**The E**ff**ect of Parental Economic Expectation on Gender Disparity in Secondary Education in Ghana: A Propensity Score Matching Approach**

#### **Prince Donkor 1,2,\*, Ya Ding <sup>1</sup> and Gideon Adu-Boateng <sup>3</sup>**


Received: 27 September 2019; Accepted: 20 November 2019; Published: 27 November 2019 -

**Abstract:** Ghana, like most sub-Saharan African countries, continues to face gender disparity at the higher levels of the educational hierarchy. This paper seeks to investigate whether gender disparity in senior secondary schools in Ghana is influenced by the economic expectations that parents have for their children's education. Using data from Ghana Living Standard Survey round 6 (GLSS 6), the study employs Propensity Score Matching in its analysis. Intra-household income inequality was used as a for measure parental expectations of the economic returns of education. The results revealed that, on the average, Ghanaian parents expect their male children to reap more economic benefits from education than girls. This attitude culminates in higher investment in boys' education to the disadvantage of their female counterparts at senior secondary schools. It is therefore recommended that appropriate policies should be implemented to ensure that the barriers that prevent women from occupying high-earning positions in the labor market are expunged. With this, parents will believe that girls can have the same economic opportunities as boys and hence will invest equal resources in children's education irrespective of their gender.

**Keywords:** gender disparity; education; Propensity Score Matching; Intra-household income inequality; senior secondary school; parental economic expectation

#### **1. Introduction**

Ensuring gender equality in all dimensions of life has become a top priority for many governments and international organizations. Since being brought into the international frontline by Convention on the Elimination of All forms of Discrimination Against Women (CEDAW) in 1976, there has been tremendous efforts to narrow and, if possible, eliminate all forms of gender inequalities especially in education [1]. Equal access to education, for boys and girls, is not only a fundamental human right; it also has economic, social, civil, and political benefits [2]. It is not surprising that gender equality in education features in two of the Sustainable Development Goals (SDGs). Goal 4 of the SDGs aims to ensure equitable and inclusive education of quality and also promotes lifelong educational opportunities for all. Goal 5 also seeks to empower all girls and women and achieve gender equality.

Equal access to education, for both genders, benefits both the current and future generations [3]. It is evident in the literature that education is associated with a higher flow of income, so education for all is a beacon of hope for alleviating poverty, especially among the marginalized (and women in particular). There is a ripple effect on the younger and unborn generations, as educating women

lowers fertility and mortality rates and other forms of inequalities for their children, thereby helping to create a sustainable planet [4,5].

It also fosters economic growth and development [6]. In this contemporary world, nations that want to maintain their competitiveness and comparative advantage focus on education for all as a tool. Educating all potential members of the labor force, be they boys or girls, men or women, is the engine of economic growth and development, as it ensures the efficient utilization of all economically active citizens. A knowledge-based labor force is a crucial determinant of the pace and sustainability of a country's economic transformation [7–10].

Equality in education is the agent of change for a more civilized global economy. When the citizens of an economy are well educated, there tends to be a reduction in the crime rate, domestic violence and other forms of gender inequalities. It also improves the social, political, and civic participation of the populace [8,11]. Education is therefore the new game-changer that underpins the achievement of most of the Sustainable Development Goals stipulated by the United Nations [12].

Even though education is seen as the fundamental solution to all forms of gender inequalities, girls and women continue to be disadvantaged in this sector. Despite the reduction in the number of girls out of school by 40% compared to the 1994 figure since the inception of the Millennium Development goals, girls still constitute the higher percentage of school dropouts [2]. Women account for about 60% of the world's illiterates. The UNESCO Institute for Statistics (UIS) estimates that more than 17 million girls will never enter a classroom to receive formal education [13]. These figures are alarming and represent a vast number of untapped and wasted human resources.

The educational disparity between boys and girls worsens in most sub-Saharan African and South Asian countries. Women and girls continue to lag behind their male counterparts at most levels of the educational hierarchy. More than half of the world's 58 million school-aged children who are not in school are girls, and about 75% of them are from these regions [13]. There is an average 10% completion gap between boys and girls at the primary level in sub-Saharan Africa, and the advantage goes to boys. Though there has been an increase in the secondary and tertiary school enrolment of girls, a substantial gap still persists [2]. This accounts for the high levels of domestic violence against women and girls, as well as the low political, social, and civic patronage of women in these regions [14]. The high level of untapped resources also explains the state of under-development of most of these countries and their non-convergence toward their developed counterparts [15].

There are numerous reasons why women and girls are discriminated against in terms of education in the sub-Saharan African region. Prominent among them are division of domestic chores based on gender, child labor, early marriage, and early pregnancy, among others [16].

In sub-Saharan Africa, children, especially girls, are an essential source of labor at homes. Due to cultural values and beliefs that consider male children as superior and therefore should not partake in many of the household activities, girls perform most of the domestic chores. A bulk of the household chores, such as cooking, washing, hauling water, and caring for younger siblings and sick family members, are performed by girls, while boys usually run errands. These activities are time-consuming and physically demanding, which affect girls' academic performance and consequently impede them from schooling [17–19].

Early marriage and its associated consequence of early motherhood also increase the dropout rate of girls in this region. Most girls, especially those in the rural areas, are forced to marry early once they reach puberty or become sexually active. One of the reasons for early marriage in sub-Saharan Africa is the desire to prevent the negative tag associated with unmarried girls who get pregnant. Families that have more girls also gain economically from the bride prices or dowries paid for their daughters. This situation is more pronounced among poverty-stricken families with more children to feed. Parents are therefore not willing to invest their scarce resources in the education of a girl-child who will in a short period be married off to another family. In addition, their new roles as wives, mothers, and caregivers for the extended families of their husbands after marriage exert pressure

on them. The time-consuming nature of these new roles contributes to the low enrolment and high dropout rates of girls from schools [18,20].

Girls who get pregnant or become mothers at an early age are often stigmatized and ridiculed by their teachers, friends, and schoolmates, which makes it difficult for them to return to school after childbirth. Some school do not even permit pregnant pupils or early mothers to be in class, as they are seen as bad influences on the rest of the students. In Ghana, girls who are pregnant are mostly not allowed to write the Basic Education Certificate Education (BECE) which is the prerequisite for senior secondary education. Even when schools allow these mothers to return, their re-entry depends on getting caretakers for their children, which they often cannot afford [18,21,22].

Another significant determinant of gender disparity in education is parents' expectation of the economic returns on the educational investments they make in their children. The economic worth and value of boys and girls are not genetically or biologically determined; it is placed on them by society. Most parents in sub-Saharan African countries (including Ghana) have different expectations for the economic value of their children's education. In most cases, higher value is placed on boys' education. This is so because, in Ghana for instance, men continue to earn more income than women in the labor market. Even when both sexes have the same educational background, men tend to be dominant in the high-paying jobs [23]. Most employers offering top-position jobs prefer men to women because of the high rate of absenteeism among the latter due to maternity leave, the need to care for sick family members, etc. [24]. Furthermore, the patriarchal nature of Ghanaian society gives men an upper hand over their female counterparts on the job market [25]. Men, therefore, gain more economically from education than women. If, in a household, a husband earns more than his wife, it will confirm to the parents that men have better economic value from education and hence it is worth investing more in their sons' education than their daughters.

Even though the parental economic expectation for girls from high-income families may exceed that of boys from low-income families, within each family, parents have higher expectations for sons than daughters due to the patriarchal nature of the Ghanaian labor market. Also, the magnitude of intra-household income inequality causes different levels of expectation. If parents of the same educational background work in a sector that is highly male-dominated, and as a result the gap between their incomes is very wide, the difference between the economic expectation for their sons and daughters will be higher than those of parents who work in a less male-dominated sector and have lower income inequality.

This paper aims to investigate whether gender disparity in senior secondary education in Ghana is influenced by the economic expectation parents have for their children's education. The study adds to the literature by providing an alternative measure for the economic expectations parents have for the education of their male and female children. A lot of studies that included this important variable in the analysis of education used qualitative measures [26–30]. This paper, however, employs a quantitative measure of parental economic expectation for education. The authors believe that intra-household income inequality between the father and mother is a good measure of this variable.

Another significance of the study is that it applied the Propensity Score Matching (PSM) method to the analysis of education. Economics provides a quantitative approach in analyzing the behaviors and decisions of economic actors by mostly estimating the causal relationship between variables [7]. However, most of these causal claims are plagued with endogeneity which nullifies the potency of the causal estimates [31]. One solution to eliminating the effect of endogeneity from causation is the PSM. This method is therefore used to find out whether there is a significant relationship between parents' expectations of the economic value of boys' and girls' education, measured by intra-household income inequality, and gender disparity in senior secondary education in Ghana.

The rest of this paper is as follows: Section 2 discusses gender inequality in Ghana's educational system, while the Section 3 reviews existing literature. Section 4 looks at the methodology used, Section 5 explores the dataset, Section 6 details the analysis of the data, and the last section discusses the results and concludes the study.

#### **2. Gender Inequality in Ghana's Educational System**

Ghana's successes in the educational sector in the initial years after the nation's independence in 1957 was marred by the economic recession it experienced in the early 1980s. The Structural Adjustment Programs (SAPs) the nation implemented in the early 1980s necessitated a drastic reduction in government expenditure in all sectors of the economy including education. Coupled with the introduction of the user fees in education, school enrolment for both boys and girls plummeted at all levels. For instance, girls' enrolment in primary school for the eligible age group dropped from 71% in 1980 to 68% in 1983. Those who enrolled in the secondary schools also fell to 28% in 1983 from the 1980's figure of 31%. A similar observation was made for boys. Primary school enrolment for boys declined from 89% in 1980 to 87% in 1983. The percentage of boys at the eligible age group who enrolled at the secondary schools decreased from 51 in 1980 to 47 in 1983 [32].

The end of the SAPs was marked by great achievements in the educational sector in general and for girls in particular. The period after the SAPs was met with the global advocacy for gender parity in all spheres of life. Ghana committed itself to ensuring gender equality in all aspect of the economy. It ratified several human right treaties, including the Convention on Rights of the Child (1989), the World Declaration for All (1990), the Beijing Declaration and Platform for Action (1995), and the Dakar Framework for Action (2000), which are against discrimination of all forms [2]. The 1992 constitution of Ghana also has a provision that makes education, especially at the primary level, free and compulsory for all children [33].

Several reforms and interventions have been made by the government to achieve Universal Primary Education (UPE) and also eliminate the disparity between boys and girls in the educational sector. Paramount among these interventions is the implementation of the Free Compulsory and Universal Basic Education (FCUBE) in 1997. To ensure the success of the FCUBE, the Girls' Education Unit (GEU) was created under the Ghana Education Service (GES). Its aim was to ensure that there is equal enrolment of girls and boys in basic education and also to reduce the percentage of girls who drop out of both primary and lower secondary schools. In addition, it was tasked to improve the rate of girls' enrolment in the senior secondary schools. Also, the government established the Capitation Grant in 2005, which abolished the user fee system and made education easily accessible to all [33].

These reforms catapulted the enrolment rate for both boys and girls at all levels of the educational ladder and even put girls at the advantageous position at the primary school level. The gross primary school enrolment of boys shot up from 74.7% in 2008 to 87.47% in 2014 before dropping to 82.87% in 2018. Same can be said about girls as their number increased to 87.81% in 2014 from the 2008 figure of 74.34% and afterward declined to 84.33% in 2018. Thus, the enrollment rate for girls in primary schools in recent years outweighs their male counterparts. It is worth noting that girls also outperform boys in terms of the progression rate to lower secondary schools. Figures from the World Bank's World Development Indicators (WDI) shows that 97.81% of girls progressed to the lower secondary school while the rate for boys stood at 92.66% in 2017 [34].

