Causal Inference in Unrepresentative Databases: An Analysis of Vitamin E Supplementation in the National Alzheimer’s Coordinating Center Uniform Dataset

Abstract

AbstractBackgroundSelf‐selection into research studies can lead to samples that are unrepresentative of the target population. For example, the National Alzheimer’s Coordinating Center (NACC) Uniform Data Set is often used to identify risk factors of Alzheimer’s disease and overrepresents highly educated and non‐Hispanic White participants, relative to the US population. Sampling weights that are related to the self‐selection of enrolling in the NACC database can be used to estimate causal effects in the US population of older adults. We applied this strategy to estimate the effect of vitamin E supplementation on functional activities using NACC data and estimated sampling weights as proof of concept. A meta‐analysis found mixed evidence about the effect of vitamin E on functional outcomes in cognitively normal older adults (Browne et al. 2019).MethodsWe used baseline data for NACC participants who enrolled from 2005‐2021, were cognitively normal, and 65 or older. The NACC database is funded by NIA/NIH Grant U24 AG072122. NACC data are contributed by the NIA‐funded ADCs. We estimated the probability of self‐selection into NACC relative to a nationally representative survey. We used the corresponding sampling weights to estimate weighted propensity scores and weighted propensity‐adjusted estimates of the effect of vitamin E supplementation on the Functional Activities Questionnaire (FAQ) and compared the results to unweighted models.ResultsAfter excluding 809 participants missing the FAQ, there were 14,358 participants. When incorporating sampling weights to obtain inference about the US population, we estimated that participants who take vitamin E supplements have a 0.29 point lower (95% CI: ‐0.54, ‐0.04) FAQ score, on average, compared to participants who do not. When failing to account for sampling bias and omitting sampling weights, we estimated that they have a 0.01 point lower (95% CI: ‐0.11, 0.09) FAQ score, on average (see figure).ConclusionWhen accounting for sampling bias we estimated that participants who take vitamin E supplements perform better on the FAQ compared to those who do not, but this effect was attenuated when failing to account for sampling weights. Incorporating estimated sampling weights allows for better isolation of causal effects in the population of interest.