Using data from the United States Sentencing Commission, we examine how judicial biases may have influenced sentences during the era of the Federal criminal sentencing guidelines. Our utility maximization model of judicial sentencing preferences leads to a partially censored ordered probit model that accounts for mass points in the sentencing distribution that occur at the upper and lower guideline limits and at sentences involving no prison time. Our results indicate that racial- and gender-based discrepancies exist, even after controlling for circumstances such as the severity of the offense and past criminal history.