Beyond traditional academic degrees: The labor market returns to occupational credentials in the United States

Despite their growing popularity, social science's understanding is still evolving regarding the role that occupational credentials play in preparing students for the labor force, in the production of human capital more broadly, and in how employers interpret these credentials as evidence of competencies when making hiring and salary decisions. In the United States, research efforts have been led in part by the Interagency Working Group on Expanded Measures of Enrollment and Attainment (GEMEnA), a federally commissioned group tasked with developing and validating measures of the participation in and credentialing of education and training for work, including metrics that measure the attainment of occupational credentials. Prior to GEMEnA, federal surveys in the United States had disparate approaches for asking sample members about occupational credentials, with some asking about them in survey modules focused on educational attainment and school enrollment, and others asking about them in survey modules focused on job training. Without standardized, systematic metrics in federal surveys, it was not possible to reliably study credentials across occupations at the national level. By creating these new “gold standard” metrics, GEMEnA has laid the foundation for social scientists to embark on new research in the areas of educational attainment and workforce development.

Pre-GEMEnA attempts at estimating the labor market returns to licenses and certifications yielded mixed results. Kleiner et al. analyzed an array of cross-sectional nationally representative surveys and found that wages were between 10% and 18% higher among those with licenses when compared to those without (Kleiner and Krueger, 2010; Kleiner and Krueger, 2013; Kleiner and Vorotnikov, 2017). In these surveys, the estimated returns to certifications were substantially smaller (Kleiner and Krueger, 2013; Kleiner and Vorotnikov, 2017). In contrast, however, research using the Education Longitudinal Study of 2002, which tracked a nationally representative cohort of high school graduates from the class of 2004, identified an earnings premium of between 14% and 25% from holding a certification among young adults (Albert, 2017).

Lacking data that included direct measures of occupational credentials, Redbird (2017) pooled data from the 1983–2012 CPS (prior to its implementation of GEMEnA's measures in 2015) and used state laws regarding licensure requirements for specific occupations to determine whether or not workers in states that required licenses for their occupations earned more than their counterparts holding the same occupation in states that did not require licenses. She found no association between state licensure laws and wages. The treatment effect identified under these conditions is a very specific one: the returns to having a license because the state requires one. There may, however, be strong returns to obtaining a license or certification in a state or in an occupation where they are not required, as the receipt of the credential may serve to distinguish the human capital of credential-holders in hiring or promoting processes in those states or industries.

One of the first surveys to incorporate GEMEnA's measures was the 2012 Survey of Income and Program Participation (SIPP), which is a nationally representative household-based survey collected by the U.S. Census Bureau. Once these new metrics were added to the SIPP, it was estimated that 21.6% of adults in the country held a currently active certification or license, with rates of receipt higher among those with more advanced traditional academic degrees such (such as bachelor's degrees) than those with high school diplomas and associate degrees (Ewert and Kominski, 2014). Using this data, Gittleman et al. (2018) found that adults with licenses were more likely to be employed, and if employed, had 7% higher wages than their peers without licenses. Gittleman and Kleiner (2016) used the National Longitudinal Study of Youth to identify individuals who switch into or out of occupations that require licenses in their state of residence, and from that, estimate a fixed effects model of the return to switching into a license-required occupation. They found the wage growth from such a switch to be between 2% and 7%. Finally, Ingram (2019) used data from the CPS, which also included the occupational credential questions per the guidance of GEMEnA, to estimate a propensity score model of the licensure earnings premium. He additionally leveraged state variation in licensure rates to estimate a model using metropolitan statistical areas (MSAs) spanning state borders, estimating wage returns of between 4% and 8%.

Taken together, these studies highlight the potential for occupational credentials to improve labor market outcomes of workers in the United States. While informative, these studies have a number of limitations. The analyses conducted by Kleiner et al. used cross-sectional data with low response rates, and so had a limited ability to account for selection and may be affected by non-response bias. Redbird (2017) and Albert's (2017) analyses used data from nationally representative surveys with larger samples and higher response rates but were conducted prior to the development of the GEMEnA measures, and in the case of Redbird's (2017) study, direct measures of licensure were not available, and were thus inferred. Gittleman et al.'s (2018) analysis benefits from the strong survey properties of the SIPP and the inclusion of the GEMEnA measures, but used cross-sectional Ordinary Least Squares (OLS) regressions which did not control for selection or omitted variable bias. Consequently, their estimated earnings benefits likely reflect some dimensions of positive selection into occupational credential programs. Gittleman and Kleiner (2016) meanwhile use the National Longitudinal Survey of Youth (NLSY) and are able to control for time-invariant omitted variables through individual fixed effects, but, like Redbird (2017), do not observe actual licensure status (instead inferring it from their occupation and state, and the laws for that state and occupation). As a consequence, they were unable to identify returns to licenses in states that do not require them.

