Minimum wages and gender wage gaps. Evidence from the UK and Ireland.

Transcription

1 Minimum wages and gender wage gaps. Evidence from the UK and Ireland. Olivier Bargain Karina Doorley Philippe Van Kerm November 2011 Preliminary Version Abstract Since women are disproportionately in low paid work, they should bene t the most from minimum wage policies. To check this prediction, we exploit the introduction of a National Minimum Wage (NMW) in Ireland (2000) and the U.K. (1999). Using panel data (Living in Ireland Survey and the British Household Panel Survey), we estimate a exible model of wage distributions following the work of Chernozhukov et al (2009). We use estimates for men and women, before and after the introduction of the NMW to construct counterfactual distributions of wages based on a xed distribution of covariates for women, aimed at separating out explained, human capital e ects, and residual discriminatory di erentials. We apply simple statistical measures of discrimination based on wage CDF s to estimate the gender wage gap before and after the introduction of the NMW. We perform several robustness checks ( xing the population using panel data to avoid e ects related to changes in labour market participation, correcting for the bias engendered by selection into employment etc.). We also suggest an extrapolation exercise which examines the e ect in the U.K. if its NMW had been as e ective as that in Ireland. Country comparisons show that the e ect of the minimum wage is not uniform. There is hardly any correction of the gender pay gap in the UK while a larger e ect is observed in Ireland, with modest spill-over e ects further up in the distribution. Reasons for this di erence are discussed and policy implications relating to the link between the gender wage gap and NMW s are drawn.

2 1 Introduction There has been a well-documented decline in the gender wage gap since the 1980 s. Blau and Khan (1997) postulate that the closure of the gender wage gap is attributable to improvements in women s occupational status and experience, to enhancements in women s unmeasured labor skills and to a decrease in discrimination against them. Generally, we expect the average wages for women to be less than those of men due to lower labour market attachment and lower experience caused by childbirth and childrearing. However, there is still a tangible gender wage gap that cannot be accounted for by observables such as education and experience. This has been estimated at about one-third by Blau and Khan (2003) and at between 10 27% by Arulampalam et al (2007). Albrecht et al (2009) account for selection into employment and nd that as much as two-thirds of the gender wage gap is unexplained in the Netherlands. In this paper, we focus on the impact of the minimum wage on gender wage inequality. Because the minimum wage mainly a ects the bottom of the wage distribution, this will allow us to estimate its e ect on the sticky oor. However, there may also be spillover e ects to the rest of the distribution that will be of interest to policymakers. A number studies have been performed in this area. For example, Blau & Kahn (2003) test for the impact of the relative level of the minimum wage on the gender gap using several years of micro data from 22 countries. They nd a negative correlation between the gender gap and the "bite"of minimum wages, which measures the impact of the NMW compared to the average wage. DiNardo et al (1996) nd evidence that the decline in the real value of the minimum wage explains a substantial proportion of rising wage inequality in the U.S. between 1979 and Ganguli and Terrell (2006) nd that the most important reason for the reduction of the gender wage gap at the bottom of the distribution in the Ukraine was the fact that the minimum wage was set at a high level in 2003 and this raised the wage oor for more women than men. Robinson (2002), using quantile regression, nds no evidence that the NMW in the UK a ected the gender wage gap in the lower part of the wage distribution. In a seperate study in 2005, Robinson uses regional variation in the "bite" of the NMW in the UK combined with a di erence-in-di erence approach and nds some evidence of a narrowing of the gender pay gap by 1 or 2 ppt in regions where women comprise a relatively large share of the low paid, and where the regional "bite" is larger. We study the relatively recent introduction of a NMW in Ireland in 2000 and in the UK in These countries set NMW s with di erent "bites" within a year of each other and, so, make an interesting case study. Using survey data, we employ a exible model of wage distributions to construct counterfactual distributions of wages based on a xed distribution of covariates for women in each country. We estimate the gender wage gap before and after the introduction of the minimum wage, at every point in the wage distribution, separating out explained, human capital e ects, and residual discriminatory di erentials. We can thus deduce what e ect (if any) the NMW in each country had on the gender wage gap at the bottom of the distribution, as well as identifying spillover e ects further up in the distribution. Our contribution to the literature is as follows. Firstly, this is one of the rst studies to examine the e ect of the NMW on the gender pay gap in Ireland, a country with a strong history of gender inequality 1

3 on the labour market due to a combination of cultural and religious ideals and a relatively weak economy until the Celitc Tiger years in the 1990 s 1. The inclusion of the U.K. allows for an interesting comparison between these neighbouring countries, who have di erent histories and, interestingly, di erent NMW levels. Comparing these countries also allows us to perform an extrapolation exercise to provide further insight into the drivers of the gender wage gap at the bottom of the wage distribution. In addition to this, we use distribution regression, a methodology recently used by Chernozhukov et al (2009) to analyse wage inequality in the US, which is seldom exploited in this literature, and suggest a simple correction for selection bias. Distribution regression allows us to model entire counterfactual distributions of wages, in order to pinpoint the gender wage gap before and after the introduction of the NMW at every point in the wage distribution. It is better suited to our study than the more commonly used quantile regression methodology for a number of reasons. First, as pointed out by Chernozhukov et al (2009), there is a considerable amount of rounding in wage variables, which makes it highly discrete. Second, a linear model for the conditional quantile function may not provide a good approximation to the conditional quantiles near the minimum wage, where the conditional quantile function may be highly nonlinear. The distribution regression approach does not su er from these problems. It provides a direct way to identify changes in the gender wage gap at speci c wage levels i.e at the NMW level in each country. Additionally, the construction of counterfactual distributions for women using distribution regression does not raise the issue of rank invariance as it would using quantile regression. 2 Institutional Background 2.1 The Minimum Wage Table 1: The "bite" of the NMW in the UK and Ireland 1 McGinnity and Russell (1998) nd a traditional (and unequal) gender division of labour in Ireland compared to other European countries. 2

