A major challenge for analysis of data from observational and survey studies is dealing with model mis-specification. A common reason for model mis-specification is the violation of the independence assumption. Model mis-specification is frequently due to the inclusion of variables that are correlated with the error terms (serial correlation) or due to variables omitted from the study. The application of standard regression models to such data could lead to over inflated results, i.e. erroneous results, and misleading conclusions. Longitudinally designed studies make substantial improvements and provide an additional handle to control omitted variables. However, even with longitudinal data, model mis-specification could occur because of the nature of observations, i.e. surveys often include objectively as well as subjectively measured variables. Subjective variables are responsible for model mis-specification, therefore, compounding the problem further. One solution to such problems is the application of instrumental variables. The instrumental variable method is seldom used with social survey data. The main criticism is the arbitrary selection of variables as instruments. Longitudinal data, because of its temporal structure, provide natural instruments. In this paper, a pragmatic strategy for analysis is proposed that utilises the nature of the data (subjective/objective) and a combination of methods within a longitudinal modelling framework to correct for model mis-specification. These applications are illustrated by using recurrent continuous morale in old age from a longitudinal survey of the elderly. The results suggest a strong presence of heterogeneity effect, i.e. current levels of morale appear to be individual-specific and independent of its previous levels