Structural change and female labour force participation
A re-examination and extension of the feminization U-curve
Kars Willemsen
Rijksuniversiteit Groningen
Abstract: A considerable amount of research suggests that female labour force participation is
1. Introduction
The research concerning Female Labour Force Participation (FLFP) received a great impulse after the publication of ‘Woman’s role in economic development’ by Ester Boserup in 1970. The content in the book does not solely focus on the fairness perspective of equal opportunities for women, but also elaborates on the opportunities for the entire society with increases in female participation. These benefits range from increases in the supply of labour and the number of taxpayers supporting a welfare state, to a decrease of social and financial risk for women. The societal benefits mentioned in the book have become a popular topic of research for many sociologists and economists.
Following the revolutionary ideas from Boserup, the research into the relation between participation of women in the labour force and economic development took off. The academic literature to date indicates a positive effect of gender equality on economic development (Knowles, 2002; Klasen and Lamanna, 2009; Elborgh-Woytek et al., 2013). The literature examining the inverse relation however, lacks proper empirical foundation and is often a topic of conflicting insights. The inverse relation implies that differences in economic development might explain differences in gender inequality. With respect to this standpoint, two main views are regularly debated. On one side the Neo-classical approach, which assumes a strictly positive relation between economic development and participation of women in the labour force. On the other side there is the academic literature which argues that specific levels of economic development are associated with specific levels of female labour force participation (FLFP) resulting in a U-shaped female labour supply curve. This therefore does not necessarily imply that with economic development more women are participating in the economy. The latter viewpoint has received most attention in the last decade and has proven to be the most fitting to describe the relation between economic development and FLFP. This paper therefore sets out to re-examine and extend the analysis of the U-curve hypothesis by using different data sources, sample groups, time-periods and geographical scales.
The female labour supply U-curve is based on the theory that at early stages of economic development, proxied by GDP per capita, FLFP is rather high after which it drops when the economy further develops. After a certain threshold point of economic development, the FLFP goes up again resulting in the upward part of the U-curve. In early economic development stages, countries are mostly agriculture-based and women can combine reproduction activities with working on the land. Moreover, the incomes earned in agriculture are not high enough to live of one income, which means everyone has to work. In this phase of economic development FLFP is relatively high (Çağatay and Özler, 1995). In figure 1 the (red) U-curve is displayed for 16 developed economies for the period 1890-2005. The left side is characterized by relatively low GDP values and high FLFP rates. Highlighted is the process for Portugal and Norway, the countries showing the extremes for GDP values (Norway high, Portugal low).
agriculture allowed women to perform less heavy tasks during their period of pregnancy, but in manufacturing, the workforce had to perform the heavy tasks within the factory (Olivetti, 2013). In general, urbanization and industrialisation are positively associated with development, but are negatively related with FLFP (Çağatay and Özler, 1995). This relates to the middle section of the U in figure 1. Clearly from 1890 till 1950, the FLFP dropped for both Portugal and Norway whereas the GDP per capita has increased, similar to what is expected following the U-curve hypothesis.
Sources: Data on FLFP from the ILO and IHS database, data for GDP per capita from the Maddison Project database
Figure 1: Female labour force participation for 16 developed economies, 1890-2005
As economic development proceeds, opportunities for women increase and FLFP goes up. Olivetti (2013) argues that a shift in the structure of the economy is the main contributor to the increase in FLFP. The manufacturing sector decreases whereas the service sector expands which provides more equal opportunities for men and women. The disadvantages faced by women in the manufacturing sector due to differences in physical capabilities are not apparent in the services industry. Therefore, employers do not necessarily prefer men anymore due to convergence of male and female productivity levels (Ngai and Petrongolo, 2017). This can be seen in figure 1 in the right part of the U curve where GDP per capita and FLFP both rise. For both Portugal and Norway, the FLFP increased with an increase in GDP per capita which is in line with what is expected.
In short, this paper examines the U-curve hypothesis by focussing on the role of structural transformation as a driver of FLFP. The analysis is based on the empirical foundation created by various scholars who examined the role of structural transformation as a driving force for changes in FLFP (Çağatay and Özler, 1995; Goldin, 1995; Luci, 2009; Olivetti, 2013; Lechman and Kaur, 2015). Since the literature so far has not provided a generalizable
Portugal (1890) Portugal (2005) Portugal (1950) Norway (1950) Norway (1890) Norway (2005) 0 .2 .4 .6 .8 F L F P 7 8 9 10 11
Log GDP per capita
FLFP Fitted values
historical period
conclusion, this paper sets out to discover whether the U-curve holds under various conditions. By changing the sample of countries, the period of analysis, the number of controls and the scale of analysis, a more complete picture can be drawn. Since, the U-curve hypothesis is about changes over time in a given country, panel datasets are used which allows for tracking of several countries over time.
The starting point of this paper is the research done by Olivetti (2013). Her research is based on FLFP for 16 developed economies between 1890 and 2005 and on FLFP for a larger world sample for the period 1950-2005. The first step in this paper is to run a similar approach to allow for comparison. By using a historical dataset for 16 developed economies for the period 1890 till 2005, similar as the period and countries examined by Olivetti, a comparison base can be established. To extent on the research by Olivetti, an education control is added to minimize the potential omitted variable bias. This increases the likelihood of finding the contribution of structural change by excluding the potential role of education in explaining FLFP.
In addition, a dataset using a more recent time-frame is used to assess if the U-curve also holds for a more present analysis. In a paper by Lechman and Kaur (2015), a similar approach is used for 162 countries over the period 1990-2012. The panel dataset for this paper includes more than 170 countries for the period 1990-2015. This dataset and analysis will be referred to as the ‘present’ dataset/analysis throughout the paper. To extent the work by Lechman and Kaur (2015) controls are added for education, maternal health, urbanization and fertility, based on several pieces of literature studying potential determinants of FLFP (Bloom et al., 2009; Dunkelberg and Spiess, 2007; Angeles, 2010; Madsen et al., 2018).
A further contribution to the literature is the third analysis based on regional data. A panel data analysis on a NUTS 2 level1 is performed with data for 260 regions from 21 European countries in the period 2000-2016. Regional analyses in this area of research have, so far, been absent, but Eurostat provides reliable regional data which allows for running and interpreting the U-curve regressions. According to several scholars invested in regional research in Europe (e.g. Esteban, 2000; Beugelsdijk et al., 2017) there is good reason to believe that there is very strong variation within countries that remains invisible when solely focussing on a country scale. Therefore, changing the level of analysis may reveal new insights in the relation between structural change and FLFP.
For the historical analysis I find some evidence for the U-curve hypothesis, but the U becomes more muted once time-fixed effects are included. Finally, when the education control is included the significance is reduced up to the point where statistical significance cannot be determined. The present analysis shows strongly significant results supporting the U-curve hypothesis even after including country and time-fixed effects. The present analysis allows for inclusion of many controls to minimize the presence of potential omitted variable bias. After including the controls, the results remain statistically significant, though more muted. The regional analysis supports the story established by the present dataset. The U-curve is apparent from the results, but since data for control variables is absent these results require further investigation.
