "tough on crime" to voters, the substantive result of the legislation was decidedly not "tough on crime."
Responding to public dissatisfaction with perceptions of leniency towards criminal defendants by the courts, state legislatures and the U.S. Congress subsequently introduced legislation designed to reduce judicial discretion in sentencing convicted criminal defendants. One motivation for curtailing discretion came from liberal voters and activists upset that ethnic minorities often received harsher penalties than white defendants convicted of the same crimes. Thus, one purpose of sentencing guidelines was to reduce sentencing disparity among convicted defendants. However, a second motivation was more politically salient: contemporary polls showed that over 80 percent of respondents of all political parties thought courts were "soft on crime." Thus, legislators sought to increase (or create the impression that they were increasing) sentences served by those convicted of crimes.
At the federal level, this effort culminated in the passage of the Sentencing Reform Act of 1984, which was signed into law by President Ronald Reagan on October 12, 1984, just prior to the presidential election. The Sentencing Reform Act created the United States Sentencing Commission, which had the responsibility to promulgate sentencing guidelines. These guidelines would create a range of determinate sentences for all categories of federal offenses, eliminating much of the sentencing discretion previously enjoyed by the federal judiciary. Although lately the Supreme Court has backed away from the ruling (see Blakely v. Washington 2004; United States v. Booker 2005) the Court initially affirmed the constitutionality of the Sentencing Reform Act in 1989 in Mistretta v. United States.
However, in a curious aftermath of legislation designed to toughen the response of courts to crime, scholars have noticed that federal judges' conviction rates have fallen since the 1980's (Leipold 2004). This unexpected trend raises the possibility that the outcome of the Sentencing Act is opposite of what Congress intended. They wanted the courts to get "tough on crime" by removing judicial discretion in sentencing, but instead ended up with fewer convictions - which is decidedly not "tough on crime" politically.
In this manuscript we investigate the connection between sentencing guidelines and aggregate conviction rates. More specifically, we ask whether the Sentencing Reform Act of 1984, and the subsequent affirmance of the constitutionality of the Act and Sentencing Commission guidelines by the Supreme Court in Mistretta in 1989, directly contributed to this drop in federal conviction rates. Toward this end, we posit that when guidelines are in effect, there exist some defendants who the presiding judge believes to be guilty, but deserving of a lighter sentence than that mandated by guidelines. Under some circumstances, the judge may choose to acquit these defendants rather than subject them to (what the judge believes is) an overly harsh sentence. Since few if any of these defendants would have been acquitted in the absence of sentencing guidelines, aggregate conviction rates are lower under a guidelines regime than they would be otherwise. Thus, this theory predicts a drop in acquittal rates once guidelines take effect.
We test these predictions by gathering data from all federal criminal bench trials from 1970 through 1995 and modeling the probability of a conviction. While the number of trials is small compared to the overall number of cases, they are the standard against which trial attorneys judge any plea deal: if they think they can do better than the deal in a trial, often they reject the deal. Thus, anything which fundamentally alters the probability of conviction at trial will also fundamentally alter the structure of plea bargaining in all cases.
We show that a significant drop in the probability of conviction in a given case occurred in 1989, the year federal guidelines came into effect. This effect is evident even after controlling for other case-level and systemic factors, and is robust to a variety of model specifications. Thus, we conclude that the Sentencing Act of 1984 caused a permanent and sustained reduction in judicial convictions of criminal defendants in federal criminal trials.
Sentencing Guidelines and Judicial Reaction
The implementation of sentencing guidelines has influenced the judicial system in many ways, both at the national and state levels. Most notably perhaps are the effects that guidelines' proponents predicted. Prior studies show that guidelines increase sentence severity for defendants who are convicted of crimes (Stolenburg and D'Allessio 1994, Engen and Steen 2000, Helms and Jacobs 2002), sentence uniformity for disparate groups of defendants who are convicted of the same crime (Miethe and Moore 1985, Stolenburg & D'Allessio 1994), and aggregate incarceration rates (Nicholson-Crotty 2004, Marvell and Moody 1996, Stolzenberg & D'Alessio 1996) In addition, there have been efforts to document unanticipated effects as well, such as changes in prosecutorial behavior (Engen and Steen 2000, Knapp 1987, Miethe and Moore 1985, Miethe 1987).
