Latent Class Analysis of Complex Sample Survey Data

Abstract

High fruit and vegetable intake is associated with decreased cancer risk. However, dietary recall data from national surveys suggest that, on any given day, intake falls below the recommended minima of three daily servings of vegetables and two daily servings of fruit. There is no single widely accepted measure of “usual” intake. One approach is to regard the distribution of intake as a mixture of “regular” (relatively frequent) and “nonregular” (relatively infrequent) consumers, using an indicator of whether an individual consumed the food of interest on the recall day. We use a new approach to summarizing dietary data, latent class analysis (LCA), to estimate “usual” intake of vegetables. The data consist of four 24-hour dietary recalls from the 1985 Continuing Survey of Intakes by Individuals collected from 1,028 women. Traditional LCA based on simple random sampling was extended to complex survey data by introducing sample weights into the latent class estimation algorithm and by accounting for the complex sample design through the use of jackknife standard errors. A two-class model showed that 18%% do not regularly consume vegetables, compared to an unweighted estimate of 33%%. Simulations showed that ignoring sample weights resulted in biased parameter estimates and that jackknife variances were slightly conservative but provided satisfactory confidence interval coverage. Using a survey-wide estimate of the design effect for variance estimation is not accurate for LCA. The methods proposed in this article are readily implemented for the analysis of complex sample survey data.

KEY WORDS:

• Categorical data
• Cluster sample
• Design effect
• Dietary propensity scores
• Jackknife
• Latent class model
• Sample weight