LATEBREAKER
Causal Inference
Are you positive? Outcome-dependent positivity violations in small populations: an example study of substance use by gender identity Dominique Heinke* Dominique Heinke Ryan P Kyle Joseph AC Delaney Bridget Whitney Sarah Mixon Robert J Fredericksen Michelle Kipke Catherine Lesko Kenneth H Mayer Michael J Li Brian Mustanski Luis Parra Geetanjali Chander Heidi Crane Lydia N Drumright
Accurately describing the health burdens of small populations is important yet challenging; many adjustment methods fail with sparse data and knowledge of best practices are limited. We evaluated the performance of two adjustment methods appropriate to small and complex samples when estimating substance use prevalence among transgender women relative to cisgender men and women.
We estimated prevalence ratios (PRs) for self-reported use of specific substances comparing trans women to cis men and trans women to cis women in an HIV collaborative with data from multi-site HIV clinical cohorts (CNICS, JHHCC) and a community cohort (RADAR). To adjust for age, race/ethnicity, study site, and year of cohort entry and of most recent self-report, we calculated PRs using 1) stabilized IPW: weighting gender groups to be similar on these factors; and 2) matching: selecting ≤2 cis men or women that was similar to each trans woman on these factors.
In the combined CNICS/JHHCC cohorts, 275 of 277 trans women matched to ≥1 one cis man and 241 matched to ≥1 cis women. In the RADAR cohort, 231 of 232 trans women matched to 1 cis man. Among more commonly used substances (nicotine, cannabis, crack/cocaine) IPW and matching provided consistent estimates in all analyses (Figure). While for less commonly used substances, both methods yielded consistent estimates for trans women vs cis men, in some instances positivity violations in the IPW analyses for trans vs cis women induced substantially divergent estimates. For example, IPW and matched estimates for methamphetamine use in trans women vs cis men were consistent in all cohorts (Figure), but there was a >2-fold difference between IPW (PR=4.48, 95%CI: 1.52-10.03) and matched (PR=1.58, 95%CI: 1.04-2.40) estimates for trans vs cis women.
Both matching and IPW can be effective adjustment approaches in small and complex samples, but IPW requires checking each outcome for positivity violations, while matching may result in loss of data.
