Introduction

Beginning with Badgett (1995), researchers have accumulated evidence of disparities in the labor market outcomes between homosexuals and heterosexuals. Research has found gay men are paid less than heterosexual men (Klawitter 2015). The evidence of a gay wage gap exists across different datasets and is robust to various methods for identifying who is gay (Klawitter 2015).Footnote 1 There is inconclusive evidence of wage differentials for lesbian women, with differences in fertility and selection into the labor market potentially explaining the differences (Klawitter 2015).Footnote 2 In addition to the evidence of disparities in pay, there is consistent evidence from resume correspondence studies that heterosexuals are preferred by hiring managers to homosexuals (Bailey et al. 2013; Mishel 2016; Tilcsik 2011).

Historically, the policy implemented in the United States to reduce these disparities has been to make it illegal for an employer to discriminate against individuals based on their membership in a protected group. When employment non-discrimination acts work as intended, the labor market outcomes of the protected group gradually improve, as appears to have happened for black men (Collins 2003; Donohue and Heckman 1991; Landes 1968; Neumark and Stock 2006). In other cases, however, the increased protections may make the protected group relatively more expensive to hire and terminate. Therefore, employers may reduce the number of employees they hire from the protected group (Bloch 1994), as may have happened for women and older workers (Beegle and Stock 2003; Lahey 2008; Neumark and Stock 2006).Footnote 3

Despite an increasing number of states passing laws to protect gay men and lesbian women, there is little research on how effective these policies have been at reducing disparities. The research on LGBT non-discrimination acts has found mixed evidence for the effectiveness of employment protections at the state level (Baumle and Poston 2011; Klawitter and Flatt 1998, 2011; Martell 2013).Footnote 4 In each of these papers, the authors assumed that all laws were identical, but there evidence of wide variation in how these state laws are constructed. When comparing state laws, the differences appear on three issues: who is protected under the law, how a complaint is resolved, and what damages and remedies are available for plaintiffs. Within these three groups of differences, there are thirteen provisions over which states differ. Table 1 details the areas where the laws diverge. In this paper, I focus on damage availabilityFootnote 5, employer size minimumsFootnote 6, and the statute of limitationsFootnote 7.

Table 1 States with each legal provision in its ENDA
Full size table

These legal differences are important because they determine the expected cost of discriminating for employers. The expected cost of discriminating can be increased either by increasing the probability that an employer is sued or by increasing the cost of being sued. The probability that an employer is sued potentially increases when employees are given longer to file a complaint and they can recoup their attorney’s fees in a successful lawsuit (which lowers the cost of a lawsuit). Stronger damage provisions increase the expected cost of discriminating since employers who do discriminate face stiffer penalties if caught. Despite being limited to studying the effect of a state law being stronger than the federal law, previous work on non-discrimination laws has shown that heterogeneity in the laws can impact the effect.Footnote 8 Jolls and Prescott (2004) used state-level variation in disability discrimination laws to show that the negative employment effects of the Americans with Disabilities Act (ADA) were primarily due to “reasonable accommodations” requirements and not firing costs. Neumark and Button (2014) showed that stronger state-level laws against age discrimination might have reduced the hiring of older workers during the Great Recession.

In this paper, I use a differences-in-differences-in-differences methodology to identify how the passage of ENDAs impacted the labor market outcomes of gay men and lesbian women. Since ENDAs are designed to benefit gay men and lesbian women, the first difference is a comparison of the labor market outcomes of gay men and lesbian women relative to gay men and lesbian women in states that have not passed a law. Because there may be non-ENDA related factors that affect wages and employment, I use the difference in labor market outcomes between gay men and lesbians and heterosexuals in the same state to account for changes in these differentials that are unrelated to the passage of an ENDA. The triple difference estimation results in a comparison of how the labor market outcomes between heterosexuals and homosexuals in a state changed after the passage of an ENDA relative to the differences in states that did not pass an ENDA. My results indicate that after an ENDA had been passed, hourly wages for cohabiting gay men rose 2.7%. There was no significant effect on annual income from wages, employment, or hours worked for gay men. For lesbian women, the results are less positive. The passage of an ENDA had no significant effect on the wage differentials for lesbian women, but was associated with a 1.7% decline in the employment and a 0.733-hour decline in hours worked.