4 Policy responses to gender wage inequality can generally be classed in three categories. Firstly, equal pay policies attempt to redress the disequilibrium between male and female wages. However, a lack of e ective enforcement of such policies contributes to the problem of unequal pay. Secondly, equal opportunities policies, including childcare arrangements and family leave are also important to enable women to have more continuous employment patterns. However there are drawbacks to their use. On the one hand, such policies may raise the relative earnings of women compared to men by encouraging the preservation of their ties to particular rms and, hence, increasing the incentives of employers and female workers to invest in rm speci c training. On the other hand, the existence of such policies could increase the incidence and/or duration of temporary labour force withdrawals among women, raising the gender wage gap for the a ected group. The incremental costs associated with mandated leave policies may also increase the incentives of employers to discriminate against women. Arulampalam et al (2007) show that the gender pay gap is higher at the top of the wage distribution in countries with more generous family-friendly policies, suggesting that the negative e ect dominates at the top of the distribution. Conversely, the gender pay gap is lower at the bottom of the distribution in countries with generous family-friendly policies suggesting that the positive e ect of family friendly policies dominates at the bottom of the distribution. Finally, wage policies can be used to combat gender wage inequality. Wage inequality in the population as a whole has been shown to be strongly positively correlated with the gender wage gap (Gupta et al, 2006). Studies of the impact of minimum wage on the wage distribution invariably nd that such regulation compresses the bottom of the distribution, reducing the sticky oor e ect, which is the tendency for the gender wage gap to be higher at the bottom of the wage distribution than elsewhere (Blau and Kahn, 2001; Ganguli and Terrell, 2006). High wage oors may also encourage women s labour force participation, by making the xed costs of work more a ordable. Additionally, if the return to experience is increased, women have more incentive to remain employed during periods of childbearing and childrearing. NMW s exist in most OECD countries. They have been accused of reducing labour demand, particularly for young, low-skilled workers. For example, Neumark et al (2004) in a study using US data nd that workers who initially earn near the minimum wage experience wage gains in the wake of an increase in the NMW. However, their hours of work and employment rates decline and the combined e ect of these changes on income can be negative for low-wage workers. Conversely, some studies suggest that the employment e ects of NMW s are small and, in some cases, may even be positive (Card and Krueger, 1994; Dickens, Machin and Manning, 1999). Regardless, NMW s are popular tools for reducing poverty, halting the exploitation of low-paid workers and increasing incentives for people to look for, or remain in, paid employment. Although this is not always an intended consequence, they can also tackle pay di erentials between men and women. In April 1999, a National Minimum Wage (NMW) was introduced to the UK labour market. This was the rst time that the UK had an economy wide minimum wage. The old industry-based Wages Council system that used to regulate pay in some sectors was abolished in 1993 by John Major s government, 3

5 who argued that the wages councils reduced employment, athough there was no clear evidence that the system had cost jobs (Machin and Manning, 1994). Following its election in 1997, Tony Blair s Labour Government committed itself to introducing a NMW and a minimum wage of $3:60 per hour was introduced in April 1999 for those aged 22 or older, with a lower youth rate of $3 per hour for those aged inclusive (those aged less than 18 were not covered). The initial rate was purposely set at a modest rate, as it was deemed best to start low rather than setting it too high and provoking adverse employment e ects. The Low Pay Commission expected that around 9 percent of workers would be directly a ected and those workers would, on average, receive a 30 percent boost to their pay (Metcalf, 1999). These numbers have subsequently been revised down, with it seeming more likely that about 6 percent of workers wages were raised up to the minimum (Dickens and Manning, 2003). This, of course, still amounts to a potentially large impact on the labour market. Moreover, it is also clear that there is systematic variation in who bene ted from the minimum wage introduction. Metcalf (1999) notes the increased importance of the minimum wage for part-time female workers. Of the workers he estimates to be directly a ected by the introduction of the minimum wage, around 55 percent were women working part-time. According to Metcalf (2008), who survey the literature relating to the employment e ect of the British NMW, there is little or no evidence of any employment e ect. Dolton et al (2011) corroberate this in a study of employment and inequality in the UK over the decade since its introduction. They nd that the average employment e ect over the entire period is neutral, although there are small but signi cant positive NMW e ects from 2003 onwards. A regression discontinuity approach is used by Dickens et al (2011) who nd a 2-4 ppt increase in the employment rate of low skilled individuals at age 22 in response to the NMW. However, they use pooled data from July 1999 to March 2009, so it is unclear whether or not the e ect was tangible immediately after the introduction of the NMW. A national minimum wage of IE$ 4:40 per hour was rst introduced in Ireland on 1st April 2000 under the National Minimum Wage Act Prior to this, minimum wages in Ireland were set by Joint Labour Committees. However the wages speci ed in these agreements were often low and covered less than a quarter of the workforce. Furthermore the level of enforcement was quite weak. In 1999 a Commission was set up to oversee the introduction of a NMW in Ireland. This Minimum Wage Commission recommended that the initial rate be set at around two thirds of median earnings, representing around IE 4.40 per hour (O Neill et al, 2006). O cial gures suggest that the minimum wage directly bene ted approximately 163,000 workers, or 13.5% of the total workforce. Figures from the Economic and Social Research Institute (ESRI) indicate that 17% of female workers and 11% of male workers earned less than the IE$4:40 rate at the time of its introduction. There is little evidence in the literature either relating to the e ectiveness of the Irish NMW in tackling the gender wage gap or to any employment e ects it may have had. McGuinness et al (2003), using an employer-employee matched dataset nd that the Irish NMW wage improved the relative position of part-time females only. The discover that the wage penalty for being employed in a company with a large proprotion of NMW earners was less for part-time women than for part-time men, while there was no di erence for full-time workers. O Neill et al (2006) nd that the minimum wage may have had a statistically signi cantly negative e ect on employment for the small 4