Still another area investigates how sentencing guidelines influence judicial behavior on the bench. These studies strongly indicate that even though guidelines limit judicial discretion, they do not eliminate it altogether. Judges still have options available to them when trying cases, and they can use these options to impose their preferences on outcomes. For instance, Kramer and Ulmer (1996) find that, in a significant portion of cases, judges' hand down sentences which are outside the range mandated by the guidelines. Other scholars argue that these departures from the guidelines are strategic, insofar as judges use them systematically to pursue broader goals. Albonetti (1997) posits that lower-than-mandated sentences are frequently "awarded" to defendants who assist the government in the prosecution of other defendants. Additionally, judges may not even need to depart from the guidelines to impose their preferences on outcomes. Knapp (1987) finds that a disproportionate share of cases in which judges identify an "aggravating factor" - which results in a longer sentence - were property crime cases, for which guidelines mandated more-lenient sentences than had typically been given prior to their enactment.  This shift indicates that at least some judges believed that the mandated sentences were too lenient, and acted to lengthen sentences.
One area of judicial discretion which has not been looked at in the context of sentencing guidelines is case disposition. While guidelines impose limits on sentencing, they do not regulate disposition at all: in bench trials judges have complete discretion over whether a defendant is convicted or acquitted. Despite their silence on the matter, we argue that guidelines alter judges' incentives such that dispositional outcomes are affected. In particular, we posit that sentencing guidelines lead judges to be less likely to convict defendants.
Judicial Behavior, Outcomes and Constraints
Judicial behavior is influenced by many factors which are external to the facts and law of a particular case. Perhaps the most prominent of these influences are judicial attitudes and ideology (e.g., Segal and Spaeth 1993). The influence of ideology has been accepted as influencing, to some degree, outcomes even at the district court level by most political scientists (Rowland and Carp 1996, but see Ashenfelter, Eisenberg and Schwab 1995). Although most also agree that hierarchical and other constraints often obviate the influence of ideology (Segal and Spaeth 1993). Evidence has also been found that electoral considerations (Huber and Gordon 2002) alter judges' behavior. In addition, judges' personal characteristics, such as their race and gender, are significantly related to sentencing outcomes (Muhlhausen 2004; Huber and Gordon 2002). Finally, several scholars argue that the separation of powers system under which judges act take into account the actions of the other branches of government in order to obtain preferred outcomes (Epstein and Knight 1998; Ferejohn and Shipan 1991; Eskridge 1991).
We posit that, like these factors, the presence of sentencing guidelines also influences in-court behavior. In particular, we hypothesize that judges who operate under sentencing guidelines are more likely to acquit than judges who do not. On the federal level, this hypothesis is supported to some degree by the fact that judicial acquittal rates have fallen since the implementation of federal sentencing guidelines in 1984 (Leipold 2004). Figure one shows the overall yearly conviction rates in federal bench trials from 1970 through 1995.
Figure One here
As one can see judicial conviction rates started at slightly over 60% and then trended with a gradual increase in conviction rates to around 80% through the 1970's and early 1980's. However, starting in the late 1980's through the end of our observation period in 1995 the conviction rate has decreased to around 50%, lower than ever before. Whenever one attempts to explain a broad national trend such as displayed in Figure one there are likely to be multiple causes.  However, we argue that one cause is the presence of sentencing guidelines. Guidelines bring down conviction rates because they alter the relationship between what a judge feels is the most appropriate outcome in a case, and what the judge legally can do.
