Case-mix adjustment through inverse probability weighting to compare health indicators across territories or providers – an application in the agency for health protection of Milan

• Published in: BMC health services research Vol. 25; no. 1; pp. 1106 - 12
• Main Authors: Mazzali, Cristina, Andreano, Anita, Magnoni, Pietro, Salvatori, Andrea, Testa, Deborah, Milanese, Alberto, Zanfino, Adele, Russo, Antonio Giampiero
• Format: Journal Article
• Language: English
• Published: England BioMed Central Ltd 19.08.2025; BioMed Central; BMC
• Subjects: Aged; Diagnosis-Related Groups; Electronic health records; Epidemiology; Female; Health Status Indicators; Humans; Inverse probability weighting; Italy; Logistic Models; Male; Medical care; Medical records; Medical research; Medicine, Experimental; Middle Aged; Quality Indicators, Health Care; Quality management; Risk adjustment; Risk Adjustment - methods; Statistical models; Italy; Inverse probability weighting; Quality indicators health care; Risk adjustment; Electronic health records
• ISSN: 1472-6963; 1472-6963
• DOI: 10.1186/s12913-025-13203-9

The Epidemiological Unit of the Agency for Health Protection of Milan (ATS) calculates several indicators, at hospital and at patients' residence area level (group level of analysis), with monitoring and programming scopes. Outcome indicators are usually influenced by differences in subject case-mix; therefore, adjustment methods are applied to compare each group level with mean ATS values. Inverse probability weighting (IPW) is explored as an alternative to multivariable generalized linear model (GLM), to overcome some limitations of the latter method. To implement IPW, a multinomial logistic model, with group level as dependent and subject characteristics as independent variables, was used to estimate patient weights, which were subsequently stabilized and truncated. Checks on IPW assumptions and covariates balance were implemented both in a quantitative and in a graphical way. Comparisons on adjustment performed with fixed effects and random intercept multivariable GLM were performed and exemplified using three outcome indicators. IPW assumptions were satisfied for all the indicators, and covariate balance presented minor issues for group levels with the lowest number of cases/events. Case-mix adjustment performed with multivariable fixed effects GLM showed a tendency to overestimate raw values and was characterized by broad confidence intervals in the three case-example indicators. Adjusted values produced using random intercept multivariable GLM suffered from a shrinkage towards the mean effect, particularly evident in the indicators with the lowest number of cases. IPW-adjusted estimates differed from raw values only when substantial differences in covariates distribution were present. Provided that appropriate checks are implemented, IPW adjustment is applicable in the context of healthcare quality evaluation and can be easily conveyed to healthcare managers for effective dissemination.