Pharmacoepidemiology
Frankenstein Surveys in Pharmacoepidemiology: Fusing Multiple Surveys Together to Produce Generalizable Estimates from Rare Populations Karilynn Rockhill* Karilynn Rockhill Joshua C. Black Elizabeth Bemis
Surveys are a critical tool in pharmacoepidemiology for population assessment of use patterns and drug effects. Behavioral surveys often utilize online recruitment and are plagued by selection bias and may not be generalizable to other populations of interest. We propose a new model for connecting multiple surveys to optimize data availability, reduce error, enhance inference, and reduce cost. This is accomplished through intentional design elements and correcting for differential selection forces across samples, which can be expressed in a DAG (Figure). We demonstrate this approach using a behavioral surveys designed to measure psychedelic drug use; a rare behavior (<5%) in the general population making a tool to detect differences difficult.
The Approach – This approach requires intentional study design including a large, population-representative survey (‘anchor’, Z’) and then connects targeted, smaller surveys to it (Z’’). When fused, interpretable estimates are created (X and Y) that are otherwise not possible through one survey alone. The anchor collects data to generate estimates to its target population and measures all hypothesized selection forces for itself (S’) and the target survey (S’’). The targeted survey measures more in-depth content for a given topic; it must measure all selection forces and be powered to detect differences in its subpopulation. These two surveys can be fused together, using statistical transport weights, to make generalizable estimates on the focused topic to the to the target population.
The Data – Data from the targeted survey among adults using ≥1 psychedelic drug in the last year (n=2,306) and our nationally-representative drug use survey (n=28,679) will be shown. We explored 2,048 bias-correction models. For example, frequent mental distress was estimated at 41.9% in the sample alone, and 47.0% in the fused data, a 5.1% change. The average residual bias was reduced to 2.7% across demographics, 10.5% across health metrics, and 4.8% across other substance use. Comparison to independent external population estimates will be provided.
