Thesis Submitted in Partial Fulfillment of the Requirements for the Degree of Master of Science in Economic Development and Globalization (MSc ED&G)
Women and Water: The Role of Water in Determining
Female Labor Force Participation
A Macro-Level Study of Low- and Middle-Income Countries
ii
Abstract
This paper analyzes the effect of water access and water quality on female labor force participation (FLFP) in 109 low- and middle-income countries between 2000 and 2017. As women are over-proportionally involved in household activities such as water fetching and caretaking, they benefit particularly from improvements in water infrastructure. Taking a macro-level perspective, a panel-data analysis identifies a U-shaped water access-FLFP nexus. Jointly with primary school enrollment, improvements in water access increase FLFP. Water quality also exercises a positive effect on FLFP, with increasing returns as availability of uncontaminated water increases. This effect is weaker in countries with higher life expectancies. The findings highlight the importance of providing enabling conditions and decreasing labor force barriers to allow women to convert better health and freed time into economic activity.
iii
Table of Contents
1. Introduction ... 1
2. Literature Review and Main Research Questions ... 3
2.1 Women and Economic Development ... 3
2.2 Women, Water, and Labor Force Participation ... 5
2.3 The Benefits of Going Macro ... 8
3. Data and Methodology ... 8
3.1 The Dependent Variable: Female Labor Force Participation ... 9
3.2 The Variables of Interest: Water Access and Water Quality ... 10
3.3 Control Variables ... 11
3.4 Descriptive Statistics ... 13
3.5 Empirical Model ... 15
4. Empirical Analysis ... 16
4.1 Regression Results ... 16
4.2 Income Group Subsamples ... 18
4.3 The Endogeneity Problem and Solution ... 20
4.4 Marginal Effects and Turning Points ... 23
4.5 A Dynamic Approach: Difference GMM ... 26
4.6 Robustness ... 27
5. Discussion of the Results... 29
6. Conclusion ... 34
References ... 36
iv
Figures
Figure 1: Trends in Water Access, Water Quality, and FLFP over time ... 13
Figure 2: Water and FLFP ... 14
Figure 3: Marginal Effect of Water Infrastructure Improvements on FLFP ... 25
Tables Table 1: Summary Statistics ... 13
Table 2: Correlation Matrices ... 14
Table 3: Fixed Effects Estimations ... 17
Table 4: Access Model Estimations of the Income Subsamples ... 18
Table 5: Quality Model Estimations with Income Dummies ... 19
Table 6: 2SLS Estimations ... 22
Table 7: Turning Points and Marginal Effects in the Water Access-FLFP Relationship ... 24
Table 8: Turning Points and Marginal Effects in the Water Quality-FLFP Relationship ... 24
Table 9: Dynamic Panel Data Estimation with Difference GMM Estimator ... 26
Table 10: Robustness: Dependent Variable GenderGap ... 28
v
Abbreviations
AIC Akaike-Information-Criterion
BIC Bayesian-Information-Criterion
FE Fixed Effects
FLFP Female Labor Force Participation
GDP Gross Domestic Product
GMM General Method of Moments
ILO International Labor Organization
IV Instrumental Variable
JMP Joint Monitoring Program for Water Supply, Sanitation, and Hygiene
LFP Labor Force Participation
LIC Low-Income Countries
LMIC Lower Middle-Income Countries
MLFP Male Labor Force Participation
OECD Organization for Economic Co-operation and Development
POLS Pooled Ordinary Least Squares
RCT Randomized Controlled Trials
RMSE Root-Mean-Square Error
SDG Sustainable Development Goals
SNA System of National Accounts
2SLS Two-Stage Least Squares
UMIC Upper Middle-Income Countries
UNICEF United Nations Children’s Fund
WHO World Health Organization
WASH Water, Sanitation, and Hygiene
WDI World Development Indicators
1
1. Introduction
"No country can ever truly flourish if it stifles the potential of its women and deprives itself of the contributions of half of its citizens."
Michelle Obama
Currently, 785 million people do not have access to basic drinking water services and at least two billion rely on drinking water from a source which is contaminated with feces (JMP, 2019). In some areas, the average daily water fetching time amounts to two or even three hours – a time burden that predominantly falls onto women (Gross et al., 2018). Simultaneously, gender equality is recognized as a fundamental human right and is often considered foundational for a prosperous and sustainable world, whereas gender gaps remain a prevalent issue in numerous areas. In many countries, women suffer legal discrimination and face socioeconomic barriers such as limited access to education or to financial services, to mention only a few. In nearly all countries, female labor force participation is lower than male participation. Yet, female labor force participation is broadly considered an important growth engine, bearing high macroeconomic gains. Especially in developing countries, water infrastructure concerns and women’s issues are largely interconnected. Therefore, attaining gender equality and improving water infrastructure are key goals of the United Nations’ 2030 Agenda and the central role of women in water provision, safeguarding, and management is one of the four internationally recognized principles of water management (ICWE, 1992; UN, 2015). Given that the burden of household and water fetching chores falls disproportionately on women, deficient water infrastructure increases gender inequality and consumes much of women’s time. Additionally, water shortages, the manual transportation of water vessels, and the use of contaminated water especially harm women’s health. As such, deficient water infrastructure severely impacts women’s lives and livelihoods and hinders them from achieving their true potential, for instance in terms of labor force participation.
2
better water quality increase FLFP. The water access-FLFP relationship is thereby primarily shaped by time-saving dynamics. In consideration that further enabling conditions are required to allow the translation of free time into labor market activities, I additionally hypothesize that a lower childcare burden resultant from children attending primary school strengthens the positive impact of water access on FLFP. As water quality is expected to affect FLFP primarily through related channels, I further analyze the hypothesis that in a better health-environment, water quality improvements exercise a stronger positive effect on FLFP.
3
2. Literature Review and Main Research Questions
It is widely known that economic development, modernization and structural change affect women and men differently (Ray, 2007). A breadth of literature has covered the role of women in the economy and in economic development.1 This chapter provides an overview of research related to women’s labor force participation and especially its interplay with water infrastructure. Additionally, the benefits imminent to the macro-level approach of this research will be discussed. Yet, since the literature is vast, this review does not aim to be exhaustive.