Lastly, Ingram's (2019) propensity score matching analysis of the CPS was able to take advantage of a large nationally representative survey with high response rates and the inclusion of the GEMEnA measures. While an important contribution, matching estimators are only able to match based on observed characteristics, and it may be that there are differences between those with credentials and those without credentials based on unobserved characteristics (e.g., ability, motivation, etc.). The magnitude of the bias that may arise from selection on unobserved factors is unknown; if large, it would lead to greater worries regarding Ingram's estimates, while if small there would be less of a concern. Additionally, propensity score estimators require a common support, and thus constrain estimation to narrow and perhaps unique segments of the sample where there are available matches on observed characteristics. This provides an ATT estimate. This is certainly of interest, and in this paper we present it as one estimate from our model. However, ATT limits generalizability to the broader population (which would be captured from the ATE, which our method additionally estimates). Our paper offers an alternative strategy to estimating the returns of occupational credentials that circumvents these limitations.

In our analysis, we build on this growing body of research by analyzing data from the 2015 and 2016 CPS, which includes a large nationally representative sample and is characterized by high response rates, and employ the occupational credential questions per the guidance of GEMEnA. To limit potential bias owing to selection, constraints, omitted variables, and measurement error, we use a local peer influence instrument via the within-CBSA credential rates of local individuals sharing the same sociodemographic characteristics as instruments, and we include the inverse mills ratios (IMRs) in the second stage of Heckman regressions that predict wages. In using the first 2 years in which the GEMEnA measures were included on the CPS and incorporating IVs to attenuate possible selection bias in estimating the effects of occupational credentials on labor market outcomes, our study improves upon past research that attempts to understand how the provision of licenses and certifications can directly benefit workers.

A distinctive quality of licenses and certifications is that they can serve as “capstones” on top of traditional academic degrees, which in turn collectively signal occupation-specific qualifications to prospective employers. The value of these signals likely varies depending on the level of quality of educational skills facilitated through traditional academic degree programs, which are more clearly understood markers of human capital. As mentioned earlier, licenses and certifications are less prevalent among those in the sub-baccalaureate labor market than among those with a bachelor's degree (Ewert and Kominski, 2014). Additionally, bachelor's degrees convey a more comprehensive set of skills and capabilities than associate degrees or high school diplomas. Therefore, we hypothesize that occupational credentials serve to differentiate high-quality sub-baccalaureate job applicants, more so than for those with bachelor's degrees, which would result in potentially larger returns to these credentials.

To illustrate, consider two hypothetical recent college graduates. The first has an associate degree in business administration and is considering a job as an administrative assistant in a marketing consulting firm. While there are no licenses required to be an administrative assistant, the job applicant might opt to acquire a computing certification (e.g., a Microsoft certification or a Cisco certification) to enhance their hiring prospects. The second has a bachelor's degree in business administration and is seeking a job as a portfolio manager at the same marketing consulting firm. Similar to the first applicant, this second applicant has acquired a computing certification for an entry-level job they held while working their way through college. In the situation of the associate degree holder, the certification may serve to differentiate the applicant from the rest of the pool of low-skill workers aiming for the administrative assistant position. In the situation of the bachelor's degree holder seeking a portfolio manager position, the certification is likely less relevant to the employer than their bachelor's degree. Therefore, the returns to the associate degree holder's certification should be higher than the returns to the bachelor's degree holder's certification, holding industry/occupation and education constant.

In addition to heterogeneity in returns to occupational credentials by the attainment of traditional academic degrees, we also explore heterogeneity by gender. Educational attainment has been increasing among women, in tandem with a college earnings premium that is larger for women than for men (DiPrete and Buchmann, 2006). Despite this growth, sizeable wage and employment gaps by gender remain (e.g., Goldin and Rouse, 2000; Blau and Kahn, 2017). In particular, women face substantial discrimination in the hiring process in part because employers believe female applicants are more committed to family than their jobs (e.g., Blau and Kahn, 2017; Correll et al., 2007; Coate and Loury, 1993) and in part because employers believe female applicants are less capable of performing the tasks required for the job (Coffman et al., 2021; Reuben et al., 2014). Note that presumptions about job commitment appear to extend to single women during their childbearing years (e.g., as in Petit, 2007). As our context is the United States, we primarily describe studies based in the United States that identify employers’ reluctance to hire women and mothers. However, we note that this is not a United States-specific phenomenon (e.g., for France, see Coudin et al, 2018; for Canada, see Javdani and McGee, 2019), although an audit study in the Netherlands found no evidence of gender or parenthood discrimination (Mari and Luijkx, 2020).

It is possible that occupational credentials on women's resumes could attenuate these sources of discrimination by signaling their commitment to their careers (countering the stereotype of early exits by potential mothers) and enhanced workplace competencies (countering the stereotype of lower capability). The evidence to date suggests this might be the case. For example, Blair and Chung's (2018) analysis of the SIPP documents the potential of license acquisition by women to reduce gender wage gaps. Similarly, Law and Marks (2009) find that, historically, occupational licensure led to increased employment in skilled and licensed fields for female workers. Therefore, we hypothesize that occupational credentials will bolster the labor market prospects of women more than for men.