1 The classification used by Eurostat to identify regions is referred to as the Nomenclature of territorial units for
Overall, I conclude that there is some support for the U-curve hypothesis, but that it depends on the model, the timeframe, the estimation methods, the sample group, the controls, the data sources and the geographical level of analysis. Clearly, the U-curve hypothesis does not find any strong generalizable support in this paper.
The paper is structured as follows. Firstly, the current state of the literature is explained in the literature review. Secondly, the historical data is examined and compared with the results from the paper by Olivetti (2013). Thirdly, the present dataset is used to analyse whether the U-curve holds in present times. Finally, the regional data is researched to examine within country differences and to find whether the presence of a U-curve is dependent on a geographical scale.
2. Literature review
One of the first researchers suggesting the U-shaped female labour supply curve was Sinha in 1965. Following this discovery many sociologists and economists have tried to re-examine this U-shaped relation due to its importance for future academic literature, but mostly due to its importance for policy makers who are interested in the trade-off between growth and gender equality. Since 1965 most scholars have argued that there is strong evidence to support the U-curve hypothesis, however several recent papers argue for caution in interpreting the often-ambiguous results. Moreover, the rationale behind the U-curve appears to be different for most academic literature.
Most work before 2000 focused on cross-section analyses to re-examine the female labour supply curve. Pampel and Tanaka (1986) use a cross-country model with data for two time points, 1965 and 1970, in which they find a curvilinear effect between development and female labour force participation (FLFP). They examine both the FLFP as a share of the total workforce and the FLFP as a percentage of the total female population. The focus is solely on the group of 45 to 59-year old women in the labour force since they want to minimize the confounding effect that differences between countries in fertility might have on the correlation between FLFP and GDP per capita. Intriguing in their approach is the variable ‘energy use per capita’ which should capture economic development. They consider this measure a more general measure of development that can summarize changes in the structure of the economy. In this area of research this is highly uncommon, but it demonstrates the possibility of using other development indicators.
be flawed. They use log GNP per capita and log GNP per capita squared to assess whether the relationship has a minimum, therefore signalling a U-curve relation. For this to hold true, the coefficient of log GNP per capita must be significantly negative and the coefficient of the square must be significantly positive. Even though the authors mention that the U-curve hypothesis cannot be rejected, according to the results from their dataset it should be rejected since their coefficients point to an inverted U-shape.
The results of the cross-country academic research up to this point might hint towards a U-shaped relation, but several contradicting results (such as Çağatay and Özler) and issues with the fit of the model in the U-curve hypothesis make that the relation is far from established. The imperfectness of the fit of the model relates to the fact that the feminization U-curve is about changes over time in a given country, which makes a cross-sectional analysis imperfect to examine the effect of structural transformation on FLFP.
After 2000 most academic papers seriously consider these data issues and measurement problems and extended the research with models based on panel data, therefore using both cross-country and time-series regressions. Luci (2009) specifically refers to Çağatay and Özler (1995) and Goldin (1995) to point out the flaws in the existing literature concerning FLFP. She points out that, potentially, endogeneity problems alter the results. Income measures may not only be exogenous, but could also be endogenous in the model, suggesting a two-way causation. Luci takes these endogeneity issues into account and uses GMM estimation in a panel dataset with data for 184 countries from 1965 to 2004 to see if there is a quadratic relation between log GDP per capita and FLFP. She confirms the ‘feminisation U’ hypothesis and relates this to economic structural transformation. She argues that in early stages of development the role of women is marginalized after which in later phases of transformation female participation increases again. In a paper by Tam (2011), International Labour Organisation (ILO) estimates are used to re-examine the U-curve hypothesis for 134 countries over the period 1950-1980. In addition to using time- and country-fixed effects, endogeneity issues are treated using a dynamic panel data model. In contrast to static panel data models, this dynamic model includes lagged levels of the dependent variable as independent variables to address the problem of endogeneity. Since the lagged levels cannot be considered endogenous, a static panel data model with fixed effects would prove to be inconsistent. To allow for the inclusion of lagged levels of the dependent variable Tam follows the Arellano and Bond (1991) method, where first differences are taken from the regression equation to eliminate fixed effects and allow for dynamic panel data. In all estimations the U-curve hypothesis remains valid and proves to be an intertemporal relationship.
in the history for currently advanced economies, but that these findings cannot be used as guidance for currently developing economies. Admittedly, the paper by Gaddis and Klasen (2013) structures its main critique around the declining portion of the U-curve, where development leads to decreases in FLFP due to the rise in manufacturing.
Ngai and Petrongolo (2017) prefer to consider the increasing portion of the U-curve as they argue that the services industry has become the most important sector in developed economies. They do not specifically focus on the relation between GDP per capita and FLFP, but they examine the rise of services and its driving force, structural transformation. They argue that women have a comparative advantage in producing services and that the rise of the services economy has led to higher relative wages and market hours for women.
Lechman and Kaur (2015) consider the critique of Gaddis and Klasen (2013) concerning the relevance for developing countries and use more recent data to re-examine the relation between FLFP and economic development. Their sample consists of 162 countries in the time-period 1990-2012. This provides various insights into the U-curve hypothesis in shorter, more current time periods. They find that the U-curve relation is still apparent for their sample, but that it does not hold when they split the sample into different income groups. This, in their opinion, is an argument against the validity of the argument that structural transformation affects FLFP which supports the arguments of Gaddis and Klasen. However, the structural transformation arguments outlined in existing literature argue that FLFP fluctuations are dependent on income and that therefore the U-curve does not need to hold for a specific income group, but rather for the group including all countries of various income levels. By only considering a group with low incomes, you might only be referring to the left (downward sloping) side of the U-curve whereas solely focussing on the group with high incomes might centre the attention to the right (upward sloping) side of the U-curve.
In addition to the note of Gaddis and Klasen (2013) about the differences in historically developing economies and currently developing economies, Olivetti (2013) adds that the timing of a country’s transition to a modern path of economic development affects the shape of women’s labour supply. Therefore, currently developing economies face different conditions compared to developed countries. To obtain more reliable results she investigated the relation between economic development and FLFP using two different datasets. Firstly, a historical dataset for 16 developed economies over the period 1890-2005 with 5-year intervals to compare their experience with the established experience of the United States. Secondly, a dataset for over 150 countries in the period 1950-2005 with 5-year intervals in which she also controls for education by adding the log of male to female years of schooling. Both these analyses confirm the existence of a U-shaped relation although slightly more muted for countries developing post 1950. Concludingly, Olivetti argues that it would be highly interesting to focus on within country differences in FLFP for future research. She supports this with the fact that there is still a substantial amount of regional variation in economic structure (Kim and Margo, 2004), FLFP and earnings (Olivetti and Petrongolo, 2014).
using historical, present and regional data to find whether the U-curve hypothesis is sample, period or scale dependent or that a generalizable conclusion is possible.
