Count data models for outpatient health services utilisation

Area of Science:

• Health Services Research
• Biostatistics
• Econometrics

Background:

• National survey data often exhibits excess zeros and skewed distributions, common in healthcare utilization counts.
• Traditional count data models may not adequately capture these characteristics, leading to potential biases.
• Understanding healthcare utilization patterns is vital for public health policy and resource allocation.

Purpose of the Study:

• To identify the most appropriate statistical model for analyzing outpatient healthcare utilization in Malaysia.
• To compare the performance of various count data models, including zero-inflated and hurdle models.
• To determine factors influencing adult outpatient healthcare utilization using Malaysian National Health and Morbidity Survey 2019 (NHMS 2019) data.

Main Methods:

• Employed six count data models: OLS, Poisson, Negative Binomial (NB), Zero-Inflated Poisson (ZIP), Marginalized-Zero-Inflated Negative Binomial (MZINB), and Hurdle.
• Utilized instrumental variable selection based on Andersen's model.
• Model selection was based on criteria like log-likelihood, AIC, BIC, goodness-of-fit, RMSE, R2, and Vuong tests.

Main Results:

• The Zero-Inflated Model (ZIM) demonstrated the best fit, evidenced by the smallest information criteria values (AIC, BIC) and log-likelihood.
• Significant factors associated with outpatient visits included state of residence, ethnicity, household income quintile, and health needs.
• Data exhibited overdispersion, characterized by excess zeros (90%) and skewed positive tails, confirming the suitability of ZIM.

Conclusions:

• The Zero-Inflated Model (ZIM) is recommended over Ordinary Least Squares (OLS) for modeling outpatient healthcare utilization in Malaysia.
• ZIM offers better interpretability and appropriate assumptions for survey data with excess zeros.
• Accurate modeling of healthcare utilization is essential for informing health policy and improving healthcare access.