The results also show the benefits of the law were dependent on the structure of the law. For gay men in states where successful plaintiffs cannot be awarded damages, there was no effect of an ENDA. Gay men in states that only allow compensatory damages experienced an increase in annual wages of 12.3% and an increase in hourly wages of 15.5%. Gay men in states that allow both compensatory damages and punitive damages experienced an increase in annual wages of 7.9% and an increase in hourly wages of 7.4%. In states with longer complaint periods, I find small increases in the employment of gay men. After controlling for the differences in provisions, the average ENDA decreased the annual income of lesbian women by 11.4%. Allowing for punitive damages decreased the hours worked for lesbian women by an additional 0.850 hours.

Data

The data used in this paper comes from the 2008 through 2014 American Community Survey (ACS) 1-Year Samples, the 1990 U.S. Census 5% Sample, and the 2000 U.S. Census 5% Sample (Ruggles et al. 2010). A longitudinal database of all state laws was created using information from The Williams Institute at UCLA, the Government Accountability Office, and state laws (Government Accountability Office 2013; Sears et al. 2009; Sears and Mallory 2011).

To identify gay men and lesbian women in the United States, the Census collects information on householders and the relationships of everyone in the household to the householder. A same-sex couple is identified when the gender of the householder and the gender of the unmarried partner (or spouse) of the householder are the same. There is no information on single gay men and lesbian women in the Census data or ACS, only cohabiting gay men and lesbian women. Also missing from the sample are gay men and lesbian women in a household where one of the partners is not the household head (such as living with one’s parents). Therefore, the sample in the analysis is restricted to comparisons between cohabiting individuals.

The sample used in the analysis begins with all adults older than 18 who claim to be the householder, spouse, or unmarried partner. In this paper, I focus on comparisons between cohabiting same-sex couples and married heterosexual couples. Cohabiting gay men and lesbian women are identified in the sample if they are cohabiting with an individual of the same gender.Footnote 9 Once I identify cohabiting same-sex couples, I restrict the sample to individuals over the age of 22 and under the age of 65. The age of 22 is selected to avoid school-aged individuals.

In the data, heterosexual couples are miscoded as same-sex cohabiting couples if the sex of one of the individuals is miscoded. Even though miscoding of sex is one of the least common errors made on Census forms, due to the small size of the gay and lesbian population, any miscoding in the heterosexual sample has the potential for increasing the number of gay men and lesbian women in a significant way (O’Connell and Golding 2006). When looking at the Census forms, researchers have that the miscoding is concentrated in the group of individuals who claimed to be married on their Census form and found smaller amounts of miscoding in the sample of same-sex couples who claim to be unmarried partners (O’Connell and Golding 2006; O’Connell and Loftquist 2009). Work on the wording of relationship questions in the ACS resulted in a large decline in the number of miscodings in the 2008 and subsequent surveys. Because before 2004 it was not possible for same-sex couples to be married, I exclude any gay or lesbian that had their marital status recoded in 2000 from married to cohabiting. This is a conservative portioning of the cohabiting gay and lesbian population, but it reduces the measurement error.

The other concern is that the passage of ENDAs laws may induce gay men and lesbian women to move to these states. Previous research has found mixed evidence that gay men and lesbian women migrate in response to the passage of pro-gay laws. Ueno et al. (2014) found that gay men were no more likely than heterosexual men to migrate to a different state, but lesbian women are more likely than heterosexual women to move. When gay men and lesbian women do move, there is inconclusive evidence that gay rights laws influence their choices. Colvin and Riccucci (2002) found no evidence that when they migrated that cohabiting gay men and lesbian women were more likely to move to a state that had passed an ENDA. Beaudin (2017) found evidence that cohabiting gay men and lesbian women were more likely than cohabiting heterosexuals to move to states with marriage equality. Because Colvin and Riccucci (2002) was estimated using a five-year window (while Beaudin (2017) used a one-year window), the results may not be as useful in evaluating the risk of current migration biasing the results. To avoid the migration problem completely, I remove any individual who has migrated within the past year from the ACS data.Footnote 10 This means that whenever possible (i.e., the latter half of the sample), the effect of an ENDA will only be identified off of the wages of individuals who have been in the state for more than a year.