3. Methodology
This section addresses the main methodology for the historical, present and regional analyses, since the approach is nearly identical. To allow for comparison with the results from Olivetti (2013) I use a model based on the approach used in that paper. Since the data originates from the same sources, results for the basic regression must be very similar2. The following regression is used
𝐹𝐿𝐹𝑃it = 𝛼𝑖 + 𝛼𝑡 + 𝛿1𝐿𝑜𝑔𝐺𝐷𝑃𝑝𝑐it + 𝛿2𝐿𝑜𝑔 𝐺𝐷𝑃𝑝𝑐it 2 + 𝛾𝐿𝑜𝑔𝑚𝑓𝑠𝑐ℎ𝑜𝑜𝑙 + 𝜀it (1)
where ‘i’ stands for country and ‘t’ for the year. ‘FLFPit’ represents the Female Labour Force Participation rate, ‘LogGDPpcit’ is the logarithm of GDP per capita and ‘LogGDPpcit2’ is the
squared value of log GDP per capita. ‘Logmfschool’ represents the logarithm of male to female years of schooling. To consider fixed effects I use ‘αi’ to capture country fixed effects and ‘αt’ to capture time fixed effects. The panel data analysis allows me to control for unobserved or unmeasurable variables. For example, the country fixed effects allow me to control for factors such as cultural beliefs or religion that do normally not change rapidly over time. This approach mitigates the concern of omitted variable bias. The time fixed effects capture the impact of aggregate time-series trends. By not controlling for time-fixed effects, the influence of aggregate trends is visible in the results which hides the impact of economic development on FLFP. For example, a country might show increases in economic development in a certain period in which it also shows increases in FLFP, which makes it easy to state that economic development caused the increase in FLFP. However, it might be that a certain shock that is inherent to that period caused the FLFP to increase which has nothing to do with economic development. By including time-fixed effects these aggregate shocks are absorbed which allows for a clearer picture of the actual effect of economic development on FLFP.
To further ease the concern of omitted variables I add an education control to the regression. In the paper by Goldin (1995), the importance of the education gap between men and women is stressed. Goldin finds that increases in female education are important determinants of the rise in FLFP both over time and across countries. If this holds true for the data used in this paper, then it must be included to alleviate some of the concern about the omitted variable bias. Eventually, the goal is to find the relation between economic development and FLFP, excluding the impact of the gender gap in education. Olivetti therefore includes the logarithm of male to female years of schooling to control for education for the period 1950 till 2005. She found that at all levels of economic development, FLFP is lower when women have fewer years of schooling relative to men. Since the education control was important for the more recent period in Olivetti’s research, it is also included in the analyses for this paper. In addition to Olivetti, the education control is also added for the historical period.
2 Data used may slightly differ due to different releases of data. For example, this paper uses data from the
Excluding it from the historical analysis could potentially lead to an omitted variable bias, therefore, the log of male to female years of schooling is included3.
4. Historical analysis
4.1 DataThe datasets combined for the historical analysis allow for a sample of 16 developed economies over the period 1890-2005. The sample consists of Australia, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Netherlands, Norway, Portugal, Spain, Sweden, United Kingdom and the United States. The model used in this section is based on the approach of Olivetti (2013) who investigated the relation between FLFP and economic development for a longer (historical) period and compared this with data for a more recent period.
The data set for FLFP is copied from the Olivetti (2013) paper who uses data from the International Historical Statistics database (IHS) for the pre-1950 period and data from the International Labour Organization (ILO) for the post-1950 period. The data for FLFP in this paper is therefore identical to the data in the Olivetti paper. Olivetti (2013) has constructed this dataset using population numbers and economically active population numbers by gender for both the pre-1950 period (IHS) and the post-1950 period (ILO). The datasets in general overlap which simplifies merging them. The observations where values differ between the two datasets are dropped. In these cases, the inconsistency is either due to differences in the perception of the economically active population or due to differences in population counts. Olivetti follows the ILO definition of being economically active which states that an individual is economically active if he/she is working for pay or profit at any time during the specific reference period, whether he/she receives wages or not. This definition allows for inclusion of unpaid family farm workers, workers in family businesses and own-account traders.
The analysis of FLFP starts in 1890 where for 8 of the 16 countries data is available. Most of the other countries have data available from 1900 whereas only Ireland and Germany join the analysis in 1910 and 1925 respectively. The data allows for analyses using both 10- and 5-year intervals of which the 5-year intervals are used in this paper to maximize the number of observations and to follow the common practices in the feminization U literature (Gaddis and Klasen, 2013; Olivetti, 2013). In figure 4.1 (left panel) the FLFP histogram shows the distribution of the FLFP observations. The bunch of the observations lies between 0.3 and 0.6 with a striking boundary around 0.6 after which there are very few observations.
Data for the values of GDP per capita are retrieved from the Maddison Project Database (release 2018). The definition used in the Maddison database is ‘Real GDP per capita in 2011US$’. The database allows for yearly cross-country comparisons for a period up until 1 AD. However, since FLFP data is available from 1890 with 5-year intervals these data boundaries are also applied to GDP per capita. In figure 4.1 (right panel) the distribution of Log GDP per capita is displayed. The distribution is rather wide and cannot be considered a normal distribution. This is not necessarily surprising since data for several countries is examined for a period over 100 years. In this long period, income has increased dramatically for most
3 Unfortunately, for the regional analysis a comparable measure to control for education cannot be established.
countries which explains the wide variation between the smallest and largest observations. Large differences over time and between countries make a normal distribution a rather unlikely outcome.
Sources: ILO and IHS data for FLFP (Olivetti data) and the Maddison Project database for GDP per capita.
Figure 4.1: Historical density analysis (1890-2005).
Historical education data is retrieved from the Barro-Lee dataset. This dataset contains information about male and female education for a period starting in 1820 and ending in 2010. For this paper male and female years of schooling are extracted from 1890 till 2005 using 5-year intervals in accordance with FLFP data. Summary statistics for all variables are provided in the appendix.
4.2 Results
In table 4.1 the results are provided for the historical analysis for the 16 developed economies. Column 1 only includes country fixed effects whereas column 2 uses both country and time fixed effects and in column 3 the education control variable is included. In column 4 the regression is executed using male labour force participation (MLFP) as the dependent variable. This last column is added to exclude the possibility of both FLFP and MLFP being correlated with economic development. MLFP could be considered a placebo group, because almost all men aged 15 or older work which holds true at all levels of economic development in this sample of countries (Olivetti, 2013). If the U-curve would also exist for MLFP then this would imply that the U-shaped relation is not specific for female labour supply.
As expected the sign for the logarithm of GDP per capita is negative while the sign for the square is positive in column 1-3, representing the U-curve. The numbers for column 1 and column 2 are very comparable with the results from Olivetti who also finds that once country and time fixed-effects are included the U-curve is more muted and clearly loses part of its significance as can be seen in column 2.
the MLFP is examined in column 4. There is no evidence of a U-shaped labour supply curve for men according to the lack of significance of the coefficients. This is in line with the findings of Olivetti (2013) who also found insignificant results for the MLFP.
