Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’

Social Science Research xxx (2017) 1e14

Contents lists available at ScienceDirect

Social Science Research journal homepage: www.elsevier.com/locate/ssresearch

Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’ Michael R. Smith a, *, Sean Waite a, Claire Durand b a b

McGill University, Canada Universit
e de Montr
eal, Canada

a r t i c l e i n f o

a b s t r a c t

Article history: Received 24 August 2014 Received in revised form 24 November 2016 Accepted 14 March 2017 Available online xxx

In this paper we use data on consecutive cohorts of recent graduates from community colleges or community college-like institutions to address the following questions about the gender earnings gap: i) What was the trend in the gender earnings gap for these recent graduates? ii) What role in the observed trends in the gender earnings gap was played by occupational demand? iii) How and to what extent did the domestic division of labour contribute to the gender earnings gap in this young sample? We find that the gap fell then rose, that occupational demand played a role in these shifts, and that the domestic division of labour did indeed contribute to the gap in this young sample. Furthermore, our results point to a process of cumulation of factors that increase the earnings gap which has both substantive and methodological implications. © 2017 Published by Elsevier Inc.

Keywords: Gender Community colleges Earnings

1. Introduction Women earn less than men in the US, Canada, and other countries. A very large body of research has established a number of reasons for the current and past gap. Those reasons can be grouped into three major categories: i) the domestic division of labour; ii) differences in human capital, in particular in the form of educational choice by gender; and iii) employer decisionmaking. With respect to these factors: i) marriage and childcare tend to disadvantage women; ii) while they have caught up and exceeded the educational attainment of men, they continue to acquire less work experience and, at the university level, have been overrepresented in a number of fields of study associated with lower earnings; and iii) employers are thought to devalue women's work and there is some evidence of processes that discourage women from remaining in some betterpaying jobs and, sometimes, of reassignment to lower paying ones. Recent discussion of gender earnings differences has been framed by trends in them. In both Canada and the US, at the beginning of the 1980s, the annual earnings of full-year, full-time women were slightly more than 60% of those of their male counterparts; by the beginning of the 1990s they exceeded 70% (Baker and Drolet, 2010; Blau and Kahn, 2000, 2006). Since then, in both countries, the convergence has slowed or stopped altogether, depending on time period. Both the narrowing of the gap and its failure to continue rapidly narrowing raise interesting questions. Why did the gap shrink? There are two possible answers. Women may have acquired characteristics the absence of which previously contributed to lower earnings. Depending on the time period studied, fewer years of education and work experience contributed to lower female earnings (Blau and Kahn, 2000; Drolet, 2011). The gap in educational attainment has

* Corresponding author. Department of Sociology, McGill University, 855 Sherbrooke Street West, Montreal, Quebec, H3A 2T7, Canada. E-mail address: [email protected] (M.R. Smith). http://dx.doi.org/10.1016/j.ssresearch.2017.03.003 0049-089X/© 2017 Published by Elsevier Inc.

Please cite this article in press as: Smith, M.R., et al., Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’, Social Science Research (2017), http://dx.doi.org/10.1016/j.ssresearch.2017.03.003

2

M.R. Smith et al. / Social Science Research xxx (2017) 1e14

disappeared and in work experience narrowed. At the upper end of the educational distribution, England (2010) has emphasized convergence in fields of study; women have increasingly qualified for entry into the well-paid occupations of law, medicine, and management. And, the magnitude of the effect on the earnings gap of a particular factor may have fallen. Drolet (2011: 9) shows that much of the reduction in the gender earnings gap from 1988 to 2008 in Canada was produced by a reduction in the size of the negative effect of factors of the sort listed above. Why might some factors reduce female earnings less than in the past? England, Gornick, and Shafer (2012: 4) point out that at the upper end of the educational distribution the higher pay associated with jobs in law, medicine, and management, along with assortative mating, make it possible to pay for childcare. That reduces the size of the negative effect of the domestic division of labour. Another possible mechanism originates with the occupational distribution. There is an earnings premium to unionization. Men have been more likely to be unionized than women. Unionization tends to closely tie pay to work experience (Zangelidis, 2008). The decline in male unionization may, then, have reduced the effect of work experience on male earnings. We have clear evidence that the narrowing of human capital differences contributed to the fall in the gender earnings gap and that factors like those discussed in the previous paragraph advantage the earnings of men less than they did in the past. Less well explained is the fact that the gender earnings difference stopped narrowing. In this paper we explore that issue, in doing so identifying and estimating the effect of some of the factors that continue to explain the earnings gap. In addressing these questions we focus on those graduating from Canadian community colleges or similar institutions e that is, those who completed a post-secondary education but not a university degree. 2. Some sources of the gender earnings gap Consider the three factors used to explain the gender earnings gap summarized above. Start with educational choice. Studying law, medicine, or management leads to higher average earnings because demand in those fields is more robust than in, say, psychology or biology, where female enrolment has been substantial. England and others are correct to underline this field of study shift. However, two qualifications are in order. One is that there are fields in which there has been a robust demand and in which women remain underrepresented; another is that occupational demand is variable. The relative robustness of demand across occupations is unlikely to remain constant over time; occupational demand can both rise and fall. This is well illustrated by a set of jobs at the lower end of the educational distribution. One of the factors contributing to the decline in the gender earnings gap from the 1980s to the 1990s was a fall in the demand for some manual employees (Blau and Kahn, 2000: 85e86). Employment in manufacturing, resource extraction and, for some periods, construction, fell. The jobs in question were relatively well-paid and primarily occupied by men. The potential to either narrow or widen gender earnings differences of this sort of shift in occupational demand is recognized. When governments have budgetary difficulties they cut back hiring, sometimes laying-off employees. Women are overrepresented within government employment so, as the essays in Karamessini and Rubery (2014) emphasize, this sort of downsizing is likely to worsen the gender earnings gap. Implicitly, periods of growth in government employment would have narrowed it. But, we would argue, this mechanism is not sufficiently recognized. For the purposes of this article, of particular interest are STEM (science, technology, engineering, mathematics) fields of study. Women's presence in these programs has increased but not by nearly as much as in programs in law, medicine, and management. What is important is that the demand for people with computer-related STEM training exploded in the early 1990s when the dot-com boom began (Wang, 2007; Senn, 2000). In the US that translated into a very large increase in the computer scientist share of STEM graduates (less than a third to more than a half), a doubling of enrolments in the field, and a heavy resort to immigrants to meet scarcities. Not surprisingly, the earnings premium of computer science graduates relative to all bachelor's recipients rose, from about 25% to over 40% (Bound et al., 2015). Computer scientists provided the software the new e-commerce required. This, in turn, caused a surge in the demand for the equipment on which the new software operated and for electrical engineers which, in Canada, led to recruitment of immigrants with those skills (Picot and Hou, 2009). The dot-com boom was followed by a dot-com bust. The demand for the relevant skills did not, however, collapse. As Bound et al. (2015) show, after 2000 the pay premium to computer science fell relative to the very end of the 1990s, fluctuated thereafter, but remained higher than it had been at the beginning of the 1990s, even during the post Great Recession years. Here is a likely cause of the failure of the earnings gap to continue to fall after the 1980s: a marked shift in demand in favour of occupations where males are overrepresented. This very large effect began with the dot-com boom but then persisted, at a lower level. Now consider together the two other sources of female earnings disadvantage: the domestic division of labour and the choices of employers. These substantially overlap in their effects. Devaluation, Cohen and Huffman (2003: 884) tell us, originates in the fact that “employers err cognitively by not seeing women's contribution to … profitability” and underreward skills “such as nurturance” that women are thought to disproportionately possess. It is likely that some significant part of these views originates in the share of women in the domestic division of labour. Their domestic obligations make employers think that they are less productive (contribute less to “profitability”) and associate them with a skill that, rightly or wrongly, employers are often not seeking. Reskin and Padavic (1988) provide some evidence of this. So, two kinds of evidence of the effect of gender differences in domestic work have been provided: the earnings disadvantage associated with women's occupational concentration (Levanon et al., 2009) and direct evidence of lower earnings Please cite this article in press as: Smith, M.R., et al., Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’, Social Science Research (2017), http://dx.doi.org/10.1016/j.ssresearch.2017.03.003

M.R. Smith et al. / Social Science Research xxx (2017) 1e14

3

associated with domestic responsibilities e for example, Budig and England (2001) finding that family-related employment disruptions were associated with lower pay and an abundance of evidence that parenthood reduces women's earnings but not men's (e.g., Cooke, 2014).1 Or, at least, both kinds of evidence exist for the US. Canada is a bit different. There is a motherhood penalty in Canada (Drolet, 2002). However, Baker and Fortin (2001) showed that, at least for the period they studied, the occupational concentration of women in Canada had little or no effect on earnings (depending on model specification). In explorations of the sources of women's lower earnings this reduces somewhat the urgency for further research on occupational concentration by gender in Canada. An interest in effects of the domestic division of labour remains. With many data sets these can be estimated directly using information on marital status, numbers of children, and childcare-related employment interruptions. Another operationalization of a probable effect of the domestic division of labour is hours of paid work. In many e particularly unionized e jobs additional hours lead to higher pay. Baker and Drolet (2010) show that differences in hours of paid work account for a large proportion of the gender earnings gap. But hours matter to earnings beyond paid overtime. A successful career in some occupations, particularly in the early years, may require long hours of work. There has been a significant amount of research on the effect of hours of work on lawyers’ careers and earnings (e.g., McNabb and Wass, 2006). Similarly long hours are often required of consultants, accountants, and doctors. Conversely, part-time work is rarely conducive to career success. We know a number of things about gender and hours of work. Marriage reduces hours of work of women but not of men (Abroms and Goldscheider, 2002). Work/life conflicts lead women to want to cut their hours of work. They do not have the same effect on men (Reynolds, 2005; Maume, 2006). Women are more likely than men to want to work fewer hours and less likely to want to work more hours (Reynolds, 2003: 1183). That fewer hours worked contributes to the earnings gap is clear. Cha (2013) and Cha and Weeden (2014) have documented the contribution to it of a lower incidence of long hours of work among women. We conclude that hours of work effects should, when possible, be considered in analyses of the gender earnings gap.