Prior evaluations of the returns to credentials have relied on regressing labor market outcomes on credential status as well as a broad set of individual controls, including education level and in some cases, individual fixed effects. However, these approaches may be biased, for all the same reasons prior literature has noted that similar regressions of returns to traditional academic degrees are known to be biased (Card 2001). These reasons include the omitted variable bias of not observing and thus failing to control for factors that select individuals into license programs (such as ability, interests, career goals, etc.), bias from selection on heterogeneity in the anticipated returns to license, credit constraints, and so forth. Measurement error of employment and earnings would further bias results. This necessitates an approach that can tease out the true returns to credentials either across the distribution (such as expressed through MTEs) or some average of the population, such as for the entire group (ATE), for those that get credentials (ATT), and for those that do not get credentials [Average Treatment Effect on the Untreated (ATU)], or the treatment effect for those affected by the instrument [the Local Average Treatment Effect (LATE)].

An IV method is one approach we can use to estimate these parameters. As discussed, the existing literature has not, to date, incorporated an IV when estimating the returns to occupational credentials. However, the literature on returns to formal education, such as bachelor's degree attainment, have used IVs extensively. Some commonly used IVs include the distance between home and colleges (Card, 1995; Doyle and Skinner, 2016), changes in tuition costs and financial aid availability (Velez et al., 2019), and changes in mandatory schooling age thresholds (van Huellen and Qin, 2019; Balestra and Backes-Gellner, 2017; Oreopolous, 2006); see Card (2001) for a review of this literature. For occupational credentials, however, these previously used instruments are less useful. Credentials can be acquired in several unobserved locations, including some through online learning, rendering geographic distances less relevant. Tuition costs for these programs vary tremendously across states and over time, but there is not a readily available cost database for the universe of these programs, nor documentation for what credential each person has in our data, let alone where they acquired it. Additionally, there are often no mandatory age requirements for credentials, and no requirements of necessary training for groups (that is, that individuals must attend training classes, whether or not they want to work in that area).

Lacking guidance of previously used and well-established IVs from the returns-to-schooling literature that could be applied to returns to credentials, we introduce an instrument that is “local peer influence” based, working on the assumption that peer groups that have higher rates of credentialing may increase the propensity of individuals in that peer group to pursue and acquire credentials. The use of this instrument is motivated by research which shows that net of sociodemographic and academic characteristics, peer groups influence academic achievement (Calvó-Armengol et al., 2009; Hanushek et al., 2003) and college enrollment (Fletcher, 2012, 2015). In our study, local peer groups are defined as the people that live in the same CBSA and are the same gender and race/ethnicity (American Indian, Asian, Black, Hispanic, White, or other, where all racial groupings are for the non-Hispanic), have the same educational attainment, and are within the age band of 5 years younger to 15 years older.

We use the non-symmetric band for age on the assumption that most individuals look to their peers for signals of appropriate behaviors, with greater weight toward those older than themselves who are further along in their schooling and careers. We tested several other age bracket options: a narrow band (2 years younger to 5 years older), a broad band (10 years younger to 30 years older), and no age restriction. We selected the age band we did as it minimized the mean squared predicted error of our model (using the first stage regression including other covariates).

Eqs (1) and (2) present the regression specifications we use as our second-stage regressions for the two outcomes, the probability of being employed (emp_ijst) conditional on being in the labor force and log hourly wages (lnwage_ijst) conditional on being employed. (1) empijst=α+βCredi+Xiγ+ρ1PeerWageij+ρ2LocalEarnij+ϕunempjt+λlfpartjt+ψs+θt+εijst {emp}_{ijst} = \alpha + \beta {Cred}_i + {X_i}\gamma + {\rho_1}{PeerWage}_{ij} + {\rho_2}{LocalEarn}_{ij} + \phi {unemp}_{jt} + \lambda {lfpart}_{jt} + {\psi_s} + {\theta_t} + {\varepsilon_{ijst}} (2) lnwageijst=α+βCredi+Xiγ+ρ1PeerWageij+ρ2LocalEarnij+λ1potexpi+λ2potexpi2+ϕ1unempjt+ϕ2lfpartjt+ψs+θt+κIMRijst+εijst \matrix{{{lnwage}_{ijst} =} \hfill & {\alpha + \beta {Cred}_i + {X_i}\gamma + {\rho_1}{PeerWage}_{ij} + {\rho_2}{LocalEarn}_{ij} + {\lambda_1}{potexp}_i + {\lambda_2}{potexp}_i^2} \hfill \cr {} \hfill & {+ {\phi_1}{unemp}_{jt} + {\phi_2}{lfpart}_{jt} + {\psi_s} + {\theta_t} + \kappa {IMR}_{ijst} + {\varepsilon_{ijst}}} \hfill \cr}

The primary regressor of interest is Cred_i, an indicator for holding a credential (separately, licenses or certifications). In addition, employment and wages are functions of individual characteristics including race/ethnicity, gender, educational attainment, and age (X_i). The outcomes are also functions of local labor market conditions. Therefore, we control for the county's unemployment rate (unemp_jt) and the labor force participation rate (lfpart_jt), and include state fixed effects (ψ_s). It is important to control for these measures of labor market conditions, as these are correlated with decisions to credential and the outcomes, as well as the IVs themselves (Cameron and Heckman, 1998, 2001). We also control for time (i.e., month and year of survey) fixed effects (θ_t), and for log wage, we include a quadratic in potential labor market experience (potexp_i) and the inverse mills ratio (IMR_ijst) so as to adjust for the selection into being employed. For the log wage regression, we use a Heckman selection mechanism by including the IMR for the probability of employment. For the IMR, our excluded instruments are the triple interactions among gender, marital status, and having dependents. We hypothesize that there are differences in employment along the intersections of these demographic variables that are not fully explained by the subgroups alone. For example, reservation wages (impacting both labor force participation and wages conditional on employment) may be higher for married women with dependents than for married non-parents, given current cultural norms regarding childcare. As described above, we instrument credential status using the local peer influence instrument, included in the first stage.