(1) (2) (3) (4)
VARIABLES FLFP FLFP FLFP MLFP
Log of GDP per capita -1.034*** -0.775* -0.520 -0.022
(0.266) (0.391) (0.445) (0.299)
Log of GDP per capita squared 0.0604*** 0.0437* 0.0286 0.0014
(0.0141) (0.0207) (0.0238) (0.0159)
Log male/female years of schooling 0.0884 (0.0571) 0.267*** (0.0508) Constant 4.739*** 3.728* 2.619 0.951 (1.247) (1.802) (2.038) (1.395) Observations Number of countries 239 16 239 16 239 16 229 16 R-squared 0.574 0.648 0.658 0.822
Country fixed effect Time fixed effects
Yes No Yes Yes Yes Yes Yes Yes
Sources: For FLFP the International historical Statistics and the International labour organization. For GDP data the Maddison Project Database. For education data the Barro-Lee database.
Clustered Robust Standard Errors: *** Significant at 1% level ** Significant at 5% level * Significant at 10% level
Table 4.1: Female labour force participation and economic development for 16 developed
economies, period 1890-2005
To conclude the historical analysis, there is some evidence for the U-curve hypothesis, however once fixed effects and the education control are included the relation between economic development and FLFP is not clear anymore. To investigate this relation further, more countries are included for a more recent timeframe in the next section.
5. Present analysis
5.1 DataSo, to extend the historical analysis until 2015 would have implied combining data for FLFP from the ILO, the IHS and the World Bank. These do not always use the same definitions and therefore potentially create more measurement issues. Since Olivetti (2013) has already made various alterations with the historical FLFP data, merging it with the World Bank data would be highly precarious.
FLFP data is extracted from the World Bank and is defined as the percentage of the female population aged 15 and older that is economically active out of all females aged 15 and older. The World Bank allows for two measures of FLFP, one is a national estimate and the other one is the ILO estimate. Since the historical analysis uses the ILO measures, I proceed by using these to remain consistent.
In figure 5.1 (left panel) the histogram of FLFP is displayed which shows a close to normal distribution. The differences between the present and historical distribution are remarkable. The spread in distribution is much wider in the present data where many observations show values higher than 0.8 whereas the historical data barely shows any observations exceeding 0.6. The differences might be due to minor discrepancies in definitions, however a more reasonable explanation is the difference in period and sample. The highest values of FLFP in this dataset are for African countries4, which are not included in the historical dataset. This observation is in line with the expectation that countries at lower levels of economic development (certain African countries) have higher values for FLFP. In table A4 in the appendix the FLFP are compared for the different analyses using matching countries and time periods. According to this table the FLFP measures are nearly identical for the historical and current dataset which strengthens the confidence in the reliability of both the historical FLFP measures and the present FLFP measures. A high correlation would imply that they also vary together, which is the case for the FLFP measures with a correlation higher than 0.99. Data for GDP per capita is retrieved from the World Bank and is based on purchasing power parity. This implies that GDP per capita measures are converted to international dollars (2011$) using PPP rates. In figure 5.1 (right panel) the distribution of log GDP per capita is shown. The spread is larger compared to the historical dataset. Logically it takes higher values than for the historical dataset since it includes later years and more rich countries. On the other hand, there are also lower values for GDP in the present dataset compared to the historical data. In the historical dataset the sample consists only of countries that are nowadays referred to as developed countries. Several countries are poorer these days than those developed economies were around 18905.
Education data comes from the Barro-Lee dataset. Male and female years of schooling averages are available for long periods of time in 5-year intervals. The measure of years of schooling refers to the average years of schooling the currently economically active population has enjoyed during their life. These measures are from the same data source as the historical analysis and are therefore characterized by the same definition. In both the historical and present analysis, the log of male to female years of schooling is used.
4 Burundi, Rwanda, Tanzania, Mozambique and Madagascar show the highest values of FLFP. At some point in
the period 1990 till 2015 it has exceeded 0.85 for all five countries.
5 Countries such as Liberia, Mozambique, Congo, Ethiopia, Burundi and Rwanda all show lower GDP values
The extent of data available by the World Bank allows for inclusion of more control variables for analyses with recent time frames. The additional controls are based on prominent literature concerning potential drivers of FLFP (Dunkelberg and Spiess, 2007; Bloom et al., 2009; Angeles, 2010; Madsen et al., 2018).
Bloom et al. (2009) research the impact of fertility on FLFP and find a large negative effect of fertility on FLFP. The fertility rates are from the world bank, specifically the world development indicators (WDI). It is defined as the total number of children that would be born to a woman if she were to live to the end of her childbearing years and bear children in accordance with age-specific fertility rates of the specified year. In accordance with the control variable for education, the logarithm is used.
Sources: Data for GDP and FLFP are retrieved from the world bank, specifically the World Development Indicators (WDI) database. The FLFP measure is a modelled6 ILO estimate.
Figure 5.1: Present density analysis (1990-2015)
Dunkelberg and Spiess (2007) examine the direct impact of maternal health and children’s health on FLFP in Germany in a two-year period. They find that poor maternal health has a significant negative effect on female labour force participation. As a proxy for maternal health, the measure lifetime risk of maternal death is used. The data comes from the world bank, again from the WDI. It is defined as the probability that a 15-year-old female will die eventually from a maternal cause assuming that current levels of fertility and mortality (including maternal mortality) do not change in the future, taking into account competing causes of death. Since it might present a strong determinant of FLFP it is included in the present dataset as a control. Just as for the other control variables the logarithm is used in the regression.
Madsen et al. (2018) investigate the drivers of the great fertility decline in developing countries after 1960. They find that urbanization is negatively related to FLFP which is in accordance with earlier research by Angeles (2010) who also investigated demographic transitions and its drivers. Data concerning the urbanization measure are again from the WDI. It captures the people living in urban areas divided by the total population. The logarithm of this percentage is used for the regression. Summary statistics for all variables are provided in the appendix.
6 The modelled ILO estimate refers to the econometric models the ILO uses to predict values for gaps in the
5.2 Results
In table 5.1 the results of the regression analysis are presented. In column 1 country-fixed effects are included, in column 2 time-fixed effects are included and in the 3rd column the logarithm of male to female years of schooling, fertility rates, maternal deaths and urbanization rates are included as controls. Column 4 and 5 have the sole purpose of functioning as a check of robustness. Column 4 excludes transition countries from the analysis. According to Gaddis and Klasen (2013), results supporting the U-curve around the year 1990 could stem from the inclusion of the transition countries: Estonia, Hungary, Poland, Slovak Republic, Slovenia, Japan and South-Korea. Gaddis and Klasen argue that such a one-time historical event may lead to false conclusions about the actual effect of structural change on FLFP. These transition countries were faced with rapidly rising unemployment and ended the policies promoting female employment. To verify if this influences the results for this dataset, the transition countries are excluded in column 4. Finally, in the fifth column the male labour force participation (MLFP) is considered as a robustness check to assess whether there might also be a relation between economic development and MLFP.