3. Community college education and its outcomes England (2010) argued that the gender earnings gap narrowed because at the top of the earnings distribution women moved into high paying jobs requiring a university education; it stopped narrowing because they did not move into the better-paid manual jobs mainly occupied by men for which a secondary school diploma has been the normal entry requirement. Nor did men move into the poorly paid jobs mainly occupied by women with a secondary school diploma. Her analysis focusses on the top and the bottom of the educational distribution. But the educational system is more than its top and bottom. There are also community colleges or community college-like institutions preparing their graduates for a “middle range” of jobs (Skolnick, 2004: 41), certainly not to be lawyers and doctors. It is common to assign university and community college graduates to the same post-secondary category (e.g., Zhang, 2009) but in Canada and elsewhere employers expect somewhat different qualities from the graduates of each level. Canadian employers, for example, seek community college graduates from programs that “are quite specific and directed to the needs of industry” (Flavell, 2012), traits that are less present in university programs. Gender differences in the earnings of community college graduates have particular interest for several reasons. First, there is a relative abundance of studies of university graduates (e.g., Davies et al., 1996; Finnie and Wannell, 2004; Morgan, 2008; Weinberger, 2011; Li and Miller, 2012; Livanos and Pouliakos, 2012). Second, most of Canada's gender earnings convergence occurred below the university level (Drolet, 2011: 5). Precisely where below that level the convergence occurred is worth examination. Third, occupations which could once be accessed with a secondary school diploma now require more. The US Department of Labor reports that Automotive Service Technicians and Mechanics require a postsecondary diploma of some sort.2 So do most other ‘technician’ jobs, many of which would once have allowed entry to those without even a secondary school diploma. Better paid manual work often requires some sort of college education. Fourth, two year colleges are what most distinguish the Canadian educational system. Its graduation rate at this level is about twice the OECD average, higher than any other country. Colleges provide about half of the graduates from Canadian post-secondary institutions (OECD, 2012: 91, 94). Fifth, this educational level is particularly relevant to women. They have continued to opt into less remunerative community college fields of study (Boudarbat, 2008). At the same time, community college has been the final educational level for a larger proportion of women than men (Boothby and Drewes, 2006). The form of college training differs somewhat across Canada. Ontario colleges have largely confined themselves to vocational education (Clark et al., 2009: 152e163). Nova Scotia Community College, which now has 13 campuses, originally offered trades training but reforms in the 1990s expanded its programs to include “the more comprehensive applied arts and technology programs that characterize more modern community colleges” (Crocker and Usher, 2006: 43). Colleges in Alberta and British Columbia have for some time been more heterogeneous. Within their systems trades, applied arts, pre-university,

1 The research showing a negative association between maternity and earnings is large. Zhang (2009) points out that part of the association is likely to be produced by the fact that measured and unmeasured characteristics influence both fertility (including age at first birth) and earnings. Her point is welltaken. Our data do not allow us to estimate these effects. (We have a single level of education, our data do not lend themselves to a fixed-effects specification.) We do acknowledge the importance of the point. 2 (http://bls.gov/ooh/installation-maintenance-and-repair/automotive-service-technicians-and-mechanics.htm).

Please cite this article in press as: Smith, M.R., et al., Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’, Social Science Research (2017), http://dx.doi.org/10.1016/j.ssresearch.2017.03.003

4

M.R. Smith et al. / Social Science Research xxx (2017) 1e14

and baccalaureate programs can all be found (Clark et al., 2009: 155). Quebec's CEGEP system is different from all the others
, 1992). Inserted between secondary school and university, the largest part of its enrolment (Denis and Lipkin, 1972; Inchauspe is in two year general academic programs that may either lead directly to the labour market or university. A smaller proportion of CEGEP students is enrolled in three year vocational programs, similar to those offered in community colleges. At the most general level, then, we seek to identify persistent sources of differences in earnings by gender as well as sources of changes in the differences. Specifically, we focus on the possible role of occupational demand and of the domestic division of labour including, in the latter case, the effect of hours of paid work. As England makes clear, the mechanisms associated with gender differences in earnings tend to vary across educational levels. We focus on a relatively neglected level e the graduates of community colleges or community college like-institutions.

4. Data, methods, and research questions The master files of Statistics Canada's National Graduates Survey (NGS) allow us to address research questions related to the issues discussed above in a way that few other data sets can match. Because the research questions we address are intimately tied to the characteristics of the data we first discuss its characteristics and then the research questions and methods used to address them. The NGS collected panel data from consecutive cohorts of graduates of post-secondary institutions - colleges, universities, and trade schools. We use six cohorts: 1986, 1990, 1995, 2000, 2005, and 2010.3 Our sample is of young people, newly entering the labour market.4 For our purposes this has a number of advantages. One is that the effects of occupational demand are likely to show up most clearly for new entrants. Once hired, administrative procedures substantially determine earnings which, therefore, respond sluggishly to changes in demand (Polachek and Siebert, 1993: 271). The second is that there is reason to think that in many well-paid jobs employer demands for long work hours are disproportionately present in early careers. For each cohort a sample was drawn in the year of graduation and initial information on the graduates collected. Sample members were interviewed on labour market outcomes and other factors two, or exceptionally in the case of the 2010 cohort, three years after graduation. The first four cohorts were also interviewed five years after graduation; the last two were not interviewed a second time. We have, then, data on the early careers of cohorts spanning 22 years (1988e2013) that allow cross-sectional and panel analyses for the first four and cross-sectional analyses alone for the last two. We can compare early career gender earnings differences across cohorts and for the first four cohorts we can analyse within-cohort changes over three year periods. The total NGS sample is about 30,000. A number of exclusions appreciably reduce our usable sample size. First, we removed university graduates because our focus is on the intermediate, college, educational level. Second, we excluded trade school graduates because they were not always interviewed at the second interview point. Third, we excluded those living in the North or who migrated to the US. Their situations are generally very different from those living in the ten provinces but their numbers are too small to be included with controls. Because in our earnings models we want to assess the effects of gender for those fully engaged in the labour market we also exclude from the relevant tables those who were studying fulltime. There are other inclusions and exclusions that are table- and analysis-specific. We discuss those shortly.5 Appendix Table 1 provides a list of the variables we use. We are particularly interested in our respondents' fields of study, marital status, hours of paid work (a response to a question on usual hours of paid work per week), and whether or not they have children. Our covariates are age at graduation (which allows us to control for effects of delayed studies), whether or not the college graduate had previously acquired a university degree (some university graduates subsequently enrol in a college to acquire a vocational qualification), geographic mobility (some research suggests that a reluctance to be mobile reduces women's earnings), whether or not self-employed, and work experience. For all survey points there is a measure of months employed since graduation. For the second survey point (for the 1986 to 2000 cohorts only) there is also information on whether or not the respondent had been continuously employed with the same employer - tenure. Given the importance of work experience in earnings determination we included months employed in all our models and both variables for the four survey waves when they were available. That means that there is a difference in specification across the models. However, dropping the tenure variable has no appreciable effect on any of the coefficients that particularly interest us. As is the case in most studies of this kind, including one using the same data set (Boudarbat and Connolly, 2013), our dependent variable for most of our tables is the logarithm of earnings. However, rather than the hourly wage rate we use annual earnings. Annual earnings yield a larger gap than the hourly wage rate (Baker and Drolet, 2010). That is not our reason for focussing on them. Rather, total earnings imply more consumption; and we are interested in the effect of hours of work e both on the amount of earnings and changes in earnings. Annual earnings and their growth will respond to hours of work in the ways discussed in Section 2.

