We provide four tables of results, with two dependent variables — quality-adjusted cost per mile and labor share — across four samples. The first sample includes estimates with the full 48 conterminous states (49 in the case of labor share). The second and third models reduce the sample by removing those states with a $100,000 or lower threshold and then those with a $1,000,000 or lower threshold of projects covered by the prevailing wage. Our fourth sample includes just those six states that changed their legislation during the observed period. In these samples, we test the more fully parameterized panel model with several controls.[*] Implicit in this is the assumption of exogeneity of the timing of the prevailing wage changes. This is our Kessler and Katz approach.
In our two-way, fixed-effects estimate, we provide three specifications of our full, 48-state sample. The inclusion of the time-fixed effect reduces the availability of several control variables from our panel model through collinearity. The three specifications provided for the full sample of both our cost and labor share models are designed to illustrate the robustness of our estimate.
All of our models contain the spatial autocorrelation correction recommended by Pesaran and report standard errors treated by a panel version of White. We begin with the large sample of our cost model for the years 2004-2019, which are the limit of those available in our data sources. The dependent variable is in natural logarithm form.
We employed two methods of treating the potential endogeneity of quality measures of state road systems. The first was to include it as an implicit measure of cost, by altering the road miles by that measure. The second was to conduct a first stage estimate of cost per mile with the quality-ranking share as an exogenous variable. We then used the predicted cost per mile as a dependent variable.
The point estimate of the second approach yielded modestly higher point estimates, but these were not statistically different from the first approach. We prefer and report the more conservative estimate. Notably, our cost measure is not contract cost, but public spending per quality-adjusted mile of road. The transportation literature typically focuses on contract costs as measured by the bidding and payment process to a private sector contractor. The public finance literature treats spending as cost in these types of applications. We default to the public finance approach because individual contract data are not available.
In these four specifications estimating the cost effect of prevailing wage on road construction and maintenance costs (adjusted for quality), we report fairly consistent effects. On our prime variable of interest, we report that the presence of a prevailing wage law raises quality-adjusted construction costs by 11.3% to 14.3%.
We chose several options to treat the heterogeneity of state prevailing wage laws. Given that the high project threshold limits are in states that did not change their prevailing wage law, a separate dummy for these states yielded effectively identical results to the full sample with a prevailing wage dummy. In our second and third specifications we omitted those states with thresholds above $1,000,000 and above $100,000. The statistical variation in the coefficients of interest was trivial.
Finally, we test the model only on those states that changed their prevailing wage laws during our sample period. The point estimate was higher in this estimate, but it was not statistically different from the other estimates. The absence of different coefficients on the remaining variables boosts our confidence in the exogeneity assumption offered by Kessler and Katz. We discuss the range of effects in more detail below.
We did not anticipate effects from the suspension of federal Davis-Bacon rules in the wake of Hurricane Katrina, and these empirics support our priors. Also as expected, our roadway usage data (vehicle miles traveled) affected costs. More miles traveled increases the cost per mile. Higher federal road share tended to reduce state and local spending, an effect with several possible causes. Neither of these latter two variables are surprising or of particular interest in this study.
We note that our findings appear very similar in both method and result to Kessler and Katz and Vitaliano. Both of these studies introduce econometric models of the effects of prevailing wage laws, and each finds that state laws increase costs of road construction and maintenance. Our findings also are close to the magnitude of those reported by Vedder, Clark and Kersey, though our methods differ substantially.
Our second estimate evaluates the effect of eliminating prevailing wage legislation on the quality-adjusted cost of roads in a two-way, fixed-effect model. The identifying variation in this model is the group of six states that changed their laws during our available sample period.
The parameter of interest here is the prevailing wage, which remained nearly constant across the increasingly spare specifications. These estimates offer cost increases of having a prevailing wage legislation of 8.5% to 8.9% per quality-adjusted road mile, in the point estimates. The third model here is the sparest of difference-in-difference models with the abbreviated specification of equation (1) above:
Here, ρ is the difference-in-difference estimator of interest. Due to the variation in timing of this legislation, the coefficient value ρ should be interpreted as the weighted average of the treatment effects, or the effect of prevailing wage law changes for each time period in which a change occurs.
The challenge of the variation in timing identified by Goodman-Bacon is that individual weights of the treatment effects may result from comparison between cross sections and control (non-treatment) variables that were treated early and those treated later. One way to assess this is to conduct a Bacon Decomposition to ascertain what share of the weighted effects results from these non-control comparisons. We do this across two specifications and find that between 90% and 95.5% are nonnegative weights, or derive from control and treatment comparisons.
In order to provide one more potential test, we also conducted a two-period difference in difference (2004 and 2019), defining the control group as those states which maintained a prevailing wage law across the sample, and the treatment group as those which changed their laws between these dates. This estimate provided similar results, of 5.6% cost reduction per quality-adjusted road mile for those states that changed their prevailing wage legislation. However, spending varies considerably from year to year due to lumpy investments, so this point estimate should be viewed as evidence of the overall robustness of the modeling rather than a cost estimate.