Column 1 shows the expected results with a negative sign for the logarithm of GDP per capita and a positive sign for the square. This is in line with the expectations associated with the U-curve hypothesis. In column 2 it becomes clear that including time fixed effects does not change the significance of the results, but it does add greatly to the explanatory power of the model represented by the R-squared statistic. Research by Lechman and Kaur (2015) uses a more or less similar timeframe (1990-2012) and number of countries (162) and finds the same results. This implies that the additional years and countries in this analysis do not change the current state of knowledge, but rather support the findings by Lechman and Kaur. Even though these findings support the U-curve, Lechman and Kaur fail to include controls to address potential omitted variable bias.
The results in column 3 show that once controls for education, fertility, maternal health and urbanization are added, the relation still holds, but that the significance has become more muted. Apparently, the control variables take away some power of GDP per capita in explaining FLFP. Olivetti (2013) addresses the issue of omitted variable bias by including a control for education the same as in this research. She finds for the period 1950-2005 that, once fixed effects are included, the education control does not have a significant impact, which is comparable to the results presented in table 5.1 (column 3). She also finds that the U-curve hypothesis remains valid after including fixed effects and a control. In addition to the education control this analysis includes more potential drivers of FLFP since various pieces of literature argued for their importance (Dunkelberg and Spiess, 2007; Bloom et al., 2009; Angeles, 2010; Madsen et al., 2018).
find a negative relation between poor maternal health and FLFP. One explanation could be the proxy chosen for this regression. It could be that the variable maternal deaths does not adequately capture the impact of maternal health. The coefficient for urbanization is significant on a 5% level and has the expected negative sign. According to Madsen et al. (2018) and Angeles (2010) increases in urbanization lead to decreases in FLFP, which is supported in this dataset.
Since the significance of the U-curve has become more muted after inclusion of the controls, it could be that the controls and the GDP measures partially explain the same relation with FLFP thus causing some multicollinearity issue. This multicollinearity issue implies that GDP per capita could be predicted by the other independent variables. To evaluate whether this is the case, the correlations of GDP per capita with the control variables are examined. The correlation of the controls with the GDP estimates is relatively high, between 0.6 and 0.75. This makes it more difficult to assess whether the changes in FLFP are due to changes in GDP or in one of the controls7. Since the significance of log GDP and log GDP squared remains, the reliability of the U-curve hypothesis is not heavily affected.
(1) (2) (3) (4) (5)
VARIABLES FLFP FLFP FLFP FLFP MLFP
Log GDP per capita -0.257*** -0.210*** -0.179** -0.203** 0.0302
(0.0698) (0.0650) (0.0887) (0.0934) (0.0446)
Log GDP per capita squared 0.0161*** 0.0108*** 0.00896* 0.01055* -0.00164
(0.00408) (0.00386) (0.00538) (0.00572) (0.00260)
Log male/female years of schooling -0.00373
(0.0422) -0.00318 (0.0428) -0.0540* (0.0294) Log fertility -0.0497** -0.0528** 0.0265* (0.0237) (0.0254) (0.0152)
Log maternal deaths 0.0209* 0.0186 0.00450
(0.0125) (0.0133) (0.00732)
Log urban population (%) -0.0570** -0.0617** -0.0183
(0.0278) (0.0282) (0.0170) Constant 1.498*** 1.481*** 1.646*** 1.751*** 0.694*** (0.297) (0.274) (0.363) (0.380) (0.188) Observations Number of countries 984 176 984 176 667 141 636 134 667 141 R-squared 0.077 0.207 0.234 0.2469 0.282
Country fixed effects Time fixed effects
Yes No Yes Yes Yes Yes Yes Yes Yes Yes Sources: Data for GDP per capita, FLFP, maternal deaths and fertility rates are retrieved from the World Bank, specifically from the World Development Indicators (WDI). The education data is retrieved from the Barro-Lee dataset.
Clustered Robust Standard Errors: *** Significant at 1% level ** Significant at 5% level * Significant at 10% level
Table 5.1: Female labour force participation for the period 1990-2005
7 The controls also show relatively high correlations ranging from -0.47 to 0.91. The highest correlation is
To address the issue with inclusion of former transition countries put forth by Gaddis and Klasen (2013), column 4 excludes these countries leaving a sample of 134 countries. The results are very similar to the regression with the full sample, which implies that it does not matter much whether the transition countries are included or not. This finding supports the theory that structural change is an important driver for FLFP and that it is not based on the inclusion of a few potential outlier-countries. This is in complete contrast with the findings by Gaddis and Klasen (2013) who find that in the OECD sample the U-curve is entirely driven by the former transition countries. The disparity may stem from differences in data sources, differences in countries in the sample, differences in age groups or differences in the timeframe.
As expected, there is no U-shaped relation between the log of GDP per capita and MLFP (column 5). The GDP measures have the opposite signs as requires and would therefore relate to an inverted U-curve, however the results are not significant, therefore pointing to the absence of a real relationship. This is in line with findings from Olivetti (2013) who suggested a male labour force placebo group and also finds no relation between GDP per capita and MLFP. The hypothesis about the relation between economic development and FLFP finds some proof in these results (column 1 and 2), but the significance drops when the controls are added (column 3). The next section zooms in further and explores this relation using regional data.
6. Regional analysis
Olivetti (2013) argues in her paper that her findings concerning the U-shaped female labour supply in the US could benefit from a thorough analysis on a lower geographical scale to determine if cross-state variations matter. She points out that there is a substantial amount of regional variation in economic structure as well as regional variation in FLFP that could be exploited. Her suggestions for future research are based on regional research by Kim who invested considerable time on studying regional or even urban differences in the US (Kim, 1998; Kim, 1999; Kim and Margo, 2004). Kim points out that the most compelling reason for studying different geographic areas at different scales is that models that may have explanatory power at a national level, may not apply at all at lower geographical levels. Several other scholars have also discussed the importance of regional data to consider within country differences. Following regional research conducted by (amongst many others) Gennaioli et al. (2014) and Beugelsdijk, Klasing and Milionis (2017) there is good reason to believe that there is very strong variation within countries and often comparability between regions from different countries. Regions within The Netherlands could differ wildly in terms of FLFP or GDP, but may be very similar to other regions in Germany or Belgium.
