Project Summary
Availability of large longitudinal datasets describing elderly populations with schizophrenia treated in usual care
settings present opportunities to expand the limited evidence on outcomes of antipsychotic drug treatment for
this population and to learn what works in the real world: which drugs, in what sequence, combination, or
intensity, for whom (what racial/ethnic groups, in what social circumstances), and at what risk. While this
objective is not new, advances in machine learning and causal inference could improve inferences, and thus
generate evidence to answer these questions. Leveraging data generated in usual care settings, we will (a)
translate novel statistical methods to assure distributional balance on observed confounders using high-
dimensional longitudinal data with multiple competing antipsychotic drugs (multi-valued treatments) and
longitudinal treatment patterns (treatment regimens); (b) utilize robust non-parametric or semi-parametric
methods; and (c) extend tree-based approaches to simultaneously model effectiveness and safety outcomes to
fill evidence gaps. We will link racially/ethnically diverse cohorts of elderly publicly-insured adults with
schizophrenia utilizing antipsychotics to geographical indicators of social contextual factors– upstream social
determinants of health (SDH) such as household income and crime rates— that are known to influence
treatment adherence and other health behaviors. Aim 1 applies causal effect estimation of the index
antipsychotic drug prescribed using weighted semi-parametric or non-parametric methods that (a) depend on
high-dimensional confounders and (b) may be moderated by patient race/ethnicity and area-level SDH. Aim 2
identifies and characterizes frequently observed treatment regimens that may differ by race/ethnicity and SDH.
Aim 3 estimates effectiveness and safety of the treatment regimens identifed in Aim 2, and determines if
race/ethnicity or SDH modify treatment effectiveness. Aim 4 estimates the impact of treatment regimens on
each individual effectiveness and safety outcome simultaneously, making use of within-patient outcome
dependencies. Our proposal has high