6 number of rms most severely a ected by the new legislation i.e. with a high proportion of expenses devoted to low wage workers. Table 1, shows the "bite" of the NMW in each country and the proportion of people earning less than it in the year before its introduction, according to the LII and BHPS datasets. Our calculations are similar to the o cial statistics, giving us con dence in the chosen datasets. 2.2 Other Policies In any policy analysis such as this, it is important to ensure that the e ect picked up is due to the policy in question and not due to other policies implemented at the same time. The period was one of large growth, declining unemployment and generous budgets in Ireland. For example, between 1998 and 2000, successive budgets increased the level of the Family Income Supplement 2, the Lone Parent Allowance and Child Bene t, decreased the income tax rate of both the higher and lower tax brackets and increased tax free allowances for all household types. In the UK, the "New Deals" for lone parents, young people and over-25s were introduced in 1998 and and the Working Families Tax Credit was introduced to replace the Family Credit in October 1999 (Brewer and Browne, 2006). As most of these measures a ect net income, and not gross income except through employment e ects, they should not have any e ect on our estimation of the e ect of the NMW on the gender wage gap. We may, however, when testing for employment e ects of the NMW, nd that these are confounded with the employment e ects of the policies mentioned above. For example, Brewer and Browne estimate that the labour market participation of lone mothers increased by around 5ppts between 1999 and 2002 due to the WFTC, which is, arguably, the most signi cant of any of these policies. 3 Data We use Living in Ireland (LII) survey data for the years 1999, 2000 and 2001 and the British Household Panel Survey (BHPS) for 1998, 1999 and 2000 in our analysis. We restrict our sample to people observed in 1999 and 2001 in Ireland and 1998 and 2000 in the UK. We restrict our sample to those aged between 22 and over 64 years of age, as under 22 year olds are not eligible for the NMW in the UK. We also drop those still in education. The sample size for the two years of interest is 12; 604 in Ireland and 20; 274 in the UK. Excluding students drops 8% of the Irish and 11% of the UK sample. The age restriction causes us to lose a further 25% of the Irish and 22% of the UK sample. When we drop individuals who are observed twice in any given year and the Scottish and Welsh refreshment samples in the UK data, we are down to 61% of the original Irish sample (7; 686 observations) 58% of the original British sample (11; 782 observations). Of these, we observe 4; 385 workers in Ireland and 7; 499 workers in the UK for the two years in question. An issue speci c to the Irish data is the "refreshment" sample of 1; 515 households that were added to the survey in 2000 because of some attrition over the life of the survey. To tackle this issue, we present 2 The Family Income Supplement is a tax credit for families at work on low pay. 5

7 Table 2: Variable means for men and women in the UK and Ireland before and after the introduction of the NMW 6

8 results both with and without the refreshment sample for Ireland. We normalise hourly wages 3 to their level during the year of the introduction of the NMW (so Irish wages are normalised to 2000 while British wages are normalised to 1999), using Consumer Price Indices. We de ne ve samples. The rst (all) is composed of everyone who has a wage in either period. This is our baseline sample and our main results relate to this sample. For robustness checks, we de ne a number of subsamples which will be dealt with in section 7. 4 We show in table 2 summary statistics only for the all sample. The main changes in the sample composition are an increased hourly wage and an increase in the average age of the population. There is also a decrease in the marriage rate in Ireland and a decrease in the proportion of temporary contracts in both countries. Results in the remainder of the paper also pertain the all sample, although we present the results of the other samples in the robustness checks in section 7 and in the Appendix. 4 Distribution Regression Recent research into the gender wage gap has increasingly focused on more global methods than the evaluation of gender wage di erences at the mean. There has been a surge of methodologies extending the decomposition of di erences at the mean to decomposition of the whole distribution. Firstly, Jenkins (1994) suggested using Generalised Lorenz and Concentration curves to derive summary indices for discrimination which accounted for the entire distribution of wages. DiNardo et al. (1996) then proposed kernel density estimates of counterfactual densities, estimated by re-weighting functions, which depend on the distribution of covariates between populations. Fortin and Lemieux (1998) divided the wage support into small intervals and estimated the probability of being in each wage interval using an ordered response probit model. Once estimated, this model was used to predict counterfactual distributions. Bonjour and Ger n (2001) considered wage distributions as if they were duration distributions, estimating a exible wage hazard function to recover the corresponding conditional wage distribution from the estimated parameters. Finally, the most widely used method to estimate distributions in the presence of covariates in recent years is Quantile Regression (QR, e.g. Buchinsky, 1994, 1998) followed by Machado and Mata s (2005) simulation algorithm to carry out counterfactual comparisons of densities. Chernozhukov et al (2009), inspired by Foresi and Peracchi s (1995) use of counterfactual distributions to model excess returns on nancial markets, recently formalised procedures for inferring how policy interventions a ect the entire marginal distribution of an outcome of interest. It is this distribution 3 Hourly wages are constructed from the current gross weekly wage and usual hours per week in LII and gross monthly pay (including overtime etc.), standard weekly hours and paid overtime hours per week in BHPS. LII: y = piwageg1 za23f (paygu) jbhrs+(1:5jbotpd) BHPS: y = 4 all_nr is a subsample of Irish data only, which discards the refreshment sample previously alluded to. same1 consists of all those who are observed both before and after the introduction of the NMW and who work at least part-time ( 15 hours per week) in at least one period. same2 consists of all those who are observed both before and after the introduction of the NMW and who work at least part-time ( 15 hours per week) in both periods. Finally pooled takes a pooled sample of two years before and two years after the NMW instead of one. 7