6.1 Data
The dataset uses data from Eurostat for 260 regions from 21 European countries. In the previous analysis, the sample consisted of all countries around the world for which data is available. Since regional data concerning participation rates is not available for all these countries, or suffers from accuracy issues8, the focus is on the reliable data from Eurostat. The countries are selected based on data availability for the period 2000 till 2016. Countries that only had regional data available from or after 2008 are excluded, since these countries miss data for (over) half the period of this research. For example, Slovenia and Norway are excluded since these regions have missing data until 2009. Moreover, countries with less than 2 sub-regions are excluded since these regions would represent the entire country. For example, Luxembourg is excluded since the only available region classification was the entire country. Finally, regions that belong to European countries, but are located outside of Europe are excluded as well. These are from France; Guadeloupe, Martinique, Guiana, Réunion and Mayotte, from Portugal; The Azores and Madeira and from Spain; The Canary Islands. The final sample consists of regions from: Austria, Belgium, Bulgaria, Croatia, Czech Republic, Denmark, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Netherlands, Poland, Portugal, Romania, Slovakia, Spain, Sweden and the United Kingdom. The classification used by Eurostat to identify these regions is referred to as the Nomenclature of territorial units for statistics (NUTS). The information used in this paper is based on regions from the NUTS 2 database which includes regions for the application of regional policies. All NUTS 2 regions have a population size in between 800,000 and 3,000,000.
Regional data of FLFP is not directly available, but Eurostat does provide for female employment (between 15 and 64 years of age) and population numbers. The FLFP rate is therefore calculated as the total female employment from 15-64 divided by the total female population from 15-64. This approach differs slightly from the data sources used for the historical and current dataset where they consider all females above the age of 15. Since Eurostat only provides employment numbers between age 15 and 64, both the female employment and the female population numbers are based on ages 15 to 64. In figure 6.1 (left panel) the density histogram of the constructed FLFP is shown. Striking is the fact that the figure appears to be very skewed to the left with a few observations even lower than 0.2. It is clearly different from the FLFP figures in the present and historical analysis, but it still has the largest bunch of the observations between 0.4 and 0.7, which strengthens the belief in the appropriateness of the FLFP measurement using employment and population data from Eurostat for women between 15 and 64 years old.
In table A4 in the appendix the regional FLFP measures are compared to the measures from the historical and present database. Apparently, the measures are consistently higher in the regional dataset. The intuition behind this is not hard to follow. The regional dataset uses a, more or less, similar number as a numerator, but a very different value for the denominator. The female labour group of 15+ (historical and present) and 15-64 (regional) are very similar since the largest group of working population is within the 15-64 range. As for the denominator, the female population of 15+ is of course a larger group than the female population between 15
and 64. Taken together, in the regional analysis a, more or less, similar number for the numerator is divided by a smaller denominator, leading to higher values for FLF
Sources: Eurostat data on female population, female employment and GDP data
Figure 6.1: Regional density analysis (2000-2016)
The difference in use of databases, definitions and measurement methods clearly alters the levels of the measures. However, the correlation between the regional and the historical dataset is around 0.97 and the correlation between the regional and the present dataset is around 0.9. Since the correlation between the FLFP measures for the different analyses is this high, the data is still useful to analyse the relation between economic development and FLFP.
Sources: Eurostat data on female population, female employment and GDP data
Figure 6.2: Female Labour Force Participation in Europe, 2000-2016
Germany, the range of FLFP across regions is very high. This supports the importance of analysing the U-curve at a regional level, where at a country level within-country differences would be missed. The lowest FLFP can be found in the South and East of Europe whereas in the Centre and North several German, Dutch, Belgian, British and Scandinavian regions are at the upper end of the FLFP distribution. Comparison of the two figures is rather tricky since the legends are adapted for the specific FLFP values for both years. For example, a region with a FLFP of 0.65 in both 2000 and 2016 will switch from dark blue in 2000 to a lighter blue in 2016 due to different groups in the legend. However, what you can conclude from these figures is that the FLFP distribution for the entire sample has remained quite stable over the period 2000 till 2016 with at the low end the Eastern and Southern European regions and at the upper end the central and Northern regions. Overall, an increase of mean FLFP can be observed from 0.52 to 0.59 over the period 2000 till 2016.
For almost all regions GDP per capita data is available for the period 2000-2016. In the Eurostat database they refer to the measure as the purchasing power standard per inhabitant. In figure 6.1 (right panel) the histogram of log GDP per capita is displayed to examine its density distribution. The distribution is slightly skewed to the left, but appears to follow a rather normal distribution. No major differences with previous analyses can be observed. Summary statistics for all variables are provided in the appendix.
6.2 Results
In table 6.1 the results for the regional regression are presented, where column 1 includes region fixed effects and column 2 includes both region and time fixed effects. Column 3 excludes the crisis year 2008 to evaluate if any drastic differences have occurred due to this one-time shock. Column 4 provides an extra robustness check where the dataset is used for a country level analysis to assess whether the geographical scale matters for the results. The results of column 1 and 2 are in line with previous analyses, where the Log of GDP per capita has a negative sign and the square a positive sign, confirming a U-shaped curve.
A comparison with previous analyses reveals several noteworthy elements. Firstly, the significance of the GDP coefficients remains significant on a 1% level for the regional and the present data after controlling for region/country- and time-fixed effects. The historical analysis, however, becomes more muted when the time-fixed effects are included. Secondly, the coefficients of the regional analysis lie further away from zero than in the previous analyses. Thirdly, the education control is absent in the regional analysis. In previous analyses it became evident that the U-curve is more muted when the education control is added. Since it is absent in this regression we cannot derive any conclusions about the potential role of the education gap in this section. The absence of the education control also makes running a regression for MLFP less relevant since results may be completely altered if the control would be included.
scale of analysis would not matter, the results should be the same for the regional and the country analysis since the countries are composed of these regions. The sign of the coefficient of GDP (-) and the sign of the coefficient of GDP squared (+), match the U-curve analysis. However, in contrast to the analysis using regions, these results are not significant. Clearly, the results change when the geographical level of analysis changes.
(1) (2) (3) (4)
VARIABLES FLFP FLFP FLFP
except 2008
FLFP countries
Log GDP per capita -1.603*** -1.448*** -1.457*** -0.774
(0.240) (0.255) (0.258) (0.678)
Log GDP per capita squared 0.0884*** 0.0785*** 0.0790*** 0.0432
(0.0124) (0.0135) (0.0136) (0.0351) Constant 7.726*** 7.158*** 7.207*** 3.979 (1.163) (1.204) (1.218) (3.276) Observations Regions 4,208 260 4,208 260 3955 260 355 21 R-squared 0.376 0.407 0.411 0.4552
Region/country fixed effects Time fixed effects
Yes No Yes Yes Yes Yes Yes Yes Sources: Eurostat, regional data for GDP, female population and female employment.
Clustered Robust Standard Errors: *** Significant at 1% level ** Significant at 5% level * Significant at 10% level
Table 6.1: Female labour force participation, a regional analysis (2000-2016)
7. Discussion
Throughout the paper it becomes clear that there are certain potential limitations to this work that ought to be discussed in more detail. The main points to be considered are firstly, endogeneity issues, of which the major concern is a potential causality matter. Secondly, the consistency throughout the historical, present and regional analysis requires some additional clarification. Thirdly, multicollinearity issues in the present dataset are to be acknowledged. Fourthly, the exclusion of control variables for the regional analysis ought to be discussed. Finally, there are some notes concerning structural change and the measure GDP as a development or structural change indicator.
the independent variable does not only influence the dependent variable, but it is also (partially) determined by the dependent variable, posing a reverse causality issue. In this area of research, the economic rationale finds some ground for both ways that causality can run. The rationale for structural change (economic development) leading to different levels of FLFP has been extensively discussed, but the effect of FLFP on economic development also finds some ground in the literature.
