Information identified as archived on the Web is for reference, research or recordkeeping purposes. It has not been altered or updated after the date of archiving. Web pages that are archived on the Web are not subject to the Government of Canada Web Standards. As per the Communications Policy of the Government of Canada, please contact us to request an alternate format.
A wide range of methodologies have been used to estimate the impact of minimum wage legislation, with the focus being on measuring the employment impact. This section can be skipped by those who are interested only in the bottom-line evidence presented later, rather than the methodologies used to generate that evidence.
The most common earlier methodology, prior to the advent of large cross section and panel data sets, was time series regression. The analysis usually involved regressing the ratio of employment to the population on a set of control variable as well as a measure of the minimum wage. Control variables included such factors as a measure of the business cycle, school enrolment and the share of youth in the population (to capture cohort size effects). The minimum wage variable is usually constructed as an employment weighted average of the minimum wage relative to the average wage in each industry. It is also usually weighted by the percent of the industry that is covered by the minimum wage. This minimum wage measure is sometimes referred to as the Kaitz index after its originator. The analysis is sometimes done for different age, gender and race groups. Twenty six of such studies in the United States are reviewed in Brown, Gilroy and Kohen (1982), 28 in Brown (1999) and 29 in Card and Krueger (1995, p. 180-82). Such time series analysis, however, suffers from the difficulty of controlling for the myriad of other factors that change over time and that can affect employment opportunities.
Aggregate cross section studies usually used the state or metropolitan area in the United States as the unit of analysis. Variation in the minimum wages tended to come from three main sources:
Similar to the time series analysis, the ratio of employment to the population was used as the dependent variable. Control variables included: the ratio of the youth to the adult population to control for cohort size effects that could otherwise affect wages; unemployment rates; family income; and urban size. As with the time series analysis, however, it is often difficult to fully control for the effect of these other factors that could influence employment opportunities.
Obviously combining time series and cross section data provides richer data sets and this has been a more common procedure in the 1990s. Brown (1999, p. 2125) cites seven of such studies, with some containing more than one set of estimates based on different data sets. Control variables are similar to the ones used in the time series and cross section studies, but also include state and time fixed effects. The minimum wage variable is sometimes constructed similar to the Kaitz index – the maximum of the federal or state minimum, adjusted for coverage (e.g., Neumark and Wascher 1992) and sometimes as the fraction of the relevant population that is affected by a minimum wage increase (Card 1992a, b; Card and Krueger 1995). Obviously, having both time series and cross section variation in minimum wages helps identifying their impacts, albeit controlling for the full range of factors that can also influence employment opportunities poses a challenge.
A more novel and recent approach involves estimating the transition probabilities of individuals from employment to non-employment based on panel data that follows the same individual over time. Essentially, the “treatment” group of individuals “at risk” of being affected by a minimum wage increase are identified as individuals whose wage is “bounded” by or lies between the old minimum wage and the new minimum wage at a point in time. Their transition probability or probability of being employed in the next period is then compared to various control groups who are not “at risk” of being affected by the minimum wage increase. Essentially a regression equation is estimated where the dependent variable is coded 1 if the person is employed in period t, zero otherwise, based on data of persons who were employed in period t-1. A series of control variables in period t-1 were used that could affect the transition probability of being employed in period t, given that they were employed in period t-1. The minimum wage variable is essentially a dummy variable coded 1 if the person was in the “treatment” group of individual at risk of being affected by the minimum wage in that their wage fell in between the old and new minimum, and coded 0 for persons in the control group who were not at risk of being affected by the minimum wage increase. The coefficient on the minimum wage variable gives the effect of being at risk of being subject to a minimum wage change (compared to the control group that is not at risk of being affected) on the probability of being employed in period t, conditional upon having been employed in period t-1, and after controlling for the effect of other variables that could affect that transition probability.
A minimum wage gap measure can also be used, constructed as the difference between the individual’s actual wage and the minimum wage for those “at risk” of being affected by the minimum wage increase in that their wage fell in between the old and new minimum. The gap would capture the extent to which an individual’s wage would have to increase to be in compliance with the new minimum wage. The larger the potential increase (and hence the greater the cost to the employer) the less likely the person would be expected to be employed in the next period. The minimum wage gap or potential wage increase would be set to 0 for persons in the control group whose wage would not be affected by the minimum wage increase.