9 regression technique that we make use of in this paper. In practical terms, this involves running a series of probit models at each point in the wage distribution, separately for men and women for each time period (before and after). The dependent variable is binary and takes the value of 1 if the individual has an hourly wage below w, where w takes the value of each point of the wage distribution seqentially, and 0 otherwise. These models are used to predict the probability that an individual is located below an hourly wage w in the distribution, as well as predicting what this probability would be if the individual was compensated as if they belonged to a di erent gender group or time period. We employ an Oaxaca- Blinder style decomposition to the marginal wage distributions of men and women before and after the introduction of the NMW to identify what the wage gaps in each time period are, and if they have changed in the after period, all else held constant. More formally, we are interested in the change in the conditional distribution of wages for men and women observed before and after the introduction of the NMW, given explanatory variables such as age, education etc, holding the marginal distribution of these covariates constant. We recover estimates of the marginal distribution by integration of the conditional distributions over job and human capital characteristics. We denote F k;m (w) the marginal wage distribution where the superscript refers to the conditional l;n wage distribution and the subscript refers to the covariate distribution. The conditional wage distribution can be that of women (k = f) or men (k = m) before (m = b) or after (m = a) the introduction of the NMW and the covariate distribution can also relate to women (l = f) or men (l = m) before (n = b) or after (n = a) the introduction of the NMW F k;m l;m (w) can be recovered by integration of the estimates of conditional distributions over job and human capital characteristics. Z Z F k;m l;n (w) = F k;m (wjx; c) h l;n (x; c) dcdx (1) h j where F k;m (jx; c) is the conditional cumulative wage distribution function for human capital characteristics x and job characteristics c and h l;n is the density distribution of human capital and job characteristics. Taking the example of F f;b f;b (w), which is the marginal wage distribution of female before workers, paid as female before workers, sample estimates are obtained by replacing F f;b (jx; c) by estimates ^F f;b (jx; c) from equation (1), and by Monte Carlo integration over our sample of N female workers before the introduction of the NMW 5 : Nf;b ^F f;b f;b (w) = X i=1 ^F f;b (wjx i ; c i ) (2) The separation of conditional distributions and distribution of characteristics o ers a straighforward way to create counterfactual marginal wage distributions. For example, Nf;b ^F m;b f;b (w) = X i=1 ^F m;b (wjx i ; c i ) (3) 5 Individual sampling weights are omitted from this expression for notational clarity, but they are used at all estimation stages. 8

10 is a counterfactual distribution that represents the distribution that would be observed among female workers before the introduction of the minimum wage if the conditional wage distributions among male workers had prevailed over the female distributions. 6 The gender gap in pay before introduction of the minimum wage is then given by the di erence between the counterfactual distribution and the observed distribution: ^ DF b (w) = = ^F f;b f;b (w) N f;b X i=1 m;b ^F f;b (w) (4) ^F f;b (wjx i ; c i ) ^F m;b (wjx i ; c i ) (5) The gender gap in pay after introduction of the minimum wage can be written analogously ^ DF a (w) = = ^F f;a f;a (w) N f;a X i=1 m;a ^F f;a (w) (6) ^F f;a (wjx i ; c i ) ^F m;a (wjx i ; c i ) (7) The total impact of the minimum wage is then given by DDF ^ (w) = DF ^ b (w) ^ DF a (w) (8) We expect the impact to be stronger for w below the minimum wage. One issue with this approach is that the NMW may have had side-e ects on female employment on top of e ects on wages for example, with changes in work hours or occupations. To sort this out, we further factorize DDF ^ (w) into a pure price e ect that re ects the NMW impact on the relative compensation of men and women, and an employment e ect, through in uence on the employment structure of women (we do not consider impacts on human capital in the short time span covered). To do so, we construct additional counterfactual marginal distributions that would be observed if the prices after introduction of the NMW were applied to the sample of women with job and human capital characteristics before the NWM: Nf;b ^F m;a f;b (w) = X ^F f;a f;b (w) = We can decompose the total change as: i=1 Nf;b X i=1 ^F m;a (wjx i ; c i ) (9) ^F f;a (wjx i ; c i ) (10) ^ DDF (w) = P DF ^ (w) + EDF ^ (w) (11) 6 Observe that, in line with the whole literature on gender wage di erentials, we use the counterfactual distribution as a device to quantify the magnitude of the wage di erences between men and women. Causal inference based on a claim that the counterfactual distribution truly represents the distribution that would be observed if women were paid like men would require (implausible) assumptions about the absence of general equilibrium or feedback e ects of the change in the conditional wage distributions onto the distribution of covariates (human capital and job characteristics). 9