To derive this hypothesis, we borrow from economics the idea that people make decisions rationally; that is, when they are faced a decision to make they identify the alternatives available to them, rank them in order of preference, and choose the alternative that they most prefer. In more technical language, we assume that judges are rational utility maximizers, and that the decisions they make while on the bench reflect this decision-making approach. This assumption has long history in the literature on judicial behavior, and has been employed in a wide range of studies on the topic from law, economics and political science beginning with Murphy's now classic The Elements of Judicial Strategy (1964) through Knight and Epstein's The Choices Justices Make (1998) to studies on opinion assignment and content (e.g. Maltzman, Spriggs and Wahlbeck 2000; Corley, Howard and Nixon 2005), certiorari (Boucher and Segal 1995; Caldeira, Wright and Zorn 1999) to many other topics including from litigation (Posner 1993) to constitutional law (Stearns 1995).
We depart from the economics literature on the point of what contributes to utility. Â Traditionally, utility represents personal value, broadly conceived. Thus, the assumption that actors maximize utility when facing a choice set requires them to select the choice which increases their own personal value the most. Judges do not ordinarily derive value from choosing one trial outcome rather than another. Rather, in the judicial politics rational choice literature utility often takes the form of policy preferences. In this vein, judges "decide to decide," vote on cases, or assign opinions to obtain a policy outcome which is as close to their ideal as possible.
However, broad policy decisions are rarely if ever at issue in criminal trials, the subject of our analysis. As a result, we employ a third conception of utility which is grounded in the notion that judges care about the law and fairness. Indeed to use the term justice embodies these concepts. Because of this we assume that judges have preferences over the fairness of case outcomes. Â Toward this end, we conceive of the range of case outcomes (including final disposition and sentencing) as being arrayed along a single dimension, severity. A judge believes that a given outcome represents an appropriate level of severity of punishment (including the possibility of no punishment if the judge believes the defendant to be not guilty); this represents his or her "ideal point." If judges cannot obtain their preferred outcome, they will choose the available outcome which is closest to their ideal point (i.e., judges have single-peaked, symmetric preferences). This model is similar to those employed in studies of political behavior which assume that different policy choices can be arrayed along a single dimension (see e.g. Ferejohn and Shipan 1990; Segal 1997).
Toward this end, we examine a hypothetical case in which a judge is presiding over a bench trial, believes the defendant is guilty, and is constrained by the guidelines in terms of sentencing options. Under these circumstances, the sitting judge can take one of three views toward the mandated sentence or sentence range. First, the judge might agree that the sentence is appropriate, given the charges, personal characteristics of the defendant and other related factors. Second, the judge might believe that the sentence is too harsh for the given circumstances; third, the judge may believe that the sentence is too lenient for the given circumstances. In all three cases, the judge has preferences over the sentencing outcome, but in only one of them - when the judge agrees with the mandated sentence - can the judge legally impose his or her most-preferred outcome.
If the judge prefers a sentence which is more lenient than the one mandated by the guidelines, he is left with two options: convict and subject the defendant to an overly-harsh sentence, or acquit even though the judge believes the defendant to be guilty. From the point of view of the judge, both options are sub-optimal: acquittal sets a guilty defendant free (i.e., is too lenient), while conviction subjects a defendant to more punishment than the judge believes is warranted (i.e., is too harsh). Importantly, though, neither option is inherently preferable than the other, and the judge might reasonably choose either one. Which option the judge chooses in any particular case will depend on case- and judge-specific factors. However, if judges acquit a significant fraction of the time, there exist a set of cases in which the dispositional outcome is acquittal, but would have been conviction in the absence of sentencing guidelines.  The emergence of this set of cases causes the overall conviction rate to be lower under sentencing guidelines than they would be without the guidelines, all else equal.
