Female Labor Force Participation in Iran: A Structural. Analysis

Female Labor Force Participation in Iran: A Structural Analysis Mahdi Majbouri Babson College Economics Division 231 Forest St., Babson Park, MA 02457, USA email:mmajbouri@babson.edu phone: 781-239-5549 fax: 781-239-5239 July 28, 2013 1

Female Labor Force Participation in Iran: A Structural Analysis Abstract In the last three decades, women s education levels in Iran have consistently increased while fertility rates have fallen rapidly. But female labor force participation (FLFP) rate remains at low levels. This paper shows that despite the low rate, FLFP follows the same patterns as the economic theory predicts and have been documented elsewhere. Moreover, a structural estimation, controlling for selection, depicts that urban women s labor supply is upward-sloping and quite elastic (1.7) with respect to wages; a surprising result with important implications for researchers and policy makers. Keywords: Female labor force participation, Structural estimation, Iran JEL: J22, O53 1

1 Introduction Women s education in Iran has consistently increased in the last decades. In 1980, for instance, an average Iranian woman aged 20 to 30 had 3.5 years of education, while her counterpart in 2010 obtained 10 years of education (data source: Barro and Lee, 2001). At the same time, fertility rate fell from about 6.5 to less than 2 children per woman (Abbasi-Shavazi et al., 2009). However, female labor force participation (FLFP) rate is still low, at less than 30%. This paper shows that despite the low rate, FLFP in Iran has the same characteristics as predicted by economic theory. Moreover, a structural estimation explains that it has a very elastic response to wages. Using an annual household survey between 1990 and 2004, I first estimate a reduced form regression of a standard static model of Labor force participation (LFP) to understand the factors affecting the long-term decision to participate in the labor market. Results show that FLFP in Iran is correlated with the same factors as predicted by economic theory and documented for other countries. For instance, women with high school education and higher are significantly more likely to participate in the labor force; assets have a negative correlation with participation; women aged 30 to 50 have higher participation rates 1 ; presence of more teenage girls in the household increases participation of older women in the household, while presence of adult males reduces it. These results are robust across all years in data. Moreover, this study provides the first structural analysis of labor force participation of women (and men) in Iran. Correcting for selection, I estimate a structural model to calculate wage elasticity. I find that urban women have an upward sloping labor supply curve. Moreover, the labor supply is quite elastic with respect to wages (about 1.7). Urban married women have an elastic labor supply (1.96), while never-married urban women have a relatively inelastic one (0.81). 2 These results are 1 the age profile is concave 2 Urban men, on the other hand, have a backward bending supply curve as their elasticity is about -0.34. 2

robust to specification and have important implications for policy makers. The rest of the paper is organized this way: In Section 2, I review the theory of LFP and discuss the estimation methods used to apply the theory to the data. Section 3 reviews the two data sets used in this study in detail. Following that, in Section 4, I provide the empirical evidence for the static model both in terms of reduced form and structural estimation. The results are consistent with the predictions of the static model and the evidence for other countries. In section 5, I explain the implications of the results. 2 Labor Force Participation in Theory The economic theory has extensively modeled the factors affecting the decision to participate in the labor market (for a review of this literature see Killingsworth and Heckman, 1986). A set of these models approach the question as if there is a single lifetime period ( static models). In these models, the individual has already been endowed with a set of characteristics and her decision to participate is usually affected by these predetermined characteristics and will not change. For example, her level of education and family background will not change. These models, which are usually applied to cross-sectional data are useful for understanding long-term decisions. In this section, I describe the canonical static model of labor force participation for a household. Consider a household as a single entity that maximizes her utility function as follows max U i (L i f, L m i, X i ; E i ) (1) S.t. px i + W f i Lf i + W m i L m i W f i T + W m i T + Y i 0 L f i, Lm i T 3

in which L f i and L m i are the leisure time of the female and male members of the household i, and X is a composite good. The exogenous elements of this model are Y i, non-labor income, p, the price of the composite good X, and W f i and Wi m, the hourly wage rates of female and male members. E i is a measure of all household characteristics that shape the household i s utility function. Solving for the optimal leisure time, one finds the structural model as follows: L j i = L i(w f i, W m i, Y i, p, E i ) (2) in which E i contains household variables that shapes the utility function especially with respect to labor force participation. It has an observed component V i as well as an unobserved part e i. 3 If the utility function is quadratic, it can be shown that L j i is a linear function and can be written as L j i = ρ + ϕ V i + ϕ f W f i + ϕ mw m i + r Y i + e i + v j i j = m, f (3) Moreover, normalized hourly wage rates (from now on, wages) are functions of predetermined (exogenous) characteristics such as schooling level and age. Wages can be specified as W f i = w f i (Zf i, uf i ) (4) W m i = w m i (Z m i, u m i ) (5) in which Z f i, and Zm i are the vector of observable characteristics of female and male members, such as schooling, respectively. u f i, and um i are mean zero constant variance disturbances. If the disturbances for wages and tastes in Equations (3), (4), and (5), i.e. v j i and uj i (j = f, m), are correlated, the OLS estimates of ϕ f and/or ϕ m are biased and inconsistent. To solve this problem, 3 Note that, one can normalize wages by the price of the composite good, p, in the utility maximization problem and obtain L j i which does not include p. 4