2.1 Women and Economic Development
Economic development and FLFP are closely interrelated. Scholars regularly emphasize the growth potential of FLFP and estimate that gender gaps in employment result in losses of between ten and 20 percent of a country’s per capita Gross Domestic Product (GDP) (Aguirre et al., 2012; Cuberes & Teignier, 2012). Yet, a second body of research states that causality goes the opposite way and highlights the impact of economic growth on women’s economic activity.2 There are different opinions concerning the strength and direction of this relationship.
While women’s access to education and political participation, among many other developmental indicators, is expected to increase throughout a country’s development path, the implications for FLFP appear to be ambiguous (Duflo, 2011). Attempts to shed light on the issue consider time allocation models and assume that a woman primarily decides between market-oriented activities, whether this is for wage or at the family farm, and household activities (Ilahi & Grimard, 2000). A large body of literature analyzes the variation in FLFP across countries (Goldin, 1994; Mammen & Paxson, 2000; Luci, 2009; Gaddis & Klasen, 2014; among many others). Textbook-related models, which assume competitive labor markets and utility maximizing behavior, predict an increasing FLFP throughout all stages of economic development and determine discrimination as a temporary phenomenon (Luci, 2009; Becker, 2010). Conversely, an application of the specific factors model concludes that technological progress or increases in trade resulting from economic development typically discriminate against women and predicts decreasing FLFP (Norris, 1992). Another approach, largely known as the Feminization U, recognizes the presence of non-linear developments. According to this hypothesis, the relationship between economic development and FLFP follows a U-shape due to structural change, fertility dynamics, and changes in education (Sinha, 1967; Boserup, 2007; Gaddis & Klasen, 2014). It predicts high levels of FLFP at low income levels as, assuming a dominant agricultural sector, women tend to work in household businesses or on family farms in proximity of the home. Such work is often unpaid, and activities are mainly at the subsistence level (Goldin, 1994; Mammen & Paxson 2000; Gaddis & Klasen, 2014). With economic growth, FLFP is expected to decline as markets expand and technological progress enters the country (which causes the increase in incomes). The arising industrial sector is considered to be a male-oriented work environment so that a resulting increase in male earnings, as well as
1 For a discussion about the complexity of gender in development and the Gender and Development versus Women
in Development debate, see Ray (2007).
2 In line with Duflo (2011), I emphasize that both sides of the development-empowerment relationship are crucial
4
the productivity increase in the economy, exercise a negative income effect on female labor supply (Gaddis & Klasen, 2014).3 The growing manufacturing sector comes in hand with decreasing agriculture activity and increasing urbanization. This reduces the number of family businesses such as farms and lowers economic opportunities for women in absolute terms (Mammen & Paxson, 2000). Thereby, the hours women work might not change significantly, as women’s labor is rather “bought by the family” (Godin, 1994, p.1) and thus conducted at home whereas the men leave the house for (remunerated) work. With further income increases in a country, women can access the emerging white-collar service sectors, meaning that FLFP is expected to increase again in absolute terms but as well relative to men. This is supported by more narrow gender gaps in education, lower fertility, better access to childcare, and more flexible working options (Mammen & Paxson, 2000; Gaddis & Klasen, 2014). At this advanced stage of development, the substitution effect, which is related to higher potential female wages and thus rising opportunity costs of staying at home, dominates the initially stronger income effect (Psacharopolus & Tzannatos, 1989; Mammen & Paxson, 2000; Gaddis & Klasen, 2014). Aside from a country’s economic performance, other factors influence a women’s participation in the labor market. The list of such determinants of FLFP is large, so that this analysis particularly discusses the indicators deemed to be most significant by the literature and considered to be most relevant in a macro-level context. Firstly, a rather intuitive determinant of FLFP is education. While in the long run, increased education rates are expected to positively affect FLFP, short-run effects rather reflect the tradeoff between a lower care burden when children are at school and a lower burden of household chores when children stay at home and help with such tasks (Eckstein & Lifshitz, 2011; Steinberg & Nakane, 2012). Additionally, women’s labor supply decisions are assumed to be affected by the situation in their household, which also impacts macro-level dynamics. In a traditional household setting, the employment status of the husband is found to be relevant. On the one hand, higher male labor force participation (MLFP) could lower the incentive of women to work, and thus exercise an income effect on women. On the other hand, a higher MLFP could provide incentive for women to also become economically active to strengthen their say in society or in their households. Additionally, facing high MLFP, women could take over agriculture-related practices, especially for achieving their subsistence needs (Mammen & Paxson, 2000; Klasen et al., 2019). Another frequently mentioned determinant of FLFP is life expectancy, the impact of which is ambiguous. Proxying a country’s healthcare, it suggests a positive effect as higher levels of general health are expected to enable economic activity. Contrastingly, high life expectancy could also imply a higher elderly care burden and cause a negative impact on FLFP (Mammen & Paxson, 2000; Altuzarra et al., 2019). Additionally, if retirement age does not increase with increasing life expectancy, a negative FLFP-effect may result (Bussmann, 2009; Altuzarra et al., 2019). Lastly, the institutional environment, especially (in)equalities in legal rights, both de jure and de facto, are largely found to affect female labor force participation (Gonzalez et al., 2015; Jain-Chandra et al., 2018).