3

Two earlier cohorts were also sampled. Because of problems of comparability with the earlier samples we start with the 1986 one. There are some older graduates in the sample. But the overwhelming majority of the sample is young and we control for age. Fields of study are defined by Statistics Canada's Classification of Instructional Programs. Comparability across the analysis period required some recoding. 4 5

Please cite this article in press as: Smith, M.R., et al., Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’, Social Science Research (2017), http://dx.doi.org/10.1016/j.ssresearch.2017.03.003

M.R. Smith et al. / Social Science Research xxx (2017) 1e14

5

While earnings is the dependent variable in most of our analyses, in one table we focus on labour force status: employed, unemployed, or not in the labour force (NLF). Earnings tend to rise with work experience and education. Exiting the labour force for reasons other than education is likely to reduce subsequent earnings. For our earnings equations we exploit the distinct character of our data with both cross-sectional and dynamic analyses. The cross-sectional analyses use Oaxaca-Blinder decompositions (Blinder, 1973; Oaxaca, 1973). Our “two-fold” version of the procedure treats male coefficients as the non-discriminatory referent (Jann, 2008).6 An advantage of this approach is that it divides the gender difference in pay into explained and unexplained components by estimating separate models for men and women. This, in turn, provides information on the relative effect of different factors on the pay of each gender; it does not average coefficients across genders as would be the case if we used a gender dummy in single equations. A further advantage of the procedure is that it yields information on the contribution to the gap of both characteristics of respondents and the returns to those characteristics. Thus, for example, it provides estimates of the effect on the earnings gap of gender differences in hours of paid work as well as the size of any effect of hours. All p-values are two-tail Huber-White robust estimates (White, 1980). For the panel analyses we use hybrid models. Since our panel data only provide two data points for each cohort the inclusion of a case-specific dummy would substantially reduce the power of significance tests (Clark and Linzer, 2012: 7e8). This poses a problem for the use of fixed-effects specifications. Nor would a random-effects specification be satisfactory. Random-effects coefficients combine within- and between-case associations. But, for some of our explanatory variables, we want to separate these. We do so with a hybrid random-effects model. It includes terms that summarize the effect of the independent variables over the two points in time as well as selected variables that are introduced as change scores (Allison, 2005: 32e38). Specifically, we introduce marital status, children, and hours of work as both between- and within-effects. Note that Allison suggests that the between-effects estimated with these models should be treated with caution. In what follows we focus on within-effects. These coefficients are “identical to those for the classic fixed-effects estimates” (Allison, 2005: 33). One final methodological issue required addressing. For a young sample like this there is a great deal of educational mobility. Two years after graduation some sample members will have enrolled in university or another college program; five years after graduation they may have dropped out of their educational program and be employed. Or they may be employed at the first survey point and enrolled in an educational program at the second. In our panel analyses we are interested in trends in earnings between the two survey points. Consequently, those who were enrolled in an educational program at one or both of the survey points were excluded from the panel analyses. We briefly return to this in the Conclusion. We are able to take a different approach in the cross-sectional analyses of the first four cohorts, for each of which we have data for two survey points. First, we present analyses of those who were employed at both points after college graduation which reduced our number of cases. Then we present the same analyses for those who were only in the sample at the second data point. We do not include in this second analysis those who, following college, completed a university degree. We do include the appreciable number who started university after college but subsequently dropped out and failed to complete their degrees: about 15% of those enrolled in a university education abandon post-secondary education altogether; a further 15% switch to another educational level - mostly a return to college (Finnie and Qiu, 2008). By each second survey point a larger proportion of respondents is available so the analyses presented in Table 2 rest on more cases. Conversely, our panel analyses return to the smaller sample sizes of the cross-sections which are confined to those present at both sample points of each cohort. We would underline here that we attempt to maximize the robustness of our results by combining panel and crosssectional analyses and, for the latter, applying the analyses to differently constituted samples. We use these data and methods to address the following research questions. 1. England (2010) reports that in the US the earnings gap converged at the upper part of the educational distribution. Drolet (2011) reported that in Canada it converged lower down in the educational distribution. For this intermediate educational level, has the gender earnings gap converged, diverged, or remained the same? 2. It is recognized that shifts in occupational demand sometimes affect the gender earnings gap. We argue that the importance of this factor has been underestimated. The effect on the gender earnings gap of the tech-boom has been generally neglected. The possibility that the demand for university graduates may have spilled over to those with less education has not been considered at all. Is there evidence of occupational demand effects on the relative earnings of this community college or community college-like sample? 3. There is ample evidence of the effects on the gender earnings gap of the domestic division of labour. For this young sample we might expect the effects of marriage and childbirth to be modest or absent. But hours of work influence earnings directly and, sometimes, are required for career advancement, particularly at the beginning of a career. What evidence of effects of the domestic division of labour is there in this sample?

5. What happened to the earnings gap? Tables 1 and 2 provide estimates of the earnings gap across cohorts. Results for all cohorts are reported in Table 1. But the sample composition differs between the first four and the last two. The first four contain all those who were employed at both

6

We have also run the analyses using a three-fold, pooling, procedure. The results are not markedly changed.

Please cite this article in press as: Smith, M.R., et al., Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’, Social Science Research (2017), http://dx.doi.org/10.1016/j.ssresearch.2017.03.003

6

M.R. Smith et al. / Social Science Research xxx (2017) 1e14

Table 1 Raw differences in log earnings, difference explained by field of study and other controls: All college graduates. 1986 cohort

1990 cohort

1995 cohort

2000 cohort

2005 cohort

2010 cohort

1988

1992

1997

2002

2007

2013

1991

1995

2000

2005

Model 1: Raw Difference Model 2: Explained by field of study Model 3: Explained by hours, experience, tenure, marriage and children Model 4: Explained after all controlsa


0.176***
0.239***
0.138***
0.212***
0.311***
0.359***
0.209***
0.268***
0.196***
0.275*** 0.057
0.016
0.013
0.043**
0.058**
0.052**
0.055*
0.057*
0.085***
0.142***
0.030**
0.065***
0.050***
0.075***
0.128***
0.101***
0.078***
0.100***
0.100***
0.053** 0.021


0.072**
0.055**
0.101***
0.125***
0.117***
0.088***
0.119***
0.141***
0.148***

N

5174

5174

4299

4299

4609

4609

4032

4032

6384

4259

*p-value<0.05; **<0.01, ***<0.001. Exclusions include: unemployed or inactive at either time periods, residents of Yukon, North West Territories, and the United States, those who cite school as the reason for working less than 30 h per week and those with earnings less than $5000. Significance estimated with robust standard errors. Earnings adjusted to 2005 dollars. a Controls are: province of residence; field of study; previous post-secondary education; experience in months since graduation; continuoiusly employed with same employer between surveys; hours worked in ref week; marital status; age (quadratic); children; mobility since graduation or interprovincial mobility between interviews.

Table 2 Raw differences in log earnings, difference explained by field of study and other controls: All college graduates.

Model Model Model Model N

1: 2: 3: 4:

Raw Difference Explained by field of study Explained by hours worked, experience, tenure, marriage and children Explained after all controlsa

1986 cohort

1990 cohort

1995 cohort

2000 cohort

1991

1995

2000

2005


0.254***
0.009
0.089***
0.082***


0.219***
0.014
0.074***
0.091***


0.355***
0.036*
0.117***
0.127***


0.274***
0.061*
0.118***
0.147***

6396

5207

5677

4679

*p-value<0.05; **<0.01, ***<0.001. Exclusions include: unemployed or inactive at follow-up survey, residents of Yukon, North West Territories, and the United States at follow-up survey, those who cite school as the reason for working less than 30 h per week at follow-up and those with earnings less than $5000 at follow-up survey. Significance estimated with robust standard errors. Earnings adjusted to 2005 dollars. a Controls are: province of residence; field of study; prevous post-secondary education; experience in months since graduation; continuoiusly employed with same employer between surveys; hours worked in ref week; marital status; age (quadratic); children and interprovincial mobility between interviews.

panel interview points. For the last two cohorts only cross-sectional information was collected. Consequently, we do not know who would have been excluded had a second panel survey taken place. That complicates comparisons between the first four cohorts and the last two; still, the broad trends in the table are, we would argue, striking, and the trends remain the same if only first survey respondents are included (tables available on request). Note that Tables 1 and 2 contain: i) raw differences by gender; ii) the amount of the differences explained by field of study without controls for other variables; iii) the amount of variance explained by a set of possible domestic division of labour indicators without controls for other variables; iv) the total explained by all variables in the models. To produce correct interpretations from a semi-log specification of percentage effects we adjust the coefficients in the tables using the formula in Thornton and Innes (1989).7 Focussing on the first survey points for the first four cohorts and the single survey points for the last two, the first line of Table 1 shows that women earned 16% less than men in 1988, this had fallen to 13% by 1992, then the difference rose abruptly to 27% in 1997, falling to 19% in 2002. The rather differently constituted sample in 2007 produced about the same difference as in 2002, then a 24% gap in 2013. As we saw in Section 1, in aggregate, gender earnings differences tended to fall during the 1980s then approximately stabilized. The results from out series diverge from this pattern. Because the sample size and composition of the cases in Table 1 are influenced by the fact that respondents have to be employed at two sample points, Table 2 tests for robustness It uses the larger sample of all those employed at the second survey point. The results are very similar to those in Table 1. The gap fell from 1991 to 1995, then increased substantially in 2000, falling back in 2005 but remaining at a higher level than had been the case for the two earlier cohorts, five years after graduation. Table 1 contains another interesting pattern. For the four first panels for which data at two points in time are available, the earnings disadvantage was always larger at the second than the first survey point. (All the differences are significant.) We return to this in the Conclusion. In answer to our first research question, then, our results suggest that from 1988 to 2013 the gender earnings gap displayed all possible patterns of change: it narrowed then widened, then fluctuated at a level between the crest of 1997 and the trough of 1992.

