Estimation of Diffusion Processes from Incomplete Data

Event-history analysis of the diffusion of practices in a social system can show how actors are influenced by each other as well as by their own characteristics. The presumption that complete data on the entire population are essential to draw valid inferences about diffusion processes has been a major limitation in empirical analyses and has precluded diffusion studies in large populations. The authors examine the impacts of several forms of incomplete data on estimation of the heterogeneous diffusion model proposed by Strang and Tuma. Left censoring causes bias, but right censoring leads to no noteworthy problems. Extensive time aggregation biases estimates of intrinsic propensities but not cross-case influences. Importantly, random sampling can yield good results on diffusion processes if there are supplementary data on influential cases outside the sample. The capability of obtaining good estimates from sampled diffusion histories should help to advance research on diffusion.