11 ^ DDF (w) = h ^F f;b m;b f;b (w) ^F f;b (w) h ^F f;a m;a f;b (w) ^F f;b i+ (w) ^F f;a f;b (w) ^F m;a f;b (w) ^F f;a f;a (w) ^F m;a f;a (w) i (12) e ect, where the rst term captures the price e ect, ^ EDF (w); of the NMW. ^ P DF (w); and the second term captures the employment 5 Results We plot the predicted distribution of wages for men and women in each time period against the actual distribution and nd an excellent t for our model (see gures 11 and 12 in the appendix). Tables 6 and 7 in the Appendix show the coe cients on our explanatory variables at four points in the wage distribution: the NMW and the 25th, 50th and 75th percentiles. The interpretation of the coe cients in distribution regression is slightly di erent to regular probit regressions. For example, the negative coe cient on age at the 25th percentile of the female before group in Ireland indicates that, as age increases, women are less likely to be located in the lower quartile of the distribution in the year before the NMW. All the coe cients have the expected sign. We depict our results using predicted and counterfactual cumulative distribution functions of wages for men and women in both countries. We show, in gures 2 and 1, three distributions for each country showing predicted wage distributions (F BF B; F AF A; MBMB and MAMA), counterfactual distributions where covariates are xed to female (M BF B; M AF A), and counterfactual distributions where covariates are xed to female before (F AF B; MAF B). The curves on the graphs are read as follows: The distribution curve KMLN = F k;m (w) shows the wage distribution using the coe cients of the group l;n k; m and the characteristics of the group l; n (where k; l = male or female and m; n = before or after) In each of tables 1 and 2, the CDF for female wages lies above that for male wages. Additionally, the CDF s for men and women before lie above those for men and women after. This is more pronounced at the bottom of the wage distribution, where the NMW is at work. The counterfactual wage distributions, MBF B and MAF A which show the distribution of wages if females we paid as males, are lower than the predicted distributions for females, F BF B and F AF A; while the counterfactual distribution, MAF B, which shows the distribution of female wages before if they were paid as males after, is also lower than F AF B, which shows the distribution of female wages before if they were paid as females after. These ndings are all consistent with our expectation that men are paid better than women and that wages are higher after the introduction of a NMW. We use the simple measures, described above in equations (8) and (12) to depict the e ect of the NMW on the gender wage gap in a more intuitive fashion. Figures 4 and 3 show the counterfactual gender wage gaps before and after the introduction of the NMW, and the di erence between them. It is clear that the gap varies throughout the distribution and by country. Importantly, we note that the 10

12 Figure 1: Predicted and Counterfactual CDF s of male and female wages in Ireland before and after the NMW Figure 2: Predicted and Counterfactual CDF s of male and female wages in the UK before and after the NMW 11

13 Figure 3: Gender wage gap decomposition and its change over time in Ireland Figure 4: Gender wage gap decomposition and its change over time in the UK 12

14 Figure 5: The change in the gender wage gap over time in Ireland Figure 6: The change in the gender wage gap over time in the UK 13

15 gender wage gap in the UK is about half the magnitude of that in Ireland at each country s NMW level, although it the gaps become similar throughout the rest of the distribution. Figures 5 and 6 show the DDF, the PDF and the EDF with 95% bootstrapped con dence intervals. To recap on these di erences in wage gaps, the DDF shows the e ect of the NMW on the counterfactual wage distributions of men and women. This is decomposed into a price (PDF ) and employment e ect (EDF ). They are interpreted as follows: A value of 1 indicates that there is a 1 ppt reduction in the di erence between a woman s probability of being paid below w and a man s probability of being paid below w. The only signi cant e ect of the NMW on the gender wage gap in the UK appears to be a small negative employment e ect in the upper half of the wage distribution. In Ireland, there is up to an 11 ppt reduction in the di erence between a women s probability of being paid below around $5:20 compared to a man s probability of being paid below $5:20. This is mainly due to the price e ect of the NMW, as seen in the PDF, but there is also a small positive employment e ect. 6 Robustness Checks and Additional Results To complement the results presented earlier, and to ensure that they are not the result of di erent before/after samples or selection bias, we suggest a number of robustness checks in this section. 6.1 Oaxaca-Blinder model Using the distribution regression framework, we can also summarize the e ects identi ed at speci c levels of w on mean wages, as is more traditionally looked at. Mean wages and counterfactual mean wages are recovered easily from marginal distributions and everything follows from there, for example: f;b f;b f;b = (F f;b ) = Z 1 0 wdf f;b f;b (13) which can be estimated from the marginal distribution estimates by numerical integration K ^ f;b f;b = X 1 2 (!g +! g 1 f;b )( ^F f;b (!g ) g=1 ^F f;b f;b (!g 1 )) (14) where f! 1 ; : : : ;! K g is a grid of points on the domain of de nition of wages at which we evaluate the marginal distributions 7, and! 0 f;b = 0 (where ^F f;b (!0 ) = 0). A comparison of how distribution regression at the mean performs compared to a standard Oaxaca- Blinder decomposition at the mean is depicted in table 8. The overall gaps at the mean, as well as the explained and unexplained componants are roughly the same whether we use distribution regression or the standard Oaxaca-Blinder decomposition at the mean. We look at these gaps at a number of other points in the distribution (p25; p50 and p75) for comparison. Interestingly, the wage gaps are greatly reduced at all points of the distribution in Ireland after the introduction of the NMW, while there is no real change in the U.K. This is in line with what we observe in gures 3 and To ease computation, we start the grid at approximately 2:5 in national currency in each country and stop it at 22. This encompasses about 95% of the wage distribution in each country. 8 (note that p75 in Ireland falls around a log wage of 2 2:5) 14