The main line of argument behind this is threefold. Firstly, by excluding women from the labour force a country restricts the so-called talent pool to solely male talent. Klasen and Lamanna (2009) argue that the unutilized female talent can explain economic growth differences between countries over the time-period 1960-2000. They show that the Middle-East, North-Africa and South-Asia suffer from low female participation rates compared to East-Asia. Secondly, women’s labour force participation increases family income and potential for increased savings which in turn results in higher capital stock per worker and therefore output growth. The paper by Galor and Weil (1996) uses fertility rates to explain economic growth. Their main argument is that when women’s relative wages (compared to men) rise, their fertility goes down due to higher opportunity cost of having children. This decrease in fertility raises the available capital per worker which in turn boosts output. Thirdly, if gender gaps in education are large, investment in female education leads to greater macro-economic growth. Knowles et al. (2002) explain that if gender gaps are large in education, female participation boosts output growth more than male participation due to diminishing returns to human capital.
One way to lessen potential concerns about endogeneity is by using a dynamic model with the application of generalised method of moments (GMM) which is suggested by several pieces of literature in this field (Luci, 2009; Gaddis and Klasen, 2013; Lechman and Kaur, 2015). The paper by Gaddis and Klasen provides the most extensive explanation of the dynamic model. They include an additional (instrumental) variable, namely the lagged level of FLFP to test whether FLFP is determined by its past values. Fixed effects are problematic to apply in this case since it leads to a dynamic panel bias. To allow for the inclusion of lagged values of the dependent variable, Gaddis and Klasen have used a common method for dynamic models, namely GMM estimation. They explain that the most important benefit of a dynamic model is that it allows treating the GDP variables as endogenous which is not possible in the fixed effects model. This paper focusses on a slightly different model where lagged values of FLFP are not included and therefore GMM estimation is not required and appropriate. However, the results of this paper and the paper by Gaddis and Klasen provide comparable conclusions. They argue that the U-curve sometimes finds some credibility, but that it is not always robust across estimation methods and sample groups. In the analyses in this paper it becomes clear that the U-curve hypothesis also finds some credibility, but that it depends on the sample group, the level of analysis, the timeframe, data sources and control variables. In both cases there is no generalizable conclusion that supports the U-curve.
all these measures is very high which makes comparison still valid. However, if data had been available for the female group of 15+ this would have made for a more consistent comparison. Another option would have been to follow the measurement methods of the regional analysis by looking into present data for FLFP for women between 15 and 64. However, data concerning participation rates for the historical and present analysis is not available for the complete sample for the age group 15-64, but solely for the group 15+ and older. Moreover, if this data would be available, this would make comparison with the Olivetti paper less sensible and reliable since she uses the age group 15+ for her analyses.
In the present dataset several control variables are included to minimize the omitted variable bias. This, however, has led to multicollinearity issues between the GDP measures and the control variables. It is therefore more difficult to determine which changes in FLFP are due to changes in GDP measures and which changes are caused by the control variables. After inclusion of all the controls, the GDP measures remain significant, though muted, which strengthens the belief in the credibility of the U-curve hypothesis for the present analysis.
Fourthly, the control variables used for the present dataset are missing for the historical and the regional dataset. Data currently available has allowed for inclusion of education measures in the historical dataset (Barro-Lee), but unfortunately there are no comparable measures available on a regional level. The regional level results should therefore be interpreted with caution, since there might be some omitted variable bias. Furthermore, comparable measures for fertility, maternal health and urbanisation are not available for the historical and regional analysis and are therefore absent.
Finally, some notes about structural change must be placed. Gaddis and Klasen (2013) point out that GDP is an imperfect measure to capture structural change. Firstly, because GDP measures are often dependent on international price comparisons and PPP revisions which created some inconsistencies in their analysis. Secondly and most importantly, GDP levels are not directly linked to levels of economic development. It could very well be that a country has a high level of GDP per capita, but is still highly dependent on agriculture/manufacturing or that a country with a low level of GDP per capita is dependent on the services industry9. They argue that industry specific (agriculture, manufacturing, services) value added and growth rates are much closer to the original idea of the U-curve hypothesis with structural change as a driving force. Although I acknowledge that value added by industry might present a closer alignment with the U-curve hypothesis, the aim of this paper was to re-examine and extend the original U-curve literature while adhering to the common practices in this area of research. The use of GDP per capita at PPP is the common practice and was therefore required for this investigation. Moreover, values for GDP per capita are available for a historical period, a period with a more recent timeframe and on various geographical scales whereas value added by industry is only available for the more recent timeframe. Even though the use of value added by industry was not the aim of this paper, it does suggest that there may be other proxies for structural change that are more in line with the U-curve hypothesis.
One last point to discuss is the appropriability of a short-time period to investigate the
9 An example of a country that has a relatively high GDP per capita level and is still highly dependent on
U-curve hypothesis. Since the argument to support the U is based on structural change, one might argue that a timeframe of 25 years (present analysis) or even 16 years (regional analysis) is too short to examine the process of structural change of a country. However, several countries/regions have shown considerable variation over this period in both GDP per capita and FLFP. The substantial increases in GDP per capita10 are a sign of economic development which therefore argues in favour of also considering structural change for relatively short timeframes. Even though it is still relevant to consider shorter timeframes, I do argue that the ideal situation would be to examine a long timeframe with a world sample to capture the total process of structural change. This presents an interesting avenue for future research where data availability poses the most pressing issue since FLFP data for most countries in the world is so far only available for a relatively recent timeframe.
8. Conclusion
This paper set out to contribute to the existing literature concerning the U-shaped female labour supply curve. The aim of the paper is to examine the impact of structural changes in the economy on female labour force participation. This research contributes to the existing literature by focussing on both a historical and a more recent period. In extension, a regional analysis in Europe provides for an innovative scale of research in the currently available literature. Moreover, literature concerning other determinants of FLFP is consulted to find relevant controls to mitigate the issue of potential omitted variable bias where possible.
The historical data consists of 16 developed economies over the period 1890-2005. The results show that there is some evidence for the U-curve hypothesis, but that this evidence to some extent disappears when time dummies are included. Thus, once the aggregate trend is excluded by including time-fixed effects, the U becomes more muted. The additional education control further weakens the relation between the GDP measures and FLFP, until there is no statistical significance present. The present data expands the number of countries to over 170 for the period 1990-2015. The results confirm the U-curve hypothesis even after controlling for time- and country-fixed effects. After including controls for education, fertility, maternal health and urbanization the U becomes more muted, but some significance of the GDP coefficients remains. The regional dataset greatly increases the number of observations, including 260 regions from 21 European countries for the period 2000-2016. The results appear to support the U-curve hypothesis with great confidence even after controlling for time- and country-fixed effects. Unfortunately, the data does not allow the inclusion of other potentially important drivers of FLFP as control variables. Since these have proven to mute the relevance of the U-curve for the historical and present data, a robust conclusion about the presence of a female labour supply U is difficult to establish. Key for further investigation on a regional level is obtaining data on drivers of FLFP. Once fitting control variables are available on a regional level, a complete comparison can be drawn between a regional level analysis and a country level analysis. This way it is possible to determine whether differences in geographical level of analysis matter for the support of the U-curve hypothesis.
