Do You Receive a Lighter Prison Sentence Because You Are a Woman or a White? An Economic Analysis of the Federal Criminal Sentencing Guidelines

See www.ussc.gov/2009guid/TABCON09.htm

United States Sentencing Commission: Fifteen Years of Guidelines Sentencing (2004, 86).

Our raw dataset has around 380,000 observations. Our final dataset contains around 45,000 observations. Of the 340,000 observations that are lost in the data cleaning process, around 150,000 are a consequence of often missing data on the type of defense counsel. In the appendix, we confirm that our main results are robust to the inclusion or exclusion of these observations. Around 66,000 observations are lost on account of being non-citizens or of citizenship status being missing. We also explore the consequences of including these observations in our estimation in the appendix. Our age, gender, race, and ethnicity restrictions lose another 30,000 observations. In addition, around 19,000 observations involved substantial assistance departures and 19,000 included drug sentences with mandatory minimum sentences (and there were about 10,000 observations involving both; these observations were dropped for reasons discussed later).

The appendix also presents results confirming the robustness of our results to the inclusion of sentences handed down after a conviction by trial.

These figures were calculated using the OFFTYPE and OFFTYPE2 variables for the years 1996–1998 and 1999–2002, respectively.

Referred to by XFOLSOR and CRIMHIS, respectively.

For details on the construction of these variables, please see the following documents on the USSC’s website: http://www.ussc.gov/training/sent_ex_rob.pdf; http://www.ussc.gov/training/material.htm

This is the BASEHI variable in the raw data.

Though not a full interaction model, Mustard does offer interactions between the race, ethnicity, and gender indicators and the offense level and criminal history.

Many of these control factors, such as education and access to a private defense attorney may differ across race and gender as well. To the extent that differences in these factors result in an unequal distribution of resources across race and gender, our results represent a lower bound of unjustified differences in sentence length as a whole. However, our intention is to estimate how much of the sentence gap can be attributed to judicial bias. With the possible exception of the criminal severity score (discussed at more length elsewhere), the sentencing judge should have no influence over our control factors. Thus, they are appropriate to condition on if our intention is to estimate judicial bias in sentencing.

Full results from these linear estimates are available upon request from the authors.

We have allowed a total of 100 iterations before forcing convergence. In our main estimation, all models converged before reaching the maximum number of iterations, with the exception of the estimation for White females. In total, for the 13 robustness checks that we have run, the pooled model achieves convergence each time, as does the model for Black and White males and all but one time for Hispanic males. Convergence is achieved 9 of 13 times for Black females and for Hispanic females. We continue to have troubles achieving convergence for White females, where convergence is not achieved 10 of 13 times. However, our key finding of favoritism toward White females vis-à-vis White males is upheld in the estimations where convergence is achieved. These three cases correspond to a larger dataset (with drug minimum sentences and substantial assistance cooperators included).

In some instances, our model was not able to estimate standard errors on some coefficients for White females, one of the smallest demographic groups that we examine. However, to test our main hypothesis of whether there exists discrimination under the guidelines, inference will come from bootstrapped standard errors that do not require the calculation of standard errors on each of the individual parameters for each group.

In addition to race, sex, and national origin, the guidelines also exclude factors such as socioeconomic status, family ties and responsibilities, and (with only limited exceptions) age and education.

In Appendix Table A1, we report our expected sentences by group and parameters used. The table also shows the likelihood of being in each of six regimes (1) 0 sentence, (2) strictly positive but below lower end of guidelines, (3) at lower end of guidelines, (4) within guidelines, (5) at upper end of guidelines, (6) above upper end of guidelines, the expected value of the sentence for individuals that fell into each regime, as well as an overall expected value. Please note that the overall expected value is not necessarily equal to a weighted average of the expected values of individuals within each regime, due to non-linearities. Also note that our expected values are only able to be calculated for observations for which there is a non-zero minimum guideline, as the expected value calculation in the first regime for these observations would involve a zero in the denominator. Thus we have excluded observations that do not meet this criteria. This may limit the external validity of our study, but allows us to consistently study the gaps between groups of the set of observations that do meet this criteria in our data.

Here, the departure is exactly at the threshold below which we would consider the departure to be “large.” It is important to take note of this mass point, just as we have for the mass points at the edge of the guidelines: there is a non-differentiable change in the marginal utility of changing the sentence at this point. Thus, there will be individual sentences for which the judge might prefer a slightly lower sentence, but is not willing to pay the increased utility cost. Here, we do note that the choice of is somewhat arbitrary in the sense that there is not a natural threshold at which there should be a statutorily driven increase in the cost of departure. However, should exist, our theory tells us that there would be a mass point observed. As a first approximation we pick to equal 6 months for both upward and downward departures.