We may still worry that even after controlling for this array of potential confounds, there remains a direct impact of a higher-credentialed local peer group on an individual's labor market outcomes, which would violate the exclusion restriction of the instrument: a group motivated enough to pursue credentialing may be strong in other ways that improves outcomes, even after controlling for the direct differences in group wages through X_i. To address this concern, we include as additional control variables a leave-one-out estimator for the local peer group's average wages (allowing for zeros for non-employment) and the across-demographic local average wages (PeerWage_ij and LocalEarn_ij, respectively). We argue that these control for remaining direct impacts of the peer group on the outcomes as well as the strength of the local labor market, such that any residual impact of the local peer group credential rate on an individual's own credentialing probability is the remaining pathway in which the local peer group can impact an individual's labor market outcomes, after controlling for all other variables.

For smaller regions and smaller peer groups, we adjust the definition of the instrument (PeerCred_ij) and peer wages measure (PeerWage_ij) to use the local rate aggregated across demographic groups, rather than within, still limited to the CBSA. We perform this adjustment when the sample size on which to estimate the local peer credential rate is fewer than 30 observations across the 2 years of data, which occurs for 13.7% of our observations. Additionally, we note that in the CPS, the CBSA is the MSA for areas connected to an MSA; for an individual living outside any MSA, their CBSA is functionally a state identifier that excludes all MSAs in the state.

The first stage equations model credential status as a function of all other covariates in Eqs (1) and (2) (for each outcome respectively), as well as the local peer influence measure (our IV), using a probit model. We recognize that the use of peer credentialing as an instrument is novel in the returns to education literature, yet the returns to credentialing literature is relatively nascent and with divergent estimates. For a validation check of our instrument, we construct the parallel local peer influence instrument for having an associate degree among those with less than a bachelor's degree, and estimate identical models examining the returns to an associate degree, a well-developed literature. We are thus able to compare the estimated returns to an associate degree using our instrument and our data set to the estimated returns to associate degrees published in well-established studies. As detailed later, our estimated returns to an associate degree with our instrument in the CPS compares very favorably to the estimates already established in the existing literature, supporting the use of our instrument for predicting credential receipt.

As with any IV, two key conditions must be met for valid identification of the effect of credentials on labor market outcomes. First, the instrument must be relevant, with a strong observed correlation between the local peer influence measure and the credential status of the individual. Second, the instrument must be uncorrelated with the error term and the exclusion restriction must hold, in that there is no impact of variation in the local peer influence measure on labor market outcomes except through the receipt of a credential. The latter assumption is a canonically untestable assumption. Instead, we must provide a convincing defense of its validity.

We consider potential violations of the exclusion restriction, and we argue that our specification addresses these issues. We first frame potential violations in the peer effects literature, following the logic and guidelines put forth by Manski (1993). There, Manski argues that a linear-in-means peer effects specification (as we implicitly have in the first stage of the regression) combines three effects: the endogenous effect, the contextual effect, and the correlated effect. As we go through these, we emphasize though that we do not need to isolate the causal pathway (the “endogenous effect” in Manski) for our IV to remain valid; we can have all three effects present, as long as the exclusion restriction is not violated. This combination of effects only affects the interpretation, not the validity, of our results: we are able to estimate the effect on occupational credential receipt on labor market outcomes regardless of knowing whether it is peer influence directly or another mechanism correlated with peer influence.

First is the endogenous effect, which is the desired causal impact of peers’ outcomes on an individual. This does not violate the exclusion restriction: it is our main hypothesized pathway. Second is the contextual effect, which is the impact of peers’ mean characteristics on an individual's outcomes. This may cause an issue; an individual with peers who have higher credential rates may also have peers who have higher baseline employment probabilities, earnings, income, consumption, and other relevant factors. These contextual effects may also impact the probability of credentialing (which only helps strengthen the first stage and is not problematic) but may also violate the exclusion restriction through these correlations with outcomes that do not act directly through credentialing.

We do two things to address these issues. First, we include indicators for the demographic characteristics we use to define peers (e.g., race/ethnicity, gender, educational attainment, and age), which account for those levels of contextual effects. Second, we also include in the regression (as independent variables) the peers’ average wages and peers’ employment rate. These serve to control for the most important sources of heterogeneity that contribute to the violation of the exclusion restriction. For example, the peers of an individual that have higher credential rates may have other kinds of motivation, networking prowess, and drive that improve labor market outcomes. By controlling for those, we constrict the potential ways in which peers’ unobserved qualities, as correlated with credential rates, may impact individual employment rates and wages. Thus, we would expect that not controlling for these peer labor market outcomes would lead to higher estimated impacts of credentialing on a person's labor market outcomes. While not presented in this paper, we tested it both ways and indeed found that including these controls somewhat attenuates the effects, meaning our preferred (with controls included) estimates are potentially understated.