7

The true proportional change ¼ exp(bDX)-1 where bDX refers to the estimated coefficient multiplied by a unit change in the independent variable.

Please cite this article in press as: Smith, M.R., et al., Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’, Social Science Research (2017), http://dx.doi.org/10.1016/j.ssresearch.2017.03.003

M.R. Smith et al. / Social Science Research xxx (2017) 1e14

7

6. Occupational demand and the earnings gap Our second research question asks, did occupational demand play a role in shifting the earnings gap? Together, Tables 1 and 2 suggest that it did. The second row of Table 1 shows that field of study did not contribute to the gender earnings gap of the 1986 cohort or of the first survey point of the second cohort. Thereafter it consistently accounts for part of the earnings gap, varying between four and six percentage points between 1995 and 2005, and for a much larger share for the two, nonpanel, cohorts. The pattern in Table 2 is similar, though the contribution only becomes significant for the 1995 and 2000 cohorts. This, of course, raises the question, which fields of study produced the outcome. Table 3 presents a complete analysis of the sources of earnings disadvantage. The upper half describes the effects on the earnings gap of the characteristics of men and women, the lower half the effect of differences in returns to those characteristics. Focus first on the field of study effects in the upper half of the table. There are two striking sets of results. One is for the category that contains Health. These occupations significantly reduce the female disadvantage in most years. The 2000 cohort provides the only exceptions. It is also striking that the sizes of the coefficients fall substantially from the first two cohorts to the last two - illustratively, from 9% in 1988 to less than 4% in 2007. Note that in the early part of the period community colleges provided the main training for nurses, a relatively well-paid, predominantly female, occupation. By the end of the period access to nursing increasingly required a university degree. So some well-paid jobs were lost from the community college category. Nonetheless, community colleges train for a large number of health-related occupations that are reasonably well-paid e hence the persistent though smaller contribution to reducing the gender earnings gap. The second set covers two categories: Mathematics, Computer and Information Sciences, and Architecture, Engineering and Related Technologies. Where significant these always increase the gender earnings gap. The Mathematics, etc. category only does so for the 1990 cohort; across cohorts the Engineering category uniformly does so in at least one year except for the 2000 cohort. How do these non-university programs increase earnings? They are designed to train what can best be described as technicians, whether or not that term appears in the program title. For example, there are programs in electrical engineering and in information management e the former with technician in the title, the latter without. Now turn to the bottom half of the table which describes the returns on characteristics. Few of the fields of study coefficients are significant: Health in 1995 (when the returns to women were higher) and in 2002 (when they were lower); Computer and Information Sciences in 2002 and 2007 and Engineering in 2002, all of which involved lower returns to women. The mid to late 1990s, the tech-boom years, appear not to have disadvantaged those women with diplomas in the relevant fields though they were disadvantaged in some other years. The contribution of the Health sector to reducing the gender gap and of Engineering and (less consistently) Mathematics and Computer and Information Sciences to increasing it are products of occupational demand. Those who took pre-university programs then went directly into the labour market (probably often in Arts and in Life Sciences) and those in fields leading to less well-paid occupations earned less than those in health sciences and in engineering- or computing-related fields. Beyond this, the coefficients shift over time. The main shifts are the following: health science reduced the earnings difference in most years but the reduction was smaller at the end of the period than at the beginning. Engineering contributed to the gap in most years, computer science in a couple. The engineering contribution is largest for the two last cohorts. The increasing importance of the engineering-related fields in the gender earnings gap does not precisely coincide with the tech-boom in the late 1990s. But we would not expect as clear an effect of the tech-boom on the earnings of community college graduates as on those of university graduates. Note, also, that all the engineering and computer science coefficients for characteristics became insignificant for the 2000 cohort e the one that confronted the tech bust. Interestingly, however, there is some evidence from the lower half of the table that the returns to those fields were lower for women than for men. Why that would have been the case is not clear. Overall, we provide strong evidence that occupational demand shifts influenced changes in the gender earnings gap, that the shifts were broadly consistent with an effect of a tech-boom and subsequent tech bust, and with a change in the composition of health-related occupations demanded e specifically, by the switch to requiring university training for nursing. 7. The domestic division of labour As in the case of occupational demand, we examine the effects of a group of factors (marriage, children, experience, tenure, hours of paid work) first as an aggregated effect in Tables 1 and 2 and then as separate component effects in Table 3. Again, in Tables 1 and 2 this cluster of factors is analysed without controlling other variables. Tables 1 and 2 show that the aggregate effects of marriage, children, and hours of work always significantly contribute to the gender earnings gap and, with the exception of the last two cohorts (2005, 2010), always more substantially than occupational demand. An interesting feature is that the effects of this cluster of factors is larger after 1995 than before. This raises the possibility that the effects of shifts in occupational demand may not only directly influence expectations but also the willingness of employers to adapt to family responsibilities. Family-congenial jobs were probably more likely in health care than tech companies. Table 3 shows the associations between earnings and the components of this cluster. Again, the upper half contains the effect of differences in characteristics, the lower half in the returns to characteristics. In terms of differences in characteristics, marital status never contributed to the earnings gap. With the exception of a small contribution in 2007, nor did having children. But hours of work always contributed substantially. The size of the contribution peaked in 1997 and fell a bit but remained high across subsequent surveys until 2013 when it fell back to a pre-1995-like level. We have separately examined Please cite this article in press as: Smith, M.R., et al., Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’, Social Science Research (2017), http://dx.doi.org/10.1016/j.ssresearch.2017.03.003

8

M.R. Smith et al. / Social Science Research xxx (2017) 1e14

Table 3 Share of the gender wage gap explained by specified characteristics in fully-specified modela: All college graduates. 1986 cohort

1990 cohort

1995 cohort

2000 cohort

2005 cohort

2010 cohort

1988

1992

1997

2002

2007

2013

1991

1995

2000

2005

Total Log Annual Wage Gap


0.176***
0.239***
0.138***
0.212***
0.311***
0.359***
0.209***
0.268***
0.196***
0.275***

Attribuable to Differences in Characteristics Education Arts Business, Management and Public Administration Physical and Life Sciences & Technologies Mathematics, Computer and Information Sciences Architecture, Engineering and Related Technologies Agriculture, Natural Resources, and Conservation Health, Parks, Recreation, and Fitness Age Age^2 Prev. post-secondary education under BA Prev. university degree Region Geographic mobilty Married/common-law Children Self-employed Hours worked Experience Tenure Attribuable to Differences in Returns to Characteristics Education Arts Business, Management and Public Administration Physical and Life Sciences & Technologies Mathematics, Computer and Information Sciences Architecture, Engineering and Related Technologies Agriculture, Natural Resources, and Conservation Health, Parks, Recreation, and Fitness Age Age^2 Prev. post-secondary education under BA Prev. university degree Region Geographic mobilty Married/common-law Children Self-employed Hours worked Experience Tenure Constant