Despite the successes achieved in the enrolment of girls at the primary school, they lag behind boys at the senior secondary and tertiary levels of education. The percentage of boys' enrolment at the senior secondary and tertiary schools figured around 72.73 and 18.68, respectively, in 2018 as against 71.72 and 13.53 for girls. Furthermore, the literacy rate of men in Ghana outstrips that of women. The 2014 GDHS indicated that 82% of Ghanaian men are literate compared to the 67% for women. These rates, however, indicate an improvement in the 2008 figures for men and women which were 77% and 63%, respectively [35].

Several socio-economic factors account for the gender imbalance at the higher levels of education in Ghana. First of all, girls are over-burdened with domestic chores before and after school which affect their academic performance thereby causing them to drop out of school before attaining senior secondary education. Data from the Ghana Living Standard Survey round 6 (GLSS 6) indicates that the workload at homes for girls is three times more than that of boys. The proportion of girls who are out of school due to domestic chores in 2014 stood at 13.1, while the figure for boys was 3.2 [23]. Another factor that adversely affects the enrolment of girls in higher levels of education in Ghana is early marriage. Girls of secondary and tertiary school-going age are considered ripe for marriage. Apart from the economic gains from the dowries receive on their daughters, families of these girls attain social recognition if their female children marry before giving birth [21]. About 40% of girls of secondary and tertiary school-going age in Ghana are married, and one-fifth of them are mothers. The corresponding figures for boys are ten times lower [23]. Early marriage and motherhood burden child-brides financially and also in terms of time, which hinders their schooling. Furthermore, families in Ghana tend to place more economic value on the education of boys than girls. There is still a widely held belief that boys are superior to girls and will generate more income from schooling. This assertion encourages parents to invest more in their sons' education.

Gender inequality in Ghana's educational sector has a spatial dimension. The urban centers have a narrower gap between boys and girls in terms of those who attain higher level of education compared to the rural areas. The gender disparity gap between men and women in the rural communities stood around 16.4%, which is higher than the 11% recorded in the urban centers [23]. More so, urban women are more likely to have more education than their counterparts in the rural areas. The median years of schooling by urban women in 2014 was 8.5 years, which exceeded the 5.7 years by rural women [35].

Girls residing in rural areas in Ghana are more likely to marry at an early age compared to their counterparts in the urban centers. Due to the high incidence of poverty in the rural communities, parents usually pressure their daughters to marry early to relieve them of financial burdens. Again, the rate of teenage pregnancy and early motherhood in the rural areas surpass that of the urban centers because of the low patronage of family planning services in these areas. The higher rate of dropout of girls due to early marriage and motherhood in the rural areas cause gender disparity in education to be wider in those communities than in the cities [21].

#### **3. Literature Review**

Parental expectations of their children's education play a critical role in the academic successes of their wards. Jacob (2010) [30] posited that its impact on the education attainment of children is the most pronounced among the four dimensions of parental involvement which include parental engagement in school-related activities, supervision at home and parent–child communication about school. Parental expectations, as defined by Yamamoto and Holloway (2010) [36], are realistic beliefs or judgments that parents have about their children's future achievement as reflected in course grades, highest level of schooling attained, or college attendance. It is what parents actually believe their children can achieve in education. There are numerous research findings that support a positive relationship between parental expectation and educational achievements.

Yamamoto and Holloway (2010) [36] explained four trajectories through which parents' expectations influence their children's performance at school. First, parental expectations signal to children the confidence their parents have in them in terms of their academic achievements. This confidence becomes the norm that the children strive to attain. Second, children's confidence about their own competence and capacities is boosted by the high expectations from parents and vice versa. Also, parents with higher expectations tend to be more involved in their children's education. They invest quality time and resources in their schooling. They tend to spend more time helping their children with their homework, partake in school activities, communicate more with their teachers, and provide counselling and support. Finally, the performance of teachers is influenced by the expectations that parents have for their children. If teachers perceive that parents have high expectation for their children's academic achievement, they will be motivated to work harder, as they know their effort will be complemented by parents at home.

The inclusion of parental expectations in the analysis of education is mostly in the domain of sociology and psychology. Most research in these academic fields that studied expectations of parents employed qualitative measures before re-coding them to be used in the analysis. The measurement approach includes the usage of categorical variables, constructs, and (in some cases) continuous variables.

A number of authors who studied the effect of parents' expectation of educational attainments used a categorical variable for its measurement. For instance, using the Growing Up in Ireland (GUI) survey, Banks et al. (2016) [28] explored the effect of parental expectation on the academic success of disabled children who were 13 years old. Those included in their study suffered from general learning/intellectual disability, specific learning disability (such as dyspraxia, dyslexia), socio-emotional disability, and physical disability (such as mobility, visual and hearing impairment). Children whose parents who said they do not expect their children to attain more than Leaving Certificate education had poor academic performance compared to those whose parents had higher expectations. A similar approach was used by Einglund et al. (2004) [37], who investigated the impact of parents' expectation on the educational success of 187 children from low-income families. Using a semi-structured interview, parents ranked how far they think their children would go in school. The responses ranged from "will not complete high school," which was coded as 1, to "will go to graduate or professional school," which had a code of 5. It was found out that the educational achievement of children from homes where parents have high expectations surpassed their counterparts. Studies by Kim et al. (2017) [27] and Gill and Reynolds (1999) [38] were not different from that of Einglund et al. (2004) [37].

In the case of O'Donnell (2007) [39], parents chose the probability that their children would 1) obtain a high school diploma by age 20, 2) obtain a college degree by age 30, 3) be employed by age 30, 4) be in jail by age 20, and 5) be a parent by age 20. The first three options were classified as positive expectations, and the remaining two were viewed as negative. The study used data from National Longitudinal Survey of Youth (NLSY97), and the participants included children who were aged 12–16 on 31 December 1996 and their parents. It was revealed that the positive expectations were associated with good outcomes, while the negative ones brought bad achievements.

Other studies used a construct in the measurement of parental expectation. For example, Jacob (2010) [30] examined the effect of parental expectation on the educational attainment of their wards using the Scale of Educational Aspirations and Expectations for Adolescent (SEAEA). Parents answered to 29 questions with each response ranging from strongly disagree to strongly agree. The responses were then loaded onto a single factor, and a path analysis was used for the exploration. In all, 598 parents of eighth to 10th grade students were sampled for the study. The result was the same as the other studies stated above. Weerasinghe and Panizzon (2015) [40] and Leung and Shek (2011) [41] employed a similar strategy to investigate why children from Asian lineage attain better performance in mathematics than others. Their constructs revealed that the "tiger" parents from Asia have high expectations for their kids and this accounts for their (children's) higher performance and educational attainment.

Only a few studies used a continuous variable in the measurement of parental expectations. Zhan (2006) [42] transformed the categorical variable he used to measure parents' expectation into a continuous variable. The question was "looking ahead, how far do you think your child will go in school?" The response by parents ranged from "leave high school before graduation," coded as 1, to "take further training after college," coded as 5. The author argued that the distribution of the response approached normality and there was a slight negative skew so the variable was treated as a continuous variable. The dataset used was the National Longitudinal Survey of Youth (NLSY79). The results showed that there is a direct relationship between expectations of parents and school performance. Clophus (2018) [26], on the other hand, employed the Career-Related Parent Support Scale (CRPSS) to measure parental expectations. The CRPSS contains 27 questions on a five-point Likert scale ranging from strongly disagree (1) to strongly agree (5). Parents who had a score of 27 were classified as those with the least expectations, while a score of 135 was seen as the highest. The participants included 58 males and 95 females from two high schools in Southwest Louisiana and their parents. The findings revealed that parental expectations had no significant effect on general educational successes as well as that of boys and girls.

It is worth noting that, although the expectations of parents positively impact on the educational achievement of their children, setting unrealistic target for children can be counterproductive. If children feel too much burden as a result of very high expectation from parents, they may work beyond their abilities and capabilities, which can cause emotional breakdowns, depression, etc. They may be demotivated if their relentless efforts do not achieve what their parents expected.

#### **4. Methodology**

The purpose of most evaluation studies is to make causal claims, i.e., to determine the effect of a treatment, an intervention or a program on an outcome variable. Counterfactuals are the key to understanding causal inferences. A counterfactual is simply the unobserved outcome of a studied agent. To obtain an unbiased causal estimate, the estimation technique used should be able to reconstruct a counterfactual for the observed outcomes [43–45].

Randomized controlled trials (RCTs) or experiments are the gold standards in generating counterfactuals for estimating causal effects. With this method, there is a random assignment of those who received the treatment (henceforth, the treated group/subject) and those who did not (henceforth, the untreated/control group) to either group making them similar in terms of their pre-treatment baseline characteristics. Technically, the control group serves as a counterfactual to the treated subjects in RCTs or experiments [31].

#### *4.1. Observational Study*

Most researchers in management studies use observational data. This is because large volumes of data can be obtained at a relatively cheaper cost compared to experiments. Others resort of this type of data due to ethical reasons. However, observational data often suffer from self-selection bias, which results in endogeneity [46]. Self-selection bias arises because there are often some baseline characteristics that influence some subjects to be selected into the treated group and others into the control group. There is mostly a systematic difference between the treated and untreated subjects even before the treatment is applied. The control group cannot be used as a counterfactual of the treated group since the two groups are not similar prior to the treatment. This makes the computation of the average treatment effect from the mean difference in the outcome variable between the two groups biased. The estimate will be uninterpretable and is not attributable to the treatment.

Self-selection bias threatens the internal validity of the variable or construct. It acts as an excluded or omitted variable from the model, and if not addressed, it will cause the treatment variable and the error term to correlate, leading to endogeneity [46].

#### *4.2. Propensity Score Matching (PSM)*

A good estimation technique to deal with the endogeneity resulting from self-selection biasedness is the Propensity Score Matching (PSM) introduced by Rosenbaum and Rubin (1983) [47]. This approach mimics RCTs by reconstructing counterfactuals for the treated studied agents from the observational data. For each subject in the treated class, PSM finds an observation(s) in the control group that has (have) similar characteristics to act as counterfactual(s). It does this by combining propensity score with an appropriate matching algorithm.

Propensity score, as defined by Rosenbaum and Rubin (1983) [47], is the likelihood that a studied agent will be assigned to a treatment class based on its pre-treatment characteristics. In estimating the propensity score (PS), the treatment variable (which is binary) is used as the dependent variable in the PS model, while the covariates serve as the independent variables. The logit or probit regression is then fitted using the measured values of the covariates. Once it has been computed, an agent in the control group whose propensity score is the same or close to the propensity score of another subject in the treated group is deemed as its counterfactual. It is recommended that the studied subjects should be stratified into five or more classes using their estimated PS, as this reduces about 90% of the bias in the covariate(s) between the treated and the control groups within each class [48]. The study used 1:1 matching. With this approach, each treated subject is paired to only one observation in the control group. After that, matching without replacement where an untreated agent that has been paired to a treated subject is not used again for matching was used.

Once the aforementioned procedure is done, the average treatment effect on the treated (ATT) can be computed as the mean difference in the outcome variable between treated group and their paired untreated counterparts after the treatment has been applied. Figure A1 in the Appendix A summarizes the steps used in PSM.

#### **5. Dataset**

#### *5.1. Source of Data*

The data set used for the analysis of this study is the sixth round of the Ghana Living Standard Survey (GLSS 6). GLSS 6 is a secondary data set that concentrates on the household as the main socio-economic unit. The survey used a two-stage stratification procedure to select the participating households. First of all, Ghana is divided into 10 strata according to the number of regions. Within each region, households are grouped in accordance to their place of residence, i.e., whether they reside in an urban or rural area.

In all, 18,000 households, which is a nationally representative sample, were covered in the survey. Only 16,772 out of the 18,000 households were successfully enumerated, which accounts for the 93.2% response rate. The GLSS 6 provides a detailed insight on the living conditions of the households involved. The data covers demographic characteristics, educational attainment, housing and income of household members.

The study considered all boys and girls of senior secondary/high school (SSS/SHS) age who still reside in their parents' home and have no missing value for any of the variables used. As a result, 1368 boys and 1111 girls were used in the data analysis. Separate propensity score matching models were estimated for each gender.

