Abstract:
In this paper we give a detailed explanation of how to estimate a multinomial logistic regression model using data from a longitudinal survey with multiple cohorts. We also show how to estimate the variance of the parameter estimates by using the estimating function bootstrap or any other design variance estimate available in the literature. We argue why it is more appropriate to estimate the autocorrelation matrix by quasi-least squares rather than by the method of moments or the odds ratios parameterization, and we show how to do so. We illustrate the technique by estimating a model for employment sector from the U. S. National Science Foundation’s Survey of Doctorate Recipients, and interpret the results. Additionally we present a simulated score test for assessing goodness of fit in general, and conclude that the estimated model for employment sector fits the data well.
Key Words:
Marginal model parameters; Rotating panel surveys; Replication variance estimation; Weighted Generalized Estimating Equations; Goodness of fit; Hosmer-Lemeshow test
