The thesis has two parts; one clinical part: studying the dimensions of human immunodeficiency virus associated lipodystrophy syndrome (HALS) by latent class models, and a more statistical part: investigating how to predict scores of latent variables so these can be used in subsequent regression analyses.

Part 1: HALS engages different phenotypic changes of peripheral lipoatrophy and central lipohypertrophy.  There are several different definitions of HALS and no consensus on the number of phenotypes. Many of the definitions consist of counting fulfilled criteria on markers and do not include patient's characteristics. These methods may erroneously reduce multiplicity either by combining markers of different phenotypes or by mixing HALS with other processes such as aging.

Latent class models identify homogenous groups of patients based on sets of variables, for example symptoms. As no gold standard exists for diagnosing HALS the normally applied diagnostic models cannot be used. Latent class models, which have never before been used to diagnose HALS, make it possible, under certain assumptions, to: statistically evaluate the number of phenotypes, test for mixing of HALS with other processes (differential item functioning), estimate
the sensitivity and specificity of the markers, calculate predictive values for the patients (posterior probabilities), and estimate the effect of risk factors on HALS.

By use of latent class models I found evidence for only two different latent classes of HALS: one displaying isolated peripheral lipoatrophy and another displaying both peripheral lipoatrophy and central lipohypertrophy. When patient characteristics were included the results indicated that smoking status is an essential variable in explaining why some patients do not get the central lipohypertrophy part of the syndrome. Thus, the results suggested that peripheral lipoatrophy and central lipohypertophy are interrelated phenotypes rather than two independent phenotypes.

Part 2: Latent class regression relates explanatory variables to latent classes. In this model no measure of the latent class variable is obtained, although this is often desired. I have proposed a new method for predicting class membership that, in contrast to methods based on posterior probabilities of class membership, yields consistent estimates when regressed on explanatory variables in a subsequent analysis.

There are four different basic models within latent variable models: factor analysis, latent class analysis, latent profile analysis and latent trait analysis. I have given a general overview of how to predict scores of latent variables so these can be used in subsequent regression models. Two different principles of predicting scores are shown to be superior depending on whether the latent variable is a dependent or an independent variable. Both these types of scores are extended to the situation of differential item functioning. Analytically I have showed that the scores result in consistent estimates when used properly in subsequent analysis. Further, through simulation studies the finite sample properties have been investigated by comparisons with maximum likelihood one-step estimation. The results suggested that the three-step regression yields estimates that are approximately unbiased, and only has a minor loss of efficiency compared to estimates from the standard one-step (full maximum likelihood) estimation procedure.