The third possibility in the hypothetical situation is that the judge prefers a harsher sentence than is mandated by the guidelines. We do not expect these cases to have a significant impact on dispositional outcomes. Judges are once again precluded from imposing their most-preferred outcome on the case, but the two options available to the judge - acquittal or conviction with a too-lenient sentence - are both more lenient than the harsh sentence preferred by the judge. Moreover, one of these options is demonstrably closer to the judge's preference than the other. If a judge feels that a defendant is guilty and a harsh sentence is warranted, then a conviction with a lenient sentence is preferable to an acquittal. For this reason, judges who prefer harsher sentences than those mandated by the guidelines are likely to convict anyway, and these situations are very unlikely to change dispositional outcomes in any individual case. Thus, overall conviction rates will not change because of this group of cases.
Figure 2 here
This logic is represented formally in Figure 2, a one-dimensional spatial model which represents the judge's options in the case described above. The horizontal line represents an axis of outcome severity. The left end-point represents the least-severe outcome possible (i.e., acquittal), and severity increases as the line moves to the right. The area represented in bold and bracketed by the endpoints G1 and G2 represents the sentence range mandated by the guidelines; the model assumes that conviction always results in a sentence within this range. Thus the only possible outcomes are acquittal or conviction with a sentence within that area. An unrepresented point, J, represents the judge's ideal sentence and may exist anywhere along the axis. If J is within the guidelines, this indicates that the judge's most-preferred sentencing option is legally available, and the judge convicts. If J is to the right of the guidelines, the judge once again convicts since he prefers a too-light sentence to acquittal, as discussed above. However, when J is to the left of the sentencing range, the judge's preferred outcome depends on the specific location of J: when J is closer to acquittal, the judge acquits; when J is closer to G1, the judge convicts. The point which divides the model's "Acquittal Zone" from its "Conviction Zone" is the midway point between A and G1.  Any J to the left of that point indicates a preference for acquittal over the guideline-mandated sentence; any J to the right of that point indicates a preference for the mandated sentence.
The existence of the Acquittal zone exists solely because guidelines constrain judicial discretion in determining sentences. If J were within that area in a case in which the judge had full discretion, the judge could convict and impose his or her preferred sentence. Thus, this model predicts that there will be some defendants who are acquitted under sentencing guidelines who would have been convicted in the absence of guidelines. Thus, the model's primary empirical prediction is that aggregate conviction rate for bench trials should be lower under sentencing guidelines than otherwise.
We note two caveats. The first is that judges' decisions are not actually constrained as strongly as they are in the above model. Judges do have some discretion to depart from the sentencing range provided in the guidelines, especially in the direction of more lenient sentences (United States Sentencing Commission 2003). However, such "downward departures" are permitted only under specific circumstances. According to the guidelines (sections 5K1.1 and 5K2.0), one prominent circumstance exists when the sentence is an "aberration" because of the severity of the sentence, the age of the defendant, or the lack of a criminal record; empirically, the dominant share of departures occur due to the presence of one of these aberrations.  The possibility of downward departures likely mitigates the effects of the logic discussed above as it gives judges who feel that the mandated sentences are too harsh an option besides acquittal. However, since departures are permitted in only specific cases, we posit that they exist side-by-side with, rather than instead of, excess acquittals: in cases where a judge cannot justify a departure, the logic above still holds.
Second, the same empirical result - i.e., a drop in conviction rates - would be observed under a second set of circumstances: if the cases which make it to trial under a guidelines regime are weaker cases, on average, than those which make it to trial under a non-guidelines regime. However, in order for this to be the case, defendants who accept a plea under a guidelines regime, but would not have without the guidelines, would have to have systematically stronger-than-average cases against them. There is no theoretical reason to believe this is the case.  Checks of our dataset indicated similar averages and proportions in terms of defendant demographics and case severity for both time periods.
