Parametric modelling of cost data in medical studies

Area of Science:

• Health economics
• Biostatistics
• Statistical modeling

Background:

• Medical resource costs are crucial for clinical decision-making.
• Cost data are typically right-skewed, posing challenges for standard statistical inference.
• Common methods for non-normal data are often suboptimal for analyzing skewed cost data.

Purpose of the Study:

• To explore parametric models for analyzing skewed medical cost data.
• To compare the performance of different distributions (normal, gamma, log-normal, log-logistic) in modeling cost data.
• To assess the impact of model choice on population mean cost inferences.

Main Methods:

• Fitting normal, gamma, log-normal, and log-logistic distributions (including three-parameter versions) to four example cost datasets.
• Estimating population mean costs using maximum likelihood.
• Deriving confidence intervals via parametric bootstrap and Markov Chain Monte Carlo (MCMC) methods.

Main Results:

• Skewed parametric distributions generally provide a better fit to cost data than the normal distribution.
• Poorly fitting models can sometimes yield similar inferences to well-fitting ones.
• Different parametric models that fit data equally well can lead to substantially different inferences, particularly with smaller sample sizes.

Conclusions:

• Inferences about population mean cost are sensitive to the chosen statistical model.
• Model uncertainty is significant unless sufficient data are available to accurately model the distribution's tail.
• Sensitivity analyses exploring the impact of model choice are essential when analyzing medical cost data.