RCTs, Epistemology, and the Minimum Wage
Can experimental evidence clear up the debate on the Minimum wage?
Randomized Control Trials (RCTs) are the modern gold standard of empirical knowledge in basically all scientific fields from economics to epidemiology. This is because RCTs eliminate the biases in conventional observational research that make it difficult to discern correlation and causation. For example, imagine that you are an education researcher and you are trying to evaluate the effectiveness of Montessori private schools on children’s future success. Simply observing the children from the Montessori school and comparing them to public school kids wouldn’t tell you anything. There are dozens of co-linear confounding variables and selection effects like parental income, geographic location, race, etc. Depending on the quality and breadth of the data it may be possible to control for some of these variables using econometric methods, but the results of an observational study like this will always be subject to scrutiny over selection issues, and rightly so.
Minimum wage research faces a similar problem. The vast majority of the most popular research on the minimum wage is based on observational data of US-state level variation. The exact extent of selection effects and amount of confounding variables is unknown. Additionally, measures of worker productivity, hours worked, and job conditions are often inadequate, inaccurate, or unavailable. At minimum, researchers attempt to control for the changing conditions of the overall US labor market, and differences in the absolute size and growth rates of the state’s economies. These sorts of studies have produced wide ranging estimates of the effects of the minimum wage, and several controversies. As the raging debate over slight incongruencies in research methods and controls continues, it is difficult not to wonder if there is a better way to study this problem. Enter: the RCT.
The paper I want to highlight in this post is “Price Floors and Employer Preferences” by John Horton. In this piece he conducts a randomized control trial on an online labor market, randomly assigning 4 different minimum wage levels ($0, $2, $3, and $4) to 160,000 job postings. This experimental design conveys several advantages over conventional empirical work. First, selection effects and biases based on the economic performance of the firms and the states/countries they are in are automatically controlled for by random assignment. Second, the online platform collects detailed measures on the pre-experiment attributes of all workers, the productivity of workers on the job, and the number of hours worked overall. These data are extremely important to analyzing the effects of the minimum wage but are not measured in the most popular empirical works on the topic. Finally, the computerized nature of the data leaves almost no room for measurement error.
There are four main results: “(1) the wages of hired workers increases, (2) at a sufficiently high minimum wage, the probability of hiring goes down, (3) hours-worked decreases at much lower levels of the minimum wage, and (4) the size of the reductions in hours-worked can be parsimoniously explained in part by the substantial substitution of higher productivity workers for lower productivity workers.”
The significant reductions in hours worked come from two sources according to Horton’s analysis. First, firms are economizing on now more expensive labor; the labor demand curve slopes downward. Second, the substitution of higher productivity workers meant that jobs were completed faster, so the total hours worked went down. Both of these responses to the minimum wage hurt low productivity workers: “I find that workers that had been working for less than the new platform minimum wage raised their wage bids after the platform-wide minimum wage was imposed. These same workers experienced a substantial decrease in their probability of being hired.”
Interestingly, these results are consistent with finding little to no dis-employment effect in an observational study that only measures wages and headcounts (which is what the vast majority of the most popular studies do). This is because almost all of the effects of the minimum wage came from substitution of higher productivity for lower productivity ones, which wouldn’t show up in headcounts, and reduction in hours worked, which is not measured in most conventional data sets.
This finding is also consistent with the predictions of the supply-demand model. Contrary to common understanding, the supply-demand model does not actually predict a dis-employment effect from a minimum wage at all, as explained in Brian Albrecht’s great post. The model does predict a mismatch between how much labor is demanded versus how much is supplied, with deadweight loss resulting from the prevention of positive sum trades. This need not result in unemployment, however, because the supply-demand model describes a market for hours of labor, not for jobs. In our world with large fixed costs for firing and hiring, changing the hours of labor is the natural margin for employers to act on in the face of a minimum wage.
Ultimately, the minimum wage serves as a transfer payment from low productivity workers to high productivity ones. Hardly what supporters of the policy have in mind when they ‘fight for fifteen.’
It is difficult to get reliable knowledge about the world, especially when it involves complex, homeostatic systems like the economy. For this reason, economists have traditionally relied on theory to guide them through our epistemic blind spots. The theoretical approach has been assisted and controverted by observational empirical work made possible by advances in computing. When these two approaches are at odds, as they are in the study of the minimum wage, it leads to widespread confusion, anger, and quibbling in the economics community. Deeper, more controlled, and more accurate empirical work, in the form of an RCT, can bridge the divide between these two camps. RCTs are not the only way to acquire knowledge about the world, but when they are available their results should be weighted much more heavily in our Bayesian updating than any observational work. John Horton’s work is no exception. His study confirms both the predictions of economic theory and reconciles these with the results of more basic empirical analyses.