3 Contradicting voices could state that manufacturing tasks as well involve refined manufacturing and handicraft
5
2.2 Women, Water, and Labor Force Participation
Duflo (2011) highlights that one of the key causes of gender inequality, especially in the labor force, results from time allocation. Studies such as Berniell and Sanchez (2012) or Samman et al. (2016) find the unsurprising pattern that, at all income levels, the majority of housework and care activities are done by women, which negatively impacts time spent in labor market activities. As a result, deficits in basic infrastructure provision are found to affect women more than men (Ghani et al., 2013). This is especially important in low- and middle-income countries where household tasks are more time consuming due to the deficient infrastructure and a resultant lack of time-saving appliances (Peters et al., 2016). In this context, a country’s economic growth can positively impact women’s time allocation by providing improved infrastructure and thereby freeing women’s time.4
Only rather recently have gender-specific analyses of water-related issues been given attention. There is broad evidence in the literature that women are predominantly responsible for water collection tasks, especially in areas that suffer from deficient water infrastructure (Sorenson et al., 2011; Koolwal & van de Walle, 2013; Graham et al., 2016; Peters et al., 2016).5 In households with water sources off-premises, 73.5 percent of the water collection tasks are carried out by women, compared to 16.6 percent men and 9.8 percent children (JMP, 2017). Therefore, the implications of deficient access to water are especially impactful upon women. This is exacerbated by a lack of substitutes for water and the resulting inelastic water demand (Ilahi, 2001; Koolwal & van de Walle, 2013). Lamb (2015) estimates that in rural Africa, women spend about 26 percent of their time on water collection and that, in some regions, they have to walk an average of six kilometers each trip to fetch water. It can thus be assumed that better access to water frees women’s time. Fisher (2008) argues that this generates direct and indirect employment opportunities: Directly, women are able to engage in income-generating activities that require water access and indirectly, women can use the freed time to access the labor market. However, time is only one cost of water fetching. It also results in high caloric expenditure and negative health implications. Carrying water over long distances causes back, hip, and neck injuries over the years and estimates suggest that, during the dry season, women in Africa spend around 30 percent of their daily energy intake on water fetching activities (Seaforth, 2001; Fisher, 2008; Sorenson et al., 2011). Moreover, health issues also result from the use and consumption of contaminated water. It is largely recognized that better water quality results in less water-borne and other water-related diseases. Improved water infrastructure thus affects women directly through their own health improvements and indirectly by easing the care burden as women are mainly responsible for caring for children or other ill family members (Ilahi, 2001; Fisher, 2008).6 Additionally, better water quality provides better starting
4 For the effect of electrification on FLFP see Dinkelman (2011) for South Africa and Grogan and Sadanand (2013)
for rural Nicaragua. For the household appliances-FLFP nexus see Greenwood et al. (2005) for the US and Cavalcanti and Tavares (2008) for the UK.
5 Besides, children and especially girls, are also involved in water fetching activities. Similar studies thus analyze
the link between water infrastructure and girls’ school enrollment or attainment, for example Ortiz-Correa et al. (2016) for Brazil, Nauges (2017) for Ghana, and Koolwal and van de Walle (2013) for nine water-scarce countries.
6 While it could be argued that a substitution effect incentivizes women to become economically active if a
6
conditions for children by reducing the risk of hepatitis and hookworm infestations during pregnancy. The latter ailment was found to increase the likelihood of low birthweight and lower child growth (Sultana & Crow, 2000; WHO/UNICEF, 2005). Cleaner water thus improves children’s health and, additionally, prevents them from contracting water-borne diseases during their childhood.
Combining this insight into FLFP and the water-women nexus, the question arises how FLFP is affected by water access and water quality. Associated studies have mainly focused on the opportunity costs of water fetching activities, as well as its health implications.7 Emphasizing a time reallocation hypothesis, this research suggests that women transfer their freed time and improved health into labor force activities. However, Ray (2007) argues that this implication lacks broad-based evidence. She further states that the conclusion largely depends on whether labor market opportunities exist, the ability and willingness of women to use them, and whether other enabling conditions, for instance provision of childcare, are given. This is in line with King and Alderman (2001) who highlight that freed time might be translated into childcare activities. Studies that analyze the water-FLFP nexus provide mixed results. Most of these studies assume a microeconomic perspective and make use of household survey data. Koolwal and van de Walle (2013) analyze the impact of time needed for water collection on women’s market-based work, as well as on own-farm and other unpaid family work. Aggregating household data of rural areas in nine developing countries to the community level, they find no evidence that improved water access leads to greater off-farm market-based work of women. For some countries, they find that improved water access reduces female participation in own agricultural production and other unpaid work. Focusing on rural areas in Pakistan, Ilahi and Grimard (2000) estimate how different levels of access to water infrastructure affect women’s time allocation between market-oriented activities, water collection, and leisure. The resulting effect at the community level is not statistically significant. At the household and individual level, however, they find that reductions in water collection times increase women’s engagement in market-oriented activities. Yet, this does not hold if families have private access to water. In that case, they find a negative effect on women’s labor force participation and conclude that the work burden of women with water access on the premises is lower, and that they enjoy increased free time. In studying Tangiers, Morocco, Devoto et al. (2012) find that having a household water connection comes along with time saving implications and increases quality of life and welfare. However, no impact on income-generating activities or health was determined. Aside from the microeconomic perspective, some macroeconomic studies also take water-related variables into account. Those studies usually aim at analyzing determinants of FLFP and use water variables as controls, while their main interest lies on other topics. Looking at the effect of structural change on the female employment to population ratios in 39 least-developed countries, Wamboye et al. (2015) include the percentage of urban population with improved sanitation facilities as a water-related control variable. While this variable does not directly represent water access or quality, its coefficient is positive and significant. Evaluating
7 Less attention has been given to aspects of political participation and within-household voting power, as well as
7
gender equality policies in 100 countries, Jain-Chandra et al. (2018) estimate determinants of gender gaps in employment and find that better water and sanitation infrastructure contributes to closing the employment gender gap. After all, the above studies mainly evaluate the timesaving implications of improved access to water. While water quality is also expected to play a role, studies involving the water quality-employment nexus are scarce, mainly due to data constraints (Fisher, 2008; Sorenson et al., 2011).
Based on this theoretical background, expectations concerning the water-FLFP relationship arise. Insight from previous research illustrates ambiguous effects of water infrastructure on FLFP. The water access-FLFP nexus seems to function through time saving and time reallocation channels, whereas water quality is assumed to affect women mainly via health-related channels. While the direction of the water-FLFP nexus is unclear and literature provides indications that it might be of negative nature, one would expect that under the right circumstances, women’s participation in the workforce benefits from easier water access and better water quality. Additionally, as water-related control variables in macro-level estimations display a positive sign, I hypothesize that:
H1: Improvements in water access increase FLFP at the macro-level. H2: Better water quality has a positive effect on FLFP at the macro-level.
