One way to improve two part modeling of longitudinal cost data - using latent class clustering of cost casemix groups
- Presenter:
Mon June 5, 2006 9:30-10:45 Room Alumni Lounge
Rationale: Though two part model has prevailed for the last two decades as the methodology to analyzing cost data in health economics, the advantages of using longitudinal (panel) data received little attention.
Objectives: The objective of this paper is to highlight the advantage of using a clustering technique in longitudinal two part model.
Methods: A longitudinal medical cost data of 3,260 patients for two years was first clustered by latent class cluster analysis by cost characteristics - categorical indicator variables with 5 categories of the each year’s cost, dichotomous indicator variables with 0 for zero cost and one for positive cost for each year. Latent class cluster analysis identifies cost casemix groups using the correlation structure of these cost indicator variables and these indicator variables are mutually independent in each identified casemix group. To determine the optimal number of casemix groups, Sample Size Adjusted Bayesian Information Criteria and Consistent Akaike Information Criteria was compared. For each group of identified cost casemix group, panel data two part model (random effect and fixed effect) for cost estimation was applied and then compared with the pooled data result, which only models heterogeneity in the intercept.
Results: The two results with casemix control and without casemix control were significantly different. This difference is mainly resulted from the significant differences in characteristics among the casemix groups. The higher level of heterogeneity control in latent class clustering can significantly improve two part model in longitudinal analysis.
Conclusions: When two part model is applied in longitudinal data, there is an advantage of heterogeneity control whether assuming individual heterogeneity intercept as random effect or fixed effect. This advantage can be further extended by using a clustering technique such as latent class clustering specifically on longitudinal cost variable.