The result of these restrictions on the data is that I am left with a sample of 6,128,130 heterosexual men and 44,143 gay men for whom I observe labor market outcomes. Within the sample of gay men, 26,340 live in states with no employment protections, and 17,803 live in a state with employment protections. For women, there are a total of 5,016,785 heterosexual women and 44,275 lesbian women. There are 26,170 lesbian women in states with no employment protections and 18,105 lesbian women in states with employment protections.

Information about provisions comes from state laws and reports compiled by the Williams Institute and the Government Accountability Office (Sears et al. 2009; Government Accountability Office 2013).Footnote 11 Wisconsin passed employment protections for gay men and lesbian women in 1982. In 1990, only Wisconsin, Massachusetts, and D.C. had employment protections for gay men and lesbian women. By 2000, the number of states had grown to twelve. In 2014, twenty-one states and Washington, D.C. had employment non-discrimination acts protecting gay men and lesbian women. As shown in Fig. 1, the Northeast, Midwest, and the West feature the highest concentration of these laws.Footnote 12 Table 1 highlights how different each law is by comparing the frequency of different legal provisions. There does not appear to be a pattern between the strength of the law and the geographic distribution of states that have passed them.

Fig. 1
figure 1

Note Data on state and local laws comes from Government Accountability Office (1997), Government Accountability Office (2013), Human Rights Campaign (2012), and Sears et al. (2009), and by reading the state laws themselves

Map of ENDA Laws in 2011

Full size image

Estimation strategy

The goal of this paper is to estimate the impact of employment non-discrimination acts on the labor market outcomes of cohabiting gay men. I focus on the relationship between the strength of the laws and the impact of the laws. Because the laws in question vary by state over time, I use the state and year variation to estimate a differences-in-differences-in-difference model. Since the passage of an ENDA does not impact heterosexuals, I can use the comparison of gay men and lesbian women and their married heterosexual counterparts to isolate the effect these laws have on the labor market outcomes of cohabiting gay men and lesbian women. I use a flexible model, saturated with year, state, and state-by-year fixed effects interacted with being homosexual:

$$ \begin{array}{cc} Y_{ist}=\beta_{0}+\beta_{1}Homo_{ist}+\beta_{2}(Homo_{ist} \times ENDA_{st})+I_{s}\gamma_{s} +I_{t}\gamma_{t}\\+(I_{s}\times I_{t})\gamma_{st}+(Homo_{ist}\times I_{s})\theta_{H,s}+(Homo_{ist}\times I_{t})\theta_{H,t}+\textbf{X}'_{ist}\delta +\epsilon_{ist} \end{array} $$
(1)

Y ist is the dependent variable of interest. In this paper, I look at log annual income from wages, log hourly wages, the probability of being employed, and hours worked. Homoist is a dummy for being a cohabiting gay man or lesbian woman. ENDAst is a dummy for state s having an employment non-discrimination act that protected gay men and lesbian women in year t. The vector X contains controls for demographic, occupation, and geographic variables that may affect wages or employment. See Appendix Table 9 for the full list of control variables.

Also included in the regression are state, year, and state-by-year fixed effects (γs, γt, and γst). These fixed effects will capture unobserved differences common to all observation in a state, year, and state-by-year cell. To account for the differences between cohabiting homosexuals and married heterosexuals that exist across states, state fixed effects are interacted with the dummy for cohabiting same-sex couples (θH,s). To account for common trends that affect cohabiting same-sex couples differently than heterosexual couples (e.g. declines in prejudice, business cycle fluctuations, etc.), year fixed effects are interacted with the dummy for cohabiting same-sex couples (θH,t). Due to individual preferences potentially being correlated within a state and treatment occurring at the state-level, the standard errors (𝜖) are clustered at the state-level.