In light of the literature, I additionally assume that the presence of enabling conditions allows women to more easily transfer their freed time into labor force participation (Ray, 2007; Koolwal & van de Walle, 2013). This is emphasized by Gross et al. (2018), who state that women face large barriers to enter the labor force. One such burden is childcare, whose main responsibility traditionally falls on women. King and Alderman (2001) state that time freed due to better water access might be translated into childcare activities instead of work. If children attend school, they are assumed to spend more time outside the home, which lowers the care burden for women. I therefore assume that a higher primary school enrollment rate reduces the care burden and thus enables women to translate freed time from better water access into participation in the labor force, and hypothesize that:
H3: A higher share of primary school enrollment reinforces the positive impact of water
access and on FLFP.
In this context, school enrollment primarily mirrors the provision of childcare. Besides, there is evidence that education levels play a role in defining the use and outcome of water access, for example through increased awareness of water management practices (Mangyo, 2008; JMP, 2015, 2017). This suggests that benefits resulting from better water access are higher when people, especially women, are relatively well educated. As primary education enrollment rates are assumed to be higher in countries with better education systems, they are expected to reflect a country’s educational level, at least to a certain degree (Keller, 2006). This supports the third hypothesis. Lastly, the literature suggests that water quality improvements predominantly affect women through health-related channels, so I additionally hypothesize that:
8
2.3 The Benefits of Going Macro
Before translating the hypotheses into a testable model, the question may arise why this thesis considers the water-FLFP relationship from a macroeconomic perspective. This concern is reasonable, given that water access and quality vary largely within countries. These within-country differences are not well captured by aggregated indicators (Gaddis & Klasen, 2014). Additionally, a macroeconomic perspective comes along with limitations in terms of data use, for example concerning the influence of household compositions on FLFP (Mammen & Paxson, 2000; Ilahi, 2001). Besides, wage levels differ between regions, occupations, and socio-cultural groups, among many others, and thus lose explanatory power at an aggregated level. Yet, they are an important aspect to consider in microeconomic studies. Recently, microeconomic approaches were rewarded with great attention, especially since last year’s Nobel prize for Duflo et al.’s work with Randomized Controlled Trials (RCT).8 However, a
common critique of RCT, which equally applies for microeconomic analyses in general, relates to its external validity. There are doubts whether the findings also hold elsewhere or for broader populations. Deaton and Cartwright (2018) question the transferability of RCT results, stating that social or economic structures may support causal relations in one place and block them in another. In the view of the water-FLFP relationship, the effect in a region with high labor demand is, for instance, expected to differ greatly from the effect in a region without a functioning labor market. Those aiding factors imply that a policy requires an enabling environment. Transferring the gained findings and conclusions to a broader population thus requires assumptions which limit the generalizability of the results and therefore its policy implications. In sum, while randomized experiments and micro-level approaches have greatly enrichened the empirical toolkit of development economics, the results from those studies cannot be easily used to draw general conclusions on the relationship between water infrastructure and women’s economic activity. Therefore, based on Rodrik (2008), who emphasizes complementarities between macro- and micro-development economists, this study aims at highlighting the connection and the larger convergence between the different levels of economic analysis. Given the availability of an extensive dataset of aggregated water-related variables, a macroeconomic approach is expected to provide a broader perspective on the topic and to shed light on the external validity of the issue at hand. I argue that the macro-level study contributes to the understanding of underlying mechanisms and builds the framework for further micro-economic approaches, including RCT, by highlighting overarching dynamics and processes. Based on theoretical guidance from macroeconomic studies, micro-level research and RCT can upscale their validity and contribute to further theoretical understanding (Deaton, 2010).
3. Data and Methodology
The dataset used for the empirical analysis herein is an unbalanced 18-year panel covering the period between 2000 and 2017 for 109 low- and middle-income countries. The data sample was
8 RCT are experiments that apply an intervention or treatment only to a randomly selected group. Assuming that
9
chosen following the suggestion of Peters et al. (2016) that time saving applications are especially pressing in the low- and middle-income context. The categorization of countries into their respective income groupings is taken from the World Bank classification (World Bank, n.d.). Subject to data availability, the sample comprises 24 low-income countries (LIC), 40 lower-middle-income countries (LMIC), and 45 upper-middle-income countries (UMIC). A detailed country list can be found in Appendix I. The analysis employs three main data sources: The Water, Sanitation and Hygiene (commonly known under the acronym WASH) dataset of the Joint Monitoring Program for Water Supply and Sanitation (JMP), which is an initiative of the United Nations Children’s Fund (UNICEF) and the World Health Organization (WHO); labor force participation data from the International Labor Organization (ILO); and data for a set of control variables, mainly from the World Bank’s World Development Indicators (WDI).
3.1 The Dependent Variable: Female Labor Force Participation
The female labor force participation rate is calculated by dividing the number of economically active women by the total female population in the same age group. Economic activity is thereby defined based on a one-hour rule. This means that a woman is counted as economically active if she engages in the respective task for at least one hour within a certain time period, which mostly refers to seven days (ILO, 2017). Labor force participation (LFP) data is taken from the ILO and is mainly based on nationally representative labor force surveys, population censuses, and nationally reported estimates. The definition of a country’s labor force recently underwent changes. Before 2013, it was defined in line with activities covered in the System of National Accounts (SNA) and therefore consisted of people working for pay or profit, unemployed people that are actively searching for a job, and people producing goods for their own consumption. People working in subsistence agriculture thus formed part of the labor force. Own-use services such as food preparation or care work were excluded as they do not lie within the SNA production boundary (UNDP, 1995; OECD, 1995; ILO, 2013; Gaddis & Klasen, 2014). In 2013, a resolution adjusted this definition and excluded own-use production from the labor force.9 This adjustment is largely criticized as it renders a substantial amount of
female and rural labor invisible and depends on subjective and changing production purposes. To date, many labor force surveys are still working under the pre-2013 definition of labor force and countries are assumed to not having yet taken over the changes (Klasen, 2019). Therefore, this analysis assumes that the data reflects labor force participation as defined prior to 2013.10 Nevertheless, this re-definition, once taken over, will cause an inevitable break in the labor force participation data.
