Overview of methodological issues in the study of chronic care populations.

The intent of the methodology section of this volume was to provide an overview of recent thinking about analytic techniques that can be used to study chronic care populations, particularly dementia SCUs. As discussed earlier in this article, although not unique to these populations, longitudinal analysis of chronic care populations frequently involves several problems: nonequivalent comparison groups, unbalanced designs, censoring and attrition, autocorrelation (correlated repeated measures) and heterogeneous correlations across repeated measures. The last 5 years has witnessed considerable change and growth in the development of longitudinal modelling methodologies. Although developments in modelling cut across disciplines, there has been a focus on certain types of methods in several fields of applied statistics: psychometrics (e.g., analysis of change, item response theory, Rasch modelling of binary outcomes), mathematical sociology (e.g., structural equation modelling), demography (e.g., transition modelling), epidemiology (e.g., risk modelling), and biostatistics (e.g., examination of intervention effects using marginal and random effects modelling of repeated measures). An important issue is the interrelationship between measurement and outcome assessment. Source of measurement error must be considered (see Zimmerman and Magaziner in this volume), as must bias in assessments (see Teresi and Golden in this volume). While it has long been known that poor measurement can attenuate and bias estimates of longitudinal effects, the answer is not to attempt artificial solutions through use of corrections for attenuation (unreliability). Design and sampling issues are also critical to the correct conduct of any longitudinal study. Otherwise, estimates of prevalence, incidence, and intervention effects will be in error (see Beckett and Evans, in this volume). Several modelling approaches have been reviewed. Different points of view regarding change analysis have been presented; controversy remains regarding the "best" methods for examining change (see the commentary by Rogosa in this volume). Theory building using methods such as structural equation modelling remains a staple of longitudinal analysis, although several caveats have been discussed. New or revisited methodologies (random effects modelling, of which growth curve analysis can be viewed as a special case and event history modelling) offer promise in dealing with the knotty problems inherent in the longitudinal study of chronic care populations, namely, censoring, attrition, and unbalanced designs resulting in missing data (see the commentary by Nesselroade, in this issue). However, it remains clear that no one model or approach will be optimal for all applications.