I estimate the models for men and women separately. In the analysis of wages (both annual and hourly), the sample was restricted to individuals in the labor force.Footnote 13 For the analysis of employment and hours worked per week, the sample of all adults is used.

In Eq. 1, the differential effect of being a cohabiting same-sex couple in the United States across all years is captured by β1. How this differential changes over time is captured by θH,t. The state fixed effects for cohabiting same-sex couples (θH,s) capture how the wage differential varies across states. The parameter of interest is β2, which captures how the differential between homosexuals and heterosexuals in states with ENDAs changed after the passage of a law.

To interpret β2 as being the causal effect of enacting employment protections for sexual orientation, it must be the case that there are no other factors related to changes in the gay-heterosexual wage differential in states that do and do not have LGBT employment protections. I test these assumptions in “Robustness Checks and Threats to Validity”. In addition to testing the assumptions of the DDD model, I test how robust the results are to confounding factors, specifically same-sex marriage and selection into the labor market. Results of the robustness checks are also reported in “Robustness Checks and Threats to Validity”.

The main contribution of this paper is to test how the heterogeneity of the law affects the impact of these laws. To do this, I add a series of controls in Eq. 1 to capture policy differences in the state law. I estimate the following equation:

$$ \begin{array}{ll} Y_{ist}=\beta_{0}+\beta_{1}Homo_{ist}+\beta_{2}(Homo_{ist} \times ENDA_{st})+\beta_{p}(Homo_{ist} \times ENDA_{st} \times P_{s,t})\\+\textbf{X}'_{ist}\delta +I_{s}\gamma_{s} +I_{t}\gamma_{t}+(I_{s}\times I_{t})\gamma_{st}+(Homo_{ist}\times I_{s})\theta_{H,s}+(Homo_{ist}\times I_{t})\theta_{H,t}+\epsilon_{ist} \end{array} $$
(2)

where Ps,t is a vector of policy variables. States rarely change their non-discrimination laws, but there have been amendments or court cases that resulted in changes to the laws.Footnote 14 The policy variables used here are the availability of damages (compensatory and punitive), the statute of limitations for complaints, the employer size minimum, and the ability to recoup attorney’s fees. The availability of damages and the ability to recoup attorney’s fees enter into Eq. 2 as dummy variables. The statute of limitations for complaints and employer size minimums have been normalized to the average value to make interpreting the coefficients easier. The average complaint period in a state with an ENDA is 245 days. State complaint periods are coded as the number of months relative to the average. The average size of firm size minimum for being covered by a law is five employees. For states with an ENDA, size minimums are coded as the employer size minimum minus five.

Main Results

In this section, I begin by discussing the effect of ENDAs on gay men and then move to discussing the results for lesbian women. For both groups, I first look at the average effect of the law before comparing the effects of different legal provisions. I focus on four labor market outcomes: log annual income from wages, log hourly wages, employment, and hours worked per week. ENDAs may increase the wages of gay men and lesbian women if the laws force employers to pay them the same wages they pay their heterosexual employees. These laws can also increase the employment of gay men and lesbian women if they make it easier for them to find a job. They may increase the hours worked if they allow gay men and lesbian women to obtain full-time employment or find a second job.

Table 2 reports the wage differential between heterosexuals and homosexuals by year in the sample.Footnote 15 Similar to early research using the 1990 Census (Allegretto and Arthur 2001), I find gay men earned 12.6% less than a comparable heterosexual man in 1990, but this penalty declined by 4.7 percentage points between 1990 and 2014. The decline has been most pronounced in states with an ENDA. States with an ENDA saw a decline of 10.2 percentage points and states with no ENDA saw an increase of 0.2 percentage points. Lesbian women earned 2.2% more than a comparable heterosexual woman, and this has not changed much since 1990. States with ENDAs and without ENDAs did not see as stark differences in the wage differentials as was observed for gay men. Given that the initial conditions for gay men and lesbian women are different, then we expect there to be different effects for the two groups. If the wage differential is due to discrimination, then gay men would experience an increase in wages after the passage of a law and lesbians would experience a wage decline. If the wage premium that lesbian women experience is due to unobserved differences that are not discrimination, then we would expect the laws to have zero effect for gay men or lesbians. Since before the law lesbian women are favored and gay men are disfavored, it is not clear a priori whether we should expect the laws effects to be symmetric across the two groups.