#### *5.2. Treatment and Outcome Variables*

The study used enrolment in senior secondary/high school (SSS/SHS) as the outcome variable in its analysis. Boys and girls of relevant school age who have ever been enrolled in SSS/SHS were coded as 1 and 0 if otherwise. The treatment variable, which is parental economic expectation for child's education, measured by intra-household income inequality, is obtained by subtracting the child's father income from the income of his/her mother. A positive real number, implying a higher income for the father relative to the mother, was coded as 1 and 0 if otherwise.

#### *5.3. Selection of Covariates*

It is important to choose the right covariates in modeling the propensity score. Empirical studies and Monte Carlo simulations have proved that only true confounders (covariates that simultaneously influence the treatment and the outcome variables) and potential confounders (covariates that influence the outcome variable) should be included in the propensity score model. The exclusion of true confounders from the model violates the strongly ignorable assumption and will make the ATT estimate biased [44,49].

Following the steps of Moheyuddin (2005) [16], father's and mother's educational level (years spent in school), age of the child, place of residence (whether rural or urban), and wage earned by the child were used as confounders. Other covariates used include average minutes spent on domestic chores daily, marital status, and whether or not the respondent has a child. Thus, the aforementioned variables were used as independent variables in the propensity score model where economic expectations of parents served as the dependent variable. However, to achieve a balance in the covariates between the treated and control groups within each stratum, the square of the father's education and the interaction of the father's and mother's education were included in the re-specified PS models.

#### **6. Analysis of Data**

#### *6.1. Descriptive Statistics*

This sub-section discusses the descriptive statistics for variables used in the separate PS models for boys and girls. The data confirms that parents in Ghana expect their sons to reap more economic benefits from education than girls. It was revealed that on a scale of 0 to 1, boys had an average of 0.53 in terms of expectation from parents, compared to 0.49 for girls. Table 1 displays the statistics for the continuous variables while Table 2 portrays that of the categorical variables. As shown in Table 1, the mean years of schooling of parents of girls outweighed that of boys. Fathers of girls spent an average of 8.92 years in school, while their mothers have 6.84 schooling years. The number for fathers and mothers of boys stood at 8.86 and 6.68 years respectively. Also, girls of SSS/SHS going age earned more weekly wages (₡111.74 GH ≈ \$21.4 USD) and spend more minutes on domestic chores (169 min) than their male counterparts. The respective figures for boys are ₡99.51 GH ≈ \$18.15 USD and 55 min, respectively.

**Table 1.** Descriptive statistics of continuous variables.



**Table 2.** Descriptive statistics of categorical variables.

The percentage of boys who have ever been enrolled in a senior secondary slightly surpassed proportion for girls. Table 2 showed that 29.82% of boys of the relevant age have obtained senior secondary education. The rate for girls was 29.34%. The percentage of boys who reside in urban areas (25.15%) marginally exceeded their female counterparts (24.57%). More girls than boys aged between 15 and 19 years have been enrolled in senior secondary schools. However, the proportions of boys of age cohort 20–24 and 25–29 years old who have had senior secondary education are more than girls of the same age groups. The ratio of secondary and tertiary school-going age girls who are married and have at least a child to boys is 10:1.

#### *6.2. Covariate Balance Diagnosis*

A critical component of propensity score matching is the balancing of covariates between the treated and untreated groups. If the covariates between the two groups are not balanced, the average treatment effect on the treated (ATT) cannot be computed as the difference in the means (in the case of continuous outcome variable) or proportions (in case of dichotomous outcome variable) of the outcome variable. There are several statistical tools used in checking the balance of these pre-treatment variables between the two groups. They include the t-test, the standardized difference, and other graphical displays such as the box plots and cumulative frequency function. In the event where these tools prove that there is no equality between the covariates in the treated and untreated groups, the PS model has to be re-specified to include higher-order moments of covariates and/or interactions between covariates [43,44].

The PS model of the study initially used the father's and mother's education together with residence of household, the respondent's age, and their weekly wages as the pre-treatment characteristics. After classifying the subjects into strata, the balancing diagnosis revealed that the covariates between the two groups were not balanced. The "troublesome" variable was father's education, which was not balanced in most of the strata. The square of father's education and an interaction between the former and mother's education were included in the re-specified PS model.

Results from the simple t-test indicated that the difference between means of each covariate in the treated and control groups within each stratum was not statistically significant as portrayed by Table A1 in the Appendix A. Table A1 reports the p-values of the t-test for the mean difference of each of the pre-treatment variables within each stratum. All the p-values exceeded 5% implying that there was no significant difference between the control and treated classes at 5% significant level. Balancing of covariates in the boys' PS model was achieved with six strata, while five blocks were enough for the girls' model to attain equality. Table A2, which showed the standardized difference before and after the matching, confirmed the results from the t-test (readers can resort to Li (2013) [43] for a detailed explanation on standardized difference). A standardized difference (SD) of less than 0.1 is considered negligible and hence no significant difference between the means of variable in the two groups [44]. All the variables had SD values of more than 0.1 before the matching was done. The matching, however, reduced the SD for each of the covariates to figures less than the threshold value.

Numerical analysis of balance of covariates takes into account the means of the variables and in some cases their standard deviations. This approach has a limitation as it considers only one or two dimensions of the variable. The means of the variable between the two groups may be the same but their mode, median, quantiles and other statistical summaries may differ. Linden (2015) [50] proposed the use of graphical displays to assess how equal the variable is between the two groups in terms of multiple dimensions. Since the estimated propensity score summarizes the distribution of all the covariates, the study employed the box plot and the cumulative frequency function of the estimated propensity score of the treated and untreated groups to assess the equality of the covariates. Figure A2 shows the box plots for boys and girls, while Figure A3 displays the cumulative frequency of their estimated propensity scores between the groups. The figures demonstrated that, in the case of both boys and girls, there is no statistical difference of the composite covariates (estimated PS) between the treated and control groups.

#### *6.3. Estimating Causal E*ff*ect*

After the covariates between the treated and control groups have been balanced within each stratum, the average treatment effect on the treated (ATT) was computed using the stratified matching. The formula for stratified matching is expressed as

$$ATT = \sum\_{q=1}^{Q} \left( \frac{\sum\_{i \in I(q)} Y\_i^T}{N\_q^T} - \frac{\sum\_{j \in I(q)} Y\_j^C}{N\_q^C} \right) \* \frac{N\_q^T}{N^T} \tag{1}$$

where *Q* is the number of strata used to achieve a balanced covariate, and *Y T i* and *Y C j* represent the value of the outcome variable for subject *i* in the treated group and its paired subject *j* in the control group respectively. *N<sup>T</sup> <sup>q</sup>* and *N<sup>C</sup> <sup>q</sup>* are the number of treated and control subjects in block *q*. *N<sup>T</sup>* , on the other hand, is the number of treated subjects in the entire study.

Estimates from the stratified matching showed that the ATT effect of parental expectation of the economic benefit of education on boys' enrolment in senior secondary is 0.048 while the effect on girls figured around 0.029. The estimate for boys is significant at 5%, while that of girls is 10%.

#### *6.4. Sensitivity Test*

The last step in PSM analysis is to assess the sensitivity of the causal estimates to unobserved confounders. The ideal way is to compare the ATT estimate with the results of a similar study that used an experimental data. However, such results may be unavailable in a practice setting. An alternative approach is to re-specify the PS model by dropping or adding higher-order covariates such as quadratic or interaction terms. If the original estimated effect does not differ significantly from the re-specified model, then the ATT estimate is less sensitive and hence unbiased [43].

The squared of father's education as well as the interaction of the former and mother education were dropped to re-calculate the propensity score and ATT. The estimation results after dropping those variables indicated that parental economic expectation increase boys' senior secondary education by 0.043, and the effect for girls stood at 0.027. The new estimates pointed out that the original ATT estimates are insensitive and therefore unbiased.

Furthermore, Figure A4 in the Appendix A indicated that there is an overlap in the distribution of covariates between the treated and untreated group for both genders. This validates the results obtained in Table 3.


**Table 3.** Estimation results of stratified matching.

#### **7. Discussion and Conclusions**

Ghana, like most sub-Saharan African countries, continues to bedevil with gender imbalance at the higher levels of the educational hierarchy. Among the several socioeconomic factors that account for why more girls than boys do not attain secondary education is parental expectation of the economic benefits their sons and daughters will gain from education. This study investigated the role that this variable plays on gender disparity in the senior secondary schools. The data revealed that, on the average, Ghanaian parents expect their sons to reap more economic benefits from education than their daughters. The higher expectation for boys encourages parents to invest more in their sons' education compared to their daughters. Results from Table 3 showed that parental expectation for sons increase the probability of boys' enrolment in a senior secondary school by 0.048 while the figure for girls is 0.029 [51]. The estimate for boys is significant at 5%, while that of girls is 10%. Thus, boys are 1.66

times more likely than girls to attain senior secondary education due to parents' expectation that they (boys) will gain more economically from education than girls.

The finding is consistent with the assertion in the literature that there is a positive relationship between parental expectation on education and academic achievements [26–30]. Ghanaian parents invest more monies and spend quality time on their sons' education as against that of their daughters because they believe that top positions and high-paying jobs on the labor market is male-dominated and therefore boys will gain more economically from education than girls. Also, the confidence of boys is boosted because they know their parents expect more from them academically than girls. This confidence becomes the norm, which motivates them (boys) to attain higher levels of education [36].

The study employed propensity score matching (PSM) in its analysis. This estimation technique is preferred to other methods because it overcomes the problem of self-selection bias inherent in observational data. Self-selection bias—the likelihood that some subjects will be selected into the treated group and others into the control group based on some pre-treatment characteristics—threatens the validity of causal estimates. If not solved, estimated results will be biased and uninterpretable. PSM is a good estimation method that eliminates the endogeneity that arises from self-selection bias. PSM is becoming increasingly proper because of its ability to reconstruct observational data to mimic an experiment, which is the gold standard for making causal inferences. Our estimates are therefore free from bias resulting from self-selection and hence are unbiased and interpretable. The sensitivity analysis allowed us to test how susceptible the estimates are to other cofounders. The results showed that our estimates are less sensitive to the inclusions or exclusions of other variables.

The findings from the study revealed that a plausible solution to combat gender inequality in Ghana's educational system, especially at the higher levels, is to eliminate gender imbalances at the workplace. Men continue to be over-represented at top positions and high-paying jobs. This deepens the negative traditional view parents have about the future prospects of the female children. Appropriate policies should be implemented to ensure that barriers that prevent women from occupying such positions are expunged. With this, parents will believe that girls can have the same economic opportunities as boys and hence will invest equal resources in their children irrespective of their gender.

It is, however, important that Ghanaian parents change their stereotypical notion that girls from household where there is income inequality in favor of men will earn less income from education. This is so because, if given an equal opportunity, girls from such homes can outperform their male counterparts in the pursuit of higher levels of education that can consequently eliminate the gender gap in wages on the Ghanaian labor market.

**Author Contributions:** Conceptualization, P.D., D.Y. and G.A.-B.; methodology, P.D.; validation, D.Y., formal analysis, P.D., D.Y. and G.A.-B.; investigation, P.D. and D.Y.; writing—original draft preparation, P.D.; writing—review and editing, D.Y.; supervision, D.Y.