16 6.2 Alternative sample de nitions In a further robustness check, we use a number of more restrictive sample de nitions. First of all, to deal with the issue of the refreshment sample detailed in section 4, we restrict the Irish data to those who are not part of this boost sample.(all_nr). The results, in gure 13, show that the magnitude of the price e ect is unchanged, while the employment e ect disappears. We can therefore infer that the employment e ect shown in gure 5 is a product of the refreshment sample, who are younger on average. However, the price e ect below $5 per hour remains robust to this sample selection, and is even larger, approaching 20ppt in some areas. We then de ne two, more restrictive, subsamples for each country which discard some of the sample heterogeneity, and one larger pooled sample same1 : we restrict our analysis to those observed both before and after the introduction of the NMW and who work at least 15 hours per week in at least one of the periods. This still allows us to capture some potential participation e ects of the NMW. same2 : we use only those who are observed in both periods and who work at least 15 hours per week in both. This allows to capture only the e ect of the NMW on those who were at work before its introduction. pooled: we retain all those who are observed two years before the introduction of the NMW and two years after. This gives us a larger sample to work with. Any e ects not picked up in the all sample due to the sample size should be picked up in this sample. Summary statistics relating to these samples are provided in tables 3, 4 and 5 in the Appendix. The change in the gender wage gap due to the NMW using these sample selections are detailed in gures 14, 15, 16, 17, 18 and 19 in the Appendix. In the Irish case, we nd that the magnitude of the price e ect estimated using either the same1, same2 or pooled samples is unchanged from the all and all_nr samples (although its signi cance is diminished in the case of same1 and same2 due to the smaller sample size). However, in contrast to the all sample but in agreement with the all_nr estimations, the employment e ect disappears and is, therefore, likely to be a product of the refreshment sample in this data. The zero price e ect and the negative employment e ect observed in the UK are also robust to the two restirctive subsamples, same1 and same2 and to the larger sample, pooled. 6.3 Sample selection correction Buchinsky (1998), in a seminal paper, proposed a method to correct for selection bias in quantile regressions. However, this methodology has recently been called into question by Huber and Melly (2011) due to the assumption that the errors are independent of the regressors, implying that all quantile and mean functions are parallel. Distribution regression allows for a simpler, more intuitive selection correction, 15

17 Figure 7: Total e ect, price e ect and employment e ect of the NMW on the gender wage gap in the UK (with and without selection correction) Figure 8: Total e ect, price e ect and employment e ect of the NMW on the gender wage gap in Ireland (with and without selection correction) 16

18 although similar problems may be associated with it 9. To account for any selection bias engendered by the decision to select into work, we suggest a simple correction technique. We adapt the distribution regression methodology by running a sequence of Heckman selection models, rather than a sequence of probit models. The procedure is computationally heavy so con dence intervals are not presented, but the results are compared to our baseline to determine if the magnitude of the observed e ect changes after accounting for selection bias. The coe cients from the wage and selection equations at four points in the wage distribution: the NMW, the 25th, 50th and 75th percentiles, are shown in tables 10 and 9 in the Appendix. The magnitude and sign of the coe cients in the wage equation are comparable to those observed without selection correction in tables 7 and 6 and we use non-labour income, the presence of children and, for the UK, the partner s working status, to model the probability of working. In Ireland, we nd that correcting for selection bias gives virtually the same gender wage gap correction around the NMW, of up to 11ppt. The results for the UK are virtually identical with and without correction for selection bias. 7 Country comparisons 7.1 Ireland and the U.K. In line with the main results for the U.K. (Robinson, 2002 and Robinson 2005), we found no signi cant e ect of the British NMW on the gender wage gap. However, we did nd that the introduction of a NMW in neighbouring Ireland of IE$4:40 led to a reduction in the gender wage gap of up to 11ppt up to a wage level of around IE$5:20, where the gap is de ned as the di erence between a man and a woman s probability of earning below a certain wage, w. Given that this probability gap was between 10 12ppt before its introduction, the NMW was responsible for a 40 90% reduction in this measure of wage gaps, in the hourly wage region up to IE$5:20. Given this large e ect of the NMW on the Irish gender pay gap, why do we nd no e ect in the U.K.? One reason is the fact that the the gender wage gap in the UK is much smaller than that in Ireland at the bottom of the distribution. Figures 3 and 4 show that the gap in the UK is approximately half of that in Ireland at each country s respective NMW. This gap increases sharply in the wage region after the British NMW and is similar to the Irish gap in the rest of the distribution. This raises the question of what would have happened if a higher NMW had been set in the UK. A simple extrapolation exercise (results available upon request) suggests that even exporting the Irish NMW to the UK would have had a negligible e ect on the gender wage. The reason for this can be seen in gure??, where we provide a snapshot of the bottom of the wage distribution in both countries. While the Irish NMW had a large impact on the wage distribution in its vicinity, particularly the female wage distribution, the year after the introduction of the British NMW still saw a reasonably large proportion of people, mainly women, 9 A similar problem to that pointed out by Huber and Melly (2011), or indeed another one may apply to this adaptation of distribution regression, so these results are merely presented for illustration, and may not be entirely sound from an econometric viewpoint. Further work is needed to verify the robustness of this approach. 17

19 Figure 9: Predicted CDF s of male and female wages in the UK and Ireland before and after the NMW earning below the legal limit. So while F AF A has shifted downwards around the NMW level in the UK, it has not done so to the extent that it has in Ireland, nor indeed to the extent that we might reasonably expect, given the new wage legislation. In theory, there should be no one earning below this level, although measurement error, the black market and ine ective enforcement may account for some of the observations. To see how the British NMW would have a ected the gender wage gap if it had been as e ective as the Irish NMW at increasing the lowest wages, we suggest a second extrapolation exercise. We follow Chernozhukov et al (2009) in constructing the new counterfactual distributions of wages after the hypothetical implementation of this e ective NMW in the UK in In short, we take the proportion by which the conditional distribution of wages in Ireland is reduced at the Irish NMW after its introduction, and then reduce the conditional distribution of British wages before the introduction of the NMW by that same factor, up to the British NMW level. We do this seperately for men and women and construct the same summary measures for the estimation of the change in the gender wage gap as before. Denote m uk and m ie the British and Irish NMW s. We disregard the sub- and superscripts elaborated in (1) in order to generalise, except for n = b; a which indicates which sample (before or after) is in question. The new counterfactual marginal wage distributions are constructed as follows: Fa uk (w) = Fa uk (w) if w m uk (15) Fa uk (w) = Fb uk (w) P a ie (w < m ie ) Fb ie if y m uk (mie ) (16) Figure 10 shows that increasing the e ectiveness of the British NMW to the Irish level could result 18