I employ instrumental variables to estimate wages first and then use the predicted wages for all individuals (not just those who worked) to estimate LFP (Equation (3)). This procedure is similar but not identical to the two-stage least square method (2SLS). In 2SLS, we only predict wages for those observations used to estimate the first stage, i.e. those who had worked and reported wages. In this pseudo two-stage least square, the instrument is individual education. It can be argued that education may only explain LFP through wages. In other words, we assume that education affect the decision to work only through wages. More education increases wages and increased wages in turn increase incentives to work. In Section 4.1, I utilize this method to estimate a simpler form of Equation (3). To obtain the reduced form equation, one can substitute normalized wage functions (Equations (4) and (5)) in Equation (2) and get L j i = L i(v i, Z f i, Zm i, Y i, e i, u f i, um i ) j = f, m (6) which can in turn be linearized into L j i = α + β V j i + γ Z f i + δ Z m i + r Y i + e i + w j i j = m, f (7) in which w j i is a mean zero constant variance disturbance that includes uf i and um i. This reduced form shows us what household and individual characteristics are correlated with FLFP. The results for a simplified form of this equation are reported in the beginning of Section 4 and the structural model estimates (Equation (3)) are discussed in 4.1. 5

3 Data Statistical Center of Iran (SCI) is the main organization in charge of gathering micro datasets in Iran in the past six decades. These relatively large datasets provide various statistics for policy makers as well as the general public. In this study, I use two major household datasets, 1) the household expenditure and Income Surveys (HEIS), a series of cross-sectional surveys conducted annually, and 2) the socio-economic characteristics of households (SECH), which are longitudinal. 3.1 Household Expenditure and Income Surveys (HEIS) This annual survey has been gathered since 1963 in rural areas and 1968 in urban areas but only years between 1990 through 2004 were made available to me. The surveys contain basic demographic information, ownership of assets, dis-aggregated expenditure, and income. Each year, new samples are drawn from the population. The samples are nationally representative and stratified at the urban and rural areas of each province. Sample selection follows a two-stage sampling method which has remained the same over the years. In the first stage, based on the most recent census, the total number of primary sampling units (PSUs) in each geographical block (rural or urban areas in each province) are determined, which is equal to the population in the block divided by five. 4 In the second stage, a number of PSUs in each block are chosen to be surveyed. This number depends on the population and variance of some variables of interest, such as food expenditure, in that block. Hence, households have different probability of selection. For instance, more rural households have been selected. The number of households for years 1990 through 2006 varies from 12,763 in 1993 to 36,579 in 1998 and averages at about 22,000. Data gathering process is done uniformly throughout the year so that 1 12 of the sample are surveyed each month. Until 2005, the definition of employment in these surveys was limited to a person who works 4 Each PSU corresponds to a census track and consists of five households. 6

for at least two days during the week prior to the survey. Another issue with these surveys is that hours worked was not asked until 2005. Hence, it is not possible to calculate hourly wages, a major component of labor market analysis with these surveys. This is why I also use another data set for this study. 3.2 Socio-Economic Characteristics of Households Surveys (SECH) Socio-Economic Characteristics of Household (SECH), which are panel datasets similar to HEIS but with more questions on demographics, individual characteristics, and labor market participation. For example, this is one of the few surveys that include variables such as hours worked, age at first marriage, and children ever born, to name a few. Since the 1979 revolution, three separate SECH datasets have been gathered, each time with a different set of households: 1987-89, 1992-95, and 2001-03. Waves of each panel are gathered during November of the years covered by the panel. SECH 1992-95 is chosen to be the dataset used to estimate the structural model since it contains hours worked in addition to income, providing the opportunity to compute wages. The definition of employment in SECH 1992-95 is the same as the definition used prior to 2005 for HEIS which makes these two surveys comparable in terms of employment. First wave of SECH 1992-95 is a nationally random sample of 5090 households, encompassing 172 sampling clusters (62 rural and 109 urban), each having an average of about 30 households. 5 In each cluster, selected households are neighbors living side-by-side. 4 The Evidence To begin with, let us estimate the reduced form, i.e. Equation (7), with LFP dummy, equal to one if the individual participates and zero otherwise, as the dependent variable instead of leisure 5 Almost 73% of rural clusters and 49% of urban clusters have exactly 30 households. 7

hours. Tables 1 and 2 depict the coefficients of the linear probability model of LFP for women aged 21 through 65 in each year between 1990 and 2004. In this case, e i, and w j i would be part of the error term and are assumed to be independent of the covariates. Ignoring male characteristics, the covariates include education, age, urban-rural location, female head dummy, number of female and male household members between age 15 and 18 and above age 18, an index for assets, home value, and province fixed effects. Since the dependent variable is a dummy variable, some continuous variables such as education and age were transformed into categorical variables. Dummy variables for each of these categories were defined and employed. For example, Illiterate, Primary, Mid School, High School, and College and higher are dummies representing levels of education. Age was divided into tenyear categories, 20 to 30, 30 to 40, etc., and dummy variables were used for each. For instance, I[20<Age 30] represents a dummy variable which is equal to one if age is between 21 and 30 and zero otherwise. Women aged 61 through 65 are the control group. Location is identified by a simple dummy, Urban, which is equal to one if the household lives in an urban area and zero otherwise. Numbers of teenage female and male members (between age 15 and 18) are included in the regression as these are children for whom many households, especially in urban areas, allocate considerable resources (to their education) so that they would have a higher chance of entering college. Numbers of adult female and male members (above age 18) are also included in this regression as a proxy for the household s available labor force that can be used to generate income. In the data, assets are simply measured as the ownership of appliances (fridge, stove, TV, radio, cassette player, computer, etc.), transportation vehicles (car, motorcycle, and bike), and access to utilities (electricity, piped water, gas, telephone, and in recent years internet). They are not reported in terms of their value but rather as dichotomous ownership dummies which are equal to one if the household owns the specific asset and zero otherwise. One way of measuring total asset ownership is 8

