Fusing Specialized Surveys of Rare Populations to Larger Surveys for Generalized Inference: Cross-Sectional Survey Study

Background: Mainstays of pharmacoepidemiology are large, representative, behavioral surveys, which focus on many drugs with few detailed behaviors. Smaller, targeted studies measure drug-specific patterns but without explicit generalizability assumptions; the evidence generated is narrow. Objective: In this cross-sectional survey study, we outline an estimation framework based on data fusion and combine two surveys: a representative, anchor survey and an enriched survey about psychedelic drugs in the United States. Application of calibration weighting transports estimates from the enriched survey to the anchor survey. Methods: The psychedelic-focused enriched survey was sampled twice from a commercial online panel of adults from April 19 to June 25, 2024, (n=2306; 40.4% female, 33.1 y median age) and January 24 to March 21, 2025 (n=2023; 39.6% female, 35.2 y median age). The anchor survey was sampled twice from a different online panel from March 13 to May 6 2024 (n=28,679 total; 2430 using psychedelics) and 14 August to 9 October 2024 (n=29,040 total; 2309 using psychedelics). Internal consistency (transport bias, the absolute difference between the weighted estimates from the anchor survey and weighted fused survey) and external validity (root-mean-square error, RMSE, of self-reported demographic, health, and substance use estimates to probability-based benchmarks) metrics were calculated. The methodology was applied to estimate reasons for using specific psychedelic drugs. Results: Without adjustments, the enriched surveys had lower percentages of male and White respondents, lower self-perceived health, and higher cigarette use. A total of 2048 weighting schemes were tested, with good internal consistency. Average transport biases with the final weighting scheme were: demographics, 0.09 percentage points; health characteristics, 0.35 percentage points; and substance use, 0.22 percentage points. Estimates after fusion were externally consistent with benchmarks. RMSE increased by 3.3% for demographic estimates (1.82 unweighted to 1.88 weighted); larger decreases were observed for health RMSE (7.30 to 3.38, 53.7% decrease) and for substance use RMSE (6.56 to 6.03, 8.1% decrease). Alcohol use substantially increased the RMSE, likely due to question differences (without alcohol, the RMSE decreased from 6.03 to 1.55). Using the fused dataset, recreational use of psilocybin (92.9%, 95% CI 91.1, 94.7), LSD (93.2%, CI 90.1, 96.4, and MDMA (93.3%, 91.0, 95.6) was more common than medical use (30.9%, CI 27.6, 34.2; 26.4%, CI 21.1, 31.7; and 21.1%, CI 17.5, 24.7, respectively). Conclusions: Building upon past data fusion research, this study fused two surveys for the purpose of surveillance. This methodology, termed the “fused survey design,” is a rigorous but accessible approach for surveilling rare behaviors like drug use, and we demonstrated constructs absent from anchor surveys may be measured with generalizable inference. This expands the surveillance epidemiology toolbox, giving researchers an actionable process to field enriched surveys with specialized questions that would be impractical to add to larger surveys due to space constraints and respondent fatigue.