Further criticism related to the labor force participation definition concerns the one-hour rule. This rule makes LFP a very broad measure as it places the same weight on a person that works 40 hours per week as on a person that works only one hour. It thus measures the entire employed population but sets a very low labor force entry threshold. Additionally, it bears mentioning
9 See the report and resolution of the 19th International Conference of Labor Statisticians for further information
(ILO, 2013) and Klasen (2019) for a discussion.
10 Testing whether FLFP differs before and after the resolution by adding a dummy variable confirms this
10
that labor force definitions vary across countries. Moreover, family workers and unpaid workers are often omitted in underlying surveys and censuses, and some countries apply a higher hour limit (ILO, 2011). Besides, the reference period of censuses and surveys differs across countries (ILO, 2017). The ILO estimates account for these shortcomings by harmonizing the data and controlling for differences in the scope of coverage, data source, methodology, as well as other country-specific factors and thereby aim at better comparability between countries. I thus opt for the use of ILO estimates instead of national estimates.11 The age group considered in this study are women aged 15 or above, which follows the ILO’s definition of working age (ILO, 2013). In line with ILO recommendations, I opted against the use of an upper age limit “as to permit comprehensive coverage of work activities of the adult population” (ILO, 2017, p. 13). The use of a broad age definition is also reasonable given the lack of age-disaggregated water data (see Section 3.2) and is thus assumed to minimize the resulting bias.
3.2 The Variables of Interest: Water Access and Water Quality
The JMP is often considered the main source of macro-level water data (Ray, 2007). Among others, it contains panel data for drinking water in households for the years between 2000 and 2017. The indicators thereby refer to the main household water source used for domestic purposes such as drinking, cooking, and personal hygiene.12 Data is mainly drawn from censuses, household surveys, and administrative datasets and reported per person, namely as the share of population which uses a specific water technology or service level. As the data collection process is conducted at the household level, weighting of the original data is used (JMP, 2018).13
A shortcoming concerning this dataset is the lack of broad gender- and age-disaggregated water data. In the presence of such data, results would provide more insight (Sorenson et al., 2011). Nevertheless, the JMP data is assumed to illustrate the extent of which women are affected by deficient infrastructure, given that they are largely responsible for the provision of domestic water (Ray, 2007). Gleick (2003) emphasized the need for caution when using the JMP data for cross-country comparisons due to disparities in data accuracy and data collection capacity. To minimize the effect of these disparities, I opt for the use of the JMP indicator W8: Improved water sources which are accessible on premises as a proxy for water access.14 This indicator expresses the share of the population whose main water collection point is an improved water source located within the plot, yard, or dwelling. Restricting the analysis to water on premises solves concerns raised by Sorenson et al. (2011) related to the distance to water sources expressed in minutes. They state that, even if a water source is located within 30 minutes of an individuals’ home, the ease of reaching this source may differ greatly between individuals, and
11 The ILO estimates are modelled, which implies that some observations are estimated using econometric
techniques. A more detailed explanation of the methodology can be found in ILO (2017).
12 Due to the restriction on household purposes, this analysis refrains from including other issues affected by water
infrastructure, such as agriculture practices. The availability of water for irrigation ensures food security and strengthens the ability to cope with climate shocks. This nexus certainly requires attention of future research.
13 For an extensive discussion of the methodology and the compilation of the dataset, see JMP (2018).
14 Water sources are divided into improved and unimproved sources. The explicit differentiations and
11
mention that factors such as road casualties, assault and attack risk, or the condition of the terrain are oftentimes neglected. Gross et al. (2018) add that population pressure at the water collection point increases water collection time and must be taken into account. I assume that the water on premises indicator is less affected by subjective answers in surveys or different measurement methods. By refraining from the use of indicators based on (possibly subjective) time dimensions, the estimates are thus deemed more reliable.15
The second hypothesis requires a water quality indicator, which is provided by the JMP as W10: Improved water sources which are free from contamination. This indicator expresses the share of the population whose main drinking water source provides uncontaminated water, which is defined by chemical and microbiological water quality standards. Water is counted as uncontaminated if a 100ml sample is free from Escherichia coli (E. coli) and if it complies with the thresholds of a maximum of 15µg/L of arsenic and 1.5 mg/L of fluoride (JMP, 2018).16 While the presence of such contamination proves that drinking water is unsafe, the WHO highlights that the absence of such bacteria does not guarantee safety (WHO, 2017). Additionally, water quality is reported to deteriorate between collection and use due to, for example, improper storage (JMP, 2017). Ideally, water quality should therefore be measured at both, the collection and consumption points. Nevertheless, most water samples are taken at the collection point and some samples are even collected at other locations, such as within the distribution network. Additionally, the JMP recognizes that water quality can vary greatly over time, so that regular water quality surveillance would be recommended. However, due to data collection constraints, one-time collection of data is included in the dataset if it covers a representative population sample. This might lead to an underestimation of the proportion of contaminated water as brief contamination events might not be detected whereas they do come along with serious public health implications (WHO, 2017; JMP, 2018). Lastly, not all countries have implemented water quality measurements following international standards so that nationally representative water quality data is lacking for many countries (WHO, 2017; JMP, 2019). Therefore, water quality data is only available for 47 of the 109 countries included in the analysis of this thesis (see Appendix I for a country list). Yet, studies show that water on premises tends to be of better quality, so that the water access indicator might also reflect water quality-related aspects to some extent (JMP, 2017).
3.3 Control Variables
Based on the insights from the literature, a broad range of factors have to be controlled for, including supply side determinants such as health and childcare, as well as demand side characteristics such as labor market opportunities. This section briefly presents the set of control variables used in the analysis. The extensive definition of the control variables, including the data sources, can be found in Appendix III. Firstly, to control for a country’s income level, I include the per capita Gross Domestic Product (GDPpc) as control variable, which is based on
15 Gross et al. (2018) state that asking a single question related to water collection time, which is the practice in
Demographic and Health Surveys, tends to result in an underestimation of the time needed to fetch water.
