Fitting Probabilistic Index Models on Large Datasets
share
Recently, Thas et al. (2012) introduced a new statistical model for the probability index. This index is defined as P(Y ≤ Y^*|X, X^*) where Y and Y* are independent random response variables associated with covariates X and X* [...] Crucially to estimate the parameters of the model, a set of pseudo-observations is constructed. For a sample size n, a total of n(n-1)/2 pairwise comparisons between observations is considered. Consequently for large sample sizes, it becomes computationally infeasible or even impossible to fit the model as the set of pseudo-observations increases nearly quadratically. In this dissertation, we provide two solutions to fit a probabilistic index model. The first algorithm consists of splitting the entire data set into unique partitions. On each of these, we fit the model and then aggregate the estimates. A second algorithm is a subsampling scheme in which we select K << n observations without replacement and after B iterations aggregate the estimates. In Monte Carlo simulations, we show how the partitioning algorithm outperforms the latter [...] We illustrate the partitioning algorithm and the interpretation of the probabilistic index model on a real data set (Przybylski and Weinstein, 2017) of n = 116,630 where we compare it against the ordinary least squares method. By modelling the probabilistic index, we give an intuitive and meaningful quantification of the effect of the time adolescents spend using digital devices such as smartphones on self-reported mental well-being. We show how moderate usage is associated with an increased probability of reporting a higher mental well-being compared to random adolescents who do not use a smartphone. On the other hand, adolescents who excessively use their smartphone are associated with a higher probability of reporting a lower mental well-being than randomly chosen peers who do not use a smartphone.[...]
READ FULL TEXT