Abstract
Background/Aims Comparative effectiveness research studies often require researchers to make inference about the relationship between exposure to various treatments over time and health outcomes. In these studies it can be a challenge to isolate the effect of the exposure from other concurrent treatments and health conditions. For example, in the study of the effect of different antipsychotic medications on body weight, concurrent exposures of oral steroids, insulin, pregnancy, and cancer diagnosis can have important confounding effects. The temporal relationships between these exposures may be difficult to assess in large samples; therefore we sought a way to graphically display exposures and outcomes in a single figure to facilitate rapid screening of potential cases.
Methods Researchers at Geisinger and Group Health identified all adults who a) received one or more prescription for a 2nd/3rd generation antipsychotic medication from 2004–2009 and b) had two or more weight measures in close proximity to those treatments. For each adult who had gained >15% body weight after initiating drug treatment we used SAS to produce a longitudinal plot encompassing all weight measurements, antipsychotic fills, pregnancy- and cancer-related visits, and steroid and insulin medications fills. Graphs were independently reviewed by investigators at each site and categorized as
-
confirmed case,
-
possible case,
-
pre-treatment weight loss, and
-
excluded, based on pre-specified criteria.
Where disagreement occurred the graphs were re-reviewed concurrently by both research teams to reach a consensus.
Results We identified and graphed 922 cases at both sites who met eligibility criteria and had gained >15% weight on drug treatment. Each graph simultaneously depicted 14 possible measurement categories including weight, 9 antipsychotic drug types, insulin fills, oral steroid fills, cancer- and pregnancy-related visits. After each team examined all graphs subjects were categorized as 129 (14.0%) confirmed, 155 (16.8%) possible, 83 pre-treatment weight loss (9.0%), and 555 (60.2%) excluded cases.
Discussion Using a novel graphical display, we rapidly reviewed temporal exposure information on large samples to facilitate case identification. As the temporal exposure information in these subjects was very complex, a standard SAS-programmed exclusion algorithm without graphical review would have likely induced selection bias.