0.021


0.072**
0.055**
0.101***
0.125***
0.117***
0.088***
0.119***
0.141***
0.148***


0.004 0.000
0.001


0.005 0.000
0.003


0.001
0.002
0.005


0.006*
0.001
0.002 0.000
0.011**
0.004


0.006 0.000
0.001


0.012
0.001
0.005


0.013
0.002
0.002

0.000 0.000 0.011**

0.003 0.000 0.003


0.002

0.000


0.001


0.001


0.002


0.002

0.000

0.001

0.000


0.001


0.003


0.003


0.011***
0.007**
0.003


0.002


0.001


0.004


0.003


0.004


0.029


0.052*


0.038*


0.030*


0.022


0.046**
0.008


0.015


0.100***
0.141***

0.006

0.006*

0.006*

0.007**

0.003

0.005

0.011**

0.002

0.002

0.000

0.093***
0.010 0.001 0.000

0.060***
0.006 0.001 0.000

0.069*** 0.000 0.003
0.001

0.046*** 0.011*
0.012*
0.001

0.026* 0.054*
0.046* 0.004

0.034*** 0.069**
0.066** 0.001

0.010 0.061
0.061* 0.000

0.011 0.038
0.042 0.001

0.037*** 0.070**
0.064** 0.003

0.036** 0.025
0.020 0.004

0.000 0.001 0.000 0.001
0.002 0.002
0.035*** 0.001 N/A
0.197***

0.001
0.001 0.000
0.001
0.002 0.005
0.073*** 0.001 0.000
0.167***

0.000 0.000 0.000 0.003
0.002 0.001
0.079*** 0.006* N/A
0.083***

0.000
0.003 0.000 0.001
0.002 0.006*
0.111*** 0.005* 0.008***
0.111***

0.002 0.007**
0.001 0.002
0.001 0.002
0.142***
0.001 N/A
0.186***

0.002 0.004
0.002
0.001
0.002 0.005
0.109*** 0.000 0.000
0.242***

0.005 0.004
0.001 0.000
0.003 0.002
0.095*** 0.009 N/A
0.121***

0.006 0.005 0.000 0.003
0.003 0.000
0.121*** 0.007 0.010*
0.150***

0.006 0.007*** 0.000 0.000
0.010** 0.003
0.112*** 0.008* N/A
0.054*

0.011 0.001 0.000 0.004
0.008 0.000
0.075*** 0.015* N/A
0.127***


0.001 0.007
0.011


0.001
0.002
0.021


0.003 0.006
0.018


0.003 0.007
0.014

0.002
0.001
0.007

0.000
0.007
0.006


0.004
0.005
0.043*

0.000 0.002
0.018


0.001 0.002
0.011


0.002 0.008
0.023

0.000


0.003


0.001


0.001

0.004*

0.002


0.003

0.000


0.001

0.000

0.005


0.004

0.010

0.006


0.011


0.014


0.037*


0.003


0.019**
0.003


0.019


0.024

0.015

0.025


0.020

0.004


0.100*


0.032

0.015

0.014


0.009


0.011*


0.006


0.004


0.004


0.005


0.020*

0.005


0.003


0.001

0.008 0.121
0.066
0.007

0.001
0.131 0.059
0.016*

0.008
0.821* 0.313 0.004

0.015** 0.119
0.038 0.001

0.002
0.917* 0.384* 0.006

0.006 0.171 0.022
0.002


0.014*
0.212 0.119
0.007


0.007 0.305
0.126 0.005

0.003 0.072
0.046 0.000

0.005 0.729
0.295 0.004

0.000
0.031*
0.001
0.007
0.018***
0.009 0.184
0.049 N/A
0.294

0.001 0.003
0.002
0.054***
0.016*
0.012 0.131
0.147*
0.011 0.093

0.000
0.026
0.003
0.033***
0.005 0.003 0.293**
0.032 N/A 0.214

0.000
0.017
0.001
0.032*
0.022**
0.001 0.441*** 0.019 0.022
0.632**

0.007 0.020 0.000
0.010
0.013 0.001 0.227
0.041 N/A 0.185

0.006 0.010 0.001
0.071***
0.023
0.007 0.267 0.000 0.015
0.608*

0.024** 0.001
0.001
0.007
0.025
0.002 0.141
0.174 N/A 0.248

0.020*
0.013
0.002
0.007
0.037 0.011 0.485** 0.027 0.033
0.796*

0.005 0.010
0.004
0.022
0.017
0.006 0.279*
0.100* N/A
0.211

0.016
0.015
0.006
0.012
0.035* 0.001 0.008 0.037 N/A
0.557

5174

5174

4299

4299

4609

4609

4032

4032

6384

4259

N

*p-value<0.05; **<0.01, ***<0.001. Exclusions include: unemployed or inactive at either time periods, residents of Yukon, North West Territories, and the United States, those who cite school as the reason for working less than 30 hours per week and those with earnings less than $5000. Significance estimated with robust standard errors. Earnings adjusted to 2005 dollars. a Controls are: province of residence; field of study; prevous post-secondary education; experience in months since graduation; continuoiusly employed with same employer between surveys; hours worked in ref week; marital status; age (quadratic); children; mobility since graduation or interprovincial mobility between interviews.

Please cite this article in press as: Smith, M.R., et al., Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’, Social Science Research (2017), http://dx.doi.org/10.1016/j.ssresearch.2017.03.003

M.R. Smith et al. / Social Science Research xxx (2017) 1e14

9

the distributions of hours of paid work by gender. For reasons of space we do not present a descriptive table here. The important point is not that men report, on average, three or four more hours of paid work per week than women (see Appendix Table e note that number of hours was not available for the 2010 cohort so we replace hours with whether working part-time or full-time). Underlying this average difference is the fact that women were much more likely to report working part-time and men much more likely to report working very long hours (say, 50 hours per week or more - Cha, 2013, provides evidence on differences in ‘overwork’ by gender in the US.) Even in this young sample, a pattern is present in which men accumulate more paid hours of work which either directly leads to higher earnings or is required to increase the likelihood of career success. We think that this difference in hours worked probably has something to do with the domestic division of labour. The lower half of the table e returns to characteristics - is more interesting for these variables than it was for occupational demand. First, it provides evidence that in a number of years, particularly at the beginning of the period, women's earnings were more adversely affected than men's by marriage and having children. In most years the relationship is insignificant. But, where significant, the association always implies that women were disadvantaged. Second, it reveals that, in 4 out of 10 years for which the associations were significant, hours worked tended to increase women's earnings by more than men's. Working fewer hours reduced the relative earnings of women; those women who worked longer hours were paid as much or more than men working similarly long hours. Table 5 will reveal something similar. The domestic division of labour can lead to shorter hours worked. It can also influence labour force status which, in turn, affects future earnings. Consider, in this respect, Table 4. It shows the effect of gender on labour force status. Broadly speaking, for a young sample like this employment and education tend to lead to increased earnings since both are a source of human capital. Unemployment and absence from the labour force for reasons other than education (“residual NLF”) will not normally contribute to earnings growth. Unemployment, however, requires job search which implies an effort to find employment. The table provides the log odds ratios of being in two alternative labour force statuses: employment or education versus being unemployed or out of the labour force for reasons other than education, and NLF versus employment and unemployment. The upper panel includes only male as a predictor. The lower panel describes the association after the addition of controls. Most of this young sample were in employment or education at the time of the survey. The small numbers NLF for reasons other than education reduces the likelihood of significance. Bearing that in mind, the first line shows, for the two years that are significant both with and without controls, men were appreciably more likely than women to be in employment or education. The second line estimates the probability of being NLF (for whatever reason) as opposed to employed or unemployed. This yields a significantly lower probability of male NLF in four years in each panel. From inspection of the data we know that women were no less likely than men to be in education and no more likely than them to unemployed. The difference originates in the fact that women were much more likely (even when the results are insignificant) to be NLF for other reasons including family responsibilities (figures available on request). Overall, we would argue that Tables 1e4 provide strong evidence consistent with an effect of the domestic division of labour e but an effect which at these young ages manifests itself through marriage and childcare to a limited extent. Rather, differences in paid hours of work matter greatly in producing the gender earnings gap. And, in terms of longer term effects, we already see evidence of absences from employment likely to damage future earnings. In Table 5 we further explore the effects of the domestic division of labour, along with field of study, using the four panels, two observations for each panel. Estimating hybrid random effects models yields estimates of the effects of changes in some of our variables of interest - marital status, the presence of children, and hours of work. We run the analyses separately for men and women. Earlier tables revealed evidence of only modest and erratic aggregate effects on the earnings gap of either marital status or children. We think that this is because of the relative youth of our sample. However, the hybrid model with its change scores yields stronger evidence of childbirth effects. Having a child is always associated with a fall in female earnings; it is never associated with a fall in male earnings. In two years, in fact, it is associated with an increase. Now look at the within coefficients for hours. They show that those whose hours of paid work increased also experienced larger earnings increases. Remember that Table 3 showed that, where significant, the return on extra hours of work was larger for women than men. Consistent with this, but more strongly so, Table 5 reveals that the increase in hours of work contribution to earnings growth was larger for women than for men. Nonetheless, this association between hours growth and earnings growth would tend to increase the gender earnings gap because men not only worked more hours than women; more men increased their hours of work between panel interview points. Our hybrid model, then, provides further evidence suggesting effects of the domestic division of labour at young ages. The effects of having children provide a direct measure of these effects. We would argue that differences in paid hours of work can also most plausibly be interpreted similarly. 8. Discussion and conclusion The research reported in this paper rests on three premises. First, in the explanation of trends in the gender earnings gap one factor has been insufficiently considered: occupational demand. This is not to say that it has been ignored. Rather, we think that its importance has been underestimated and its implications insufficiently considered. Second, while the domestic division of labour has a substantial presence in the relevant research the measurement of its effect has been mostly confined to family statuses. Married people tend to earn more but the premium to men is higher. Children reduce women's salaries but not men's. In addition to these clearly important factors how much people earn is influenced by hours worked - both directly Please cite this article in press as: Smith, M.R., et al., Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’, Social Science Research (2017), http://dx.doi.org/10.1016/j.ssresearch.2017.03.003

10

M.R. Smith et al. / Social Science Research xxx (2017) 1e14

Table 4 Logistic regression: Gender and post graduation labour force status - employed versus unemployed and not in the labour force; not in the labour force versus employed and unemployed of those completing a college diploma. 1986 cohort

1990 cohort

1995 cohort

2000 cohort

1988

1992

1995

1997

2000

2002

2005

2007

2013

Emp vs Unemp þ NLF 1.004 0.860 0.998 Emp þ Edu vs Unemp þ residual NLF 1.034 1.011 0.999 NLF vs Emp þ Unemp 0.750 0.816 0.537*** Male þ controlsa Emp vs Unemp þ NLF 1.125 0.901 1.214 Emp þ Edu vs Unemp þ residual NLF 1.168 1.087 1.197 NLF vs Emp þ Unemp 0.785 0.934 0.495**

1.372** 1.414*** 0.370*** 1.455** 1.445** 0.367***

1.142 1.210 0.674** 1.074 1.076 0.713

1.846*** 2.039*** 0.403*** 1.562** 1.673** 0.510**

0.706* 0.716* 1.221 0.889 0.871 0.968

1.035 1.092 0.679 1.074 1.098 0.561*

0.976 1.017 0.734 1.071 1.110 0.736

1.007 1.021 0.738 1.014 1.021 0.837

N

7204

7525

7489

6788

6763

9520

6603

1991

Male

10311 10266 7310

2005 cohort 2010 cohort

*p-value<0.05; **<0.01, ***<0.001. Exclusions include: residents of Yukon, North West Territories, and the United States. Significance estimated with robust standard errors. a Controls are: province of residence, field of study, prevous post-secondary education, marital status, age (quadratic), children, mobility since graduation or interprovincial mobility between surveys.