The third peer effect pathway is the correlated effect, which is that the person and their peers may be in the same context and environment, which may drive credentialing probabilities (not problematic, strengthens the first stage) as well as the labor market outcomes (problematic, would violate the exclusion restriction). Including the peers’ average wages, peers’ average employment rate, local unemployment rates, local labor force participation rates, and state fixed effects are important again to address this concern. For example, individuals in a peer group with a higher credential rate in an area with a high labor force participation rate and in a state with a more advanced vocational education system may have better connections to the labor market overall, or may have heightened access to education and training opportunities in their area. However, these would also impact the peers in the same way, and by separating out the peers’ labor market outcomes, we control for these types of differences and absorb those types of potential violation of the exclusion restriction.

Additionally, we note that by looking at peers both geographically and demographically in the model, we are able to leverage identification along both pathways in order to separate out the peer impact. For example, for concerns regarding violations of the exclusion restriction that may occur due to common geographical contexts, we note that our effect is in part identified by comparing individuals in the same CBSA but from different demographic groups, which then accounts for such factors. Note that we do not include CBSA fixed effects in the model, but state fixed effects, to allow for comparison of peer groups between CBSAs within the state. We additionally control for CBSA-level labor conditions such as the overall unemployment rate, as well as the unemployment rates of one's peer group. For concerns regarding violations of the exclusion restriction that may occur from demographic group commonalities—e.g., young Black men facing certain labor market hurdles across the country that are not specific to one area—we rely on the identification of the net effect after including those demographic group controls at the national level. We are thus comparing within a demographic group and between CBSAs within the same state to leverage remaining variation in the credential rate between these groups. Together, the inclusion of peer group labor market outcomes as additional controls along with state fixed effects, local labor market conditions, and demographic group indicators, provides us with a reasonable claim that the remaining variation in the credentialing rate of peers is likely to only impact an individual's labor market outcomes through the change in the probability that the individual is also credentialed.

Given our parametric formulation of the outcome variables with covariates, our interpretation of the estimates of LATE fall under the limitations noted by Blandhol et al. (2022). They find such models will estimate averages of not only compliers (the typical interpretation of LATE), but also always and never takers, some of which may have negative weights. Thus, we must exercise caution when interpreting IV estimators, as they are unlikely to be the LATE of compliers. The MTE approach not only yields a LATE-type estimator from the IV, but also the ATT and ATE given the distribution of the instrument in the population, which are our primary focus. This comes at the price of additional parametric assumptions but relaxes the burden of interpretation of LATE noted by Blandhol et al. (2022).

We use the instruments in the MTE estimation framework summarized in Cornelissen et al. (2016) (see also Carneiro et al., 2011). We apply the parametric Roy model of MTE, which allows us to estimate additional parameters other than LATE, by leveraging the continuous nature of our IV. Intuitively, the MTE estimates several LATEs along the entire distribution of the continuous IV. This allows for repeated estimates of a LATE at different levels of what the literature calls the (unobserved) distaste for treatment. These MTEs can then be aggregated up as weighted averages along that distribution to estimate, for example, the ATE by adjusting to the sample population level given the distribution of the IV in the sample. This represents a major advantage of the MTE estimator: LATE is not likely to be highly policy-relevant in our case (and may not even be LATE, per Blandhol et al. 2022), given our instrument, as policy-makers likely care about the returns not merely for individuals that would be motivated to pursue a credential only with sufficient peer support, but rather for the population more generally. Using the MTE model, we are able to estimate the ATE, ATT, ATU, and LATE of occupational credentials, and map out the returns across the MTE curve as a function of the distaste for treatment. As described in Cornelissen et al. (2016), the MTE estimator is given by: (3) MTE(Xi=x, UDi=p)=∂E(Yi|Xi=x, P(Zi)=p)∂p MTE({X_i} = x,\,{U_{Di}} = p) = {{\partial E({Y_i}|{X_i} = x,\,P({Z_i}) = p)} \over {\partial p}} where U_Di is the percentile of the unobserved distaste for treatment, Z_i is our continuous IV, and p is the probability of receiving a credential. We present the MTE curves as a function of U_Di for the average observable characteristics X_i. The estimates are weighted averages of the MTE across certain populations (e.g., across the treated group for the ATT). The support of Z_i varies by the credential of interest. Figure A1 in the Supplementary Appendix presents the histograms of Z_i for each of the two cases (licensure and certifications), for the treated and control groups. In both cases, there is full support of the treated group's distribution among the control group. Anticipating the first-stage findings below, the distribution for the treated group is shifted to the right of the distribution for the control group. Additionally, the licensure distribution has a wider support, ranging from between 0 to 0.6 (with some outliers above 0.6). The certification distribution has a narrower support, ranging from 0 to just above 0.1, again with some outliers above 0.1.