to define a new variable that is the sum of these dummies. In this method, technically, there is no difference between owning a car or owning a radio as both increase the new defined variable by one unit. Instead of giving the same weights to all assets, one can give different weights to various assets based on how much the ownership of an asset signals the wealth of the household. Sahn and Stifel (2000) use factor analysis to find weights for each type of asset. Their main argument is that weights should be determined by the data that is being used. In this method, assets owned by many households get small or negative weights. The fewer people owning an asset, the larger the weights will be, as it shows that the asset is more valuable. As a second approach, I followed their method and constructed the asset index using weights computed for households in 1992. Interestingly, the results are qualitatively similar using both of these approaches. Therefore, I only report the results using the asset index computed by Sahn and Stifel (2000) factor analysis method. Owned home value, the other covariate in the reduced form regression, is the answer to the question, If you wanted to rent this house, how much would be the rent? It is a proxy for the value of the house (or wealth in general) owned by the household. Province fixed effects were also used to account for any province-level unobservable factor, especially province labor market. 6 As Tables 1 and 2 show, generally, women with primary and middle school education are not likely to work more than illiterate women (the control). 7 As education increases to high school and higher, participation increases significantly. The coefficient of high school is around 0.2 before 1997 and decreases to about 0.14 afterwards. College education coefficient is almost two times that of high school before 1997 and becomes three times that afterwards. Generally, it fluctuates around 0.5. 8 6 Since these are fixed effects, the robust-heteroskedastic standard errors are corrected for correlation within province. 7 Although, there are some discrepancies. For instance, in 1994, 1999, and 2003, women with primary education were slightly more likely to work than illiterate women. The same was true for women with middle school education in 1991, and 1994. In 1998 the opposite was the case. But, interestingly, the coefficients are small and it can be safely inferred that primary and middle school education may not change the likelihood of a woman working. 8 There is a decline in the coefficients of high school and college in 2004. 9

These coefficients are particularly large, especially compared to any other coefficient reported. For instance, they show that women with tertiary education are about 50% more likely to work than illiterate women. The education profile is increasing and convex. There are five age groups: 21 to 30, 31 to 40, 41 to 50, and 51 to 60, and 61 to 65 (the control). As the coefficients show, not surprisingly, women aged 61 to 65, i.e. the control, are less likely to participate in the labor force than other age groups. Moreover, consistent with the evidence from other countries and the economic theory, they show that the age profile is generally concave. In other words, women aged 31 to 50 are more likely to participate than any other group. Being in an Urban area decreases the probability of participation, since women in the rural area usually work on their family farm; an opportunity not available to most urban women. Interestingly, having more teenage females in the households (females aged between 15 and 18) increases the likelihood of participation for women slightly, while more teenage males does not change this likelihood. This is particularly interesting. As households, especially in urban areas, invest more on the education of their children in this age group, we expect that the households try to acquire more resources, both money and time, for this investment. On one hand, households need more income to be able to pay for the tuition of quality high school education as well as future college tuition, and on the other hand, they may like to spend more time with their kids improving the quality of their education through home schooling. The former will entice mothers to participate in the labor force while the latter persuades them to stay at home. But there are two more reasons to entice women to work when they have more teenage daughters than sons: First, households are saving more for their daughters future dowry when they are at this age, and second, teenage females in the household may contribute to home production and make more free time available for their mothers. Presence of one more adult female in the household increases the likelihood of women working by about three percent (more than twice the coefficient of teenage females). This is not large nor 10

surprising, as more adult females in the household would increase the number of people who can potentially contribute to home production and hence provides more free time to more female members to contribute to labor market. In addition, since these adult females are at the age of marriage, saving for their dowries is a strong motivation to work for all members of the household. Interestingly, more adult male members have negative correlation with FLFP. The coefficient is almost as large as the coefficient for adult females but in opposite direction. Adult males in the household usually contribute to household income by working in the labor market. Since households with more adult males may have more sources of income, they are less in need of income brought by adult females and hence fewer women would work in such households. Consistent with theory, the asset index has a negative coefficient. Using equal weights for each type of asset does not affect the results qualitatively. On the other hand, although marginal effects of owned home value are negative, they are quite small and insignificant. 9 This might be because of selection. Rich households with large home value are a selected group with different unobservable factors. Women in such households are likely to work more because of those unobservable factors that are also correlated with wealth. So the estimated coefficient of home value would become smaller. Overall, the results show that FLFP characteristics in Iran are quite similar to those of other countries (See for example Pagan, 2002). Moreover, although none of the marginal effects represents a causal relationship, they follow the predictions of a standard simple economic theory. 4.1 Estimating the Structural Model for LFP In addition to estimating the reduced form, one may be interested in measuring the impact of direct economic factors that affect FLFP, i.e. Equation (2) or its linear form, Equation (3), which are derived directly from economic theory. One of the main predictors of participation according to this model 9 The coefficients become significant when I do not employ province fixed effects. 11