16 Further information on drinking water quality can be found in WHO (2011). More information on the JMP
12
purchasing power parity (PPP) and measured in 2011 international dollar. Additionally, the share of a country’s rural population (PopRural) is used to control for rural-urban disparities in water infrastructure and water needs, and their impact on women’s economic activity (Cavalcanti & Tavares, 2008).17 To introduce control variables that reflect labor demand effects, the share of manufacturing value added in a country’s GDP (Manufacturing) is added to the model. This variable is expected to mirror structural change-related developments (Gaddis & Klasen, 2014). As FLFP is also subject to fluctuations in a country’s general employment situation, I additionally include the unemployment rate (Unemployment). Turning towards the labor supply side, the employment situation of men is expected to affect women, especially in countries with predominantly traditional family settings (Mammen & Paxson, 2000; Cavalcanti & Tavares, 2008). Therefore, male labor force participation rate (MLFP) is employed as a control variable. To ensure consistency with the FLFP indicator, the age group of 15 years and older was used. Furthermore, childcare and education are expected to determine a women’s economic activity (Jain-Chandra et al., 2018). I therefore include primary education in year t, proxied by net primary enrollment rates, as a control variable (ChildEducation). Given this short-run perspective, the education indicator particularly refers to the education rates of children instead of the education of women or of society overall. The indicator thus represents a childcare aspect and mainly reflects the tradeoff between a lowered care burden and less household support if children are at school (Eckstein & Lifshitz, 2011; Steinberg & Nakane, 2012).18 It can, however, also be assumed that the educational system is better in countries with high primary enrollment rates, which suggests a higher possibility for adult education and potentially higher education of women (Keller, 2006). Additionally, the literature suggested the inclusion of health-related control variables, so that female life expectancy (LifeExpectancy) was added to the model. This indicator proxies the quality of a country’s healthcare system. Besides, the restriction on female life expectancy allows to specifically focus on women’s own health and thus on conclusions how better personal health affects FLFP.19 Lastly, institutions and norms are supposed to affect demand and supply of female labor in a country (Jain-Chandra et al., 2018). To account for this, I add the perception of a country’s institutions (Institution) as computed by the Worldwide Governance Indicators (WGI) as a control variable. Given that this indicator reflects subjective perceptions, it is expected to primarily mirror de facto legal rights and institutional performance. The indicator ranges from 0.5 to 5.5, whereas institutional strength improves with increasingly positive values.20
17 Due to lower population density in rural areas, the costs of (water) infrastructure provision are expected to be
higher (Cavalcanti & Tavares, 2008). Besides, geoclimatic variables such as latitude or precipitation are expected to affect the supply and demand of water infrastructure. These factors are neglected due to data constraints.
18 Differences in daily hours of schooling and impediments such as teacher absenteeism are thereby neglected.
Yet, being essential for social mobility, attention in future studies is recommended (Moses et al., 2017).
19 As female and total life expectancy are logically highly correlated (r = 0.99), the life expectancy coefficient as
well captures general health and therewith the health of other household members.
20 Originally, the six WGI indicators range from approximately -2.5 to 2.5, with zero representing the global mean.
13
3.4 Descriptive Statistics
Having selected the variables used in the model, a first look at the data provides further insight. Table 1 displays summary statistics for the 109 sample countries.
Table 1: Summary Statistics
Variable Obs Mean Std.Dev. Min Max
FLFP 1166 50.82 16.82 6.88 87.68 WaterAccess 1166 62.58 30.12 0.83 99.84 GDPpc 1166 7833.53 5882.10 613.11 40368.08 ChildEducation 1166 86.96 12.80 26.83 99.92 LifeExpectancy 1166 69.38 8.69 45.06 82.56 PopRural 1166 50.73 19.68 8.37 91.75 Manufacturing 1166 12.81 6.58 0.26 50.64 MLFP 1166 74.76 8.63 43.79 91.11 Unemployment 1166 8.31 6.64 0.32 37.25 Institution 1166 2.59 .52 1.10 4.05 WaterQuality 536 66.09 26.44 6.60 99.30
It becomes clear that there is a large variation in the data. FLFP ranges between 6.88 and 87.68 percent, with a mean of 50.82 percent. Turning towards the water access variable, the range is even wider. The share of the population that counts with a main water source on their premises lies between 0.83 percent and 99.84 percent. This points towards striking differences in water access within the dataset. For the water quality sample, the observations are reduced to 536 due to the aforementioned data limitations. The summary statistics for the water quality indicator was added to Table 1. The statistics for the other indicators in the water quality sample can be found in Appendix IV. On average, 66.09 percent of the population receives uncontaminated water from their main water source. Moreover, the higher mean of MLFP compared to FLFP indicates the presence of a gender gap in labor force participation.
Figure 1: Trends in Water Access, Water Quality, and FLFP over time
Own illustrations based on ILO and JMP data. Panel A and C based on 1,166 observations, Panel B based on 536 observations.
14
observation, LICs show the highest share of FLFP and UMICs the lowest. This supports the predictions of the Feminization U hypothesis. Considering the overall levels and the trends, it seems likely that LICs and LMICs are located at the left side of the U-curve, whereas UMICs appear to be located on the right side given their increasing FLFP over time.
Figure 2: Water and FLFP
Own illustrations illustrations based on ILO and JMP data. Panel A based on 1,166 observations, Panel B based on 536 observations.
A first insight into the water-FLFP relationship results from plotting FLFP against the water access and water quality variables (Figure 2). Both panels illustrate a negative correlation between the respective water variable and FLFP. Additionally, the scatterplots reveal the presence of a nonlinear relationship, which suggests the addition of quadratic water variables to the testable model. The correlation matrices (Table 2) confirm the negative correlations between the water-related variables and FLFP.21 This first insight is puzzling, as it contradicts the hypotheses underlying this thesis.