Our attempt at spatial modeling, designed to capture network effects of road construction and maintenance was statistically significant, but small. LeSage and Dominguez argue that this provides a spillover effect, but in this case the dollar effect is negligible, even for the state with the highest total road spending. This variable appears to have captured some cost link between states not controlled for by the Pesaran procedure. However, the estimated cost per mile effect is economically negligible.
The Hurricane Katrina suspension was not statistically meaningful, and vehicle miles traveled offered somewhat larger effects, but not important in scale. The point estimate of the largest variable found that an additional million vehicle miles traveled increased cost per mile by $1.83. Again, this is statistically but not economically significant.
The following estimates turn to the effect of prevailing wage on labor share. Recall that prevailing wage legislation exists to cause supra-normal wages (above market equilibrium) in states with these laws. This potentially has a range of effects on other variables, including the choice of the mix of occupations to hire within a construction firm as well as the share of capital and labor employed in construction. Our interest here is in determining the adjustment along the labor-capital margin that may be caused by changes to state-level prevailing wage legislation.
We begin with the full model, which, unlike the previous cost samples, includes Washington, D.C., because the data process associated with rating road quality is not part of this estimate. We approach this as we did with the cost models above, testing two full samples with different prevailing wage coding, then testing a homogenous sample. For the remaining models we follow the same approach as in the cost estimates above, reporting only three specifications.
This table reports results from our full model, which assumes exogeneity of the timing of prevailing wage legislation changes. The first model is of 48 conterminous states and the District of Columbia. The second and third models exclude the high-threshold prevailing wages states. The final model includes only those states that changed their prevailing wage law during the sample period.
In each of these models, the prevailing wage variable was both economically and statistically meaningful. Also, these models enjoyed the expected direction of effect, reducing the labor share by 6% to 18%. Neither the Hurricane Katrina nor vehicle-miles-traveled variables are statistically meaningful, nor do their point estimates rise to a meaningful level. The federal road share of construction is negative and meets traditional levels of statistical significance. These controls are not directly part of our research questions.
This specification requires exogeneity in the timing of the prevailing wage legislative change, the Kessler and Katz assumption. While we view this as highly plausible, we also relax that assumption in our two-way, fixed-effects model illustrated below.
For both sets of dependent variables we conduct one further robustness check, the exclusion of Louisiana due to the enormous inflow of road construction dollars in the wake of Hurricane Katrina. In these set of models (not shown), the prevailing wage coefficient point estimates remain similar to those reported above, but the level of statistical significance changes. For the cost model, the results were nearly identical. In our labor share estimate, in model 1, the value falls slightly inside the traditional level of statistical significance, while both models 2 and 3 fall outside that level.
As previously mentioned, the dependent variable of labor share in this estimate has a weakness which we cannot fully overcome. There is no apparent reason why this variable would be correlated with prevailing wage, but we cannot fully exclude this consideration. This is an endogeneity bias risk that we must consider when evaluating these findings.
Our robustness tests provide some evidence that project cost thresholds affect the labor share, but the impact is very modest. Excluding Louisiana, due to its extreme labor share post-Hurricane Katrina, casts doubt on the statistical certainty of the labor share estimates, but not on the point estimate.
Taken together, these concerns motivate us to reject the inference that changes to prevailing wage laws alter the labor share of construction. This may be due to measurement error, or due to a longer time path for such adjustments to occur. Also, it may simply be true that prevailing wage legislation does not change the market wage enough to alter the input mix on road construction. It may also simply be a matter of too few observations. Firms may not adjust quickly to labor cost changes, so detecting the effect may require more observations than we currently have.
We have considerably more confidence in our cost models. In these models the data quality is higher, and we have nearly identical results across models for each of the four different samples of states. Using the high and low point estimates, we provide the cost savings of ending state prevailing wage legislation in the six states that did so during our sample period. As yet another robustness check, we estimate the cost models 1-3, using only these six states in a sample. This is a small sample panel, but the only meaningful difference was a much higher coefficient (roughly 26% higher costs with prevailing wage).
Using the coefficients from the larger samples, we estimate the cost savings per mile due to the elimination of a prevailing wage law in the six states. Since these estimates use a common coefficient for all states, the savings differentials come from the differences in state-level costs per mile and the differences in assessed road quality. However, these estimates hold these two values constant. It is possible that the termination of prevailing wage legislation through legislative action has road quality effects, a question we leave to further research. These results appear in appear in Graphic 9.
[*] These models treat prevailing wage differently. The first codes all prevailing wage states identically, the other omits the high-threshold states. Our second sample omits the high prevailing wage threshold states, leaving us more homogenous sample of state applications of prevailing wage laws.