is wages. However, estimating the effect of wages on participation is difficult as we do not observe wages for those who do not participate. Moreover, the error terms in Equations (3), (4), and (5), v j i and u j i (j = f, m) may be correlated with each other making OLS estimates biased. To overcome these issues, as explained in Section 2, I employ education as an instrumental variable to predict wages for everyone (not just those who worked) and then, use those predicted wages instead of the actual wages to estimate Equation (3). As discussed in Secton 2, this is different from 2SLS. 10 In order to implement this pseudo 2SLS method, I need to correct for selection on wage variable in the first stage, as wages are not observed for everyone. So, in the first stage, I estimate a two-step Heckman selection model of whether wages are observed for an individual. Estimating the structural model has the following steps: 1. In the first step of this Heckman selection model, inverse Mills ratios can be computed using a probit model in which household characteristics are the selection identifying variables. These are similar regressions as ones depicted in Table 1, except that they are probit. 2. In the second step, I estimate the log of hourly wages using education as explanatory variables and correcting for selection by inverse Mills ratios. This becomes like the first stage of a 2SLS. 3. Using this regression, I then predict the log of wages for all women regardless of participation status, and substitute these predictions instead of wages in the second stage, i.e. Equation (3). In order to satisfy the exclusion restriction in the second stage regression, We assume that education does not predict labor force participation directly, but only through wages. This is a common assumption for a static model in which education is an endowment. Using predicted wages in the second stage would give us incorrect standard errors. Therefore, 10 If it was a 2SLS method, wages would be predicted only for those observations used in the first stage, i.e. those who reported wages. In the second stage also, only those predictions would be used. In other words, we would have run the labor force participation regression only for those who worked. This is not possible, as the dependent variable in the second stage would be constant and equal to one. 12

bootstrapping is used, for the whole procedure, to obtain correct standard errors in the second stage. First, a bootstrapped sample is drawn from the data. Then, the three steps, described above, were implemented on this sample. This procedure is repeated a thousand times, each time with a new bootstrapped sample. The results were used to compute the correct standard errors for each regression. Resampling is implemented based on the clusters. Since we can only calculate hourly wages from the SECH 92-95, I am using the pooled data for all four years to estimate this structural model. I run this pseudo 2SLS for women and men in rural and urban areas. The first step regression of Heckman selection model is similar to the ones reported in Table 1, except that it is probit rather than OLS. Therefore I do not report them here. 11 Table 3 shows the second step of Heckman selection model which is like the first stage of a 2SLS. Table 4 reports the second stage in which I estimate the partial elasticities of wages on labor force participation. In all these regressions, standard errors are corrected for correlation inside clusters. Table 3 reports the regression of log of real wages on education, crisis dummy variable, time trend, and age dummies, using the inverse Mills ratio to correct for selection. 12 The crisis dummy is equal to one for the years 1994 and 95, when a minor economic crisis happened, and zero for the other two years, i.e. 1992 and 93. This is like a first stage of a two-stage least square. Education dummies are instruments for wages and hence excluded from the second stage, i.e. LFP regression in Table 4. Education coefficients are increasing as the level of education increases. Age profile is concave with its maximum usually at ages above 50. The χ 2 statistics for testing the joint significance of instruments are significant in all regressions except for rural women. When significant, the statistics are larger than 10, the threshold below which instruments are generally considered weak. Using these estimates, I predict wages for all individuals between the ages of 21 and 65, whether 11 These regressions are available upon request. 12 This regression is estimated jointly with the Probit Regression using Stata heckman command to get correct standard errors. 13

they worked or not. These predicted wages are used in the structural model, i.e. Equation (3), instead of log of wages. Other covariates are crisis dummy, time trend, and variables used in the reduced form regressions (Table 1) except education dummies. A linear probability model with cluster random effect is used to control for correlation inside clusters. Table 4 depicts these regressions. The main coefficient of interest, coefficient of log of real wages, is different across women and men of urban or rural areas. Interestingly, women s and men s wage rates in rural areas are not a predictor of their participation status. However, for women in the urban areas, the hourly wage has a large significant coefficient. For every one percent rise in wages, the likelihood to participate increases by 0.3 percentage point. Average LFP rates and elasticities computed at the average LFP are reported at the bottom of the table. They are particularly quite large for urban women as one percent increase in wages increases LFP by 1.7%. Urban men with higher wages, on the other hand, are likely to work less. One percent increase in wages decreases urban men s participation by 0.3%. Interestingly, These results are quite similar to the evidence found for the United States. Note that men s wages in rural areas may not be predicted accurately, since household members working as unpaid family labor contribute to household income, but this income is reported as the income of only one individual in the household, i.e. household head. Thus, the hourly wages for heads in such households are overestimated. Age profile is concave for men in both areas but is almost flat for women. Similar to the reduced form regressions, more adult females and fewer adult males in the household increase participation. On the other hand, more teenage females and males do not change it. Assets are negatively correlated with urban women s participation but positively correlated with men s. Home value has generally negative correlation with participation for almost everyone. These results, for the most part, are consistent with the expectations and comparable to evidence presented in Tables 1 and 2. Overall, we observe that not only does the reduced form follows the economic theory, but also the 14

structural labor force participation model gives similar results that have been found for the developed countries, including the United States (Killingsworth and Heckman, 1986; Heckman, 1980; Dooley, 1982; Stelcner and Brestaw, 51; Franz, 1981; Franz and Kawasaki, 1981; Renaud and Siegers, 1984). After all, leaving the low labor force participation aside, there is little difference between Iran and other countries, even the developed ones. In both urban and rural areas of Iran, there is a significant difference between the married and unmarried women s participation rates. The effect of marriage may be the result of discriminatory and/or non-discriminatory factors. But whatever caused it, one may be inclined to redo the structural estimation for these two groups separately. Here, I divide the pooled panel data into two groups: currently married and never married people. I then re-estimate the three steps, i.e. the Heckman selection, first stage, and second stage, for each sub-group. The coefficient of log of wages in the second stage regressions for never-married and currently married women are reported in Table 5. 13 Interestingly, the results for currently married and never-married women are quite similar to the results for the whole sample. The coefficient of log of wages is insignificant in rural areas regardless of marital status, but they are positive and large for urban women whether the individual has never married or is currently married. However, the elasticities are smaller for never married urban women. It seems that never-married women act more like men than married women do. The high wage elasticity for married women is contradictory to the marriage lock theory which argues that marriage limits the responsiveness of women to economic forces. This is a new and interesting result in the literature of FLFP in Iran. 13 Other coefficient in this regression as well as the Heckman selection model and the first stage regressions are available upon request. 15