Table 2: Correlation Matrices
(A) The Access Model
Variables (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (1) FLFP 1.000 (2) WaterAccess -0.535 1.000 (3) GDPpc -0.315 0.647 1.000 (4) ChildEducation -0.212 0.594 0.393 1.000 (5) LifeExpectancy -0.460 0.804 0.586 0.623 1.000 (6) PopRural 0.357 -0.605 -0.661 -0.370 -0.579 1.000 (7) Manufacturing -0.342 0.285 0.285 0.220 0.232 -0.268 1.000 (8) MLFP 0.419 -0.195 -0.237 -0.123 -0.152 0.221 -0.008 1.000 (9) Unemployment -0.294 0.228 0.193 0.179 0.063 -0.197 0.044 -0.398 1.000 (10) Institution -0.122 0.401 0.281 0.337 0.301 -0.201 -0.040 -0.189 0.346 1.000 N= 1,166
(B) The Quality Model
Variables (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (1) FLFP 1.000 (2) WaterQuality -0.404 1.000 (3) GDPpc -0.143 0.605 1.000 (4) ChildEducation -0.060 0.402 0.407 1.000 (5) LifeExpectancy -0.302 0.681 0.586 0.684 1.000 (6) PopRural 0.329 -0.654 -0.694 -0.356 -0.586 1.000 (7) Manufacturing -0.288 0.272 0.259 0.187 0.316 -0.364 1.000 (8) MLFP 0.341 -0.453 -0.126 -0.138 -0.128 0.154 0.161 1.000 (9) Unemployment -0.335 0.537 0.167 0.170 0.356 -0.321 -0.116 -0.420 1.000 (10) Institution -0.085 0.315 0.289 0.302 0.397 -0.315 0.121 -0.255 0.237 1.000 N= 536 21 The life expectancy indicator displays a high correlation with the water-related variable, which points towards
15
3.5 Empirical Model
To analyze the water-FLFP nexus, FLFP is the dependent variable. For the first hypothesis, water access is the main variable of interest. The resulting testable model will be denominated Access Model and evaluates the effect of having access to water on premises on FLFP:
(1) Access Model:
𝐹𝐿𝐹𝑃𝑖𝑡 = 𝛽0+ 𝛽1𝑊𝑎𝑡𝑒𝑟𝐴𝑐𝑐𝑒𝑠𝑠𝑖𝑡+ 𝛽2𝑊𝑎𝑡𝑒𝑟𝐴𝑐𝑐𝑒𝑠𝑠𝑖𝑡2 + 𝛽3𝑊𝑎𝑡𝑒𝑟𝐴𝑐𝑐𝑒𝑠𝑠𝑖𝑡∗ 𝐶ℎ𝑖𝑙𝑑𝐸𝑑𝑢𝑐𝑎𝑡𝑖𝑜𝑛𝑖𝑡+ 𝛽𝑛𝑋𝑖𝑡+ 𝜑𝑖 + 𝜇𝑡+ 𝜀𝑖𝑡
𝐹𝐿𝐹𝑃𝑖𝑡 denotes the female labor force participation rate of country 𝑖 in year 𝑡. The direction
and intensity of the effect of water access on FLFP is expressed by the coefficient 𝛽1. In line with the first hypothesis, I expect 𝛽1 to be positive and significant. While 𝛽1 tests for the linear effect of water access on FLFP, Figure 2A suggests that the relationship might be of non-linear nature. This is considered by adding the square of 𝑊𝑎𝑡𝑒𝑟𝐴𝑐𝑐𝑒𝑠𝑠𝑖𝑡 which is reflected in the coefficient 𝛽2. A statistically significant 𝛽2 would imply that the marginal effect of having water on the premises changes at different levels of water access. To assess the third hypothesis, an interaction between 𝑊𝑎𝑡𝑒𝑟𝐴𝑐𝑐𝑒𝑠𝑠𝑖𝑡 and the primary education indicator 𝐶ℎ𝑖𝑙𝑑𝐸𝑑𝑢𝑐𝑎𝑡𝑖𝑜𝑛𝑖𝑡 has been added to the model. The related coefficient, 𝛽3, expresses the joint effect of water access and primary school enrollment on FLFP. A statistically significant 𝛽3 would imply that
the marginal effect of water access on FLFP depends on the level of primary education and vice versa. A positive sign of 𝛽3 would thereby support the third hypothesis that children’s education and the resultant reduced childcare burden serve as enabling conditions, allowing women to transfer their newly-freed time into labor force activities. Additionally, a set of control variables denoted by 𝑋𝑖𝑡 has been added to the model. This vector consists of the control variables
introduced in Section 3.3, namely GDP per capita, primary education, female life expectancy, the share of rural population, manufacturing value added, MLFP, unemployment, and institutions. The addition of the control variables has been evaluated by jointly considering model selection criteria and the number of observations, as illustrated in Appendix V. Joint significance of the independent variables justifies the model specification.22 As a cross-country analysis might suffer from unobserved, country-specific factors that limit or boost FLFP, I control for such factors by adding country fixed effects 𝜑𝑖 to the testable model. This implies the use of a fixed effects model for the estimation, where time-invariant, country-specific factors are assumed to be captured by a country-specific intercept. A Hausman test confirms that the fixed effect approach is preferred over the alternative of a random effect specification.23
Aside from country-specific unobservables, the outcome may underlie variations that happen over time but are not attributed to the explanatory variables. Therefore, year fixed effects 𝜇𝑡 are
added to the model, which control for the influence of (global) year-specific factors influencing FLFP. 𝜇𝑡 captures unobserved changes or common shocks between years that are not specifically attributed to a country, such as global recessions. A test of joint significance of the
22 For an overview of model selection criteria, see Hill et al. (2012), p. 236 – 238.
23 Refer to Hill et al. (2012), Chapter 15, especially 15.5, for further information. As the Hausman test does not
16
year fixed effects confirms that this addition is reasonable. Lastly, 𝜀𝑖𝑡 is the error term. Robust standard errors are used and clustered at the country level to control for possible heteroskedasticity and serial correlation.
