Energy Transition and Rural Hospital Closure: Leaving Rural America Behind?

Energy production

We rely on the U.S. Dairy Association (USDA's) comprehensive, national oil and gas production data set from the USDA's Economic Research Service (USDA ERS). ²⁶ This county-level data set covers the 2000–2011 period, the period wherein unconventional oil and gas extraction methods created a boom in domestic production. We expect some degree of temporal lag between the oil and gas boom and rural hospital closure. That is, it is unlikely that a sudden boom in production will create the conditions to allow a rural hospital to continue operations. However, it is plausible that long-run oil and gas production might generate the conditions—population growth, tax revenue, and economic development—that might make hospital closure less likely.

Both oil and gas production are extremely skewed. Only one-third of all counties had oil or gas production during the 2000–2011 period, meaning that most counties have a score of 0 on both variables. A common method to correct skewed variables is log transformation, but this method results in missing data when predictors have a true zero. To address this problem and avoid dropping cases, we took the square root of both variables and implemented a method described by Hosmer and Lemeshow by creating binary variables to indicate whether a county had oil and gas production. ²⁷ That is, our models include both a square root transformed interval-level predictor for oil and gas production and a binary indicator of whether the county hosted any oil and gas production.

For wind energy, we accessed a shape file developed by U.S. Geological Service researchers. ²⁸ From this shape file, we used the geoinpoly user-written program in Stata 15 ²⁹ to generate a count of active wind turbines within the borders of each county from 2000 to 2010. A limitation of these data is that, for about 9% of turbines, the year that it became active is unknown. We assumed that these turbines were active over the 2000–2010 period, although it is possible some came online after 2010. We suspect, but cannot confirm, that missing or inaccurate data are probably more common in the older regions of the data set because it was constructed from historical documents. As with other predictors, this variable was extremely skewed—56% of counties did not host any wind turbines, whereas some counties had thousands. We again took the square root of this variable and created a binary variable for counties that had wind turbines and those that did not. Both the binary indicator and the square rooted count are included in our regression models described hereunder.

As we noted previously, the ongoing transitions in the U.S. energy system have led to a significant decline in the coal sector, a traditional economic livelihood in parts of the rural United States. We accessed data from the Energy Information Agency to identify counties that had experienced the shuttering of a coal mine from 2000 to 2010. ³⁰ For our purpose, any mine that went from active status (e.g., producing coal and workers employed) and transitioned to nonproducing is counted as “closed.” We use a binary indicator wherein 0 represents county did not experience a coal mine closure and 1 represents mine closure. Descriptive statistics for all variables are displayed in Table 1.

Firth logistic models

In Table 2 we provide log-odds coefficients, standard errors, and model fit diagnostics for our firth logistic models. The model includes all energy production variables, as well as county socioeconomic variables and our indicator of Medicaid expansion. Energy production appears to have mostly null effects on the likelihood of hospital closure—oil production, gas production, and coal mine closure do not approach statistical significance. The lone exception is wind power—as the number of wind turbines within 100 miles of a county centroid increases, that county is less likely to experience a hospital level closure. Counties with greater population density are more likely to have a closure, and counties in states that expanded Medicaid coverage are more likely to retain their hospitals. In model 2, we dropped the predictors that were not statistically significant in model 1. Although the coefficients for the retained predictors are roughly similar, the model fit statistics (i.e., Akaike information criterion, Bayesian Information Criteria, and McKelvey and Zavoina R²) have all been damaged, suggesting that the excluded predictors did somewhat improve model fit.

Spatial linear probability model

Direct, spill-over, and total effects derived from the spatial LPM are displayed in Table 3. This model uses a spatial contiguity matrix that assumes a first-order spatial autocorrelation in the error term and allows for neighboring counties (whether sharing a border or vertex) to influence one another. In general, this model has reproduced the null results of the firth logistic model—none of our energy production variables reach statistical significance, and all have substantively small coefficients. As population density increases, a hospital closure becomes slightly more likely and rural hospitals are more likely to survive in states that expanded Medicaid coverage.

Robustness checks and diagnostics

After estimating our models, we performed a series of robustness checks. First, we implemented alternative transformations of oil and gas production, including a cube root transformation and splitting each variable into quartiles. These new variables produced similar null results. We also reran our models for subset of counties based on the rural–urban continuum codes provided by the USDA ERS. ³⁵ The modest to null effects of energy production were reproduced, even when excluding urban counties. Multicollinearity was assessed through variance inflation factors, none of which exceeded 4, indicating that multicollinearity is not a problem in our models.