**Funding:** This research received no external funding.

**Acknowledgments:** I would like to express my profound gratitude to my wife, Doris Gyekye, Kwasi Sarfo-Adu, and Faustina Amponsah Partey for their painstaking effort in proofreading this article.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **Appendix A**

**Figure A1.** Steps for estimating treatment effect. Source: Li (2013) [43].

> 0 .2 .4 .6 .8 1 propensity score treated untreated

2

4

kdensity \_pscore

6

8



**Table A2.** Standardized difference of covariates before and after matching.


**Figure A2.** Box plots of estimated propensity score for treated and untreated class.

0

> 0 .2 .4 .6 .8 propensity score treated control

1

2

3

kdensity \_pscore

4

propensity\_treated propensity\_untreated

.2

.4

.6

.8

1

−

−

−

− − −

− − −

−

− −

propensity\_treated propensity\_untreated

−

− −

0

.2

.4

.6

.8

**Figure A3.** Cumulative density function of estimated propensity score for treated and untreated class.

**Figure A4.** Distributional overlap of covariates between untreated and untreated groups.

#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **Gender and Academic Rank in the UK**

#### **Georgina Santos 1,\* and Stéphanie Dang Van Phu 2,3**


Received: 19 November 2018; Accepted: 5 May 2019; Published: 5 June 2019

**Abstract:** This paper fills in a research gap in what concerns gender and academic rank at UK universities, where women are not far from reaching the 50% share of all academic and research staff, but not even close to reaching such a share at (full) professorial level. Using an ordered logit model and the results of a survey conducted in 2013 with 2270 responses from academics from all fields of knowledge at the 24 Russell Group universities, we find three consistent results. First, being a woman has a negative and significant association with academic rank, except for the case when parenthood is timed with career considerations in mind. Second, the percentage of time spent on teaching and teaching-related activities has a negative and statistically significant association with academic rank. This association is more pronounced in the case of women, who spend a higher percentage of their working time on teaching and teaching-related activities than men, as do those in lower academic ranks. Since women tend to be in lower ranks, the percentage of time spent on teaching and teaching-related activities may be considered both a cause and a result of the gender gap. Third, we find a positive and significant association between the number of children under the age of 18 years and the academic rank of both men and women, as long as babies were timed with career considerations in mind, and a non-significant association when they were not. A possible explanation for this is unlikely to be that children have a positive impact on academic rank, but rather that they arrived after a certain rank had been secured. We conclude with some policy recommendations to help reduce the gender gap.

**Keywords:** gender discrimination; academic progression; women faculty; female professors; maternity penalty; gender gap

#### **1. Introduction**

In 2011/2012, 44.5% of all the academic staff employed at UK Higher Education Institutions were female, yet only 20.3% of professors, which is the highest academic rank in the UK, were women [1]. Focusing on the 24 Russell Group universities in the UK, which are research-intensive universities, in 2011/2012, 40.7% of all academic staff at these 24 universities were female (a share somewhat lower than that at all UK universities) and from all professors only 18.9% were women [2].

Although all universities in the UK value diversity and are committed to equality of opportunity, women are under-represented at senior academic grades. If current trends continue, it will be decades before gender equality at professorial level is reached.

Using an ordered logit model and a new rich and detailed data set, which we collected in 2013, with 2270 observations of academics of both genders at all levels in all fields of knowledge at the 24 Russell Group universities in the UK, we contribute to the literature by examining the association between gender and academic rank, controlling for a number of variables, including but not limited to, respondent's year of birth, number of children, responsibility for household chores, academic degrees, number of publications, grants, percentage of working time spent on teaching and teaching-related activities, and main area of research. This is timely and relevant, given that the last empirical quantitative study to include UK-based academics of all fields of knowledge was conducted in the year 2000; the results of that study are reported in [3].

We find some results in line with previous work conducted for other countries or for specific fields of knowledge and some novel ones. First, being a woman has a negative association with academic rank, even after controlling for year of birth (i.e., age), marital status, responsibility for household chores, area of research, timing of babies, number of children under the age of 18 years, holding a PhD or not, percentage of working time spent on teaching and teaching-related activities, and a number of research productivity variables. The only case where the variable gender is not significant is when only men and women who timed their children with career considerations in mind are included in the sample. Importantly, we also find that the percentage of time spent on teaching and teaching-related activities, which is higher for women than for men, is negatively associated with academic rank. In addition, and this can be seen as our most important and novel finding, there is a positive association between the number of children under the age of 18 years and the academic rank of both men and women, as long as babies were timed with career considerations in mind. A possible explanation for this is unlikely to be that children help academic career progression, but rather that they arrived after a certain rank had been secured. Timing of children seems to be crucial.

The paper proceeds as follows. Section 2 reviews the most recent and prominent literature on the topic, which, apart from [3], lacks any quantitative study specifically designed for the UK. Section 3 explains how the data were collected. Section 4 presents the model. Section 5 discusses the results. Section 6 concludes and gives some policy recommendations.

#### **2. Previous Work**

The potential explanations for the gender imbalance in academia tend to fall under two categories: (a) Women work fewer hours than their male counterparts because their time constraints are more stringent, and as a result progress at a slower rate than men, with a lower percentage making it to the grade of professor; and (b) Women are discriminated against, and inadvertently, denied opportunities that could give them access to high rank positions.

The time constraints hypothesis argues that women need to, want to, or choose to devote time to raising their children and/or taking responsibility for household chores, whilst their male counterparts devote this time to productive work or leisure. The idea is essentially that women with responsibilities for housework and childcare have less energy available for remunerated work than men have, and this affects their job opportunities and productivity [4]. Some authors argue that many high-end jobs require virtually complete commitment to work, and go on to assert that more men than women are prepared to devote themselves to work so fully [5,6]. As a side note, some also hold controversial views regarding innate cognitive and temperamental differences between men and women [5,6]. This topic, however, falls under the remit of sociology, psychology, biology, and related sciences, and is therefore not discussed in the present study.

It has also been argued that women, especially those with children, face more family-work -balancing challenges than men [3,7]. A number of studies carried out in different Schools at MIT [8] and a European Commission report [9] also found that family and career tensions were greater for women than for men.

The association between marriage and children and academic rank, salary, and research productivity, however, is far from clear. One study finds that children have a negative effect on academic careers of women and a positive effect on academic careers of men [10]. On somewhat similar lines, another study finds that marriage and young children (under 6 years of age) reduce the probability that women get a tenure-track job [11]. Two further studies find a non-significant association between marriage and promotion, and a positive and significant association between children and male promotion but a negative, albeit non-significant, association between children and female promotion (in the humanities) (p. 400 [12]) and (p. 51 and p. 62 [13]). Another study finds a positive and significant association between young children and male economists' promotion chances and a negative association between marriage and children and female economists' tenure chances [14]. On the other hand, academics who have older children (aged 6 to 17 years) have been found to have a greater chance of getting tenure, relative to academics without children in this age range, regardless of their gender, probably because children trigger the need to secure ongoing employment [11]. There may also be selection effects because these children were under the age of six years when their parents were completing their doctorates or securing tenure-track positions, and academics, especially women, who manage to do all that whilst simultaneously caring for young children may be especially good at managing their time and the demands of work and family or may have received more support from their partners (p. 400 [11]).

Another study, in turn, finds that having children and having a spouse or partner employed at the same institution are unrelated to tenure and rank among women faculty but having children has a positive association with both tenure and rank for men, who also benefit from being married in terms of their academic rank (p. 301 [15]). Other research finds a positive association between being married or living with a partner and salary [16].

One point that a number of studies find is that academic women are less likely to be married with children, relative to academic men [3,10,13–15], or they are more successful if they delay or forgo marriage and children [11]. It is not clear, however, whether this is a decision women make because they fear that by having children they will jeopardize their careers, even though in reality having children may have made no difference, or whether thanks to the decision of not having children they were able to progress, something they would have not been able to do had they had children. Although intuition would point towards a negative impact of children on the academic progression of women, and this is supported by solid microeconomic theory such as that presented in [4], the evidence, as shown above, is far from conclusive.

Publications are typically considered a key factor for academic progression. In general, publications have a positive association with rank and promotion [13,14,17–20], although there is also some evidence that male economists on tenure-track positions get tenure regardless of their publications (p. 203 [14]). At the same time, on average men produce more publications than women, and this is found across different disciplines [12,14,17,21–28], although the results reported in [13] suggest very small differences.

Women spending more time with their children than men do, especially when they are of preschool age, could potentially be linked to lower publication rates [7,25]. One study, for example, concludes that untenured male economists become substantially more productive after having a first child but female economists with two and three children have, on average, a research record reflecting a loss of two and a half years and four years of research output, respectively, by the time all of their children have reached their teens [28].

On the other hand, a review examining the relationship between marriage, children, and research productivity concludes that there is no evidence of a negative effect of family factors on the research productivity of women (p. 18, p. 99 and p. 189 [29]), in line with [14,27,30,31]. Interestingly, though, another study finds a positive relationship between having children and research productivity for female economists but no relationship for male economists [32]. This same study also finds that women with children are more productive than women without children, as well as some evidence of self-selection that may explain this counterintuitive result: only the most productive women dare to pursue an academic career and have children at the same time [32].

Grants are also typically considered important for promotion, and indeed there is a positive association between grants and promotion [18]. Blake and La Valle [3], whose study actually focuses on grant applications, find that in the five-year period prior to their survey, from those who were eligible to apply, women were less likely than men to have applied for grants, with 56 per cent applying in contrast to 67 per cent of men (p. 36), and women with children were also less likely to have applied for grants than men with children, with 50 per cent applying in contrast to 62 per cent of men (p. 104). Having said all that, Blake and La Valle find that the success rate for grant applications is virtually the same for men and women and conclude that there is no gender bias in the awarding processes (p. 37 [3]). The main difference between men and women, they argue, "lies in application behaviour rather than in success once applications have been made" (p. 37 [3]). This finding of no gender differences in the outcomes of grant applications is in line with [33–35], but in contrast with [36–39].

Notwithstanding all of the above, lower grant application activity and lower number of publications in absolute and relative terms may be explained not just by time constraints due to housework or childcare but also by time constraints imposed in the very workplace, for example, with higher teaching or administrative workloads [3,18,20,22,23,25,27]. Higher teaching or administrative workloads on women could be the result of subtle discrimination. Needless to say, very rarely is there any blatant open discrimination in academia but a theme that emerges from the literature is that there may be forms of (sometimes unconscious) discrimination that are concealed, almost unnoticeable, and therefore harder to identify. Examples of studies which point towards this unconscious bias against women include [10–13,15,17,19,40], all of which find a gender gap in academic rank or salary, which remains unexplained after controlling for credentials, productivity and/or family circumstances, amongst other variables. One study, however, finds unexplained differences in promotion to tenure in some disciplines, but discrimination in favour of women in engineering [14]. Bias in grant awarding has also been found, as mentioned above, in [36–39].

Given the importance that the hypotheses of time constraints and workplace discrimination have received in the literature, we concentrate on these two perspectives as prime suspects to help explain the low representation of women in higher academic ranks. Despite the rich literature on gender and academic progression, this is the largest quantitative study to have been carried out for the UK case since Blake and La Valle's in 2000 [3].

#### **3. Data**

We conducted a questionnaire amongst male and female academics, which can be found in Appendix A, and is virtually the same as that conducted by Blake and La Valle in 1999/2000 [3]. After piloting it, it went live and was open for responses from 29 May to 1 July 2013.

The sample was drawn from the 24 Russell Group universities in the UK, which were arranged in alphabetical order. The Research Excellence Framework (REF) in the UK is the system used for assessing the quality of research in UK higher education institutions. Submissions to the REF in 2013 were made in 36 units of assessment, or fields of research. Up to ten out of the 36 REF areas, which are listed in Appendix B, were randomly chosen for each of the 24 universities. The departmental websites representing the randomly selected REF areas were then used to identify all members of academic and research staff. In some cases, REF areas include more than one area, which meant a number of departments were contacted. For example, REF area 4 includes Psychology, Psychiatry and Neuroscience. If that area was randomly selected for a university, staff at all three departments were contacted if all three were represented at the institution in question. If not all departments were represented, then those that were, were the ones contacted. If an area was randomly selected for a university but had no presence at that university, another number between 1 and 36 was randomly selected. Typical cases include the London School of Economics and Political Science and Imperial College London, which are institutions with some degree of specialization where many of the 36 REF areas are missing.