20 in a narrowing of the gender wage gap of up to 6ppt; below the level of the NMW. Figure 10: E ect of British and Irish NMW s on distributions of wages in the UK To put this into the context of the literature, the two main studies relating to the UK nd ambiguous e ects of the NMW on the gender pay gap. Firstly, Robinson (2002) nds no evidence that the British NMW narrowed the gender wage gap at the bottom of the distribution, concluding that the eradication of gender wage inequality in the country will need to come from reducing the occupation and skills gap. In a seperate study in 2005, Robinson nds that the overall gender pay gap narrowed by around 1 2ppt in regions where the bite of the NMW was large and/or where women comprised a relatively large share of the low paid. However, at the bottom of the distribution, she reports that the NMW worked in favour of men who, on average, were paid much further below the minimum than women. We can now add to this debate with evidence that the negligible e ect of the British NMW on the gender wage gap was also due to the disproportionate number of women still earning less than the legal threshold after its introduction in the UK. 7.2 Other countries Our results for Ireland are consistent with ndings in the broader literature. Chernozhukov et al (2009), using distribution regression, nd that most, if not all of the increase in wage inequality in the bottom half of the U.S. wage distribution between 1979 and 1988 was due to the declining value of the real NMW and that this increase was larger for women (33%) than for men (11%). Their results con rm what DiNardo et al (1996) had found for the same period in the U.S. Ganguli & Terrell (2005) nd that the introduction of a NMW in the Ukraine decreased the coe cient gap between men and women by around 35% in the rst decile group between 1991 and its introduction in 2003, but increased it further up in 19

21 the distribution. 8 Conclusion National minimuim wages can be contraversial tools for redistribution but, as we have shown in this paper, they can be e ective if implemented properly. Using a new methodology, we tackle the issue of the e ect of NMW s on the gender wage gap. We nd strong evidence that the NMW can reduce the gender wage gap at the bottom of the distribution, with some spillover further up in the distribution. However, we nd discrepencies in how e ective a tool it is in di erent countries. 9 Bibliography References [1] Albrecht, J., A. Van Vuuren. & S. Vroman, "Counterfactual distributions with sample selection adjustments: Econometric theory and an application to the Netherlands," Labour Economics, Elsevier, vol. 16(4), pages , August. [2] Arulampalam, W., A. Booth & M. Bryan, "Is There a Glass Ceiling over Europe? Exploring the Gender Pay Gap across the Wage Distribution," Industrial and Labor Relations Review, ILR Review, ILR School, Cornell University, vol. 60(2), pages , January. [3] Bassett, G. & R. Koenker, Regression Quantiles, Econometrica. January, 46:1, pp [4] Blau, F. & L. Kahn, 1997, "The Impact of Wage Structure on Trends in U.S. Gender Wage Di erentials ," NBER Working Papers 4748, National Bureau of Economic Research, Inc. [5] Blau, F. & L. Kahn, "Understanding International Di erences in the Gender Pay Gap," Journal of Labor Economics, University of Chicago Press, vol. 21(1), pages , January. [6] Bonjour, D. & M. Ger n, "The unequal distribution of unequal pay - An empirical analysis of the gender wage gap in Switzerland," Empirical Economics, Springer, vol. 26(2), pages [7] Brewer, M. & J. Browne, "The e ect of the working families tax credit on labour market participation," Open Access publications from University College London University College London. [8] Buchinsky, M., "Changes in the U.S. Wage Structure : Application of Quantile Regression," Econometrica, Econometric Society, vol. 62(2), pages , March. [9] Buchinsky, M., "The dynamics of changes in the female wage distribution in the USA: a quantile regression approach," Journal of Applied Econometrics, John Wiley & Sons, Ltd., vol. 13(1), pages

22 [10] Card, D. & A. Krueger, "Minimum Wages and Employment: A Case Study of the Fast- Food Industry in New Jersey and Pennsylvania," American Economic Review, American Economic Association, vol. 84(4), pages , September. [11] Chernozhukov, V., I. Fernandez-Val & B. Melly, "Inference on counterfactual distributions," CeMMAP working papers CWP09/09, Centre for Microdata Methods and Practice, Institute for Fiscal Studies. [12] Dickens, R., S. Machin & A. Manning, "The E ects of Minimum Wages on Employment: Theory and Evidence from Britain," Journal of Labor Economics, University of Chicago Press, vol. 17(1), pages 1-22, January. [13] Dickens, R. & A. Manning, "The Impact of the National Minimum Wage on the Wage Distribution in a Low-Wage Sector," Royal Economic Society Annual Conference , Royal Economic Society. [14] Dickens, R., R. Riley & D. Wilkinson, 2011, "The UK Minimum Wage at Age 22: A Regression Discontinuity Approach", Working Paper Series, Department of Economics, University of Sussex. [15] DiNardo, J., N. Fortin & T. Lemieux, "Labor Market Institutions and the Distribution of Wages, : A Semiparametric Approach," Econometrica, Econometric Society, vol. 64(5), pages , September. [16] Dolado, J. & V. Llorens, "Gender Wage Gaps by Education in Spain: Glass Floors versus Glass Ceilings," CEPR Discussion Papers 4203, C.E.P.R. Discussion Papers. [17] Dolton, P., C. Rosazza-Bondibene & J. Wadsworth, "Employment, Inequality and the UK National Minimum Wage over the Medium-Term," IZA Discussion Papers 5278, Institute for the Study of Labor (IZA). [18] Foresi, S. & F. Peracchi, "The Conditional Distribution of Excess Returns: An Empirical Analysis," Journal of the American Statistical Association Vol. 90, No. 430 (Jun., 1995), pp [19] Fortin, N. & T. Lemieux, "Rank Regressions, Wage Distributions, and the Gender Gap," Journal of Human Resources, University of Wisconsin Press, vol. 33(3), pages [20] Ganguli, I. & K. Terrell, "Institutions, markets and men s and women s wage inequality: Evidence from Ukraine," Journal of Comparative Economics, Elsevier, vol. 34(2), pages , June. [21] Gronau, R., "Wage Comparisons-A Selectivity Bias," Journal of Political Economy, University of Chicago Press, vol. 82(6), pages , Nov.-Dec. [22] Datta Gupta, N., R. Oaxaca, L. Ronald & N. Smith (2006) "Swimming Upstream, Floating Downstream: Comparing Women s Relative Wage Progress in the United States and Denmark," Industrial & Labor Relations Review, Vol. 59, No. 2, article 4. 21