4.2 Estimating the Structural Model for Hours After estimating the elasticity of LFP with respect to wages, the natural next step is to estimate the elasticity of hours worked with respect to wages. The econometric framework is similar to what I explained in Section 4.1. Here, I regress log of hours worked (instead of LFP) on predicted log of wages. Predicted log of wages are obtained from the same regressions explained in Section 4.1 and depicted in Tables 3 for the whole sample. Since log of non-zero hours worked is used as the dependent variable in the second stage, a two-step Heckman Selection Model should be used to correct for selection. The procedure is exactly the same as what was discussed for LFP, above. Table 6 reports this structural model. The coefficient of log of wage represents the elasticity since the dependent variable is in terms of log as well. As shown, this elasticity is insignificant for all groups except urban men. This is consistent with the empirical evidence for the United States, i.e. although female participation is elastic with respect to wages (extensive margin), their hours worked is not (intensive margin). Elasticity of hours worked with respect to wages for urban men is negative and significant showing that if wages increase by one percent hours worked will decrease by 0.26%. This result is not surprising as men with lower wages particularly work over-time and usually have multiple jobs to be able to maintain a decent living standard. As wage rates increase there is less need to work more and hours worked decrease. Following the same concern explained in Section 4.1, one can divide the sample into never-married and currently married people, and redo the same analysis for hours. This analysis is reported in Table 7 for the never-married and currently married women. The results in these tables are generally similar to the evidence for the whole sample presented in Table 6. In other words, hours worked for women, both in urban and rural areas and across marital status, is inelastic with respect to wages. Overall, one can observe that the results are quite consistent with the economic intuition, and the 16

evidence from other countries such as the United States. 5 Conclusion This paper studied LFP for Iranian women (and men) by applying a canonical static model to data and estimating the reduced form as well as the structural model. From the reduced form estimations, we observed that FLFP in Iran follows predictions of the economic theory. The structural estimation showed that urban women have an upward-sloping supply curve, and their response is quite strong as the elasticity is about 1.7. The large positive wage elasticity of FLFP means that women are eager to participate more, if there are small increases in wages. This could imply that the low FLFP rate is not a supply issue, rather it could be due to low demand for female labor. The high unemployment rate for women (more than 30%) may also support this argument. Another piece of evidence consistent with it is that, as Esfahani and Shajari (2012) showed, educated women in Iran are more likely to be employers and self-employed than educated men. In this context, research and policy should be oriented towards the demand side. More research is necessary to understand the characteristics of demand for female labor and find potentials for policy interventions. However, there are several policies for increasing FLFP that can be considered: one is to create incentives, such as tax breaks, for employers who recruit women especially for high skilled jobs (which have higher wages). Although, it could be effective, this policy may not be optimal as it creates externalities and dead-weight loss. Another option is to provide economic incentives, such as subsidized loans, for female self-employment and entrepreneurship. Depending on the incentive, this policy may be costly and not quite effective (as entrepreneurship depends on many personal and environmental factors and economic incentives may have a small impact on it). 17

But one solution with almost no dead-weight loss is to speed up the process of economic development. As the economy becomes more service-oriented and industries such as finance, insurance, higher education, and health care flourish, demand for high skilled workers increases. As a result, highly educated women, whose number is growing rapidly in Iran, 14 find more positions and higher wages available for them, which, in turn, increases FLFP rate substantially. Considering these, the potentials for future research on demand for female labor are strong especially that many developing countries have similar experiences. References Abbasi-Shavazi, M. J., P. McDonald, and M. Hosseini-Chavoshi (2009). The Fertility Transition in Iran: Revolution and Reproduction. Springer. Barro, R. J. and J.-W. Lee (2001). International data on educational attainment: updates and implications. Oxford Economic Papers 53 (3), 541 563. Dooley, M. D. (1982). Labor supply and fertility of married women: an analysis with grouped and individual data from the 1970 u.s. census. Journal of Human Resources 17, 499 532. Esfahani, H. S. and P. Shajari (2012). Gender, education, family structure, and the allocation of labor in Iran. Middle East Development Journal (MEDJ) 4 (02), 1250008 1 40. Franz, W. (1981). Schatzung regionaler arbeitsangebotsfunktionen mit hilfe der tobit-methode und des probit-verfahrens unter berucksichtigung des sog. sample selection bias. Discussion Paper No. 171-81, Institut fur Volkswirtschaftslehre und Statistik, University of Mannheim, Mannheim, Federal Republic of Germany. 14 Since 2004, more than two-third of college students are women. 18

Franz, W. and S. Kawasaki (1981). Labor supply of married women in the federal republic of germany: theory and empirical results from a new estimation procedure. Empirical Economics 6, 129 143. Heckman, J. (1980). Female Labor Supply, Chapter Sample Selection Bias as a Specification Error, pp. 206 248. Princeton, NJ: Princeton University Press. Killingsworth, M. R. and J. J. Heckman (1986). Female labor supply: A survey. In O. C. Ashenfelter and R. Layard (Eds.), Handbook of Labor Economics (First ed.), Volume 1 of Handbooks in Economics, pp. 103 204. Amsterdam: North-Holland. Pagan, C., J. A. (2002). Gender differences in labor market decisions in rural guatemala. Review of Development Economics 6 (3), 428 441. Renaud, P. S. A. and J. J. Siegers (1984). Income and substitution effects in family labour supply. De Economist 132, 350 366. Sahn, D. E. and D. C. Stifel (2000). Poverty comparisons over time and across countries in Africa. World Development 28 (12), 2123 2155. Stelcner, M. and J. Brestaw (51). Income taxes and the labor supply of married women in Quebec. Southern Economic Journal 1985, 1053 1072. 19