A total of 13,556 names and e-mail addresses were manually collected. No scraping software of any sort was used at any point. These potential participants were then contacted by e-mail and invited to complete a survey online. Due to a number of people having left the departments in question but still being listed on their websites 886 mails were returned with a delivery failure notice. From the remaining 12,670 individuals, 2270 responded to the survey. The response rate was therefore 17.9%, but we still achieved our target of at least 2000 responses.

The response rate may have varied according to a number of reasons, and in order to correct for self-selection bias the data from the sample was weighted using post-stratification survey weights. Appendix B gives details of how weights were estimated to make our sample of 2270 respondents representative of the whole population of 62,637 individuals employed as academic and research staff at all 24 Russell Group universities in 2012, following the methodology proposed in [41,42].

#### **4. Model**

We use an ordered logit model to explore the variables that may be associated with the probability of a member of academic staff being appointed at a certain level. A member of staff's appointment is characterized as being separated into five ordered levels, which we call grade 6, grade 7, grade 8, grade 9, and grade 10, with different terms of contract (open-ended, on probation and fixed-term for grades 6, 7, and 8, and open-ended and fixed-term for grades 9 and 10). Grade 6, for example, is typically the entry level for a tenure-track academic member of staff, but it is also the level at which a postdoc on a fixed-term contract may be hired. Grade 10, at the highest end of the spectrum, is that of full professor. Most appointments at grades 9 or 10 are open-ended, although occasionally some are fixed-term. Very rarely, however, do they involve a probation period, and we only had two observations of grade 9 and two of grade 10 on probation, which we merged with those on open-ended contracts. This is not controversial because at UK universities those on probation are typically confirmed on open-ended contracts. The grading system across UK universities is fairly similar, as is the associated salary scale. Because each grade has an associated salary scale, grade and salary are virtually interchangeable at most departments and universities. The actual number given to a certain grade (6, 7, etc.) does not matter in itself as long as it is clearly defined.

In the survey we did not ask what grade respondents were appointed at, but rather, we asked for the title of their posts, so that these could be linked to a consistent grade scale which we defined as shown on Table 1.



Depending on personal preferences, an academic may prefer to be appointed at grade 9 on a fixed-term contract rather than at grade 7 on an open-ended contract, or vice versa. In other words, when grade and type of contract are combined, it is not possible to order all the possible combinations. Thus, an order can be established for:


Furthermore, fixed-term appointments, by definition, almost never lead to appointments at the professorial level. Thus, given that the aim of this study is to examine the association between gender and academic rank, which we also call grade, we exclude respondents on fixed-term contracts, which represent 26% of our sample, and focus on those either on probation or on open-ended contracts.

Having excluded the fixed-term contract cases, our dependent variable is grade, which ranges from grade 6 to grade 10, taking values 1 to 5 correspondingly. The type of contract can be either probation or open-ended and these two are not discriminated within this categorical variable.

We consider a number of independent variables detailed in Section 5 and use an ordered logistic model:

$$\operatorname{Grade}^\* = X' \times \beta + \varepsilon$$

where *X* is the column vector of individual characteristics and β is the column vector of coefficients to be estimated by the ordered logistic regression, with ε assumed to follow a logistic distribution.

#### **5. Results and Discussion**

All our results were computed with STATA. Tables 2 and 3 present all the variables we used and their descriptive statistics.


**Table 2.** Categorical variables and their descriptive statistics (unweighted sample).

All the regressions we report were estimated with weights, which were computed as explained in Appendix B.


#### **Table 3.** Numerical variables and their descriptive statistics (unweighted sample).

#### *5.1. Baseline Model*

Our baseline model includes gender, year of birth, number of children under the age of 18 years, and responsibility for the household chores (cooking, shopping, cleaning, washing/ironing). As it can be seen from the column reporting the results of the baseline model in Table 4, being a woman has a negative and significant association with academic rank. This is not worrying because we are not controlling for research productivity at this stage.

We also find the usual and expected result that the younger a person is, the less likely he/she is to be high up on the academic ladder, an intuitive result in line with [12,13,17].

The number of children under the age of 18 years has a positive and significant association with grade. This result holds for the whole sample but also for the subsample of men and the subsample of women separately, although for brevity, the subsample results are not reported here. Previous research found that having children is positively associated with the academic rank of men, but found that either it has a negative association with the academic rank of women [10], or the association with the academic rank of women is not statistically significant [12,13,15].

Ours is therefore an interesting result. The problem with observational data is that it is not easy to determine causality. From an intuitive point of view, it is unlikely that having children under the age of 18 years has a positive impact on academic rank and it is more likely that academics wait to have their children until they have reached a certain grade. We further investigate this issue below.


**Table 4.** Ordered logistic regression of grade on alternative model specifications.

Note: Standard errors are in parenthesis. \* (\*\*) (\*\*\*) indicate statistical significance at the 10 (5) (1) % levels.

The variable household chores has the correct sign but it is not significant. We also included a number of other variables, such as ethnicity, childcare responsibilities, and responsibility for looking after a disabled, sick or elderly friend or relative, none of which were statistically significant. On similar lines, another study finds that neither care of an elderly parent or relative nor time spent on household or childcare duties has a significant association with research productivity of faculty men or women (pp. 434–435 [30]).

We also tried marital status, but this was also non-significant, in line with (p. 400 [12]) and (p. 51 [13]). On the other hand, one study finds that having a spouse or partner employed at the same institution is unrelated to tenure and rank amongst women faculty but being married is positively associated with both tenure and rank for men faculty (p. 301 [15]), and another study finds a positive association between being married or living with a partner and salary [16].

#### *5.2. PhD, Publications, Grants and Area of Research*

Papers published in peer reviewed journals, papers published in conference proceedings, and number of grants obtained are typically seen as important for career progression in academia, and thus we included those variables in our model. We also included the variable PhD (no PhD degree, one PhD degree, two PhD degrees). In addition, we included research area in order to control for differences across different fields of knowledge. The results are reported on Table 4, under the column entitled PhD and research productivity variables. As in the baseline model, gender has a negative coefficient and is statistically significant. Given that we are controlling for research productivity, this result is very worrying and may be an indicator of discrimination against women. Being a woman per se has a negative association with grade. This is in line with findings in [10–13,15,17,19,40]. On similar lines, a study on faculty salaries, finds a negative association between being a woman and salary (p. 595 [16]).

The variables year of birth and number of children under 18 have the same sign as before and are significant. Again, the variable household chores is not significant.

Having a PhD, as expected, has a positive association with grade, although the variable is only significant at 10% in this specification. Another intuitive result, similar to that found in [18], is the positive association between the number of grants obtained in the last five years and academic rank. The reason that neither the number of papers published in peer-reviewed journals nor the number of papers published in conference proceedings in the last five years is statistically significant, even though both coefficients have the expected positive sign, is that these two variables are correlated between themselves and with the number of grants, as could have been reasonably expected. Importantly, all three variables are significant at least at a 5% level when they are included alone in the model. Another study finds that the number of publications is important for academic progression, but grants obtained are not [20], probably due to the two variables being correlated, although it does not consider this as a possible explanation for this counterintuitive result.

The reference (research) area in this and all specifications in this study is area 1 (Science). This is an arbitrary choice as any area could have been used as reference area.

The results show that for the models on Table 4 that take into account research area, relative to area 1 (Science), there are no significant differences, except for area 3 (Social Sciences), i.e., academics working in Social Sciences are likely to hold a higher rank, everything else being equal.

We also tried models which included marital status and ethnicity but none of these variables proved to be statistically significant.

In addition, we estimated a number of OLS regressions with journal publications, conference proceedings, and grants as dependent variables, and gender, area, grade, and number of children under 18 as independent variables. The results are presented in Table A6 of Appendix C. The coefficient for gender was negative and significant, albeit at 10%, for journal publications, i.e., women publish less, in line with [14,17,21–27]. For conference proceedings and for grants, the coefficient for gender was not significant. The coefficient for grade was positive and significant in all three regressions. A higher grade may "provide the level of resources and job security that serve to bolster one's level of productivity" (p. 436 [30]) or academics with higher grades may be simply more experienced and therefore more productive. The coefficient for the number of children under the age of 18 years was not significant in the journal publications or the conference proceedings regressions, in line with [27,29–31]. One study finds a positive relationship between having children and journal publications for female economists but no relationship for male economists [32]. Our coefficient for the number of children under 18, however, remained not significant even when we ran separate regressions for men and for women, although for brevity, these are not reported. The coefficient for the number of children under 18 was positive and significant in the grants regression.

#### *5.3. Percentage of Time Spent on Teaching and Teaching-Related Activities*

The percentage of working time allocated to different activities during the working day can have an impact on academic rank. Thus, we specified a model which includes the percentage of time spent on teaching and teaching-related activities, as reported by respondents. The last column of Table 4 shows the results. Gender, year of birth, and number of children all have the same signs as before and are statistically significant. The variable household chores continues to be not significant and having a PhD has the same sign as before and continues to be significant at a 10% level. The variables on research productivity have the same signs and significance as before. For the area of research, relative to area 1 (Science), there are positive differences for area 3 (Social Sciences), at a 1% level, and for area 4 (Arts and Humanities), at a 10% level, i.e., academics working in Social Sciences or in Arts and Humanities are likely to hold a higher rank than academics working in Science, with everything else constant.

The coefficient for percentage of time spent on teaching and teaching-related activities is negative and statistically significant, in line with [18]. On similar lines, another study finds that "involvement in teaching negatively affects salary" (p. 886 [43]). Either teaching does not help career progression or those in lower academic ranks are given a heavier teaching workload, or both, potentially making this a vicious circle.

We also estimated the OLS regressions of Table A6 again, adding the percentage of time spent on teaching and teaching related activities as an independent variable. The results are reported in Table A7 of Appendix C. The coefficient for gender ceased to be significant in the journal publications regression, remained not significant in the conference proceedings regression, and was positive and significant, albeit at 10%, in the number of grants regression. This is a key result because it reveals that once we control for the share of time spent on teaching, women publish as many journal papers as men and get more grants than men.

The coefficient for grade continued to be positive and significant in all three regressions. The coefficient for the number of children under the age of 18 years continued to be not significant in the journal publications and in the conference proceedings regressions, and positive and significant in the grants regression.

Crucially, the coefficient for percentage of time spent on teaching and teaching related activities was negative and significant at a 1% level in the journal publications regression. It was also negative and significant, albeit at a 10% level, in the conference proceedings and grants regressions. Although we cannot establish causality this is a very important result.

In order to understand whether the percentage of time spent on teaching affects the academic rank of women and men differently, we estimated the same model for men only and for women only, but this time we dropped the variable household chores, which was consistently not significant in Table 4. Table 5 shows the results. The variables year of birth and number of grants have the same sign as before and are significant. Number of children under 18 also has the same sign as before and is significant, but only at a 5% level for women. The coefficients for journal and conference publications continue to be positive and not significant, except for journal publications in the case of women, which is now significant. PhD is not significant any longer in the case of women. For areas of research, relative to area 1 (Science), there are no significant differences, except for area 3 (Social Sciences) in the case of men. Percentage of time spent on teaching and teaching-related activities is still negative and statistically significant in both cases but with a slightly lower coefficient for men.

In order to understand how correlated teaching is to gender, as well as to area of research and academic rank we estimated an OLS regression. Table 6 shows the results.

As it can be seen on Table 6, the coefficient for area 2 (Medicine and Life Sciences) is negative and significant and the coefficients for area 3 (Social Sciences) and 4 (Arts and Humanities) are positive and significant, implying that the percentage of time spent by faculty on teaching is lower in Medicine and Life Sciences relative to Sciences, and higher in Social Sciences and Arts and Humanities, relative to Science. Although counterintuitive at first sight, many teaching contact-hours in courses falling under the remit of Medicine, Life Sciences, and Science tend to rely on lab and class work, usually led by teaching assistants, demonstrators, and PhD students, who are on casual and fixed-term contracts, rather than on faculty. Faculty in the Arts and Humanities and in the Social Sciences, on the other hand, tend to bear most of the contact-hours with students, and hence the difference in coefficients.