(there may be pay for overtime) and indirectly (some lucrative careers require long working hours, particularly in their early years). Hours worked, we argue, is a factor contributing to gender differences in earnings levels and, possibly, trends in them. Third, existing research suggests strongly that further progress in understanding the sources of gender earnings differences will gain from close examination of different educational levels. The background to our analyses is a body of research which has shown that the gender earnings gap fell during the 1980s and into the early 1990s, then stalled and remained at quite a high level. Our results qualify these conclusions for those receiving a middle level education in Canada. Consistent with the relevant research the gap for our sample at first fell; inconsistent with that research the gap then rose appreciably from the early to the late 1990s, then it fell a bit but remained higher than it had been in the early 1990s. It is true that the results in our earnings models for the first four cohorts exclude those who were out of employment at one of the two survey points. If genders were differentially likely to be out of employment and if those out of employment at one survey point differ from those in employment that would influence our results. The fact that Table 4 showed that women were more likely to be residually NLF is relevant here. Nonetheless, we do not think that this changes one of our main conclusions: occupational demand shifts contributed to changes in the gender earnings gap. The rise in the gap approximately coincided with the tech-boom and declined somewhat after the bust, and tech-boom-related fields of study e engineering and computer science - explained a significant part of the gap in the later cohorts. This, in our view, is strong evidence of an occupational demand effect. This might, at first, surprise. After all, the jobs most closely associated with the tech-boom are computer scientist and electrical engineer. These are occupations produced by universities rather than community colleges. But there is the possibility, and our results suggest the reality, of a derived demand. Having hired computer scientists and electrical engineers, firms are obliged to provide appropriate support to their work in the form of community college-educated technicians. That seems to have happened in Canada from the mid-1990s. Our domestic division of labour effects might also surprise. Our cross-sectional earnings models revealed some but modest adverse effects of marriage and children on earnings. But other results suggest something different. First, for some of our cohorts women were more likely to be NLF for reasons other than education. The range of motivations within this category would include looking after a child for the modest proportion in our sample who had one and moving to accommodate a partner's career. Either is a form of the domestic division of labour, broadly construed. Neither is likely to favour earnings growth. Second, when entered as a change effect in the hybrid model having children always reduced women's earnings but not men's. Third, Table 3 revealed, for the early years at least, that a negative return to marriage and childcare contributed to the gender earnings gap. Fourth, and this we would wish to underline, hours of work was strongly cross-sectionally associated with earnings and contributed to the earnings difference. Just as importantly, it was equally strongly associated with growth in earnings. Working longer hours caused both men's and women's earnings to increase more; but more men worked longer hours and fewer worked part-time so the effect of hours on earnings growth tended to increase the gender earnings gap. We have difficulty imagining a reason other than the domestic division of labour for gender differences in hours worked in this young sample. Certainly that is the interpretation of hours effects in Cha (2013) and Cha and Weeden (2014). What about the focus on a single level of the Canadian educational system e community colleges? We know that community colleges produce graduates destined for a range of jobs likely to differ from those of university graduates. We also know that trends in the gender earnings gap have differed across educational levels. Narrowing in Canada, Drolet (2011) informed us, was concentrated in the lower part of the educational distribution. Some of that narrowing was caused by a decline in the demand for a range of jobs e many associated with manufacturing and mining e primarily occupied by males. At the other end of the distribution women have moved into university programs leading to lucrative careers. The processes influencing the gender earnings gap, then, have differed at the two ends of the earnings distribution: declining demand for male-dominated occupations that required little education but paid quite well at the lower end; switching by females into educational programs and occupations previously mainly occupied by men at the upper end. This leaves open the question: what has been happening in the community colleges at the middle of the distribution, where a significant proportion of Please cite this article in press as: Smith, M.R., et al., Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’, Social Science Research (2017), http://dx.doi.org/10.1016/j.ssresearch.2017.03.003

M.R. Smith et al. / Social Science Research xxx (2017) 1e14

11

Table 5 Hybrid Model of log earnings regressed on select independent variables and controls: All college graduates. 1986 cohort

1990 cohort

Men

Women

Men

Women

Men

1995 cohort Women

Men

2000 cohort Women

Year (follow-up survey)

0.172***

0.138***

0.168***

0.128***

0.254***

0.207***

0.182***

0.139***

Education Arts Business, management and public administration Physical and life sciences and technologies Mathematics, computer and information sciences Architecture, engineering and related technologies Agriculture, natural resources and conservation Health, parks, recreation and fitness

0.017
0.097 0.082* 0.123** 0.183*** 0.149*** 0.016 0.194***


0.136***
0.087
0.021 0.081 0.174*** 0.124***
0.160** 0.278***

0.038
0.243***
0.001 0.132** 0.069* 0.038
0.070* 0.136***


0.034
0.021 0.006 0.132** 0.212*** 0.115**
0.055 0.372***


0.131**
0.019
0.020 0.095 0.229*** 0.105*** 0.015 0.071


0.094***
0.050
0.017 0.261*** 0.127*** 0.094**
0.047 0.240***


0.012
0.088 0.068 0.044 0.178*** 0.162***
0.059 0.119*


0.184***
0.089*
0.016 0.092 0.018 0.135***
0.177*** 0.107***

Age Age^2 Previous post-secondary education under Bachelor' Previous post-secondary education above Bachelor's Interprovincial mobility Self-employed Experience Tenure

0.037**
0.001* 0.049*** 0.045 0.018 0.045 0.010*** 0.096***

0.027***
0.000** 0.002 0.072* 0.021
0.167* 0.007*** 0.083***

0.022**
0.000* 0.066** 0.042 0.047
0.145** 0.004*** 0.110***

0.006 0.000 0.071** 0.059*
0.015
0.167** 0.004*** 0.132***

0.048***
0.001*** 0.049 0.052 0.033
0.084* 0.003*** 0.072***

0.047***
0.001*** 0.053** 0.145*** 0.047*
0.143* 0.002** 0.124***

0.047***
0.001*** 0.061* 0.106** 0.053
0.137** 0.012*** 0.128***

0.046***
0.001*** 0.084*** 0.232*** 0.023
0.053 0.012*** 0.163***

Hours (between effect) Hours (within effects) Married/Common (between effect) Married/Common (within effect) Children (between effect) Children (within effect)

0.009*** 0.008*** 0.078*** 0.030* 0.032 0.014

0.015*** 0.014*** 0.027
0.003
0.059**
0.112***

0.010*** 0.007*** 0.121*** 0.002
0.008 0.013

0.022*** 0.015*** 0.008 0.004
0.034
0.081***

0.014*** 0.011*** 0.100*** 0.021 0.053* 0.040*

0.025*** 0.023*** 0.017
0.011
0.064***
0.047*

0.016*** 0.009*** 0.086*** 0.025 0.055 0.068*

0.024*** 0.022*** 0.036* 0.035*
0.053*
0.057***

Constant R-sq: overall

9.136*** 0.220

9.017*** 0.298

9.469*** 0.215

9.109*** 0.351

8.935*** 0.335

8.346*** 0.425

8.591*** 0.302

8.211*** 0.373

N

5260

5088

4118

4480

4644

4574

3336

4728

young Canadians complete their educations, particularly women? Evidently, since we focus on only one level we cannot speak to differences in trends across levels. But we can say that, within the middle level we have studied, the overall trend in the gender earnings gap does not match the fall-stagnate pattern reported for the aggregate. We can also say that occupational demand shifts matter at this level in a way that one might not have expected. Further research on other levels looks promising. A first implication of these results has to do with inferences about trends in the gender earnings gap. Much of the writing on the subject explicitly (e.g., Drolet and Mumford, 2012) or implicitly (e.g., England et al., 2010) assumes that those trends were produced by broad institutional changes e leading to educational choice changes, access to childcare, employer decision-making, and so on. No doubt those broad institutional changes have been important. But our results suggest that interpreting trends also requires paying attention to occupational demand e which, for any given occupation, can both rise and fall. There is strong evidence of a larger economic process of skill-biased technological change (e.g., Autor et al., 2003), substantially produced by the spread in the use of information technology. That bias has had two consequences. One is an increase in earnings inequality, as the skills for which demand has risen have tended to be associated with the university educated, who already on average earned more. Another is a rise in the gender earnings gap because men continue to disproportionately train in STEM areas (though the difference is falling). It follows from this that we should be cautious in inferring institutional changes from trends in the gender earnings gap. Shifts in occupational demand have no doubt influenced the gap. Increased public and para-public sector employment would have shrunk it in the past. Declining or stagnant public and para-public sector employment tends to reverse that trend and rising demand for those with computer and electronics skills has made a particularly large contribution to increasing the gap in the sample studied here (and most likely even more so among university graduates). Drawing any conclusions about whether the earnings gap is narrowing and why requires attention to occupational demand effects. The second implication of our results is related to the age of our sample. In our young, moderately educated, sample there is already evidence of domestic division of labour-related damage to earnings. A minority of women, but a higher proportion than men, exited the labour market because of domestic responsibilities. The hybrid model produced strong evidence of earnings disadvantage for women associated with childcare. Above all, both the cross-sectional and hybrid models show that women's earnings level and the growth in their earnings were adversely affected by the fact they were more likely to work part-time and less likely to work the very long hours required by some high-paid jobs. We acknowledge that we do not present direct evidence connecting paid hours worked to the domestic division of labour. But in the absence of a plausible alternative explanation we are inclined to think that the difference is produced by a domestic division of labour (whether couples are married or not) that disfavours women.

