Abstract:
|
While generalized estimating equations are a favorite choice for modeling longitudinal and clustered data, asymptotic properties of the popular sandwich estimator rely on small cluster sizes relative to sample size. In many modern contexts, such as digital phenotyping studies, there may be high frequency data collected on individuals over long follow-up periods leading to large cluster sizes relative to sample size. As a result, within-cluster resampling can be used to circumvent the large cluster sizes that would generally make generalized estimating equations usuable. However, within-cluster resampling generally ignores any informative correlation and relies on problematic moment-based estimators of variance. We propose an adaptation of within-cluster resampling that accounts for informative correlation and derive a corresponding analytic non-negative variance estimator. We demonstrate our methods performance through simulation and in modeling anxiety in a digital phenotyping study of schizophrenia.
|