**Table 5.** Ordered logistic regression of grade for men and women subsamples.

Note: Standard errors are in parenthesis. \* (\*\*) (\*\*\*) indicate statistical significance at the 10 (5) (1) % levels.

Importantly, the coefficient for gender is positive and significant. According to these results, the women in our sample tend to spend a higher percentage of their working time on teaching and teaching-related activities than their male counterparts. This result is in line with findings in [25,27,44].

We also find that the lower the academic rank, the higher the percentage of time spent on teaching. Since women tend to have lower academic ranks than men, the two effects may have some synergy and become an obstacle for academic progression. With that in mind, we present the results of a second regression, where a statistical interaction term, gender × grade, is also included as an explanatory variable. The coefficient of the interaction term is the difference in the effect of grade between men and women. The fact that the interaction is significant, albeit at a 5% level, indicates that the effect of grade is different for men and for women. It should be noted, however, that the variable grade is now not statistically significant, which is not a problem because adding an interaction term drastically changes the interpretation of all the coefficients, i.e., the effect of grade is now conditional on the value of gender (and vice-versa). The effect of grade is now −0.896 for men and −3.061 for women. This is obtained as −0.896 − 2.165 × 0 = −0.896 and −0.896 − 2.165 × 1 = −3.061, respectively. Put more simply, going up one grade (say from lecturer to senior lecturer or from senior lecturer to reader) reduces the percentage of time spent on teaching by a factor of 0.896 for men and by a factor of 3.061 for women. Women going up the academic ladder see the percentage of time they spend on teaching and teaching-related activities decrease more than men going up the academic ladder, everything else being equal.


**Table 6.** Linear regression of percentage of time spent on teaching and teaching-related activities on gender, area of research, grade, and gender × grade.

Note: Standard errors are in parenthesis. \* (\*\*) (\*\*\*) indicate statistical significance at the 10 (5) (1) % levels.

The effect of gender is now 9.799 − 2.165 × grade, with grade taking values between 1 and 5. It is easy to check that this effect decreases as grade increases, and becomes negative for the highest grade, 5, which is that of professor.

To summarize, the results from the second regression on Table 6 indicate that, for any given research area, 1, 2, 3 or 4, the percentage of working time spent on teaching and teaching-related activities is higher for women than for men at all grades, except for that of professor, when it is finally slightly lower, thanks to the more rapid decrease they experience, relative to men, as they progress on the academic ladder.

A higher percentage of working time spent on teaching and teaching-related activities may be an indicator of a heavier teaching load. Since we did not ask any question about the total number of hours effectively worked per year (rather than contracted), we cannot discard the possibility that men and/or those in higher academic ranks work many more hours than women and/or those in lower academic ranks, in which case the percentage of working time spent on teaching and teaching-related activities could potentially be lower even if the actual teaching load (measured for example by contact hours and number of students) were the same or higher.

Heavier teaching loads on women could be the result of subtle, probably unintentional, discrimination, which arguably, becomes less obvious as women progress academically and the percentage of time they spend on teaching and teaching-related activities decreases more than that of their male counterparts, for each grade they progress.

#### *5.4. Timing of Children*

The most puzzling result in this study is that the variable number of children under the age of 18 years has a positive association with the academic rank for both men and women, and not just for men, as previously found in [10,12–15]. Our results are more in line with [11], which finds that although young children (under the age of 6 years) reduce the chances of women getting a tenure-track job, older children (aged 6 to 17 years) have a positive association with women getting a tenure-track job and with both men and women getting tenure. Interestingly, in contrast with us, the authors find no effect of children, young or old, on men or women being promoted to full professor [11]. They argue

that the need to provide for their children motivates academics to get tenure-track jobs and tenure, but once tenure is secured there is no motivation to get a full professorship on economic grounds as they have already ensured that their children will be provided for [11].

The answer to the puzzle of this positive association between the number of children under the age of 18 years and academic rank in our results seems to be linked to the timing of children. One of the questions in the survey asked if the respondent's timing with regard to having a child had been influenced by promotion/tenure/job permanency concerns. Therefore, we estimated two regressions, one for those whose timing was influenced by career concerns and one for those whose timing was not. Table 7 shows the results.


**Table 7.** Ordered logistic regression of grade for those who timed their children with career considerations in mind and those who did not.

Note: Standard errors are in parenthesis. \* (\*\*) (\*\*\*) indicate statistical significance at the 10 (5) (1) % levels.

The coefficient for year of birth in Table 7 continues to be negative and significant. Having a PhD is not significant any longer. Journal and conference publications have the expected sign and are now significant for the subsample of respondents who timed their children with career considerations but are still not significant for the subsample of those who did not. The coefficient for number of grants continues to be positive and significant. The percentage of time spent on teaching and teaching-related activities continues to be negative but is now not significant for the subsample of respondents who timed their children with career considerations.

For the subsample of respondents who timed parenthood, the results for area of research are as follows. Relative to area 1 (Science), there are positive differences for area 3 (Social Sciences) and for area 4 (Arts and Humanities), both at a 1% level. For the subsample of respondents who did not time parenthood, relative to area 1 (Science), there are negative differences for area 2 (Medicine and Life Sciences) and positive differences for area 3 (Social Sciences), albeit both only significant at a 10% level.

Moving on to the variables gender and number of children under 18, both of which have been consistently significant in all our models, Table 7 shows what could be regarded as the most important findings in this study. The coefficient for the variable number of children under the age of 18 years, which was positive in all our previous regressions, remains positive and significant when only those whose timing was influenced by career considerations are included in the sample. However, when only those whose timing was not influenced are included in the sample, the coefficient for the variable number of children under 18 becomes not significant. Another interesting result is that the coefficient for gender, which was consistently negative and significant in all our models, becomes not significant for those who timed their babies with career considerations, and we discuss this further below.

A caveat that needs to be highlighted is that the samples are rather small in both cases because: (a) 54% of all those on open-ended contracts and 71% of those on probation did not have children under the age of 18 years at the time of the survey, (b) the already small group of respondents who did have children under the age of 18 years was split into those who timed and those who did not time parenthood with career considerations, and (c) the sample of those who timed their children would have been 27% larger and the sample of those who did not time their children would have been 22% larger if all respondents with children under the age of 18 years had disclosed their age. Dropping the variable year of birth would make the samples larger but an important control variable, significant in all our models, would be lost in that case.

The association between number of children under the age of 18 years and higher grade does not equal causality. Given that the variable number of children under 18 is significant for the sample who timed their children with career considerations in mind but not significant for the sample who did not, there would appear to be some evidence to suspect that rather than children having a positive impact, children arrived after a certain grade had been secured.

The fact that the variable gender, which was negative and significant in all our models, becomes not significant for the sample who timed their babies, could also be seen as evidence that women who timed their children secured a certain grade first, thus protecting themselves from discrimination, or at least discrimination after having children.

Timing seems to be key. This important finding implies that women may find the decision of when to have a baby excruciating because postponing motherhood could cost them not ever having children at all, as fertility declines with age. There is evidence that "women are more successful in obtaining academic careers if they delay or forsake marriage and children" (p. 401 [11]) and that academics who did not have children often regret the decision later in life when it is too late, as do those who wish they had had more children (p. 69 [10]). A qualitative study also finds that "women academics have been tailoring their personal lives to fit their professional lives" (p. 223 [45]).

The fact that men and women may need to time their reproduction per se reflects that academia is not women friendly. Furthermore, it is worth highlighting that although 50% of our (unweighted) sample of respondents were of childbearing age (42 years old or younger) at the time of the survey (i.e., 2013), 60.7% did not have any children under the age of 18 years. In England and Wales, about 20% of women are childless at the age of 45 years [46], compared to 53% in our sample. Furthermore, 15.9% reported that their decision on whether to have or not to have children had been based on career considerations, and 57% from those whose decision on whether to have children or not had been based on career considerations did not have children of any age.

For comparison purposes, 66% of the surveyed academic women in [3] did not have children under the age of 18 years even though 80% of them were 50 years old or younger, 53% of academic women in the sample in [27] did not have any children, and 42% of academic women with tenure in the sample in [10] did not have any children. In our sample 48% of women on open-ended contracts did not have children of any age.

Figure 1 shows the percentage of respondents in our sample who did not have children of any age at the time of the survey by gender and type of contract.

**Figure 1.** Percentage of respondents who do not have any children by gender and type of contract. Source: Unweighted survey responses.

Figure 1 supports our finding about timing of children with career considerations, and is in line with [45], who finds that academic women tend to time having babies for after they have secured permanency. In our sample, this conclusion also applies to men. However, for every type of contract, the percentage of men who do not have any children is lower than the percentage of women who do not have any children. For the whole sample the difference (50% versus 59%) is statistically significant at 1%.

Figure 1 also shows that the percentage of respondents that do not have any children decreases as the terms of their employment become more secure. This can be explained by two logical, intuitive reasons. One reason could be simply responsible parenthood, which concerns the consideration of the factors that have a bearing on whether to start a family and also, on family size. Potential parents may decide that in order to provide for the basic and also other needs of their children they would rather have a permanent job, or at least, be on the track to one. Another reason could be simply that the average age of all respondents on open-ended contracts at the time of the survey was 48 years, and so most of those respondents wanting to have children would have already had them. This age-related explanation, however, does not seem to apply fully to our sample because the average age of those on probation was 35 years, three years younger than the average age of those on fixed contracts. Despite those on fixed contracts being older, on average, than those on probation, the percentage of those with no children was higher.

Despite the caveat of "responsible parenthood" the statistics from Figure 1 are somewhat worrying and tell a story of the kind of working environment that academia is, or at least is perceived to be. This is surprising given that all universities have written policies on work-life balance, which at least on paper, support family life. Clearly, perceptions need to be changed, so that structural change can be brought about. We discuss some policy recommendations regarding this issue in the last section.

#### *5.5. Other Variables*

It is worth noting that we estimated many alternative specifications of the model, including a number of other variables. For example, as well as papers in peer-reviewed journals and papers in conference proceedings included in some of our tables, we also tried guest-edited journal volumes, chapters in books, authored books, and edited books. None of these variables was significant. We also tried variables on availability of flexible working arrangements, part-time opportunities, good career guidance, influential role models and/or mentors, support from senior colleagues, support from other colleagues, knowing the "right people" within the respondent's institution and/or outside, availability of good childcare, and support from partner/spouse. In addition, we tried variables on academic activities which respondents had to reduce involvement in and/or attendance to because they were pregnant/expecting a child and/or had preschool age children, such as committees/boards memberships, refereeing and peer reviewing, guest-editing journal volumes, being main editor of a journal, being on Editorial Boards of academic journals, invitations to present keynote speeches, lectures or chair sessions at conferences, presenting other papers at conferences, amongst others.

The variables that consistently proved to be significant in our regressions were gender, number of children under 18, percentage of time spent on teaching and teaching-related activities, and number of grants obtained.

#### **6. Policy Recommendations**

The 24 Russell Group universities have a number of policies in place already to support work-life balance and family life, including flexible working arrangements and part-time opportunities. In the UK, all employers also offer unpaid parental leave schemes to care for children under the age of 18 years. In addition, the biological father or the mother's partner (regardless of gender or marital status) is typically entitled to one or two weeks of paternity leave following the birth or adoption of a child, with at least statutory pay, and in some cases, full pay. Most Russell Group universities also offer generous maternity leave packages, with new mothers being entitled to up to 52 weeks of maternity/adoption leave, with at least the first 18 weeks being paid at 90% of their salary. Some universities have even more generous packages. In 2015, the UK government also introduced shared parental leave, which allows parents to share up to 50 weeks of leave and 37 weeks of statutory pay after their child is born. All 24 universities offer this. The uptake of shared parental leave in the UK has been low mainly due to workplace culture as well as parents' views, which see the mother as the primary caregiver, especially in the first year, and the complexity of the shared parental leave policy [47]. Another factor for the low uptake may also be financial, as in many cases the combined income is lower with shared parental leave than with the traditional maternity leave.