Table 1: Linear Probability Model of Labor Force Participation for Women Aged 21 through 65 with Province Fixed Effects, 1990-1996 1990 1991 1992 1993 1994 1995 1996 Primary 0.018 0.023 0.010 0.015 0.016* 0.008-0.002 (0.016) (0.016) (0.013) (0.019) (0.008) (0.014) (0.011) Mid School 0.009 0.029* -0.019 0.019 0.025 0.002-0.007 (0.013) (0.012) (0.017) (0.019) (0.014) (0.014) (0.014) High School 0.244*** 0.255*** 0.216*** 0.246*** 0.227*** 0.183*** 0.166*** (0.016) (0.016) (0.022) (0.024) (0.018) (0.017) (0.020) College & higher 0.531*** 0.551*** 0.553*** 0.570*** 0.538*** 0.514*** 0.503*** (0.035) (0.041) (0.028) (0.035) (0.027) (0.025) (0.031) 20 I[20<Age 30] 0.048* 0.028 0.048* 0.076* 0.053* 0.053* 0.053* (0.022) (0.020) (0.022) (0.029) (0.025) (0.022) (0.022) I[30<Age 40] 0.083** 0.065*** 0.087*** 0.130*** 0.092*** 0.100*** 0.097*** (0.023) (0.015) (0.017) (0.026) (0.019) (0.019) (0.018) I[40<Age 50] 0.073** 0.065*** 0.090*** 0.116*** 0.104*** 0.087*** 0.095*** (0.022) (0.014) (0.013) (0.017) (0.014) (0.014) (0.014) I[50<Age 60] 0.042** 0.023* 0.019* 0.063*** 0.056*** 0.036*** 0.054*** (0.013) (0.010) (0.009) (0.012) (0.010) (0.009) (0.012) Urban -0.092** -0.099** -0.097*** -0.148*** -0.108** -0.122*** -0.151*** (0.026) (0.026) (0.023) (0.033) (0.031) (0.029) (0.037) No. of females b/w 15 to 18 0.018*** 0.025*** 0.011 0.020* 0.012* 0.019** 0.009 (0.005) (0.006) (0.006) (0.008) (0.006) (0.006) (0.006) No. of males b/w 15 to 18 0.002 0.005-0.002-0.000-0.003 0.003-0.003 (0.007) (0.005) (0.006) (0.008) (0.006) (0.005) (0.005) No. of females above age 18 0.028*** 0.023*** 0.028*** 0.020** 0.026*** 0.024*** 0.023*** (0.005) (0.005) (0.005) (0.006) (0.005) (0.004) (0.004) No. of males above age 18-0.037*** -0.042*** -0.038*** -0.036*** -0.034*** -0.025*** -0.026*** Continued on next page

Table 1 continued from previous page 1990 1991 1992 1993 1994 1995 1996 (0.008) (0.006) (0.005) (0.007) (0.004) (0.005) (0.006) Asset Index -0.019** -0.016* -0.014* -0.026*** -0.022*** -0.021*** -0.018* (0.007) (0.007) (0.006) (0.007) (0.004) (0.004) (0.007) Owned Home Value 10 6-0.030 0.011-0.095-0.033-0.113* -0.158* -0.096* (0.061) (0.016) (0.110) (0.090) (0.051) (0.058) (0.042) Constant 0.122*** 0.143*** 0.150*** 0.170*** 0.135*** 0.174*** 0.185*** (0.015) (0.016) (0.019) (0.018) (0.027) (0.020) (0.019) Province FE Yes Yes Yes Yes Yes Yes Yes Observations 19705 20126 19714 13566 21628 39639 23780 21 Average FLFP 0.141 0.133 0.157 0.181 0.168 0.199 0.187 (0.003) (0.003) (0.003) (0.004) (0.003) (0.002) (0.003) Note: Dependent variable is a dummy equal to one if the individual was looking for a job or was working for at least two days in the week prior to the survey, and zero otherwise. Primary, Mid School, High School, and College & higher are education dummy variables. Illiterates are the control group. I[a<Age b] is a dummy variable equal to one if Age is between a and b (b included) and zero otherwise. Women aged 61 through 65 are the control group. Asset index is computed following Sahn and Stifel (2000). Owned home value is the answer to the questions If you wanted to rent this house, how much would be the rent? Robust heteroskedastic standard errors are in parentheses. p<0.10, * p<0.05, ** p<0.01, *** p<0.001

