Even where feasible, randomized clinical trials cannot be conducted quickly enough to answer urgent questions about treatments in the context of COVID-19. For example, there is a need to optimize treatment strategies to reduce the excess burden of COVID-19 mortality among African American and Hispanic individuals who were poorly represented in many clinical trials of COVID-19 therapies such as dexamethasone. We illustrate recently developed propensity score weighting methods to evaluate heterogeneous treatment effects of dexamethasone vs. standard of care in adult patients hospitalized for COVID-19. We investigate potential heterogeneity by the level of oxygen support, age, race/ethnicity, diabetes and prior treatment with remdesivir. We show that subgroup-specific covariate balance is achieved through a combination of machine learning methods for propensity score estimation, an augmented design matrix for the propensity score that includes subgroup-covariate interactions, and propensity score overlap weighting. We discuss how the use of novel methods enhances validity and plausibility of the observed results.