Retrospective Network Imputation from Life History Data: The Impact of Designs

Abstract

Retrospective life history designs are among the few practical approaches for collecting longitudinal network information from large populations, particularly in the context of relationships like sexual partnerships that cannot be measured via digital traces or documentary evidence. While all such designs afford the ability to “peer into the past” vis-à-vis the point of data collection, little is known about the impact of the specific design parameters on the time horizon over which such information is useful. In this article, we investigate the effect of two different survey designs on retrospective network imputation: (1) intervalN, where subjects are asked to provide information on all partners within the past $\mathrm{N}$ time units; and (2) lastK, where subjects are asked to provide information about their $\mathrm{K}$ most recent partners. We simulate a “ground truth” sexual partnership network using a published model of Krivitsky (2012), and we then sample this data using the two retrospective designs under various choices of $\mathrm{N}$ and $\mathrm{K}$ . We examine the accumulation of missingness as a function of time prior to interview, and we investigate the impact of this missingness on model-based imputation of the state of the network at prior time points via conditional ERGM prediction. We quantitatively show that—even setting aside problems of alter identification and informant accuracy—choice of survey design and parameters used can drastically change the amount of missingness in the dataset. These differences in missingness have a large impact on the quality of retrospective parameter estimation and network imputation, including important effects on properties related to disease transmission.