Many of the Russell Group universities offer subsidized childcare within campus, others offer subsidies for childcare off campus, and the UK government also offers tax-free childcare, albeit with a cap. In addition, most universities offer career guidance through appraisal schemes for men and women, and in some cases, through workshops designed by and for women specifically. As explained in Section 5.4, we tested all of these variables but they were not statistically significant, which does not necessarily imply that these policies and benefits are not important. If they were not in place, the gender gap would probably be wider. Despite all these policies and benefits, our results show that women tend to hold lower grades than men. In order to achieve structural change at the institutional level and facilitate the advancement of women in academia, we propose the following two policies, following up from the variables that were found to have an association with academic rank: transparent workload models and promotion on the basis of clear and transparent criteria.

Universities should have systems in place to allow a fair and equitable distribution of teaching (and administrative) loads amongst faculty as well as continuous monitoring of such distribution. This could be actioned through, for example, a transparent workload model where everyone can see everybody else's teaching loads, including number of courses taught, contact hours, number of students, marking, dissertation supervision, etc. Some British universities, including some in the Russell Group, have already adopted or are in the process of adopting workload models. Some are university-wide workload models and others are designed within Schools or Departments. The tariffs used vary across institutions, and sometimes, across Schools or Departments within the same institution, and are at present the subject of much debate. The tariffs of any workload model meant

for academics should be set by academics, as academics know the time it takes to prepare a lecture, mark an exam, supervise a student project, write a journal paper, prepare a research proposal, etc. In addition, promotion should be based on clear and transparent criteria. Although there are typically three criteria by which candidates for promotion are judged (research, teaching, and administration), these criteria are not equally weighted (p. 2 [7]), (p. 47 [22]). The decisive factor for promotion is research, i.e., if a candidate's research is deemed inadequate, no amount of teaching or administration will compensate for this (p. 48 [22]). If this is the path that the Russell Group Universities want to stick to then this should be made crystal clear and no claims of the possibility of promotion on the basis of teaching (or administration) excellence should be made. Guidelines should be communicated to all staff so that everyone is clear that the most important criterion for promotion is research. However, if universities are going to continue with their current (written) policies for promotion, many of which include excellence in teaching, then, these policies should be implemented in practice. Excellence in teaching, however, is difficult to demonstrate. Student evaluation, for example, could be one of the metrics, although this is frequently positively correlated with faculty evaluation (higher grades on average) of students and small class sizes [22]. Peer and other evaluations may also be controversial, so careful thought would need to be given to how excellence in teaching can be established.

Adopting these two policies will help reduce the discriminatory teaching loads on women, which is a contributor to their lack of progression, and will make promotions fairer and more transparent, with a probable outcome of having more women climbing up the academic ladder.

#### **7. Conclusions**

Using an ordered logit model and the results of a survey, which we conducted in 2013, with 2270 observations of academics of both genders at all levels in all fields of knowledge at the 24 Russell Group universities in the UK, we have examined the association between gender and academic rank, controlling for a number of variables, including but not limited to, respondent's year of birth, number of children, responsibility for household chores, academic degrees, number of publications, grants, percentage of working time spent on teaching and teaching-related activities, and main area of research.

One caveat that should be highlighted is that this study only finds associations with models that use observational data. Causal relationships cannot be identified with the current dataset. Still, the associations found are very important and can guide policy.

Our findings can be summarized as follows.

A negative association between being a woman and academic rank is indeed observed in all our models but one, when run for a small subsample of male and female academics who timed their children with career considerations in mind. In general, however, women are less likely to hold a higher academic rank even after controlling for individual characteristics using variables like respondent's year of birth, marital status, responsibility for the household chores, area of research, number of children under 18, holding a PhD or not, percentage of working time spent on teaching and teaching-related activities, and a number of research productivity variables. This result is in line with [10–15,17,19,40], all of whom also find that women tend to progress at a lower rate than men, even after accounting for variables that would capture family formation and/or academic/research achievements. We call this the gender effect. Put simply, two people who have similar, or even identical credentials and personal circumstances except for one being a man and the other being a woman, are likely to have different academic ranks, with the man having a higher rank than the woman. One explanation for this phenomenon may be discrimination against women.

Another important finding is that the percentage of time spent on teaching and teaching-related activities has a negative and statistically significant association with academic rank, in line with [18]. On similar lines, another study finds a negative association between teaching and salary [43]. Furthermore, our results show that women spend a higher percentage of their working time on teaching and teaching-related activities than men at all academic grades, except for that of professor. This is in line with [25,27,44], which also find that women tend to spend either more time or a higher

percentage of their working time on teaching and teaching-related activities, but in contrast with [3,7], which do not find differences between the genders related to absolute or relative time spent on teaching and teaching-related activities.

In addition, we find that going up one grade (say from lecturer to senior lecturer or from senior lecturer to reader) reduces the percentage of time spent on teaching more for women than for men, and so eventually, female professors spend a lower percentage of their working time on teaching and teaching-related activities than male professors.

If a higher percentage of working time spent on teaching and teaching-related activities is to be taken as an indicator of a heavier teaching load, then we can conclude that women at all ranks, except for that of professor, are being discriminated against. At the same time, relative to men, women experience a higher reduction in the percentage of time spent on teaching and teaching-related activities by going up one grade.

Another important result, which is new and has not been quantified before for the UK, is a positive and significant association between number of children under the age of 18 years and the academic rank of both men and women, as long as babies were timed with career considerations in mind. In line with [11], the reason for this is very unlikely to be that children have a positive impact on academic rank, other than triggering their parents' eagerness to achieve a certain level of job stability and income in order to provide for them. What this result is probably showing is that children arrived after a certain rank (for example, an open-ended contract) had been secured. Importantly, for the subsample of academics who timed their children, the variable gender ceases to be significant.

These findings pose a dilemma for women because the 30 s is the decade when they have two competing goals in their lives: establishing themselves in their careers having finished their doctorates, and having children. Delaying pregnancy can mean that these women are left childless as fertility declines with age, especially after the age of 35 years. Some further inspection of our data confirms our finding about timing of children with career considerations: the percentage of respondents that do not have children (of any age) decreases as the terms of their employment become more secure. This state of affairs is especially biased against women.

**Author Contributions:** Conceptualization: G.S., S.D.V.P., Methodology: G.S., S.D.V.P., Acquisition and analysis of data: G.S., S.D.V.P., Writing: G.S., S.D.V.P., Reviewing and editing: G.S.

**Funding:** The authors did not receive any funding for this work.

**Acknowledgments:** We are thankful to Pauline Card, Andrea Collins, Craig Gurney, Rob Huggins, Bob Smith and Pete Mackie from the School of Geography and Planning at Cardiff University for useful feedback on the survey questions. After going live, the survey had a bug and we are grateful to Neil Dodgson, back then at the Computer Laboratory at the University of Cambridge, now at the School of Engineering and Computer Science at Victoria University of Wellington, New Zealand, for identifying it.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **Appendix A**

In this Appendix we include the survey that was conducted in 2013, which provided the data for this study.

#### **Survey: Gender and Academic Progression**

#### WELCOME

My name is Georgina Santos and I am a lecturer at Cardiff University.

I am undertaking a piece of research to assess and understand whether there are any problems linked to Gender and Academic Progression. In 2010/11 44.2% of all the academic staff employed at UK Higher Education Institutions were female, yet only 19.8% of Professors were women (Higher Education Statistics Agency, 2012).

I would be very grateful if you could complete this questionnaire, which is essentially the same questionnaire that was conducted in 1999–2000 by the National Centre for Social Research (Blake, M. and I. La Valle, 2000, "Who applies for research funding", report published by the Wellcome Trust), although the aims and objectives of that piece of research were different from mine.

*Higher Education Statistics Agency (2012), Sta*ff *at higher education institutions in the United Kingdom 2010*/*11. https:*//*www.hesa.ac.uk*/*news*/*19-01-2012*/*sfr170-sta*ff *.*

*Blake, M. and I. La Valle (2000), Who Applies for Research Funding? Key factors shaping funding application behaviour among women and men in British higher education institutions, An independent summary report prepared for the Biotechnology and Biological Sciences Research Council (BBSRC), the Economic and Social Research Council (ESRC), the Engineering and Physical Sciences Research Council (EPSRC), the Medical Research Council (MRC), the Natural Environment Research Council (NERC), the Particle and Physics Research Council (PPARC) and The Wellcome Trust, London: The Wellcome Trust. https:*//*wellcome.ac.uk*/*sites*/*default*/*files*/*wtd003209\_0.pdf.*

#### DATA PROTECTION

For the purposes of this survey Cardiff University is the data controller. All data collected in this survey will be held securely by the survey software provider (University of Bristol) under contract and then retained by the research team working on the project "Gender and Academic Progression' at Cardiff University in accordance with the Data Protection Act (1998). Data from the survey, including answers to questions where personal details are requested, will only be used by the research team for research purposes and will not be shared with anyone outside the research team.

Participation in the survey is completely voluntary and you may withdraw at any point. You may also complete part of it and save it to complete it later.

Cookies, personal data stored by your Web browser, are not used in this survey.

#### *Background & Demographic Information*

1. What is your gender?

Male Female

	- Married Living with a partner Separated Widowed Single Other

*Sustainability* **2019**, *11*, 3171


If you have children aged 18 years or under, please tick Yes, No or Not applicable. Which of the following have you used while in your current job? *Please tick one column in each row.*


6. Do you have responsibility for looking after a disabled, sick or elderly friend or relative? *(Optional)* Yes No

Partially


Other

Please note that the following questions apply whether you are a man or a woman.

9. Please select the options(s) that best describe your situation(s). *(Select all that apply)* Are or were expecting a child before earning tenure/getting an open-ended contract/being confirmed on post until retiring age.

Have or had pre-school age children to care for before earning tenure/getting an open-ended contract/being confirmed on post until retiring age.

Are or were expecting a child after earning tenure/getting an open-ended contract/being confirmed on post until retiring age but prior to promotion to full professor.

Have or had pre-school age children to care for after earning tenure/getting an open-ended contract/being confirmed on post until retiring age but prior to promotion to full professor. None of the above

10. Please tick one box in each row.



Note: RAE: Research Assessment Exercise, REF: Research Excellence Framework.

	- Yes No

Not applicable

12. Is or was your timing regarding having children influenced by promotion/tenure/job permanency concerns? Yes

No

Not applicable

	- Senior Researcher
	- Researcher
	- Research assistant
	- Teaching fellow
	- Senior teaching fellow

Other (please specify): Please state the precise year when you obtained the previously reported grade:





#### *Current employment conditions and workload*

If you have more than one job, please answer the questions in this section for the academic/research job on which you spend most time. If you spend equal time on two jobs, answer for the one which you have held for longest.

18. Which of the following are available in your current job (whether formally or informally)? *Tick yes, if they would be available to you if you had children or you were expecting a child. We would like to hear from all respondents, even if the benefits are not applicable to you or you don't k now if they are available. Please tick one column in each row.*


19. Approximately what percentage of your time do you spend on the tasks below in an average week:


*Please record the percentage of time you actually spend on the tasks rather than contracted time. If the time for any of the tasks is none, please enter "0". If you have two jobs, please provide the detailed information only for your main job as a percentage of your total hours in that job.*


#### *Career and education history*

20. Please indicate which was your main activity in each of the last 10 academic years. Your main activity is that which you were engaged in for the longest period of time in that year. Please read all columns before ticking any. If more than one applies, please tick the one closest to the left of the grid. Please enter a tick on each row. Include years during which you were in full-time education.

Please indicate which was your main activity in each of the last 10 academic years. *If more than one applies, please tick the one closest to the left of the grid.*

#### *Sustainability* **2019**,*11*, 3171


#### *Academic qualifications*

21. Please list all your academic qualifications. For pending awards (exams taken or thesis submitted but not yet awarded), please enter "pending" in the "Year of award" column. Please give all the information requested in the column headings.