Table 2 Gay and lesbian wage differentials by year
Full size table

I begin by looking at the effect of ENDAs on cohabiting gay men in Table 3. In Panel A of Table 3, I replicate what the previous literature has done and use a single dummy variable for having an ENDA. The first column of Table 3 reports the results for log annual income, the second column reports the results for log hourly wages, the third column reports the results for employment, and the fourth column reports the results for usual hours worked per week. I find no effect of passing an ENDA on the log annual income of gay men. There was, however, a positive effect of passing an ENDA on log hourly wages. Between 1990 and 2014, the passage of an ENDA increased the hourly wages of cohabiting gay men by 2.7%. This increase in hourly wages was statistically significant at the 5% level. Similar to results found in the previous literature, the passage of an ENDA had no significant effect on the employment of cohabiting gay men.Footnote 16 I also find no statistically significant effect of the passage of an ENDA on hours worked per week.Footnote 17

Table 3 Effects of ENDAs on labor market outcomes of cohabiting gay men
Full size table

The results in Panel A suggest the declines in the wage penalty for gay men observed in Table 2 in states that passed an ENDA are due in part to the passage of these laws. In states with an ENDA, the wage penalty for gay men declined 10.2 percentage points between 1990 and 2014. The passage of ENDAs can explain 26% of this decline (2.7% divided by 10.2%).

Panel B in Table 3 reports the effect each policy had on outcomes. Among states with ENDAs, the average size of firm size minimum for being covered by the law is five employees. Decreasing the firm size minimum by one employee relative to the average minimum will increase the number of firms that are covered by the law. For all the outcomes studied here, I find no significant effect of smaller firm size minimum. I find similar results for the complaint period. When I compare the effects of the complaint period across the different outcomes, I find that a longer statute of limitations only has a significant effect on employment. Increasing the complaint period by 1-month above the average increases the employment of gay men by 0.2%.

Where I find the strongest effects is in damages. Damages have a significant effect on both annual income and hourly wages. In states that allow compensatory damages, the annual income of gay men increase by 9.4%, and the hourly wages of gay men increase by 12.7%. When a state allows for punitive damages, the annual income of gay men falls by 4.4%, and the hourly wages of gay men fall by 8.1%. When looking at the damages provisions together, there is a distinct pattern of decreasing returns to strength. Providing for damages is important to the increase in earnings of gay men since it is a credible message to discriminating firms that there is a cost to engaging in discriminatory behavior. But, firms increase the wage of gay men less as the cost of discriminating increases. In states that allow only compensatory damages, there is a 12.8% net increase in annual wages. In states that allow both compensatory damages and punitive damages, there is an 8.8% net increase in annual wages. A similar pattern exists for hourly wages.

Table 4 shows the average effect of ENDAs for lesbian women appears similar to the effects found in earlier research on non-discrimination laws for women (Neumark and Stock 2006). In Panel A of Table 4, I show there was no increase in annual income or wages after the passage of the ENDA, consistent with what earlier work on ENDAs found for lesbian women (Klawitter and Flatt 1998; Klawitter 2011). My results for the employment effects for lesbian women show that ENDAs were detrimental to the employment of lesbian women, where Klawitter and Flatt (1998) and Klawitter (2011) found no effect. There was a 1.7% decline in employment after an ENDA is passed. I also find a 0.733-hour decline in hours worked for lesbian women.