We test our theory with data from the Federal Court Cases: Integrated Data Base, available through the ICPSR (Federal Judicial Center). Our data are from 1970 through 1995. The dates selected provide sufficient time before and after both the 1984 Sentencing Reform Act and guidelines' subsequent implementation to evaluate the impact of the sentencing reform act.  In addition the data provide very different political circumstances including both unified and divided government and essentially end before the very different circumstances of the Republican takeover of Congress following the 1994 midterm elections. The collected data had over 1.3 million observations, but we selected only the cases in which a judge presided over a trial and issued a final determination (to convict or acquit) following that trial. Hence all pleas, mistrials, and dismissals without prejudice were eliminated from the analysis, as well as all jury trials. This left us with a total of 33,554 observations.
While this is small number relative to the initial number of cases, we argue that this number is both sufficient for the analyses and in no way lessens the importance of bench trials. First over 33,000 observations is a significant number of observations to allow confidence in the reliability and validity of the results. While the number of trials is small compared to the overall number of cases, they are the standard against which trial attorneys judge any plea deal: if they think they can do better than the deal in a trial, often they reject the deal. Thus, anything which fundamentally alters the probability of conviction at trial will also fundamentally alter the structure of plea bargaining in all cases. In addition, so much significant judicial research is based on subsets of cases. As an example, the Supreme Court receives over 7,000 petitions per year for certiorari, yet renders judgment on fewer than 100 cases per year. Important and insightful research has flowed from those articles and books which have solely focused on the outcomes of these Supreme Court cases.
One empirical question to address is to identify exactly when federal sentencing guidelines took effect. The U.S. Congress passed legislation creating the Federal Sentencing Commission, which was charged with devising guidelines, in 1984; the law was signed by President Reagan on October 12th of that year. However, the guidelines did not take full effect until the guidelines had been fully written by the commission, and the U.S. Supreme Court had rejected challenges to their constitutionality (United States v. Johnson; Mistretta v. United States). Thus, the law was not implemented until late 1989.
Figure 3 here
In Figure 3, we display the bivariate relationship between the adoption of sentencing guidelines and federal conviction rates in bench trials. We compare three different time periods of time and two specific dates representing the implementation of the guidelines. The first column shows conviction rates from the beginning or our data until 1983, the year before the Sentencing Reform Act of 1984, the second from the implementation of the Act in 1984 until the year before the Mistretta opinion and finally from 1989, the year of the opinion, onward. Similar to Figure two, the columns show a general downward trend in judicial conviction rates, with a noticeable drop in the judicial conviction following the Sentencing Reform Act of 1984. However the differences are most pronounced following the Mistretta opinion.
This drop represents prima facie evidence for a link between guidelines and conviction rates. Note that we are not arguing that federal sentencing guidelines are the sole cause of this drop in conviction rates. National sentencing trends are subject to a host of influences, including crime rates, caseloads, and the political orientation of judges to name but a few. Any combination of these may have been in play during this time period. Rather, we argue that the guidelines were one significant contributing factor, and that the occurrence of a significant drop beginning concurrently with the implementation of sentencing guidelines is not a coincidence. In addition, there is some argument that cases which went to trial after the implementation of the guidelines are somehow systematically weaker than cases which went to trial before the implementation of the guidelines. However, there is simply no theoretical reason or empirical evidence to support this assertion. Thus, below, we test this hypothesis more systematically.
Our primary empirical analysis consists of a multivariate estimation of the judicial determination of guilt or innocence. Our dependent variable takes on two values: 0 when the defendant is acquitted and 1 when the defendant is convicted. As a result, we employ logit as our estimation tool. Our key independent variable indicates the implementation of sentencing guidelines in October 1989: thus, it is a dummy variable, Guidelines, coded 0 for all cases heard prior to implementation, and 1 for all cases heard after. Our theory predicts that the coefficient on Guidelines will be negative and significant.