Please cite this article in press as: Smith, M.R., et al., Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’, Social Science Research (2017), http://dx.doi.org/10.1016/j.ssresearch.2017.03.003

12

M.R. Smith et al. / Social Science Research xxx (2017) 1e14

We would also emphasize that our results for people in their early 20s are likely to influence earnings later in life. Periods out of employment or education mean less human capital. Working part-time rarely does anything good for a career. Being a high-flyer in law, accounting, management consulting or, for that matter, academe often requires very long, early career, hours of work. This suggests a process of cumulation of disadvantage. The idea that a cumulation of effects may be important has been identified by a number of researchers (e.g., DiPrete and Eirich, 2006; Ferraro et al., 2008). Budig and England (2001) provided evidence of such cumulation. They showed that women's earnings fell with the number of maternity-related career interruptions. Their data came from the National Longitudinal Survey of Youth. This meant that their cases fell between the ages of 14 and 21 in 1979; the end year for their analysis was 1993 so their cases would have been between 27 and 34. Community college is usually completed by the mid 20s so most of their sample would have been several years older than most of ours. (Though, interestingly, the Appendix table shows that the average age of graduation rose appreciably across cohorts.) Many of the women in our sample would have later experienced the career interruptions the effects of which were documented by Budig and England. For some, the effects of those career interruptions would be added to the early career effects we have identified. Moreover, return to the fact that Table 1 showed that in the cohorts for which we have two data points the gender earnings gap was always larger at the second survey point than the first. All these within-cohort differences are significant. Even within the three years separating the two survey points the earnings gap had grown. The cumulation of disadvantage starts early. Why might cumulation matter? Substantively, it means that the gender earnings gap is likely to reflect a series of sources of disadvantage over a life course. This means that the earnings of women who, controlling for the usual sorts of things included in models, seem particularly disadvantaged often probably reflect a lifetime's accumulation of negative earnings shocks. Methodologically, there is a high probability that cross-sectional surveys used to estimate female earnings disadvantage will often fail to measure early experienced factors. Even if the survey used includes questions on hours worked, how likely is it, for example, that respondents of 50 accurately recollect the number of hours they worked 25 years earlier? This in turn has implications for the interpretation of residuals. It has been common to assign unexplained gender differences in earnings to discrimination (Blau and Kahn, 2006; Pfeiffer and Sohr, 2009; Weinberger, 2011; Castegnetti and Rosti, 2013). Our results suggest that they are often likely to have originated in some earlier, unmeasured, source of disadvantage. That certainly does not preclude a discrimination interpretation. After all, one can treat the differential socialization and broader family pressure on women as discrimination as England (2006: 253) has clearly done. Recent or current labour force absence or part-time work are, or where possible should be, included in models. But part of the residual is likely to include earlier unmeasured examples of these effects. Treating a residual as some unspecified estimate of discrimination may mislead. There is a range of ways in which the unexplained difference may have originated. Finally, there is the issue of levels of education. We will not rehearse the reasons that the community college level is interesting; we discussed that in some detail earlier. Here we suggest that exploring the sources of earnings disadvantage at a particular level raises questions that provoke research at other levels. We found evidence of domestic division of labour effects at this level. But we know that the decision to marry and have children is influenced by earnings prospects. Better earnings prospects tend to delay both (Xie et al., 2003). One might, then, expect to see a smaller effect of the sorts of domestic division of labour-related factors among recent university graduates and larger effects among those leaving high school. This may in turn mean that the cumulation of disadvantage is inversely related to the level of education. That is a question for future research. Funding This research was funded by the Social Sciences and Humanities Research Council of Canada (435-2013-1042) and used data made available through Statistics Canada's Research Data Centre program. Acknowledgements Enormously helpful suggestions were made by a reviewer. Appendix Table 1 Sample description. 1986 cohort 1990 cohort 1995 cohort 2000 cohort 2005 cohort 2010 cohort

Men Education Arts Social and behavioral sciences, law and humanities Business, management and public Administration Physical and life sciences & technologies Mathematics, computer and information sciences

1988 1991 1992 1995 1997 2000 2002 2005 2007

2013

0.87 4.92 7.49 19.10 2.43 8.76

0.73 5.41 5.90 19.95 1.16 5.20

2.09 3.97 12.53 20.31 2.11 9.63

1.88 5.70 14.09 18.94 1.38 8.44

1.19 4.60 5.43 18.56 1.25 13.55

0.63 7.77 6.81 20.47 1.11 9.41

Please cite this article in press as: Smith, M.R., et al., Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’, Social Science Research (2017), http://dx.doi.org/10.1016/j.ssresearch.2017.03.003

M.R. Smith et al. / Social Science Research xxx (2017) 1e14

13

Table 1 (continued ) 1986 cohort 1990 cohort 1995 cohort 2000 cohort 2005 cohort 2010 cohort

Architecture, engineering and related technologies Agriculture, natural resources and conservation Health, parks, and fitness Age at graduation (mean) No previous education Previous post-secondary education under bachelor's degree Previous university degree Experience in months since graduation and first interview Continuously employed with the same employer between interviews Geographic mobility Married/common-law Have children Self-employed Hours worked (mean)/Full-time in 2010 cohort (%) N Women Education Arts Social and behavioral sciences, law and humanities Business, management and public Administration Physical and life sciences & technologies Mathematics, computer and information sciences Architecture, engineering and related technologies Agriculture, natural resources and conservation Health, parks, and fitness Age at graduation (mean) No previous education Previous post-secondary education under bachelor's degree Previous university degree Experience in months since graduation and first interview Continuously employed with the same employer between interviews Geographic mobility Married/common-law Have children Self-employed Hours worked (mean)/Full-time in 2010 cohort (%) N

1988 1991 1992 1995 1997 2000 2002 2005 2007

2013

44.59 5.96 5.87 23.18 74.28 23.15 2.58 21.46 54.46 4.58 27.01 11.62 4.54 42.39 2630

33.12 6.83 9.42 24.29 83.84 11.85 4.31 20.59 62.88 4.24 3.74 52.01 31.99 24.50 15.77 6.11 3.73 41.61 40.71 2059

36.90 7.23 5.43 25.00 85.06 9.41 5.52 18.15 50.24 3.68 2.77 53.45 34.26 27.98 17.53 6.26 4.68 41.58 41.63 2322

44.10 5.51 5.82 25.97 80.48 9.70 9.82 22.31 61.21 6.11 5.39 54.58 38.77 31.35 23.41 8.39 4.65 42.79 40.24 1668

39.15 3.71 10.93 25.83 71.27 18.47 10.26 16.80 N/A 3.24 3.08 52.57 30.83 34.17 17.86 6.07 6.25 41.39 40.71 2493

50.53 2.59 8.54 27.67 67.56 21.08 11.37 25.77 N/A 2.88 43.70 23.58 5.17 94.86 1770

5.89 5.26 8.86 34.63 1.03 7.09 2.97 1.72 32.54 22.91 76.16 20.84 3.01 21.58 54.62 4.24 30.61 13.03 1.52 38.46 2544

7.58 5.77 13.25 33.70 0.98 4.56 2.33 2.04 29.80 24.93 84.45 10.61 4.93 21.78 69.72 4.32 2.35 55.82 42.14 27.51 21.45 2.84 2.03 37.29 36.52 2240

10.72 5.28 19.11 29.07 0.72 4.11 3.14 2.63 25.22 26.49 79.54 13.71 6.75 17.52 50.93 2.90 1.88 61.84 43.11 33.27 24.17 3.32 2.81 35.99 35.29 2287

10.71 5.85 6.91 32.89 2.00 6.81 5.80 2.27 26.77 26.77 75.20 12.70 12.10 22.94 67.57 3.70 3.11 58.80 37.63 35.87 27.72 6.00 2.59 37.06 36.63 2364

4.54 7.66 17.43 29.75 0.89 2.20 3.95 1.78 31.80 27.35 64.21 22.59 13.20 17.64 N/A 2.86 2.80 56.12 41.39 38.60 28.21 4.26 3.23 36.73 35.76 3891

