Administrative law judges' (ALJs') relative lack of formal independence has engendered worries that they give agencies a home-court advantage. We examine the 2010 Dodd-Frank Act, which allowed the Securities and Exchange Commission (SEC) to move cases into its administrative court. The problem with this policy experiment is that the SEC retains the discretion to bring cases in federal court, so it is impossible to identify which cases the policy treats. We propose a difference-in-differences design, using natural-language-processing methods to create control and treatment groups. We construct propensity scores using random-forest methods. After binning cases into likely or not likely to be affected by the courts' expansion, the difference-in-differences estimation indicates that the expansion made defendants 30 percentage points more likely to settle and 36 percentage points more likely to receive a nonmonetary penalty. There is a 24-percentage-point reduction in the likelihood of a monetary penalty.