Table 4 Effects of ENDAs on labor market outcomes of cohabiting lesbian women
Full size table

The second panel of Table 4 explores how the legal provisions influence the effect of an ENDA for lesbian women. By breaking out the effect of an ENDA by provision, I show that for the most part there is no effect of the provisions on the wages or employment of lesbian women. There are two exceptions. First, after controlling for the differences in provisions, the average ENDA now has the effect of decreasing the annual income of lesbian women by 11.4%. The specific provisions do not appear to mitigate this effect, suggesting this result is driven by correlations amongst the provisions. Second, I find that stronger damages may further decrease the employment of lesbian women. Allowing for punitive damages decreases the hours worked for lesbian women by an additional 0.850 hours.

These costs and benefits of ENDAs for gay men and lesbian women are not spread evenly across the labor force. In Table 5, I compare the effect of ENDAs on wages for those working full-time and part-time.Footnote 18 I find significant differences in how ENDAs impacted full-time and part-time workers. For gay men who were working more than 30 hours a week, ENDAs increased their wages by 3.0%-3.2%. Gay men who were working part-time (less than 30 hours a week) did not experience a wage increase after an ENDA was passed. For lesbian women, the pattern is flipped. There is weak evidence that lesbian women working part-time may have received wage increases. I find a 12.0% increase in hourly wages of lesbian women working part-time, but no significant effect on annual income from wages. The effects for full-time lesbian women are negative for both groups, but not significant at the 5% level.

Table 5 Effect of ENDA on earnings by full-time and part-time employment status
Full size table

Robustness Checks and Threats to Validity

So far the results have shown increases in the wages of gay men and declines in employment and hours worked for lesbian women as a result of passing an ENDA. This next section addresses potential threats to the validity of these results.

The differences-in-differences-in-differences methodology rests on the assumption that the unobservables are uncorrelated with the treatment. Error terms of the wage equation may not be parallel if the level of discrimination is changing faster in states with employment protections than in states without employment protections. Using responses from the General Social Survey, it is possible to calculate the percent of individuals that express prejudiced sentiments to questions about homosexuality. Research has shown that the wages of gay men are correlated with the share of individuals in a state who give prejudiced answers to questions about homosexuality in the General Social Survey (Pride 2017).Footnote 19

There is no publicly available data on prejudice at the state level, so the publicly available Census division level data is used. In the General Social Survey, there are four questions about homosexuality asked in every wave. Table 6 details the text of each question and the possible answers. The questions in the GSS ask a respondent’s feelings about sexual relations between adults of the same gender, whether they support homosexuals teaching in colleges, whether they support books promoting homosexuality to be housed in public libraries, and whether homosexuals should be able to give speeches in favor of homosexuality in public. For every Census division, I calculate the share of individuals giving the most prejudiced answer to all the questions. Since the General Social Survey is asked every two years, I impute the odd years using the mean of the preceding and succeeding shares of prejudiced individuals. I then estimate the rate at which states that pass an ENDA are growing less prejudiced relative to other states.

Table 6 Questions from the general social survey
Full size table

To test the pre-trends, I regress the share of GSS respondents giving all prejudiced answers (Prejudices,t) on dummies for the year relative to the passage of an ENDA (τt), where (τ0) is equal to the year of passage and is the omitted category in the regression. I also include state and year fixed effects (δs and δt).

$$ Prejudice_{s,t}=\beta+\sum\limits_{t=-5}^{-1}\tau_{t}+\sum\limits_{t = 1}^{5}\tau_{t}+\delta_{t}I_{t}+\delta_{s}I_{s}+\epsilon_{s,t} $$
(3)

Figure 2 shows how the pre-trends in prejudice are evolving relative to the passage of an ENDA. I find no evidence states that pass an ENDA experience faster declines in prejudice. The share of individuals expressing prejudice against the LGBT community rises slightly four and five years after the passage of an ENDA, suggesting the passage of ENDAs may result in a small backlash against the LGBT community.Footnote 20