Table 2: Linear Probability Model of Labor Force Participation for Women Aged 21 through 65 with Province Fixed Effects, 1997-2004 1997 1998 1999 2000 2001 2002 2003 2004 Primary 0.006-0.007 0.016 0.009-0.001 0.013 0.025* -0.016 (0.013) (0.011) (0.009) (0.010) (0.012) (0.009) (0.011) (0.012) Mid School 0.001-0.027 0.015 0.015-0.021 0.023 0.007-0.016 (0.015) (0.015) (0.012) (0.009) (0.014) (0.014) (0.013) (0.013) High School 0.140*** 0.119*** 0.142*** 0.141*** 0.112*** 0.151*** 0.144*** 0.089*** (0.017) (0.024) (0.017) (0.018) (0.020) (0.021) (0.019) (0.020) College & higher 0.501*** 0.456*** 0.505*** 0.472*** 0.430*** 0.477*** 0.475*** 0.394*** (0.031) (0.035) (0.022) (0.023) (0.026) (0.020) (0.025) (0.028) 22 I[20<Age 30] 0.066** 0.047* 0.039 0.043* 0.045* 0.030 0.041* 0.048** (0.021) (0.022) (0.021) (0.016) (0.018) (0.020) (0.015) (0.017) I[30<Age 40] 0.113*** 0.100*** 0.085*** 0.090*** 0.079*** 0.064*** 0.089*** 0.090*** (0.019) (0.019) (0.017) (0.015) (0.015) (0.017) (0.014) (0.017) I[40<Age 50] 0.122*** 0.110*** 0.088*** 0.105*** 0.093*** 0.070*** 0.087*** 0.102*** (0.019) (0.015) (0.017) (0.014) (0.015) (0.016) (0.014) (0.017) I[50<Age 60] 0.076*** 0.053*** 0.058*** 0.056*** 0.041** 0.036** 0.056*** 0.062*** (0.016) (0.012) (0.013) (0.013) (0.012) (0.011) (0.013) (0.015) Urban -0.139*** -0.106*** -0.106*** -0.105*** -0.110*** -0.095*** -0.076** -0.111*** (0.033) (0.026) (0.022) (0.021) (0.023) (0.022) (0.022) (0.023) No. of females b/w 15 to 18 0.012** 0.006 0.013* 0.012* 0.019*** 0.017*** 0.015*** 0.014** (0.004) (0.007) (0.006) (0.005) (0.005) (0.004) (0.004) (0.005) No. of males b/w 15 to 18-0.004-0.010-0.001-0.003 0.004 0.006 0.012* 0.008 (0.003) (0.007) (0.004) (0.004) (0.004) (0.004) (0.005) (0.005) No. of females above age 18 0.030*** 0.025*** 0.021*** 0.025*** 0.021*** 0.022*** 0.034*** 0.035*** (0.005) (0.006) (0.004) (0.004) (0.005) (0.004) (0.007) (0.005) No. of males above age 18-0.024*** -0.023*** -0.025*** -0.027*** -0.023*** -0.015*** -0.014** -0.024*** Continued on next page

Table 2 continued from previous page 1997 1998 1999 2000 2001 2002 2003 2004 (0.004) (0.005) (0.005) (0.004) (0.003) (0.003) (0.005) (0.003) Asset Index -0.030*** -0.024** -0.026*** -0.028*** -0.028*** -0.028*** -0.036*** -0.031*** (0.007) (0.007) (0.007) (0.006) (0.007) (0.005) (0.006) (0.006) Owned Home Value 10 6 0.071-0.051-0.054-0.016-0.082-0.011-0.066-2.756 (0.035) (0.048) (0.077) (0.073) (0.045) (0.033) (0.052) (4.862) Constant 0.156*** 0.154*** 0.160*** 0.151*** 0.173*** 0.133*** 0.110*** 0.171*** (0.013) (0.020) (0.018) (0.015) (0.019) (0.017) (0.020) (0.020) Province FE Yes Yes Yes Yes Yes Yes Yes Yes Observations 23960 19468 30611 29926 30318 36661 26268 28087 23 Average FLFP 0.191 0.175 0.187 0.181 0.177 0.179 0.194 0.202 (0.003) (0.003) (0.003) (0.004) (0.003) (0.002) (0.003) (0.003) Note: Dependent variable is a dummy equal to one if the individual was looking for a job or was working for at least two days in the week prior to the survey, and zero otherwise. Primary, Mid School, High School, and College & higher are education dummy variables. Illiterates are the control group. I[a<Age b] is a dummy variable equal to one if Age is between a and b (b included) and zero otherwise. Women aged 61 through 65 are the control group. Asset index is computed following Sahn and Stifel (2000). Owned home Value Owned home value is the answer to the questions If you wanted to rent this house, how much would be the rent? Robust heteroskedastic standard errors are in parentheses. p<0.10, * p<0.05, ** p<0.01, *** p<0.001

Table 3: Estimation of Log of Real Wages for People Aged 21 through 65 Controlling for Heckman Selection on Wages, Pooled data 1992-95 Rural Urban Women Men Women Men Primary 0.279 0.147 0.259 0.180** (0.179) (0.092) (0.223) (0.063) Mid School 0.242 0.280* 0.514** 0.230** (0.247) (0.110) (0.174) (0.071) High School 0.527-0.134 1.096*** 0.239** (0.274) (0.162) (0.256) (0.073) College & higher 0.907*** 1.931*** 0.644*** (0.169) (0.420) (0.099) Crisis Years 0.087 0.004 0.222 0.105* (0.213) (0.094) (0.140) (0.052) Trend -0.167-0.011-0.245*** -0.160*** (0.101) (0.050) (0.056) (0.027) I[20<Age 30] -0.454-0.317** -0.270-0.832*** (0.268) (0.110) (0.194) (0.074) I[30<Age 40] -0.244 0.025 0.222-0.529*** (0.272) (0.084) (0.212) (0.097) I[40<Age 50] -0.028-0.014 0.398* -0.317*** (0.311) (0.068) (0.200) (0.084) Constant 10.777*** 10.490*** 9.011*** 11.647*** (0.691) (0.119) (0.690) (0.123) χ 2 test for education 4.47 39.77*** 25.32*** 45.83*** Note: Dependent variable is the individual s log of real wage. Crisis Years is a time dummy equal to one for 1994 and 95, when a minor macroeconomic crisis happened, and zero otherwise. For a description of other covariates, please see Table 1. SECH 1992-95 is the data used. Education dummies are the instruments for log of real wage in the second stage in Table 4. As there were few women in rural areas with college & higher level of education, they are combined with rural women with high school education. Bootstrapped standard errors, resampled at cluster level and computed using 1000 replications are in parentheses. Cluster random effects were employed. p<0.10, * p<0.05, ** p<0.01, *** p<0.001 This χ 2 statistic tests whether education dummies which are the instruments for log of wage in the regression in Table 4, are jointly significant. 24