We are additionally interested in decomposing the net returns to wages, allowing for the effect to be a function of not only the returns to wages conditional on working, but also the returns driven by changes in the likelihood of employment. Letting w be hourly wages and emp be the employment status, and restricting all to individuals within the labor force, we note that by the law of total probability and the fact that non-workers have zero wages, (4) E[w|X]=E[w|X, emp=1]Pr(emp=1|X) E[w|X] = E[w|X,\,emp = 1]\Pr (emp = 1|X)

The overall difference in wages between those that have a credential (Cred = 1) versus those that do not (Cred = 0) can be decomposed as: (5) E[w|X, Cred=1]−E[w|X, Cred=0]=E[w|X, emp=1, Cred=1]Pr(emp=1|X,Cred=1)−E[w|X, emp=1, Cred=0]Pr(emp=1|X,Cred=0)=βWPr(emp=1|X, Cred=1)+βEE[w|X,emp=1, Cred=0] \matrix{{E[w|X,\,Cred = 1]} \hfill & {- E[w|X,\,Cred = 0] = E[w|X,\,emp = 1,\,Cred = 1]\Pr (emp = 1|X,Cred = 1)} \hfill \cr {} \hfill & {- E[w|X,\,emp = 1,\,Cred = 0]\Pr (emp = 1|X,Cred = 0)} \hfill \cr {} \hfill & {= {\beta_W}\Pr (emp = 1|X,\,Cred = 1) + {\beta_E}E[w|X,emp = 1,\,Cred = 0]} \hfill \cr} where (6) βW=E[w|X, emp=1, Cred=1]−E[w|X, emp=1, Cred=10][w|X, emp=1, Cred=1]βE=Pr(emp=1|X, Cred=1)−Pr(emp=1|X, Cred=0) \matrix{{{\beta_W} = E[w|X,\,emp = 1,\,Cred = 1] - E[w|X,\,emp = 1,\,Cred = 10][w|X,\,emp = 1,\,Cred = 1]} \hfill \cr {{\beta_E} = \Pr (emp = 1|X,\,Cred = 1) - \Pr (emp = 1|X,\,Cred = 0)} \hfill \cr}

The total returns are the sum of two elements: β_W, the wage returns conditional on working (the typically estimated return), or the intensive margin of the effect; and β_E, the wage returns to employment given being in the labor force, or the extensive margin of the effect. We estimate each of the four elements of Eq. (5) to construct the decomposition.

Given the data-driven construction of the instruments, analysis being conducted across multiple stages, use of IMR within the MTE framework, and the decomposition being a function of several parameters from separate regressions with different samples, we bootstrap all of the standard errors. We block-bootstrap at the CBSA level with 500 bootstraps to account for within-labor market intraclass correlation that would otherwise bias the standard errors.

Table 3 presents the first-stage regression results of the effect of the proportion of peers with a license on having an occupational credential. The base controls add in state, month of survey and year of survey fixed effects, the local peer group's mean wages, and other demographic variables, but excluding race/ethnicity, gender, educational attainment, and age (the demographic characteristics that define peer groups), which are added in the final “All controls” column. The All controls column shows the importance of controlling for these demographic variables independently. The coefficients in the first stage remain highly significant and large in magnitude but decrease as more controls are added. For example, among those in the sub-baccalaureate group, going from a local peer group with a 12.9% licensure rate (the mean peer group average) to one with a 22.9% licensure rate is associated with a 0.425 (4.25 percentage point) increase in the likelihood of that individual obtaining a license – a sizeable increase. The instruments are strong, with F-statistics often in the hundreds, and never below 20.

We find additional evidence in support of the instruments by exploring CBSA-level measures of racial segregation. To do so, we use a percentile transformation of the divergence index created by the University of California at Berkeley's Othering and Belonging Institute (Menendian et al., 2021). In local communities that are more racially segregated, we would hypothesize that our instrument is a stronger predictor of individual credential status – that we are more likely to have correctly represented an estimate of their peer group along racial–ethnic lines (in addition to the other factors). We find evidence of this. In regressions of the endogenous individual credential status on the instrument (not shown), the measure of local segregation, and their interaction, the interaction was positive (although relatively small in magnitude) in all four cases (sub-baccalaureate vs. BA or higher, license vs. certification), and statistically significant in three of those cases.

We next move to the second-stage results, starting with the returns to employment conditional on being in the labor force as the outcome. Figure 2 presents the MTE returns curves by level of education. These plots provide a visual depiction of selection into treatment via the instrument. These curves are plotted over the distribution of “resistance to treatment” – in our case, how unlikely someone is to obtain a credential by virtue of having a highly credentialed network. If the most positive returns are concentrated with those who easily comply with the treatment—who are very likely to obtain a credential when their peers are credentialed—then this is a case of positive selection. Positive selection is reflected in a downward slope of the MTE curve, where the most positive effects occur low in that resistance distribution. In contrast, if the graph slopes upward, this is indicative of negative selection. An upward slope means that the strongest returns are concentrated among those least likely to obtain a credential in the way our instrument suggests (i.e., influenced by one's peers in a highly credentialed peer network).