Fig. 2
figure 2

Note: Data comes from the 1990 through 2014 General Social Surveys. Prejudice is calculated as the percent of individuals in a Census division that give the most prejudice answers to questions regarding homosexuality in the General Social Survey. See Table 6 for the questions asked and the possible answers

Change in Prejudice Relative to Year of Passage for ENDAs

Full size image

The second issue potentially biasing the results is that the estimates for log hourly wages are conditional on being in the labor market. It is possible that selection into the labor market is not random. Note however that for the selection correction to an effect on the estimation of β2 the selection must occur differently in states with and without an ENDA. In the data, there appears to be little evidence of this. For both gay men and lesbian women, the rate of selection does not vary much across states with different laws. Neither does the rate of selection into the labor market appear to vary by gender. Approximately, 82% of gay men and lesbian women report earnings and 18% do not. The differences between homosexuals and heterosexuals are equally stable. The share of gay men reporting they did not work the previous year is 3% higher than married heterosexual men in states with an ENDA, in states without an ENDA the gap is 2%. In states with an ENDA, lesbian women are 14% more likely to report wages, and in states with no ENDA the gap is also approximately 14%. Therefore, while β1 is likely to be sensitive to controlling for selection (especially amongst women), the effects on β2 will be small.

To correct the estimates for selection, I use a semi-parametric estimation strategy. I regress the indicator for being employed (Eist) on all of the controls used in the baseline estimation.Footnote 21 I also include additional controls: the number of children an individual has (Kidsi,s,t) and a dummy for whether any of those children are under the age of two (Y oungi,s,t).

$$ \begin{array}{lc} &E_{ist}=\beta_{0}+\beta_{1}Homo_{ist}+\beta_{2}(Homo_{ist} \times ENDA_{st})+I_{s}\gamma_{s} +I_{t}\gamma_{t}\\&+(I_{s}\times I_{t})\gamma_{st}+(Homo_{ist}\times I_{s})\theta_{H,s}+(Homo_{ist}\times I_{t})\theta_{H,t}+\textbf{X}'_{ist}\delta\\&~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~+\alpha_{1}Kids_{i,s,t}+\alpha_{2}Young_{i,s,t} +\epsilon_{ist} \end{array} $$
(4)

Using the estimated coefficients from Eq. 4, I estimate the predicted probability that an individual would be employed (ρ). I use a fifth order polynomial of this predicted probability as a control in the wage regression to control for selection into the labor market (5).

$$ \begin{array}{lc} &Y_{ist}=\beta_{0}+\beta_{1}Homo_{ist}+\beta_{2}(Homo_{ist} \times ENDA_{st})+I_{s}\gamma_{s} +I_{t}\gamma_{t}\\&+(I_{s}\times I_{t})\gamma_{st}+(Homo_{ist}\times I_{s})\theta_{H,s}+(Homo_{ist}\times I_{t})\theta_{H,t}+\textbf{X}'_{ist}\delta\\&~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~+\rho_{i,s,t}+\rho_{i,s,t}^{2}+\rho_{i,s,t}^{3}+\rho_{i,s,t}^{4}+\rho_{i,s,t}^{5} +\epsilon_{ist} \end{array} $$
(5)

I show in Table 7 that selection did not have a significant effect on the results.Footnote 22 In column 1, I report the results for the average effect of an ENDA for gay men. The results do not appear to have been driven by selection into the labor force. The effect of an ENDA conditional on selection into employment remains similar to what it was in the baseline estimation. The effect on log annual income does not change if one controls for selection. The effect of an ENDA on log hourly wages increases slightly to 2.9%.Footnote 23 The results for lesbian women remain the same.