Table 4: Linear Probability Model of Labor Force Participation on Predicted Log of Real Wages from Regression in Table 3 for People Aged 21 through 65, Pooled data 1992-95 Rural Urban Women Men Women Men ln(wage) -0.018-0.080 0.259*** -0.303*** (0.151) (0.066) (0.061) (0.062) Crisis Years 0.082-0.010-0.055 0.026 (0.046) (0.012) (0.043) (0.018) Trend -0.018 0.001 0.059** -0.056*** (0.037) (0.006) (0.023) (0.013) I[20<Age 30] 0.107 0.016 0.024-0.034 (0.081) (0.022) (0.041) (0.049) I[30<Age 40] 0.106 0.065*** -0.031 0.109** (0.057) (0.018) (0.050) (0.038) I[40<Age 50] 0.083 0.047** -0.062 0.136*** (0.051) (0.014) (0.048) (0.030) No. of females b/w 15 to 18 0.016 0.018** -0.007 0.006 (0.012) (0.006) (0.009) (0.008) No. of males b/w 15 to 18 0.013-0.005-0.017* -0.005 (0.012) (0.006) (0.008) (0.008) No. of females above age 18 0.024* 0.009 0.030** 0.003 (0.010) (0.006) (0.009) (0.006) No. of males above age 18-0.030*** -0.026*** -0.018** -0.043*** (0.008) (0.006) (0.006) (0.006) Asset Index 0.013 0.023** -0.030** 0.018** (0.016) (0.007) (0.010) (0.007) Owned home value 10 6-0.028-0.142-0.048-0.052* (0.164) (0.088) (0.028) (0.026) Constant 0.380 1.799** -2.293** 4.363*** (1.716) (0.690) (0.726) (0.724) Observations 6345 6401 10009 10630 Average LFP 0.253 0.951 0.149 0.891 (0.005) (0.003) (0.004) (0.003) Elasticities -0.071-0.084 1.738-0.340 Continued on next page 25

Table 4 continued from previous page Rural Urban Women Men Women Men Note: Dependent variable is a dummy equal to one if the individual is participating in the labor force and zero otherwise. For a description of other covariates, please see Tables 1 and 3. SECH 1992-95 is the data used. ln(wage) is the predicted log of real wages from the first stage regression reported in Table 3. Education dummies are the instruments for log of real wage in the second stage in Table 4. Bootstrapped standard errors, resampled at cluster level and computed using 1000 replications are in parentheses. Cluster random effects were employed. p<0.10, * p<0.05, ** p<0.01, *** p<0.001 Table 5: Linear Probability Model of Labor Force Participation on Predicted Log of Real Wages for Married and Never-married Aged 21 through 65, Pooled data 1992-95 Rural Urban Married Never-Mar. Married never-mar. ln(wage) -0.055 0.015 0.235* 0.286** (0.211) (0.170) (0.093) (0.110) [1em]Observations 5197 649 7873 1202 Average LFP 0.234 0.354 0.120 0.353 (0.006) (0.019) (0.004) (0.014) Elasticities -0.235 0.042 1.958 0.810 Note: Dependent variable is a dummy equal to one if the individual is participating in the labor force and zero otherwise. For a description of other covariates, please see Tables 1 and 3. SECH 1992-95 is the data used. ln(wage) is the predicted log of real wages from the first stage regression in which education dummies are the instruments for log of real wage here. Bootstrapped standard errors, resampled at cluster level and computed using 1000 replications are in parentheses. Cluster random effects were employed. p<0.10, * p<0.05, ** p<0.01, *** p<0.001 26

Table 6: Linear Estimation of Log of Hours Worked on Predicted Log of Real Wages for People Aged 21 through 65, Pooled data 1992-95 Rural Urban Women Men Women Men ln(wage) 0.049-0.105 0.656-0.255*** (0.633) (0.073) (0.672) (0.068) Observations 5197 649 7873 1202 Note: Dependent variable is the positive (non-zero) hours worked by an individual in the week preceeding the survey. Heckman two-step selection model is used to correct for selection on Hours. The selection identifying variables for hours worked are education dummies and the first step Heckman selection model is available upon request. For a description of other covariates, please see Tables 1 and 3. SECH 1992-95 is the data used. ln(wage) is the predicted log of real wages from the first stage regression is available upon request. Education dummies are the instruments for log of real wage in this regression. Bootstrapped standard errors, resampled at clusterlevel and computed using 1000 replications are in parentheses. Cluster random effects were employed. p<0.10, * p<0.05, ** p<0.01, *** p<0.001 Table 7: Linear Estimation of Log of Hours Worked on Predicted Log of Real Wages for Married and Never-married Women Aged 21 through 65, Pooled data 1992-95 Rural Urban Married Never-Mar. Married never-mar. ln(wage) -0.061 0.105 0.355-0.108 (0.565) (0.451) (0.321) (0.139) Observations 5197 649 7873 1202 Note: Dependent variable is the positive (non-zero) hours worked by an individual in the week preceding the survey. Heckman two-step selection model is used to correct for selection on Hours. The selection identifying variables for hours worked are education dummies and the first step Heckman selection model is available upon request. For a description of other covariates, please see Tables 1 and 3. SECH 1992-95 is the data used. ln(wage) is the predicted log of real wages from the first stage regression is available upon request. Education dummies are the instruments for log of real wage in this regression. Bootstrapped standard errors, resampled at cluster level and computed using 1000 replications are in parentheses. Cluster random effects were employed. p<0.10, * p<0.05, ** p<0.01, *** p<0.001 27