Abstract P330: Excluding Participants With Missing or Incomplete Accelerometry Data: Evaluating the Potential for Bias

Abstract

Background: Investigating the effect of physical activity (PA) on health has been improved with accelerometry. However, accelerometer measurement may increase participant burden, leading some to not wear devices. Previous research suggests that participants who do not wear accelerometers may be less active and have worse health outcomes than those who adhere to measurement protocol. This selection bias may lead to underestimates of the true association between PA and CVD outcomes. Using simulations, we illustrate how the association of meeting WHO PA guidelines and hypertension (HT) might be impacted by missing accelerometry data. Methods: Based on observed prevalence of PA and HT in an ongoing cohort study, we simulated data with 2000 participants, with 40% not meeting PA guidelines and 10% having HT. We set the “true” association between not meeting PA guidelines and HT at 1.9 (OR=1.9; 95% CI=1.5, 2.5). Then, we simulated scenarios to show the impact of four different patterns of missing accelerometer data: A) Missingness independent of PA and HT; B) Missingness associated with PA only; C) Missingness associated with HT only; D) Missingness associated with PA and HT. For each scenario, we conducted complete cases analyses using simple logistic regression to assess associations between meeting PA guidelines and HT. Results: Odds ratios (ORs) in Scenarios A, B, and C were unbiased. ORs in Scenario D underestimated the magnitude of the “true” association and were increasingly attenuated as the strength of the relationship between study variables and missingness was increased. Conclusions: Accelerometer measurement may result in missing data, which can bias odds ratio estimates if the missingness is associated with both the exposure and outcome, a scenario reported in previous research. Efforts to minimize missing data during study design should be prioritized. Carefully considering the structure of missingness in analyses will inform quantification of potential bias, analytic strategies, and interpretation of findings.