#### *Publication record*

22. How many of the following have you had published in the last five years (i.e.: since January 2008)? Please include joint and single author publications, publications through consortia, articles "in press" and those available online but not on paper yet and "online only" as well. *If the answer for any category is none, please enter "0".*


#### *Other academic activities*

23. Have you been involved in any of the following in the last five years (i.e.: since January 2008)? *Please tick one box in each row.*



24. Were you included in your department's 2008 Research Assessment Exercise (RAE)? *Please tick one only. (Optional)*

Yes No Not applicable (e.g., not in the department at the time) I don't know

25. Will you be included in your department's 2014 Research Excellence Framework (REF)? *Please tick one only. (Optional)*

Yes, definitely Yes, probably No Not applicable (e.g., not in the department at the time) I don't know

#### *Attitudes*

26. Regardless of your gender and whether you have children or not, please answer the following question. Which of the following have been available to you in your academic or research career to date? If any of these are not relevant to you, please tick the 'Not applicable' box. *Please tick one box in each row.*


27. At this stage in your career, in order to gain promotion in your institution, how important is your performance in the following areas? *If any of these are not relevant, please tick the 'Not applicable' box.*


*Grants and commissioned research*

28. Have you obtained any commissioned research contracts from industry, government departments, charities, etc. in the last five years (i.e.: since January 2008)?

Yes

No

How many such research contracts have you obtained in the last five years (i.e.: since January 2008)? Please write in the number.

Number obtained: . . . . . . . . . . . .

29. Have you been awarded any grants in the last five years, i.e., since January 2008? If Yes, please fill in the table but do not include commissioned research or contracts which were covered in the previous question. Please include the last six grants on which you were named as an applicant, even if you were not named as the principal applicant.

#### *Sustainability* **2019**,*11*, 3171


30. Thank you very much for taking time to complete this survey. If you would like to add any comments about the issues raised in the questionnaire please do so below, on the understanding that we may anonymously quote part or all of what you write.

#### **Appendix B**

In this Appendix we explain why weights were needed to make our sample representative of the whole population, and how they were estimated.

#### **Weights**

The population of the study was all academic and research staff employed at the 24 Russell Group universities. Any member of the population belonging to a department linked to one of the 36 REF areas had the same probability of being invited to respond to the survey. The response rate may have varied according to a number of reasons, some of which were reported by the respondents themselves, such as for example, lack of time or concerns over privacy issues. There are no data on "lack of time" of the population, let alone "lack of time" during the weeks when the survey was live online, or how different individuals feel about disclosing personal information. Other reasons for non-response, and for which there are no population data either, include personal circumstances such as having or not having children under the age of 18 years (which may carry an inherent interest in the research in question but may also reduce the time a member of staff can afford to fill surveys in), personal tastes (i.e., liking or not liking surveys), altruism or selfishness (being prepared to collaborate with a researcher or not), etc.

The characteristics that could also influence response rates and for which there are some data, or at least proxies, on the population are gender, research area, and seniority. Needless to say, in order to correct for the potentially different response rates data on the population is essential. Thus, data for the whole population (academic and research staff at the 24 Russell Group universities) on gender, research area, and seniority was provided by the Higher Education Statistics Agency (HESA) on request, as explained below.

#### *HESA Data on Gender*

The HESA holds data on the legal sex of staff members, as opposed to the gender with which they identify [48].

#### *HESA Data on Research Area*

The HESA does not hold data on the area of research being carried out by each member of the population. However, it does hold data on "cost centers" and "staff members' qualifications". The cost centers tend to have similar cost structures for teaching and research, similar patterns for capital expenditure, academic coherence in terms of the academic disciplines of staff, and similar rates of funding for research grants and contracts. However, given the interdisciplinary characteristics of many departments across the 24 Russell Group universities, it is not unusual to see economists working in Geography departments or Schools of Business, and carrying out research in Economics, or Chemists working in Biology departments and carrying out research in Chemistry, or Physicists working in Chemistry departments and carrying out research in Physics, to name a few examples. For this reason, we decided to use the data on "staff members' qualifications", rather than the data on the number of staff associated to different cost centers. The HESA uses "academic discipline" to designate "the subject or subjects appropriate to that staff member's academic qualification", which although may "not necessarily be the academic subject in which that staff member may currently be teaching or researching" [48], has a much higher chance of being closely related to it than "cost centers".

#### *HESA Data on Seniority*

The HESA does not hold data on the grade at which each member of staff is employed (professor, reader, lecturer, etc.) but holds data on professorial role, i.e., professor or non-professor.

We grouped the 36 REF areas and the 146 different academic disciplines from the HESA in 16 areas. Table A1 shows the mapping. Table A2 shows the number of individuals in the sample and in the population in each of the 16 areas, also classified by gender and by whether they hold a professorial role or not.


**Table A1.**Our classification mapped against REF and HESA classifications.


(P3) Media studies

History

**Table A1.** *Cont.*


**Table A1.** *Cont.*

**Table A1.** *Cont.*



**Table A1.** *Cont.*


**Table A1.** *Cont.*

Source: REF website (https://www.ref.ac.uk/2014/panels/unitsofassessment/) and data provided by HESA on request.


**Table A2.**Populations and sample individuals classified by area of research, gender, and professorial role.

Source: Responses from our survey and data provided by HESA on request. Note: "M" denotes male, "F" denotes female, "Prof" denotes professorial status, "Non-Prof" denotes non-professorial status.

As it can be seen from Table A2 the shares (not actually shown since we show the actual numbers) differ between the sample and the population. The reasons for this may be linked to the likelihood of different individuals to respond to the survey. This likelihood may vary with the three characteristics in question, gender, research area, and seniority. Non-response repartition is hardly the result of a random phenomenon. Some types of individual have been over-sampled, and some have been under-sampled, and as a consequence, the distribution of these three characteristics across the sample is different from that of the population. This would introduce bias in any estimate.

In order to correct for survey non-response and reduce any potential bias in the estimates we used weights. The method essentially consists of increasing the weight of the sample respondents to take into account the population of non-respondents. The method chosen is based on the mechanism of homogeneous response within subpopulations. Therefore, the Russell Group population is assumed to be homogeneous concerning non-response within well-chosen subpopulations.

The data from the sample was thus weighted using post-stratification survey weights so that the sample would reflect the distribution of academics in the 24 Russell Group universities according to gender, professorial role (or not) and area of research.

Before applying any weights, we tested whether these were indeed needed. Using the population data provided by HESA, we created a database of 62,637 individuals representing the whole Russell Group population, and therefore containing the three characteristics (gender, area of research, and professorial role). Once we had this database we created a dummy variable Y and assigned a "1" to our survey respondents (contained in the population) and a "0" to everyone else in the population (not included in our sample). The following step was to regress the zero-one response indicator on the three variables (gender, area of research, and professor marker). We used logistic regression for this, following [41,42].

Table A3 shows the results of a logistic regression run for gender, professor marker, and area of research. Table A4 shows the results of the same logistic regression run excluding the gender variable.


**Table A3.** Logistic regression of response indicator on gender, professor marker and area of research.

Note: Standard errors in parenthesis. Gender denotes gender (male or female), professor marker denotes professorial role (either a professor or a non-professor) and area of research denotes one of our 16 areas of research. The independent variable is a dummy variable, which is the response-non-response indicator.

The results show that all three variables were significant in determining whether an individual responded to the questionnaire or not. As we can see, gender seems to have a significant impact on the probability of response to the survey, as when we break by gender, other variables have less explanatory power. The coefficient for the gender variable (coded 0 for man and 1 for woman) is positive and statistically significant, showing that women were more likely to respond to the survey. The log likelihood and pseudo-R squared are also higher when gender is included.

The conclusion from Tables A3 and A4 was therefore that non-response adjustments according to gender, area of research, and professorial role were required and for that we used weights.


**Table A4.** Logistic regression of response indicator on professor, marker, and area of research.

Note: Standard errors in parenthesis. The variables are defined as in Table A3.

Once the subpopulations were defined, the probability of response of all members of that subpopulation was assumed to be the same, i.e., constant within the subpopulation, in line with the mechanism of homogeneous response within subpopulations, as already highlighted above. In addition, this probability was assumed to be independent from the probabilities of response of all the other subpopulations.

When the size of each subpopulation is known there is no need to estimate the probabilities to respond using a logistic regression, as post-stratification estimators are better [42]. The method based on estimated probabilities of response does not allow any control over the dispersion of values. Indeed, the estimator can become very unsteady because of very under-represented types of respondents, which have high weights assigned to them. As argued in [42], the construction of homogeneous groups of respondents conveys more robustness, especially when the model of regression is not accurate.

The weights in our case can therefore be simply estimated by the ratio:

$$w\_h = \frac{N\_h}{r\_r}$$

where *h* = 1 . . . 64 and *h* denotes the 64 different strata, i.e., the 64 possible combinations of characteristics an individual can have (male, female; professor, non-professor, and one of 16 different research areas), *rr* is the number of respondents of subpopulation *h* (i.e., that were included in the sample), *N<sup>h</sup>* is the size of subpopulation *h*.

The problem we have (and we would still have even if we were to use a logistic regression to estimate probabilities of response) is that, as Table A2 clearly shows, in two cases we have zero respondents in our sample (i.e., *r<sup>r</sup>* = 0. The two subpopulations in question are female professors in Business and Management Studies and male professors in Sport and Exercise Sciences, Leisure and Tourism. Division by zero does not exist and therefore *w<sup>h</sup>* = *N<sup>h</sup> rr* cannot be computed. As a solution, we merged Business and Management Studies with Economics and Econometrics and Education with Sport and Exercise Sciences, Leisure and Tourism. The rationale behind the first merge was that many economists work in Business and Management Studies Schools and also academics doing research in Business and Management Studies and academics doing research in Economics and Econometrics share some similarities regarding training. They tend to hold first degrees, masters, and PhDs and often these are gained from departments that have both Economics and Business. The rationale behind the second merge was that the years and type of training tend to be similar. Many academics in those areas do not actually hold PhDs, but they hold postgraduate diplomas and certificates, often requiring about one year of full-time equivalent study/training.

As underlined in [42], the number of subpopulations is the result of a problematic trade-off between increasing the number of strata (ensuring great homogeneity within each stratus) and lowering the number of strata (ensuring a lower variance of the estimator). Once the problematic cells with zeros, which made the calculation of ratios impossible, disappeared, the ratios were computed.

Table A5 shows the weights computed as the ratios of the size of the subpopulation with characteristics *h* to the size of the subsample with characteristics *h*.


**Table A5.** Weights computed as ratios.

Source: Table A2.

The weight coefficients for professor and for non-professor are higher for females than for males in all research areas, except for Medical Sciences. This shows that women had been over-represented among the respondent population. Given the subject (and the results of the logistic regressions shown on Tables A3 and A4), it is not a surprise that women were more prone to respond to our survey than men.

We also note that the difference of weight coefficients between the professor and non-professor subpopulations is much stronger in the female subpopulation than in the male subpopulation. Hence, a women professor is the subpopulation most over-represented in our survey respondents.

The weights computed in Table A5 were used in all our models to make our sample representative of the population.

#### **Appendix C**

In this Appendix we present the results of the linear regressions of the research productivity variables, discussed in Sections 5.2 and 5.3.


**Table A6.** Linear regressions of research productivity variables on gender, area of research, grade and number of children under 18.

Note: Standard errors are in parenthesis. \* (\*\*) (\*\*\*) indicate statistical significance at the 10 (5) (1) % levels.

**Table A7.** Linear regressions of research productivity variables on gender, area of research, grade, number of children under 18 and percentage of time spent on teaching and teaching-related activities.


Note: Standard errors are in parenthesis. \* (\*\*) (\*\*\*) indicate statistical significance at the 10 (5) (1) % levels.

#### **References and Note**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

*Article*