We also include several control variables to account for other influences on the likelihood of a conviction. In selecting control variables, we are guided by prior studies of judicially-imposed sentences (e.g., Huber and Gordon 2004; Brace and Hall 1997; Hall 1992; Kuklinski and Stanga 1979). This study centers on dispositional outcomes rather than sentences, but it is reasonable to posit that the same factors influence both types of outcomes. Simply put, since the judiciary take mandatory sentencing into account in determining guilt or innocence they would take the factors that would be used in determining the sentence into the initial trial phase of conviction or acquittal.
In addition we used variables to control for alternative explanations. For example as previously discussed in footnote two, one explanation is selection bias. As arrests and prosecutions grew over time, there was an inevitable increase of "bad" cases. Thus our sample overemphasized bad cases that would lead to acquittal. However as we previously discussed, given no corresponding increase in dismissals or verdict set asides we do not believe this to be the case. Nevertheless, we include district wide variables for caseload and crime rate to control for this possibility. As the caseload rate goes up, one would expect this to lead to a greater number of acquittals. The crime rate can have one of two influences. The crime rate can be an indication of more arrests, and thus more "bad" arrests. If this is so, we should expect an increase in the crime rate to lead to an increase in acquittals. Conversely, more arrests might also simply lead to a greater number of convictions. We are neutral as the influence of this variable.
Also, we account for the possibility of downward departures by including control variables which indicate the presence of one of the "aberrations" which can trigger a departure. Toward this end, we included control variables for the severity of the crime (1, if felony, 0 otherwise) and if the defendant had a prior criminal record (1, if yes, 0 otherwise). In addition to these variables we also added several control variables that are more generally used to determine sentencing outcomes. Most derive from the initial database, and account for individual-level case factors. For example, we created a variable that coded whether or not the defendant was nonwhite (1, yes, 0 otherwise), and a second variable indicating whether the defendant was female (1 yes, 0 otherwise). Following the literature on sentencing, we expect minority racial status to increase the probability of conviction (Steffensmeier and Demuth 2000; Engen and Steen 2000). Moving beyond defendant demographics, two dummy variables indicate the presence of counsel: separately, private counsel indicates that the defendant is aided by a private attorney, while public counsel indicates that the defendant is represented by a public defender. The base category consists of defendants who were not represented by counsel. Both types of attorney should lower the probability of conviction, just as prior studies find that the presence of an attorney lowers the severity of a sentence. However, we expect the coefficient on private counsel to be larger in magnitude than the coefficient on public counsel.
Finally, we use a measure of judicial ideology by employing a measure that is similar to the formula developed by Giles, Hettinger and Pepper (2001, 2002), which used Nominate scores of the home state senators or of the nominating president depending upon the political circumstances of the appointment (Nixon n.d.; Howard and Nixon 2003; Howard 2008). The ideology are calculated by using a formula to match the common space nominate scores of congressional representatives who later served as federal judges. The formula uses the circumstances of the appointment, including party of the judicial nominee, party of the appointing president, the ideology of the state of the nominee as determined by the Wright, Erikson and McIver measure, whether the nominee comes from the northeast or south, and whether or not there was a unified government at the time of the appointment.
It assumes ideology remains constant from year to year and when one moves from being a congressional representative to a federal judicial position. The scores range from -.5 (most liberal to .5 (most conservative) and provide substantial differentiation between judges. Unlike the Giles, et. al. scores, this measure allows for differences for judges even if appointed from the same state by the same president in the same year. Since the data does not list the individual judge we used the yearly nominate score of the median justice for each judicial district in our dataset (Martin, Quinn and Epstein 2004; Hausegger and Haynie 2003).
There were two case-level control variables which we considered including, but decided ultimately to exclude. The first is the age of the defendant. Gordon and Huber find that the harshest sentences are given to middle-aged defendants; both very old and very young defendants received lighter sentences. As such, we intended to include age and age squared to account for a possibly similar trend in case disposition. However, this proved impossible as the variable in the Integrated Data Base which indicates the defendant's age is missing in most observations, and in all observations after 1983. Thus, including this variable would require us to not only lose most of our data, but also result in having no variation in our primary independent variable. As such, we dropped age altogether.