4.28 5.01 17.28 29.41 0.59 1.37 5.40 1.42 35.25 28.60 59.59 23.97 16.44 28.30 N/A 3.64 51.69 34.32 3.99 84.51 2489

References Abroms, L.C., Goldscheider, F., 2002. More work for mother: how spouses, cohabiting partners and relatives affect the hours mothers work. J. Fam. Econ. Issues 23, 147e166. Allison, P.D., 2005. Fixed Effects Regression Methods for Longitudinal Data Using SAS. SAS Press, Cary, NC. Autor, D.H., Levy, F., Murnane, R.J., 2003. The skill content of recent technological change: an empirical exploration. Q. J. Econ. 118, 1279e1333. Baker, M., Drolet, M., 2010. A new view of the male/female pay gap. Can. Public Policy 36, 429e464. Baker, M., Fortin, N., 2001. Occupational gender composition and wages in Canada, 1987-1988. Can. J. Econ. 34, 345e376. Blau, F.D., Kahn, L.M., 2000. Gender differences in pay. J. Econ. Perspect. 14, 75e99. Blau, F.D., Kahn, L.M., 2006. Th eU.S. gender pay gap in the 1990s: slowing convergence. Industrial Labor Relat. Rev. 60, 45e66. Blinder, A.S., 1973. Wage discrimination: reduced form and structural estimates. J. Hum. Resour. 8, 436e455. Boothby, D., Drewes, T., 2006. Postsecondary education in Canada: returns to university, college and trades education. Can. Public Policy 32, 1e21. Boudarbat, B., 2008. Fields of study choice by community college students in Canada. Econ. Educ. Rev. 27, 79e93. Boudarbat, B., Connolly, M., 2013. The gender wage gap among recent post-secondary graduates in Canada: a distributional approach. Can. J. Econ. 46, 037e1065. Bound, J., Breno, B., Golden, J.M., Khanna, G., 2015. Recruitment of foreigners in the market for computer scientists in the United States. J. Labor Econ. 33, S187eS223. Budig, M.J., England, P., 2001. The wage penalty for motherhood. Am. Sociol. Rev. 66, 204e225. Castegnetti, C., Rosti, L., 2013. Unfair tournaments: gender stereotyping and wage discrimination among Italian graduates. Gend. Soc. 27, 630e658. Cha, Y., 2013. Overworking and the persistence of gender segregation in occupations. Gend. Soc. 27, 158e184. Cha, Y., Weeden, K., 2014. Overwork and the slow convergence in the gender gap in wages. Am. Sociol. Rev. 79, 457e484. Clark, I.D., Moran, G., Skolnik, M.L., Trick, D., 2009. Academic Transformation: the Forces Reshaping Higher Education in Ontario. McGill-Queen's University Press, Montreal. Clark, T., Linzer, D., 2012. Should I Use Fixed or Random Effects? Society for Political Methodology Working Paper. http://userwww.service.emory.edu/ ~dlinzer/ClarkLinzer-REFE-Mar2012.pdf. Cohen, P.N., Huffman, M.L., 2003. Occupational segregation and the devaluation of women's work across U.S. labor markets. Soc. Forces 81, 881e908. Cooke, L.P., 2014. Gendered parenthood penalties and premiums across the earnings distribution in Australia, the United Kingdom, and the United States. Eur. Sociol. Rev. 30, 360e372. Crocker, R., Usher, A., 2006. Innovation and Differentiation in Canada's Post-Secondary Institutions. Canadian Policy Research Networks Research Report W33. Davies, S., Mosher, C., O'Grady, B., 1996. Educating women: gender inequalities among Canadian university graduates. Can. Rev. Sociol. Anthropol. 33, 125e142. Denis, A., Lipkin, J., 1972. Quebec's CEGEP: promise and reality. McGill J. Educ. 7, 119e134.

Please cite this article in press as: Smith, M.R., et al., Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’, Social Science Research (2017), http://dx.doi.org/10.1016/j.ssresearch.2017.03.003

14

M.R. Smith et al. / Social Science Research xxx (2017) 1e14

DiPrete, T.A., Eirich, G.M., 2006. Cumulative advantage as a mechanism for inequality: a review of theoretical and empirical developments. Annu. Rev. Sociol. 32, 271e297. Drolet, M., 2002. Wives, Mothers and Wages: Does Timing Matter? Statistics Canada Research. Paper No.186. Drolet, M., 2011. Why has the gender wage gap narrowed? Perspect. Labour Income 23, 5e15. Drolet, M., Mumford, K., 2012. The gender pay gap for private-sector employees in Canada and Britain. Br. J. Industrial Relat. 50, 529e553. England, P., 2006. Toward gender equality: progress and bottlenecks. In: Blau, F., Brinton, M.C., Grusky, D.B. (Eds.), The Declining Significance of Gender? Russell Sage Foundation, New York, pp. 245e264. England, P., 2010. The gender revolution: uneven and stalled. Gend. Soc. 24, 149e166. England, P., Gornick, J., Fitzgibbons Shafer, E., 2010. Women's employment, education, and the gender gap in 17 countries. Mon. Labor Rev. April. 20e29. Ferraro, K.F., Shippee, T.P., Schafer, M.H., 2008. Cumulativeinequality theory for research on aging and the life course. In: Bengtson, V.L., Gans, G., Putney, N. (Eds.), Handbook of Theories of Aging, 2nd ed. Springer, NY, pp. 413e433. Finnie, R., Qiu, H., 2008. The Patterns of Persistence in Post-Secondary Education in Canada. MESA Research Paper 2008-6. Canadian Educational Project, Toronto. www.mesa-project.org/research.php. Finnie, R., Wannell, T., 2004. Evolution of the gender earnings gap among Canadian university graduates. Appl. Econ. 36, 1967e1978. Flavell, D., 2012. Why Job Growth May Favour College Grads, Apprentices over University Students. Toronto Star. https://www.thestar.com/business/2012/ 01/13/why_job_growth_may_favour_college_grads_apprentices_over_university_students.html.
, P., 1992. L'avenir du ce
gep. Liber, Montreal. Inchauspe Jann, B., 2008. The Blinder-Oaxaca decomposition for linear regression models. Stata J. 8, 453e479. Karamessini, M., Rubery, J., 2014. Women and Austerity: the Economic Crisis and the Future of Gender Equality. Abingdon. Routledge. Levanon, A., England, P., Allison, P., 2009. Occupational feminization and pay: assessing causal dynamics using 1950-2000 U.S. census data. Soc. Forces 88, 865e891. Li, I.W., Miller, P.W., 2012. Gender discrimination in the Australian graduate labour market. Aust. J. Labour Econ. 15, 167e199. Livanos, I., Pouliakos, K., 2012. Educational segregation and the gender wage gap in Greece. J. Econ. Stud. 39, 554e575. Maume, D.J., 2006. Gender differences in restricting work efforts because of family responsibilities. J. Marriage Fam. 68, 859e869. McNabb, R., Wass, V., 2006. Male-female earnings differentials among lawyers in Britain:a legacy of the law or a current practice? Labour Econ. 13, 219e235. Morgan, L.A., 2008. Major matters: a comparison of the within-major gender pay gap across college majors for early-career graduates. Ind. Relat. 47, 625e650. Oaxaca, R., 1973. Male-female wage differentials in urban labor markets. Int. Econ. Rev. 14 (3), 693e709. OECD, 2012. OECD Economic Surveys: Canada 2012. OECD Publishing, Paris. Pfeiffer, C., Sohr, T., 2009. Analysing the gender wage gap (GWG) using personnel records. Labour 23, 257e282. Picot, G., Hou, F., 2009. Immigrant Characteristics, the it Bust, and Their Effect on the Eentry Earnings of Immigrants. Statistics Canada Analytical Studies Research Paper No. 315. Polachek, S.W., Siebert, W.S., 1993. The Economics of Earnings. Cambridge University Press, Cambridge. Reskin, B.F., Padavic, I., 1988. Supervisors as gatekeepers: male supervisors' response to women's integration in plant jobs. Soc. Probl. 1988, 536e550. Reynolds, J., 2003. You can't always get what you want: mismatches between actual and preferred work hours in the U.S. Soc. Forces 81, 1171e1199. Reynolds, J., 2005. In the face of conflict: work-life conflict and desired work hour adjustments. J. Marriage Fam. 67, 1313e1331. Senn, J.A., 2000. Electronic commerce beyond the "dot com" boom. Natl. Tax J. 53, 373e384. Skolnick, M.L., 2004. The relationship of the community college to other providers of postsecondary and adult education in Canada and implications for policy. High. Educ. Perspect 1, 36e58. Thornton, R.J., Innes, J.T., 1989. Interpreting semilogarithmic regression coefficients in labor research. J. Labor Res. 10, 443e447. Wang, Z., 2007. Technological innovation and market turbulence: the dot-com experience. Rev. Econ. Dyn. 10, 78e105. Weinberger, C.J., 2011. In search of the glass ceiling: gender and earnings growth among U.S. college graduates in the 1990s. Industrial Labor Relat. Rev. 64, 949e980. White, H., 1980. A heteroskedasticity-consistent covariance matrix estimator and a direct test for heteroskedasticity. Econometrica 48, 817e830. Xie, Y., Raymo, J.M., Goyette, K., Thornton, A., 2003. Economic potential and entry into marriage and cohabitation. Demography 40, 351e367. Zangelidis, A., 2008. Seniority profiles in unionized workplaces: do unions still have the edge? Oxf. Bull. Econ. Statistics 70, 327e345. Zhang, X., 2009. Earnings of women with and without children. Perspect. Labour Income. 5e13. March.

Please cite this article in press as: Smith, M.R., et al., Gender differences in the earnings produced by a middle range education: The case of Canadian ‘colleges’, Social Science Research (2017), http://dx.doi.org/10.1016/j.ssresearch.2017.03.003