Figure 2

For licenses, we find small, generally positive effects on employment (the curves tending to have positive values). The sub-baccalaureate curve slopes upwards, consistent with negative selection into treatment – those with the highest distaste for treatment (the far right of the x-axis) have the most to gain in outcomes, but are least likely to get the treatment due to that distaste. This negative selection into treatment results in a LATE estimate [via two-stage least squares (2SLS) or MTE] that is lower than the ATE, because restricting the estimate to only compliers omits the non-compliers with potentially more positive effects. For baccalaureate license-holding, the slope is basically flat, indicating an absence of strong selection on distaste for treatment (note the similarity between LATE and ATE). Returns are strongest for those with the greatest distaste for treatment, which includes those least moved by our instrument.

For certifications, we see much stronger differences between sub-baccalaureate and baccalaureate populations. The graph suggests negative returns to employment for sub-baccalaureate workers most incentivized by peer credentialing (with the negative selection evidenced by a strong upward slope and an ATE much larger and more positive than LATE), while for the baccalaureate population, there are positive returns for all workers but particularly for those with little resistance for treatment (positive selection, slight downward slope, LATE larger than ATE). Negative selection into treatment has been previously identified in the returns to education literature – for example, see Brand and Xie (2010), who find that non-college attenders stand to gain more from college attendance than those who do pursue a degree. Rationalizing this finding in a general population requires either belief in barriers that are systematically imposed on those who stand to benefit most (for example, credit constraints that do not allow them to take time away from work required to attain a credential, despite being poised to benefit the most from credential receipt) or misinformation concentrated among the same group. Note that in our case, this is only required in reference to those on the margin of pursuing a credential based on peer influence – the existence of individuals who are easily persuaded by peer influence but stand to gain little (or even lose) from a credential, along with their opposites. We also find positive employment effects on the untreated population among bachelor's degree holders for both credential types.

Table 4 presents OLS, ATT, and ATE estimates from the second-stage regression predicting employment, conditional on being in the labor force. We are most interested in the ATT and ATE estimates as they provide insight into how credentials benefit workers in the aggregate. Table A1 in the Supplementary Appendix replicates this table, but also includes 2SLS estimate as well as the MTE-based LATE and Average Treatment Effect for the Untreated (ATUT) estimates. We identify, on average, positive employment effects of licensure as well as positive effects of certification for sub-baccalaureate workers. The ATE across models is positive in all four models, and significant in three of the four. Of particular importance, the ATE for licenses among sub-baccalaureate workers is a 15% increased probability of employment compared to 4% for workers with bachelor's degrees. The ATE for certifications among sub-baccalaureate workers is a 37% increased probability of employment compared to 2% for workers with bachelor's degrees.

Figure 3 presents the returns to log wages, both with and without occupation and industry (at the two-digit level for Standard Occupational Classification (SOC)- and North American Industry Classification System (NAICS)-based codes, respectively) fixed effects. The sub-baccalaureate population has positive selection in all four models, indicated by the downward-sloping curve, while workers with bachelor's degrees or higher have generally negative selection. The negative selection for the bachelor's or higher models may reflect the presence both of low-earning credentialed jobs that require bachelor's degrees, such as teaching, social work, and nursing, as well as several high-earning non-credentialed jobs for individuals with at least bachelor's degrees, including business management and several STEM occupations. This is supported by the shallower slopes of the MTE curves once we control for occupation and industry.

Figure 3

Table 5 presents OLS, ATT, and ATE estimates across the models for log wages. Table A2 in the Supplementary Appendix replicates this table, but also includes LATE, 2SLS, and ATUT estimates. For both credential types and education levels, the ATE are closer to zero and do not maintain statistical significance when occupation and industry fixed effects are included, compared to when they are excluded. Without occupation and industry fixed effects, part of the identified treatment effect may include credentialed individual's increased ability to transition into higher-paying occupations compared to a non-credentialed individual. This arguably represents a more complete treatment effect, reflecting another dimension of credentials’ returns (entry into higher-paying occupations).

While the model that includes industry and occupation fixed effects only identifies the increase in wages within industries and occupations (that is, not accounting for the credential's impact on their ability to transition to higher-paying industries and occupations), the model that includes the fixed effects may better isolate the effect of credential-holding in differentiating candidates in the same industry pursuing the same job. The ATEs are largest for sub-baccalaureate licensed workers at around 0.26, or a 26% wage increase. Our estimate is somewhat larger than prior cross-sectional estimates of a 10%–18% increase (c.f. Kleiner and Krueger, 2010; Kleiner and Krueger, 2013; Kleiner and Vorotnikov, 2017). Yet the estimates with occupation and industry fixed effects are smaller than those found in the cross-sectional literature (and do not rise to statistical significance), consistent with the possibility of positive selection bias in prior estimates. However, the ATE is not significant for certifications, where the variance in quality and labor market returns is likely much higher than for licenses, leading to noisier estimates. This is true for certified sub-baccalaureate workers as well as for workers with a bachelor's degree.

Interestingly, the ATE for licensed workers with a bachelor's degree or higher is −0.19 (19% wage decrease) when not controlling for occupation or industry. We speculate that the negative ATE reflects the presence of several low-paying licensed jobs among this population (e.g., teachers, social workers) and high-paying jobs that are not licensed (e.g., management, software developers). This interpretation is supported by the fact that the same model, once occupation and industry fixed effects are added, produces a positive, albeit insignificant, coefficient. Note that in prior cross-sectional work, Kleiner and Vorotnikov (2017) find a significant certification earnings premium of approximately 9%, very similar to our (insignificant) estimate with our occupation and industry fixed effects.

