Empirical Results

The regression results found in Tables 3 – 8 support our hypothesis. Urban counties saw a higher spike in unemployment than other metropolitan counties following lockdown orders across Colorado and Utah. To test our hypothesis, we ran similar OLS regressions on counties in both Colorado and Utah. We wanted to extrapolate the relationship between unemployment rates (as a percentage) and COVID-19 restrictions by county type. We limited our regression to metropolitan counties (urban and suburban) to limit the effect different industries have on unemployment. Our models also explicitly control for other confounding variables like cumulative COVID-19 cases per capita and the labor force size in each county.

In Colorado, urban counties are associated with a 1.975 percentage point higher unemployment rate than suburban counties after the lockdown order, all else equal. To put this into perspective, 2% of the Denver County labor force accounts for over 83,000 people. This result is statistically significant at the 90% level with a p-value of 0.05373. Additionally, this model, which controls for the size of the labor force and cumulative COVID-19 cases per capita, accounts for approximately 69.2% of the variation in Colorado metropolitan unemployment rates (Table 3). The residuals vs. fitted plot for this regression showed us that the residual for high unemployment counties is more than zero (Fig. 3). This plot implies some heteroskedasticity, violating the Gauss-Markov OLS assumptions. We accounted for heteroskedasticity in Table 4 by including a model that uses robust standard errors. Even with these robust standard errors, the Colorado results are significant at the 95% confidence level (Table 4).

Tables 3 and 4: Colorado Results (Click to expand)

Running this regression model on the Utah data tells a similar story. Urban counties in Utah saw a 2.204 percentage point higher unemployment rate than suburban Utah counties following the stay-at-home recommendation. For reference, 2% of the labor force in Salt Lake County accounts for roughly 126,000 people. Similar to the Colorado regression, this result is statistically significant at the 90% confidence level. The Utah version of our model returned an adjusted R-squared value of 0.672, meaning our model explains about 67.2% of the variation of metropolitan county unemployment rates in Utah (Table 5). Looking at the plot of the residual-fitted values provided in Figure 3, we can see that the residuals are not centered at zero for all points in the data, implying some heteroskedasticity (Fig. 3). Again, we must account for the violation of the heteroskedasticity assumption with robust standard errors. Table 6 shows that our model with robust standard errors is now statistically significant at the 99% confidence level.

Tables 5 & 6: Utah Results (Click to Expand)

While the regressions on both Colorado and Utah show a statistical relationship between COVID-19 restrictions and unemployment rates, we must also consider another vital distinction between the two states. Colorado Governor Jared Polis implemented a mandatory lockdown that temporarily closed non-essential businesses. Utah Governor Gary Herbert, on the other hand, made no such order. There was no mandatory lockdown in the State of Utah. Alternatively, Gov. Herbert issued a voluntary stay-at-home recommendation, and no businesses were closed by state mandate. To test the differential effect of mandatory lockdowns on unemployment, we can run a differences-in-differences model with Utah counties as the control group and Colorado counties as the treatment group. The coefficient associated with treatment counties after the lockdown is 0.976. Urban counties in Colorado, which implemented a mandatory lockdown, are associated with a 0.648 percentage point lower unemployment rate than similar counties in Utah without a mandatory lockdown (Table 7). Both results are not statistically significant, with p-values of 0.8562 and 0.6835, respectively. We cannot reject the null that the differential effect is zero. As a matter of fact, none of the coefficients associated with the treatment group were statistically significant at the 95% level. This model explains about 72% of the variation in metropolitan unemployment rates (Table 7). Like with our other models, we cannot ignore heteroskedastic errors in our model (Fig. 5). Our model with robust standard errors asserts the proposition that mandatory lockdown orders did not have a statistically significant differential effect on unemployment rates in Colorado compared to unemployment rates in Utah with a voluntary stay-at-home recommendation (Table 8).

Tables 7 & 8: Combined Results

The results of our models have two crucial implications. First, the negative economic impact on urban counties because of COVID-19 mandates/recommendations is worse than the impact on suburban counties, which supports our hypothesis. Suburban households are not seeing the loss of income that urban households are. We suspect this is because urban areas have a higher concentration of hourly employees and service industry workers. Policymakers must consider the effect of regulations on these urban workers. These results also suggest that federal and state governments should focus on supporting these workers with progressive stimulus payments. Urban workers are suffering the effects of this pandemic at a greater intensity than workers that live in the suburbs. The second major takeaway relates to mandatory vs. voluntary COVID-19 restrictions. Our differences-in-differences models show no statistically significant difference in unemployment rates between Colorado and Utah, which had vastly different approaches to the COVID-19 pandemic. While this paper does not analyze the effect of different types of restrictions on COVID-19 cases, the results of this research imply that policymakers should take whichever approach is more effective at containing the virus without worrying about the labor effects.

Residuals vs. Fitted Plots