Better outcomes at lower costs? The effect of public health expenditures on hospital efficiency

ABSTRACT

Local health departments play a critical role in the community they serve as they are the foundation of the U.S. public health system providing services such as immunizations to the less affluent and advocating for state smoking bans. Research indicates public health expenditures improve overall health of the population. Importantly, a healthier population may lead to efficiency gains for surrounding health care providers. We use a two-stage semi-parametric Data Envelopment Analysis to estimate the effects of public health spending on the technical efficiency of the surrounding hospitals. Our results indicate hospitals operating in an area with a high level of per capita public health expenditures experience gains in efficiency of approximately 1.67 percentage points relative to hospitals in low spending areas suggesting a $20 billion in annual savings due to increased hospital efficiency. We also found that the more traditional approaches using the biased estimate for technical efficiency yielded the same conclusions with less computational burdens.

Abbreviation: WHO: World Health Organization; NACCHO: National Association of City and County Health Organization; DEA: Data Envelopment Analysis; CMS: Centers for Medicare and Medicaid Services; FTE: Full time equivalent; AHRF: Area Health Resource Files; MSA: Metropolitan Statistical Area; AHA: American Hospital Association

KEYWORDS:

• Public health expenditures
• hospital technical efficiency
• data envelopment analysis
• two-stage estimation

JEL CLASSIFICATION:

• H41
• H51
• I18
• C61
• C14

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ There were 628, 571, and 681 hospitals reporting zero Doctors (FTE) staffed in 2007, 2010, and 2012 respectively. The doctor variable, according to our contact at the AHA, is unreliable and as a result, we did not use it in our research.

² A staffed bed is one where the hospital has the labor resources necessary to service that bed in the event it is occupied.

³ The estimates provided in Table 1 are produced using the raw AHA survey data which is prior to performing any cleaning actions to ensure the data is reliable.

⁴ Our sample includes only general medical and surgical hospitals. More than three quarters of hospitals in the AHA data are General Medical and Surgical hospitals.

⁵ This analysis does not make use of NACCHO’s 2013 or 2016 National Profile of Local Health Departments as we were unable to meet the two year lag time at the time this study was conducted. Specifically, to meet the two year lag between public health and hospital data, we would need to acquire 2015 and 2018 AHA hospital data. At the time when AHA data was purchased for this study, neither 2015 nor 2018 data was available for purchase.

⁶ Prior to merging the NACCHO and AHRF data, several data decisions were made to simplify the merge. Specifically, the public health data was restricted to only those counties operating in the continental United States. Additionally, because the governing structure of local health departments varies by state, the public health data was restricted to only those health departments operating at the county level.

⁷ Prior to the merger, the hospital specific data was restricted to the continental United States as well as restricted to only general medical and surgical hospitals.

⁸ Per capita county public health expenditures are only used to determine the level (Low, Medium, or High) of public health spending each hospital has experienced two years prior. As a robustness check, we examined how the assigned public health level evolved over time for each hospital in the sample, finding the variable to be consistent. We believe transforming the expenditure data in this way will limit the potential impact of inaccurate surveys on our final results.

⁹ By definition, Algorithm 1 is embedded in the first four steps of Algorithm 2.

¹⁰ Algorithm 2 has been modified to estimate an input-oriented model as defined by Needlea and Fannin (2013).

¹¹ The technical appendix is available upon request.

¹² Once the bootstrapped values are calculated, percentile confidence intervals are constructed such that Pr$\left[-b_{\alpha }^{\ast }\le \left({\widehat{\stackrel{ˆ}{{\beta }_j}}}^{\ast }-\widehat{\stackrel{ˆ}{{\beta }_j}}\right)\le -a_{\alpha }^{\ast }\right]\approx 1-\alpha$.