We recognize that the estimated sub-baccalaureate returns to certifications (as measured by ATT) are quite large in magnitude. We suggest caution in interpreting these results, especially given how much rarer certification is in the sample (with no license); only 2.3% of the sample holds one (see Table 3). However, the results remain suggestive of potentially substantial returns to occupational certifications.

Additionally, while not reported in Table 5, the selection into paid employment modeled by the coefficient on the Inverse Mill's Ratio is consistently negative and statistically significant, ranging between −0.265 and −0.167. This may partly be explained by our not having employment measures for self-employed workers (see Footnote 4). This negative selection into paid employment is consistent with many findings in the literature [see, for example, Bollinger and Hirsch (2013)].

We next turn to the decomposition of the returns on hourly wages, with parameter estimates for the OLS and the ATE from the MTE model shown in Table 6. Here we use hourly wages in place of log wages for ease of interpretation. Note that this decomposition is for the entire sample (that is, not only for treated, untreated, or for some local treatment group); a derivation of the decomposition for such a subgroup would require additional assumptions and is left for future work. Thus, we only calculate and present the OLS and ATE estimates here.

The estimates in Table 6 show that for sub-baccalaureate licenses, approximately three quarters of the total effect comes from the extensive margin (increased likelihood of working) and one quarter from the intensive margin (higher wages). For bachelor's or higher, the only positive and significant effect is for licenses on the extensive margin: workers are more likely to be employed because of their license, but have no subsequent pay increase of significance. Note that this is the model which includes occupation and industry fixed effects, which diminishes the effect of credential receipt on wages.

We are able to construct a parallel local peer group instrument for the returns to an associate degree for the sub-baccalaureate population. Specifically, we estimate the fraction of people in the same demographic group in the CBSA that have an associate degree and repeat the MTE analysis for those with an associate degree or less. This serves two valuable purposes for our study. First, we can compare our estimate of the returns to education for traditional academic degrees already established in the literature. This allows a validation check of our instrument choice and MTE approach. Second, it allows us to compare for the same sample the returns to an associate degree to the returns to a license or to a certification, as a potential alternative educational pathway.

The OLS, ATT, and ATE estimates are presented in Table 7. For ease of comparison, we include the license and certification results from Table 5. Table A3 in the Supplementary Appendix replicates this table, but also includes LATE, 2SLS, and ATUT estimates. We estimate the return to an associate degree of 0.16 using OLS and 0.25 for ATE. Lang and Weinstein (2013), using cross-sectional OLS, estimate a return to an associate degree that ranges 0.10 to 0.18 depending on the major, and Dadgar and Trimble (2015) use individual fixed effects and find estimates between 0.02 and 0.09, putting our OLS estimate on the higher end (likely because we are unable to include individual fixed effects). More causal estimates put the return around 0.15 log points per year (see Oreopoulos and Petronijevic, 2013) for about a 30-log point increase, putting our ATE estimate of the 2-year degree slightly smaller than that indicated in this literature. This serves to validate our IV choice and modeling approach, both in range of estimates from the OLS versus the ATE estimate and the increase in coefficients in moving from OLS to ATE.

Table 7 also shows that the ATE is very similar between the return to license, return to certification, and return to associate degree, all around 25%. Of course, certain credentials will require formal schooling such as would happen through the acquisition of an associate degree. Nonetheless, these results are suggestive of the potential for occupational credentials to serve as viable alternative educational pathways leading to higher-paying jobs.

Table 8 presents OLS, ATT, and ATE estimates from models predicting employment and Table 9 presents OLS, ATT, and ATE estimates from models predicting wages. These tables are parallel to Tables 4 and 5, but with the models now stratified by gender. Tables A4 and A5 in the Supplementary Appendix replicate these tables, but also include LATE, 2SLS, and ATUT estimates. As discussed, credentials could potentially signal both career commitment and capability (countering bases for gender-based hiring discrimination identified in papers such as Correll et al, 2007; Petit 2007; Coffman et al., 2021). We hypothesize that through this mechanism, occupational credentials will bolster the labor market prospects of women, more so than for men. This is partially supported in our analysis. As shown in Table 8, we find that our employment effects are almost entirely driven by women for three of the four cases, with significant increases in employment for female sub-baccalaureate workers holding licenses or certifications as well as for female baccalaureate workers holding licenses. In contrast, the only significant employment effect for men is concentrated among certification-holders with at least a bachelor's degree.

For wages, we see the opposite story (Table 9). The only positive and significant wage effects accrue to men – specifically, to men with licenses (in either education group), and to men with a certification (among those with a bachelor's degree or higher). For women, we actually find evidence of negative returns to credentials, and these negative effects generally are not eliminated by accounting for industry and occupation. Together, these results suggest that for women, occupational credentialing matters in terms of hiring but not necessarily in terms of wages. It is plausible that female credential-holders are better-compensated along other non-wage dimensions not captured in our analysis.