23 [23] Heckman, J., "Sample Selection Bias as a Speci cation Error," Econometrica, Econometric Society, vol. 47(1), pages , January. [24] Huber, M. & B. Melly, "Quantile Regression in the Presence of Sample Selection," Economics Working Paper Series 1109, University of St. Gallen, School of Economics and Political Science. [25] Ichimura, H., "Semiparametric least squares (SLS) and weighted SLS estimation of single index models," Journal of Econometrics 58, [26] Jenkins, S., "Earnings discrimination measurement : A distributional approach," Journal of Econometrics, Elsevier, vol. 61(1), pages , March. [27] Mata, J. & J. Machado, "Counterfactual decomposition of changes in wage distributions using quantile regression," Journal of Applied Econometrics, John Wiley & Sons, Ltd., vol. 20(4), pages [28] Machin, S. & A. Manning "The E ects of Minimum Wages on Wage Dispersion and Employment: Evidence from the U.K. Wages Councils," Industrial and Labor Relations Review, Vol. 47, No. 2 (Jan., 1994), pp [29] McGinnity, F & H. Russell, "Gender Inequalities in Time Use: The Distribution of Caring, Housework and Employment Among Women and Men in Ireland," Dublin, Ireland: The Equality Authority. [30] Metcalf, D., "The Low Pay Commission and the National Minimum Wage," The Economic Journal, 109: doi: / [31] Metcalf, D., "Why has the British National Minimum Wage had Little or No Impact on Employment?," Journal of Industrial Relations June : [32] Neumark, D., W. Cunningham & L. Siga, "The e ects of the minimum wage in Brazil on the distribution of family incomes: ," Journal of Development Economics, Elsevier, vol. 80(1), pages , June. [33] O Neill, J. & S. Polachek, 1993, "Why the Gender Gap in Wages Narrowed in the 1980s," Journal of Labor Economics, University of Chicago Press, vol. 11(1), pages , January. [34] O Neill, D., B. Nolan & J. Williams, "Evaluating the Introduction of a National Minimum Wage: Evidence from a New Survey of Firms in Ireland." Labour: Review of Labour Economics and Industrial Relations, Vol. 20, No. 1, pp , March [35] Robinson, H., " Wrong Side of the Track? The Impact of the Minimum Wage on Gender Pay Gaps in Britain," Oxford Bulletin of Economics and Statistics, Department of Economics, University of Oxford, vol. 64(5), pages , December. 22

24 [36] Robinson, H., "Regional evidence on the e ect of the national minimum wage on the gender pay gap," Regional Studies, Taylor and Francis Journals, vol. 39(7), pages , October. 10 Appendix Table 3: Variable means for same1 sample (men and women observed before and after the introduction of the NMW in the UK and Ireland who work at least part-time in at least one period) 23

25 Figure 11: Actual V Predicted CDF s of hourly wages for Ireland 24

26 Figure 12: Actual V Predicted CDF s of hourly wages for the UK 25

27 Table 4: Variable means for same2 sample (men and women observed before and after the introduction of the NMW in the UK and Ireland who work at least part-time in at least one period)) 26

28 Table 5: Variable means for pooled sample (men and women observed before and after the introduction of the NMW in the UK and Ireland who work at least part-time in at least one period)) 27

29 Table 6: Coe cients of distribution regression of hourly wage at the NMW, the 25th, 50th and 75th percentiles in Ireland)) 28

30 Table 7: Coe cients of distribution regression of hourly wage at the NMW, the 25th, 50th and 75th percentiles in the UK)) 29

31 Table 8: Decompostion of wage di erentials into the endowment and coe cient gaps using distribution regression and the Oaxaca-Blinder technique 30

32 Table 9: Coe cients of distribution regression of hourly wage with selection correction at the NMW, the 25th, 50th and 75th percentiles in Ireland Figure 13: Total e ect, price e ect and employment e ect of the NMW on the gender wage gap in Ireland (all_nr sample) 31

33 Table 10: Coe cients of distribution regression of hourly wage with selection correction at the NMW, the 25th, 50th and 75th percentiles in the UK 32

34 Figure 14: Total e ect, price e ect and employment e ect of the NMW on the gender wage gap in Ireland (same1 sample) Figure 15: Total e ect, price e ect and employment e ect of the NMW on the gender wage gap in Ireland (same2 sample) 33