Predicted probabilities' relationship to inclusion probabilities.

Area of Science:

• Biostatistics
• Health Services Research
• Epidemiology

Background:

• Case-control studies are commonly used in epidemiology and health services research.
• Analyzing retrospective data for predictive probabilities requires careful consideration of sampling methods.
• Medicare's penalties for hospital readmissions highlight the need for accurate predictive models.

Purpose of the Study:

• To demonstrate that retrospective case-control data can be analyzed using maximum likelihood under a general multiplicative intercept model.
• To show how auxiliary information can be used with retrospective data to estimate response probabilities.
• To highlight the differences in predictive probabilities obtained from retrospective versus prospective data, using hospital readmission as a case study.

Main Methods:

• Application of maximum likelihood estimation to case-control data.
• Utilizing a general multiplicative intercept model for risk analysis.
• Analysis of binary Medicare data to predict 30-day hospital readmission probabilities.

Main Results:

• Retrospective data analysis can approximate prospective analysis under specific models, up to a sampling fraction constant.
• Auxiliary information enables the estimation of response probabilities from retrospective data.
• Unadjusted predictive probabilities from retrospective data differ from prospective data, as shown in the Medicare readmission example.

Conclusions:

• Maximum likelihood analysis of retrospective data can yield valid risk estimates.
• Adjustments are necessary when using retrospective data to obtain accurate predictive probabilities.
• The findings have implications for healthcare policy and hospital performance measurement, particularly concerning readmission penalties.