Control groups are utilised to control for the other factors that could affect the employment transition probabilities of persons who are not at risk of being affected by the minimum wage increase. These “baseline” effects should be “subtracted” from the effect for those “at risk” to get a pure “treatment effect” of the impact of the minimum wage. The control group could include all other persons not “at risk” of being affected by the minimum wage – that is, all persons whose wages are not bounded by the minimum wage increase. Or it could include only persons who are likely to be similar to those experiencing the minimum wage increase but who do not experience a minimum wage increase – for example, those whose wages lie in the same bound as those potentially affected by a minimum wage increase, but who are in jurisdictions that did not have a minimum wage increase. Such, “treatment” - “control” group procedures likely control for the impact of factors that could confound the impact of minimum wages better than simple time series, cross-section studies.
Another novel approach that has been used in a number of recent studies – and the ones that have generated the most controversy – involves comparisons of before-and-after employment changes in firms in jurisdiction that have experienced a minimum wage increase (the “treatment” group), compared to firms in contiguous jurisdictions that are otherwise similar, except that they did not experience a minimum wage increase (the “control” group). The minimum wage impact is simply the change in employment in the jurisdiction experiencing the minimum wage increase less the change in employment in the contiguous jurisdiction that did not experience the increase. The latter control group is utilised to control for other factors that could have affected employment changes over that period, with the implicit assumption that these factors affected the minimum wage treatment group and the control group similarly.
These are termed “natural experiments” because they are natural events that approximate random assignment into treatment and control groups – with such random assignment being the ideal experiment but one that seldom occurs in the social (as opposed to medical) sciences. In general the data is obtained from surveys of employers or from administrative payroll data. The existing studies tend to be of fast-food restaurant chains.
While such procedures have an “elegant simplicity” to them since they essentially involve comparisons of means, and they approximate random assignment, they can have potential problems. For example, other factors (e.g., business cycles) may be affecting the treatment and control groups differently. The minimum wage increase could be endogenous in that the jurisdiction that instituted the increase could have done so at a time when they felt there would be unusual prosperity and employment growth (perhaps induced by other policies they were instituting). Furthermore, these procedures generally only estimate short-run effects since they are usually evaluated just before and just after the minimum wage increase.
Another recent novel approach (Neumark 2001) employs a pre-specified research design involving a pre-commitment to the empirical specification in advance of having the data so as to avoid subsequent data mining and “author bias” in selectively reporting on specifications that yield preferred estimates. This comes at the expense, however, of having less data (and hence less precise estimates) because the researcher cannot use past data that is affected by minimum wage changes.
A more complicated methodology for estimating the wage and employment effects of minimum wage legislation involves comparing the actual distribution of employment at different wages with the hypothetical distribution that would prevail if there were no minimum wage increase (Meyer and Wise, 1983a,b). The minimum wage does four things to the wage distribution: shifts a portion of those below the minimum up to a spike in the wage distribution around the minimum wage; leaves a portion of those below the minimum because of exemptions and non-compliance; shifts a portion completely out of the distribution if they lose their jobs; and shifts the wage distribution above the minimum upwards because of spillover or ripple effects.
Meyer and Wise estimate the hypothetical wage distribution that would prevail in the absence of the minimum wage increase by making assumptions about how the actual distribution above the minimum would extend back to reflect the distribution that would prevail if there were no minimum wage change. The difference between the actual distribution and this hypothetical distribution that would have prevailed in the absence of the minimum wage increase is used as an estimate of the adverse employment effect of the minimum wage change. The viability of such a procedure, however, depends upon the assumptions about how the actual distribution above the minimum would extend back to reflect the distribution that would prevail if there were no minimum wage change.
Clearly there is a wide range of methodologies used to evaluate the impact of minimum wage legislation. Each has its pros and cons. A consensus has not emerged on any single methodology as being best, nor has a consensus emerged as to the inherent bias and direction of the bias that may emerge from each of the methodologies.
In part because of the different methodologies and the lack of agreement as what is the preferred methodology, the empirical literature (discussed subsequently) does not yield a consensus – in fact, an earlier consensus has been shattered by new results done by reputable researchers. Those new results have been hotly contested by other reputable researchers in part due to the fact that the new results – minimum wage increases, at least over the limited ranges that have occurred, do not have an adverse employment effect and may even increase employment -– go against a fundamental theoretical tenet of economics, being that wage increases will unambiguously reduce the demand for labour since substitution and output effects work in the same direction. While there has been some narrowing of the differences in the empirical literature, a new consensus has not emerged.