10 For the present period the average of GDP per capita went up by 49% For the regional analysis the average of
The literature on this topic so far presented a mixed story about the relation between structural change and FLFP. The literature supporting a U-curve (Pampel and Tanaka, 1986; Çağatay and Özler, 1995; Luci, 2009; Tam, 2011; Olivetti, 2013; Lechman and Kaur, 2015) seems to present a solid base. However, the extensive research by Gaddis and Klasen (2013) has shown that for different models and different data sources the structural change U-curve does not hold in many cases. Moreover, the literature concerning fertility, education, urbanization and maternal health as determinants of FLFP (Dunkelberg and Spiess, 2007; Bloom et al., 2009; Angeles, 2010; Madsen et al., 2018) also finds ground, pointing out that structural change may not be the sole or best predictor of FLFP. As a contribution, this paper finds that differences in the model, the timeframe, the estimation methods, the sample group, the controls, the data sources and the geographical level of analysis matter for the results. Therefore, the U-curve hypothesis appears to be dependent on all kinds of circumstances and does not find generalizable support.
9. References
* Angeles, L. (2010). Demographic transitions: analysing the effects of mortality on fertility. Journal of Population Economics, 23(1), pp.99-120.
* Arellano, M. and Bond, S. (1991). Some Tests of Specification for Panel Data: Monte Carlo Evidence and an Application to Employment Equations. The Review of Economic Studies,
58(2), p.277.
* Barro, Robert and Jong-Wha Lee, 2013, "A New Data Set of Educational Attainment in the World, 1950-2010." Journal of Development Economics, vol 104, pp.184-198.
* Beugelsdijk, S., Klasing, M., & Milionis, P. (2018). Regional economic development in Europe: the role of total factor productivity. Regional Studies, 52(4), 461-476.
* Bloom, D., Canning, D., Fink, G. and Finlay, J. (2009). Fertility, female labour force participation, and the demographic dividend. Journal of Economic Growth, 14(2), pp.79-101. * Boserup, E. (1970), Women`s role in economic development, St.Martin, New York.
* Çağatay, N. and Özler, Ş. (1995). Feminization of the labour force: The effects of long-term development and structural adjustment. World Development, 23(11), pp.1883-1894.
* Dunkelberg, A. and Spiess, C. (2007). The Impact of Child and Maternal Health Indicators on Female Labour Force Participation after Childbirth - Evidence for Germany. SSRN
Electronic Journal.
* Elborgh-Woytek, K., Newiak, M., Kochhar, K., Fabrizio, S., Kpodar, K., Wingender, P., Clements, B. and Schwartz, G. (2013). Women, Work, and the Economy: Macroeconomic Gains from Gender Equity. IMF staff discussion note
* EuroGeographics for the administrative boundaries. (Eurostat)
* Esteban, J. (2000). Regional convergence in Europe and the industry mix: a shift-share analysis. Regional Science and Urban Economics, 30(3), pp.353-364.
* Gaddis, I. and Klasen, S. (2013). Economic development, structural change, and women’s labour force participation: Journal of Population Economics, 27(3), pp.639-681.
* Galor, O. and Weil, D. (1996), The gender gap, fertility, and growth. The American
Economic Review, Vol 86, No.3, pp. 374-387
* Gennaioli, N., La Porta, R., Lopez-de-Silanes, F. and Shleifer, A. (2014). Human Capital and Regional Development *. The Quarterly Journal of Economics, 128(1), pp.105-164. * Goldin, C. (1995), The U-shaped female labour force function in economic development and economic history. Schultz TP Investment in Women’s Human Capital and Economic
Development. University of Chicago Press ; 1995. pp. 61-90.
* Kim, S. (1998). Economic Integration and Convergence: U.S. Regions, 1840-1990.
Journal of Economic History, 58 (3): 659-683.
* Kim, S. 1999. Regions, Resources and Economic Geography: The Sources of U.S. Regional Comparative Advantage, 1880-1987. Regional Science and Urban Economics, 29 (1): 1-32.
* Kim, Sukkoo and Robert A. Margo. (2004). Historical Perspectives on U.S. Economic Geography. in V. Henderson and J. Thisse, eds. Handbook of Urban and Regional
Economics, Volume 4: Cities and Geography, pp. 2982-3019. Amsterdam: North-Holland.
* Klasen, S. and Lamanna, F. (2009). The Impact of Gender Inequality in Education and Employment on Economic Growth: New Evidence for a Panel of Countries. Feminist
Economics, 15(3), pp.91-132.
* Knowles, S. (2002). Are educational gender gaps a brake on economic development? Some cross-country empirical evidence. Oxford Economic Papers, 54(1), pp.118-149.
* Lechman, E. and Kaur, H. (2015). Economic growth and female labour force participation – verifying the U-feminization hypothesis. New evidence for 162 countries over the period 1990-2012. Economics & Sociology, 8(1), pp.246-257.
* Luci, A. (2009). Female labour market participation and economic growth. International
Journal of Innovation and Sustainable Development, 4(2/3), p.97.
* Maddison Project Database, version 2018. Bolt, Jutta, Robert Inklaar, Herman de Jong and Jan Luiten van Zanden (2018), “Rebasing ‘Maddison’: new income comparisons and the shape of long-run economic development”,
* Madsen, J. B. and Moslehi, S. and Wang, C. (2018) What Has Driven the Great Fertility Decline in Developing Countries since 1960?, The Journal of Development Studies, 54:4, 738-757,
* Ngai, L. and Petrongolo, B. (2017). Gender Gaps and the Rise of the Service Economy.
American Economic Journal: Macroeconomics, 9(4), pp.1-44.
* Olivetti, C. (2013). The female labour force and long-run development: The American experience in comparative perspective. National Bureau of economic research. Cambridge * Olivetti, C. and Petrongolo, B. (2014). Gender gaps across countries and skills: Demand, supply and the industry structure. Review of Economic Dynamics, 17(4), pp.842-859.
* Pampel, F. and Tanaka, K. (1986). Economic Development and Female Labour Force Participation: A Reconsideration. Social Forces, 64(3), p.599.
* Sinha, J. N. (1967). Dynamics of female participation in economic activity in a developing economy. The Proceedings of the World Population Conference, Vol. IV: Selected Papers and
Summaries. New York: Department of Economic and Social Affairs, United Nations. pp.
* Tam, H. (2011). U-shaped female labour participation with economic development: Some panel data evidence. Economics Letters, 110(2), pp.140-142.
* World development indicators. Washington, D.C. :The World Bank.
