Measurement error methods for wearable device data, treating daily measurements as imprecise draws from long-term average
We propose using formal measurement error methods to analyze wearable device data. The data structure lends itself to maximum likelihood estimation, treating daily measurements as (perhaps correlated) replicates from each subject’s long-term average. For example, one might assume that the device-recorded sedentary time on a particular day is not exactly equal to the true sedentary time, but rather the true sedentary time plus a mean-0 error. Further, the true sedentary time on a particular day might be assumed to reflect the long-term average plus a mean-0 random error. Alternatively, considering that activity variables are usually right-skewed and almost always non-negative, it might be more reasonable to assume multiplicative mean-1 errors, perhaps distributed lognormal or Gamma. Within a maximum likelihood framework, metrics like AIC or BIC can be helpful for specifying the error structure and other parametric assumptions.
We plan to focus primarily on cross-sectional and cohort study designs, where the goal is to estimate a covariate-adjusted association relating an activity variable to a health outcome (e.g. BMI, waist circumference, cardiometabolic variables). We will compare point and interval estimates for the naïve approach to those obtained by fitting various maximum likelihood estimators, and compare metrics including AIC and BIC. We plan to run simulation studies mimicking IDATA to evaluate the consequences of ignoring measurement error and to evaluate performance of the corrective methods (e.g. validity, numerical stability, robustness to model misspecification). We may also consider measurement error methods other than maximum likelihood.
Vadim Zipunnikov, PhD (Johns Hopkins Bloomberg School of Public Health)