The Lazy Bootstrap. A Fast Resampling Method for Evaluating Latent Class Model Fit

TL;DR

The paper introduces the lazy bootstrap, a fast resampling method for evaluating latent class model fit that is flexible, computationally efficient, and applicable to various models, demonstrated through simulations and empirical data.

Contribution

It proposes a novel, rapid resampling scheme for model fit assessment that reduces computational time significantly compared to traditional bootstrap methods.

Findings

Low type I error rates across tested statistics

Comparable results to parametric bootstrap in empirical data

Computational efficiency improved by three orders of magnitude

Abstract

The latent class model is a powerful unsupervised clustering algorithm for categorical data. Many statistics exist to test the fit of the latent class model. However, traditional methods to evaluate those fit statistics are not always useful. Asymptotic distributions are not always known, and empirical reference distributions can be very time consuming to obtain. In this paper we propose a fast resampling scheme with which any type of model fit can be assessed. We illustrate it here on the latent class model, but the methodology can be applied in any situation. The principle behind the lazy bootstrap method is to specify a statistic which captures the characteristics of the data that a model should capture correctly. If those characteristics in the observed data and in model-generated data are very different we can assume that the model could not have produced the observed data. With this method we achieve the flexibility of tests from the Bayesian framework, while only needing maximum likelihood estimates. We provide a step-wise algorithm with which the fit of a model can be assessed based on the characteristics we as researcher find important. In a Monte Carlo study we show that the method has very low type I errors, for all illustrated statistics. Power to reject a model depended largely on the type of statistic that was used and on sample size. We applied the method to an empirical data set on clinical subgroups with risk of Myocardial infarction and compared the results directly to the parametric bootstrap. The results of our method were highly similar to those obtained by the parametric bootstrap, while the required computations differed three orders of magnitude in favour of our method.