Finally, we include a variable to account for the general trend in conviction rates, to control for the possibility that an observed difference between the pre and post-1989 time periods is actually the result of a more general, slower-moving trend. Because the specific prediction of our theory is that a secular change occurs at a given time, empirically accounting for this overall time trends is especially important. Thus, we present three models in which time is modeled in three different ways. Model 1 includes a logarithmic time trend. Cases decided in 1970 were coded as 1; cases decided in 1971 were coded as 2, and so on. Then we take the natural log of these codes, to account for any monotonic trend which may change as time passes. Model 2 includes a linear yearly trend. Finally, Model 3 includes both the linear time variable, as well as time squared. We include this because the overall time trend suggested by Figure 1 is not necessarily monotonic: within our dataset conviction rates first rise for a period, then fall. Thus, we believe that including time and time squared presents the strongest possible test of our hypothesis.
Table one here
Table one provides summary statistics of the independent variables. Seventeen percent of the cases in our dataset occurred after the imposition of the guidelines, and almost 40 percent of the defendants had private counsel. Given the scale of judicial ideology, the mean ideology of the judicial districts was slightly conservative, reflecting a time period dominated by Republican presidential appointments. Given the variables and the dichotomous nature of the dependent variables, we test our expectations about judicial and jury conviction rates using logistic regression, constructing a model that takes on the following form:
Conviction = ï¢0 + ï¢ 1 Guidelines + ï¢ï€ 2 Private Counsel + ï¢ 3 Public Counsel + ï¢ 4 Female + ï¢ 5 Nonwhite + ï¢ 6 Ideology + ï¢ 7 Caseload + ï¢ 8 Crime Rate + ï¢ 9 prior Convictions + ï¢ 10 Felony Charge + ï¢ 11 Time Trend + ï¥
We also clustered the data on the district courts for both models and report robust standard errors to control for the potential influence of any outlying observations (Western 1995). 
Results of all three models are presented in Table 2.
Table 2 here
Summary statistics for all three models are quite similar. Each one performs well in predicting the probabilities of a conviction. The chi-squares range from 653.1 to 674.9 (p<.001 for each). More importantly, the results confirm the primary hypothesis, that there was a significant drop in conviction rates in cases before federal judges in 1989, when sentencing guidelines went into effect. This is so even controlling for alternative explanations such as sample bias or downward departures. The test variable, guidelines, is negative and significant in all three specifications of the model, and as we shall see in our discussion of probabilities, substantively significant in leading to acquittals. This indicates that even when time trends and other explanations are accounted for in several different ways, the drop occurring in 1989 is still important. This lends considerable support to the notion that the Sentencing Act of 1984, particularly the implementation of guidelines in 1989 through the Mistretta decision, is a significant factor in the decline of judicial conviction rates.
Looking at control variables, the direction of the probabilities largely conform to previous studies and common sense understanding of sentencing outcomes. In all three specifications, the coefficient on private counsel (-.101, p < .05, model 1) is negative and significant; indicating that the probability of a defendant's conviction is lower if the defendant has secured private counsel. However, the coefficient on public counsel is not significant in any of the specifications, indicating that defendants who are defended by public defenders are no less likely to be convicted than those who have no attorney. This is no slur on the ability of public counsel but more likely recognition that these are often very hard cases with evidence clearly against the defendant. Female (1.25, p < .001, model 1) and nonwhite (.964, p< .001, model 1) are both positive and statistically significant, indicating that female defendants are more likely to be convicted than male; and that minority defendants are more likely to be convicted than white.
Turning to our alternate explanation variables Caseload is also significant in all three specifications (.172, p <.05, model 1), indicating that the probability of conviction in any individual case is systematically related to the number of cases a district court tries in any given year. As cases go up, the number of "bad" cases also goes up, in contradiction to our expectations. However, the substantive impact is quite small, and this also supports our argument that the increase in judicial acquittals is not systematically related to the sample. Caseloads may increase but, except for the negative impact of guidelines, convictions are increasing in tandem with fewer dismissals and jury verdict set asides. Crime rate is neither statistically nor substantively significant.
