Hospital utilization in mixed public–private system: evidence from Australian hospital data

Abstract

This article investigates whether patients who used a mixture of private and public hospital care have higher total hospital utilization than those who exclusively used either public or private hospital care. Using Australian hospital administrative data of heart disease patients, we found that those who used a mixture of private and public care had the highest total hospital utilization. Our findings are robust to how utilization is measured and endogeneity between utilization and hospital type choice.

Keywords:

• hospital utilization
• health care
• public and private system
• Australian hospital

JEL Classification:

• I18
• I10

Notes

¹ Barros and Olivella (2005) and Biglaiser and Ma (2007) investigate the impacts of dual practising physician behaviour on cost and quality of care.

² Examples of studies investigating the determinants of the health care utilization using survey data include, from Australia (Savage and Wright, 2003); Germany (Riphahn et al., 2003); Ireland (Harmon and Nolan, 2001); and Spain (Vera-Hernàndez, 1999).

³ A similar reasoning can be made for the case of unobserved preferences toward private and public care.

⁴ A related phenomenon is adverse/advantageous selection into private health insurance. The evidence of such phenomenon is mixed for the case of Australia (Cameron et al., 1988; Savage and Wright, 2003; Buchmueller et al., 2013).

⁵ Note that since the first stage is a multinomial model with three categories, we have two outcome categories (and a reference category) and hence two sets of response residuals, defined as the difference between the observed value (0 or 1) and the predicted probability of the category based on the estimated multinomial model.

⁶ Different radius-based measures are used to assess the sensitivity of the estimates to the instruments.

⁷ In 2004–05, the state of Victoria had 285 hospitals, and in total provided 27 789 available hospital beds. There were 141 private hospitals; most were small in size; only about a quarter of the available beds were provided by private hospitals.

⁸ ICD-10-AM is the Australian Modified version of the ICD-10 diagnosis codes, a classification system of more than 14 000 different codes for diseases, signs and symptoms, abnormal findings, external causes of injury, etc. 

⁹ We conducted several sensitivity analyses by varying both the days of utilization and the admissible survival days. The main results are broadly similar under these variations. Results are available from the authors upon request.

¹⁰ Public and private account payment types are the two largest categories of payment types, constituting roughly 90% of all payments for hospital care in the data.

¹¹ SEIFA indexes consist of four different summary measures constructed from a number of variables that represent different aspects of relative socio-economic standing of residents in a geographic area. The four measures are the Index of Relative socio-economic disadvantage, index of relative socio-economic advantage and disadvantage, index of economic resources (IER) and index of education and occupation (see Australian Bureau of Statistics (2006), for details). Only the first index is used in our empirical model. The ARIA index is a measure of remoteness jointly developed by the National Centre for Social Applications of Geographic Information Systems (GISCA) and the Australian Department of Health and Ageing.

¹² The Charlson index at FH is the score of the index based on comorbidity information at the first hospital admission for heart disease-related treatment.

¹³ For example, the quality of treatments in the first hospital admission – which is correlated with unobserved hospital effects – may determine the needs for further treatment in the future.

¹⁴ All null hypotheses were rejected at significance level 0.001 for each estimated model.

¹⁵ Estimates for the first stage are summarized in Appendix 3, and the signs of the estimated coefficients are as expected.

¹⁶ We thank an anonymous referee for pointing this out.

¹⁷ Given space constraints, only the coefficient estimates and statistical significance are shown in Table 5. Bootstrapped SEs of the estimates are available from the authors.

¹⁸ Estimation results are summarized in top two panels of the table in Appendix 4; full results are available from the authors upon request.

¹⁹ Estimation results are summarized in bottom two panels of the table in Appendix 4; the full results are available from the authors upon request.

²⁰ To obtain these nonlinear GLM estimates, we specify a distribution for and a link function which ‘connects’ the dependent variable to the explanatory variables. We first search for the ‘right’ distribution using an assumed log-link function, starting from a gamma distribution. We select the correct distribution based on the result of the modified Park test (Manning and Mullahy, 2001). Once the correct distribution has been identified, we test the assumed log-link function using the Pregibon Link test and the modified Hosmer–Lemeshow test. If the log-link function is rejected, we replace it with a power link function and search for a power coefficient that passes the two tests. In the end, we obtain a different GLM specification for each utilization measure. For cost-weighted utilization and LOS, we adopt a negative binomial specification with a square-root link function for the former and an identity link function for the latter. For the SEP, we use the inverse Gaussian specification with a power link function with power coefficient of 2.6. Given that these are nonlinear models, we evaluate the marginal effect of patient type at each observation and average the effects across all observations in the sample. As shown in the above table, the marginal effect estimates are consistent with those of the base model, except in the case of LOS where fitting a negative binomial model with the identity link function produces marginal effects that are negative and statistically significant for All-private patients. Nonetheless, the main result that Mixed patients utilized more hospital care than other patient types remains robust.