Navigating missing data in multilevel autoregressive modeling: Consequences of missingness mechanisms and temporal patterns of missings on estimation performance

This is the authors' abstract. We don't add key points for this paper.

The multilevel autoregressive (MLAR) model is a popular method for quantifying affective inertia, the resistance of affective processes to change, using experience sampling method (ESM) data. Although missed observations are common in ESM, their effect on the estimation performance of the MLAR model remains underexplored. We investigate the effect of compliance level, specific missingness mechanisms (i.e., missing completely at random versus tail-based missingness), and temporal patterns (whether or not missings occur consecutively) on the bias and variance of the fixed autoregressive effect estimator in the MLAR model. In a simulation study, we show that the presence of missing data increases both bias and variance. The bias is more severe when compliance is low and when the missingness is tail-based (i.e., when the missingness depends on the process value itself). The conclusions hold regardless of whether the random autoregressive effects are estimated or not. We corroborate these results by setting data of an empirical ESM dataset to missing following the same principles. To lower the estimation bias and variance, we recommend making design choices to promote compliance and/or increase the probability of data being missing completely at random: decreasing participant burden by using shorter questionnaires, providing adequate incentives to participants, and considering using planned missingness designs or episode-based sampling.