Our two control variables that account for the "downward departures" explanation perform as expected. Prior convictions (1.70, p< .001) have a statistically significant and a very similar and strong impact on the probability of conviction. A felony charge has a much smaller impact on conviction and was only statistically significant in model 3, and in fact is the only variable to deviate in statistical significance across all three models.
The fourth and last column in Table 2 gives some indication of the magnitude of each independent variable's effect on the probability of conviction. For continuous variables this represents the change in the probability of conviction resulting from an increase of one standard deviation in the independent variable. For dichotomous variables it represents a one unit change from 0 to 1. For each calculation, all other variables are set to their median values. The probabilities reported in Table 2, Column 4 are derived from Model 1, using the Clarify plug-in for Stata. (Tomz et al, 2003). Looking first at the key independent variable, guidelines, the magnitude of moving from a non-guidelines regime to a guidelines regime appears to be quite sizeable, as the probability of conviction drops 14%. This places guidelines among the independent variables which have the largest impact on the probability of conviction. However, this figure is somewhat misleading. Calculating the same values from Models 2 and 3, the effect is 20%, and 3% respectively. Thus, while it appears that the existence of a drop in conviction rates is robust to differences in model specification, the predicted size of this drop is not. As a result, we feel that we cannot make a firm claim relating to the size of the effect based on these data. However we do feel confident in considering these figures, 3% and 20%, as lower and upper bounds. Besides demographic characteristics, the other variables with large probabilities were prior convictions, which increased the probability of conviction by .25 in model 1 and having private counsel, which decreased the probability of conviction by .04.
As a final note, although the Guidelines were introduced in 1987, they applied only to cases in which the relevant conduct occurred after the Guidelines' adoption. Thus many of the cases in the 1987 - 1990 time period were not required to be sentenced under the guidelines. To account for this we tested our model with several different competing implementation variables, specifically five regime dummies, each of which "switches" in a different year, 1987 through 1991. In all the cases the models showed that 1990 continues to be the strongest and most important year for the sentencing change. 
Discussion and Conclusion
The Sentencing Reform Act of 1984 was a popular piece of legislation. Public opinion polls showed significant support for legislation that was "tough on crime." Large majorities in both the Senate and House of Representatives voted in favor of the law. President Ronald Reagan signed the Act into law in Mid October, 1984, just a few weeks before the presidential election. The use of mandatory sentencing guidelines would eliminate much judicial discretion leading to longer sentences for the convicted criminal defendants. The Supreme Court in Mistretta v. United States (1989) upheld the constitutionality of the 1984 Act.
However, as we have shown, in the aftermath of an act curtailing judicial discretion there was an unintended consequence - namely an increase in the acquittal rate of criminal defendants in federal bench trials. While we cannot offer clear evidence as to the magnitude of the affect on conviction rates, the robustness of the results offers clear evidence that the implementation of the guidelines led to a sustained reduction in judicial convictions. This result holds even controlling for crime rate, judicial ideology and other indicators shown to influence sentencing and thus conviction. Such results are consistent with our theory developed in figure two. When the mandated minimum sentence is harsher than the judge believes appropriate then acquittal becomes a preferred option.
Of course there are several research issues still to consider. Because the dataset does not contain the names of the individual judge we had to rely on mean ideology. An individual level measure might allow us to consider and understand the affect of ideology on acquittals as a result of the Sentencing Act. More state level analysis of mandatory sentencing and state judicial response would add the literature in this area. Finally, with the recent Supreme Court decisions of Blakely and Booker calling into question the constitutionality of mandatory sentencing our theory would predict an increase in convictions particularly if the Court explicitly overturns the Mistretta decision. Thus we need to revisit this issue again in the future.