Table 7 Robustness of results on wages to controlling for selection into employment
Full size table

The third potential concern is that at the same time that states were passing ENDAs, some states were also granting the right of same-sex marriage to gay men and lesbian women. It is possible that the passage of same-sex marriage increases the wages of gay men and lesbian women, though the previous literature has not found there is a wage premium associated with cohabiting for gay men (Zavodny 2007). Table 8 reports the results controlling for the passage of same-sex marriage. The top panel of Table 8 shows the results for men. After controlling for the passage of same-sex marriage, the average effect of an ENDA remains relatively unchanged at 2.8%. The results for employment and hours worked are still not statistically significant. The bottom panel of Table 8 reports the results for women. The passage of same-sex marriage similarly does not change the results in a significant way. There is still a negative and significant effect of ENDAs on employment and hours worked for lesbian women.Footnote 24

Table 8 Robustness of results to controlling for same-sex marriage
Full size table

Conclusion

In this paper, I explored the effect that the passage of ENDAs at the state-level had on the labor market outcomes of gay men and lesbian women. The results showed that ENDAs led to a decline in the wage gap between gay men and married heterosexual men, but also reduced the employment of lesbian women. The fact that gay men would benefit from the passage of an ENDA and lesbian women would suffer from the passage of an ENDA can be interpreted as employers reducing the importance of sexual orientation in employment decisions. Because lesbian women may be favored over heterosexual women, the passage of an ENDA reduces the benefits they experience and increases the employment prospects of married heterosexual women. The passage of an ENDA increases the employment prospects of gay men relative to heterosexual men.

When I treat all ENDAs as identical as was done in the previous literature, I find ENDAs increased the wages of gay men by 2.7% and had no effect on employment and hours worked per week. This is similar to the effects in Baumle and Poston (2011), which found ENDAs increased wages by 2.6%. The results are also similar to the magnitudes found in the General Social Survey by Martell (2013).

The key contribution of this paper is that I show the effects of ENDAs for gay men are concentrated in a handful states. The evidence suggests that strong and weak laws had different effects that reduced the aggregate effect observed in previous research (Baumle and Poston 2011; Klawitter and Flatt 1998; Klawitter 2011; Martell 2013). By looking at the state-level variation in these non-discrimination laws, I can estimate the effect of strong versus weak laws. For gay men, the effects of the law depend heavily on the structure of the law. If I look at what would be considered a strong law and a weak law separately, I find that strong laws had larger wage increases and larger employment increases. Based on the estimates from Panel B of Table 3, a strong law (e.g. a law with compensatory damages, punitive damages, attorney’s fees, a size minimum of 1, and a statute of limitations of 1 year) saw an average increase in annual wages of 12.3% and an increase in employment of 0.8%. A weak law (e.g., a law with no compensatory damages and no punitive damages, a size minimum of 15 employees, and a statute of limitations of 180 days) saw no increase in wages and a decrease in employment of 0.4%.

For lesbian women, the specific provisions of the laws have little effect on their labor market outcomes (Table 4). The average law had no effect on wages, but decreased employment by 1.7% and decreased hours by 0.733 hours. States with punitive damages saw larger declines in hours worked than states without punitive damages.

These results highlight the care that must be taken when crafting a law to protect a marginalized group. Simply using the same template as existing laws that protect employees against racial and gender discrimination may be detrimental. The most current version of the federal Employment Non-Discrimination Act to pass the Senate in 2013 had a statute of limitations of 180 days. This is lower than the average of state statutes of limitations by two months. The employer size minimum in the proposed ENDA is 15 employees. This is ten more employees than the average state law. The federal ENDA allows for compensatory and punitive damages and attorney’s fees to be awarded. Using the results from Tables 3 and 4, it is possible to calculate the effect of the federal law. For gay men, this law would have an estimated wage effect of 7.4%-7.9% (depending on whether one uses annual income from wages or hourly wages) and an estimated employment effect of -0.4%. For lesbian women, this law would decrease employment by 1.7% and decrease hours worked by 2.179 hours. These estimates suggest that policymakers must think carefully about the strength of the provisions when crafting the bill.

The effect of these employment protection laws can be very nuanced. Future research into similar laws where state laws are not constrained by federal protections, such as pay secrecy bans or transgender protections, need to take the legal differences into account when estimating the effect of the laws. Failure to do so may result in inconclusive results that are driven by the differences between weak and strong laws. My results also suggest that other policies that are implemented differently across states may experience similar differences depending on the structure of the implementation.