To address the second hypothesis that better water quality increases FLFP, the water quality indicator is the main independent variable:
(2) Quality Model:
𝐹𝐿𝐹𝑃𝑖𝑡 = 𝛽0+ 𝛽1𝑊𝑎𝑡𝑒𝑟𝑄𝑢𝑎𝑙𝑖𝑡𝑦𝑖𝑡+ 𝛽2𝑊𝑎𝑡𝑒𝑟𝑄𝑢𝑎𝑙𝑖𝑡𝑦𝑖𝑡2 + 𝛽3𝑊𝑎𝑡𝑒𝑟𝑄𝑢𝑎𝑙𝑖𝑡𝑦𝑖𝑡∗ 𝐿𝑖𝑓𝑒𝐸𝑥𝑝𝑒𝑐𝑡𝑎𝑛𝑐𝑦𝑖𝑡 + 𝛽𝑛𝑋𝑖𝑡+ 𝜑𝑖 + 𝜇𝑡+ 𝜀𝑖𝑡
The model specification is similar to the Access Model. The indicator 𝑊𝑎𝑡𝑒𝑟𝑄𝑢𝑎𝑙𝑖𝑡𝑦𝑖𝑡 reflects the share of the population in country 𝑖 whose main drinking water source in year 𝑡 provides uncontaminated water. In line with hypothesis two, the related coefficient 𝛽1 is expected to be positive. Based on indications for a non-linear water quality-FLFP relationship in Figure 2B, a quadratic water quality term was added, with implications for 𝛽2 as outlined for the Access
Model. Lastly, 𝐿𝑖𝑓𝑒𝐸𝑥𝑝𝑒𝑐𝑡𝑎𝑛𝑐𝑦𝑖𝑡, as a proxy for a country’s healthcare system, is interacted with 𝑊𝑎𝑡𝑒𝑟𝑄𝑢𝑎𝑙𝑖𝑡𝑦𝑖𝑡 to address hypothesis four. A statistically significant 𝛽3 would imply that the marginal effect of water quality on FLFP depends on women’s life expectancy and vice versa. A positive sign of 𝛽3 would thereby support the hypothesis that the positive impact of
water quality on FLFP is reinforced by a good healthcare system and better women’s health. The remaining variables are specified as outlined for the Access Model.
4. Empirical Analysis
The empirical analysis is based on an unbalanced panel with 109 countries for the Access Model and 47 countries for the Quality Model. Wooldridge (2016) argues that unbalanced panels are a common phenomenon in panel data analyses where countries are the clustered entities. He states that if the reason that data is missing for some years is not correlated with the idiosyncratic errors 𝜀𝑖𝑡, the unbalanced nature of the panel dataset is unproblematic. I argue that it is reasonable to assume that the missing data is due to data collection issues or other data constraints but is not related to a country’s share of FLFP in the respective year. Therefore, no concerns are expected to result from the unbalanced nature of the panel. In this section, I first estimate fixed effects regressions for the entire sample and for income group subsamples. To assess causality and account for possible endogeneity, I apply an instrumental variable approach in Section 4.3 and use the results to identify marginal effects and turning points in the water-FLFP relationship. Afterwards, a dynamic panel data estimator is used in Section 4.5 to account for possible persistence in FLFP. Lastly, I test the robustness of the results in Section 4.6.
4.1 Regression Results
Table 3, Column 1 presents the results for the Access Model specification using a fixed effects estimation. The F-test confirms the overall significance of the model and the R2 illustrates that
17
one percent level, this suggests that an increase in the share of people having water on their premises lowers a country’s FLFP, which contradicts the first hypothesis. This result is counterintuitive and will be addressed with more detail later on. The quadratic water access term is also statistically significant, which confirms the presence of a non-linear relationship between water access and FLFP. Its positive sign illustrates diminishing returns of the initially negative effect. With further increasing water access shares, the effect eventually turns positive and the marginal effect is shaped by increasing returns. Therefore, the findings illustrate a U-shaped relationship between water access and FLFP. Additionally, the interaction with children’s education displays a statistically significant effect on FLFP. Even if this effect is rather small, it implies that the effect of having water on premises on FLFP depends on a country’s primary school enrollment rate, where a higher enrollment rate cushions the initially negative effect of water access on FLFP. This is in line with the third hypothesis. If the initial share of water access and the enrollment rate were to be zero, a one percentage point increase in the share of people with water access on their premises would lower a country’s FLFP by 0.367 percentage points. The statistically significant coefficient of the quadratic term and the positive interaction term suggest that the water-FLFP nexus is interconnected with other topics, especially childcare, which will be further discussed in Chapter 5.
Table 3: Fixed Effects Estimations
FLFP Access Model (1) Quality Model (2) WaterAccess -0.367*** (0.109) WaterAccess2 0.002** (0.001) WaterAccess*ChildEducation 0.002** (0.001) WaterQuality 0.377 (0.317) WaterQuality2 0.002 (0.002) WaterQuality*LifeExpectancy -0.010* (0.005) GDPpc 0.0001 (0.0001) 0.0003 (0.0002) ChildEducation -0.037 (0.036) 0.012 (0.051) LifeExpectancy -0.045 (0.106) -0.211 (0.263) PopRural -0.207** (0.097) -0.118 (0.161) Manufacturing -0.012 (0.053) -0.083 (0.077) MLFP 0.484*** (0.099) 0.478*** (0.153) Unemployment 0.021 (0.066) -0.047 (0.083) Institution -0.315 (0.873) -1.136 (1.302) Constant 40.793*** (11.373) 43.197** (19.295) R-squared 0.378 0.281 Observations 1,166 536 F-test 3.837 4.647 Prob>F 0.000 0.000
18
Next, the second hypothesis will be tested. This hypothesis suggests that water quality positively contributes to a country’s FLFP. Table 3, Column 2 provides the related fixed effects estimation. The water quality coefficient is positive; however, it is not statistically different from zero which sheds doubt on its true impact on FLFP. The same applies for the coefficient of the quadratic water quality variable. The interaction term between water quality and life expectancy is significant at the 10 percent level and exercises a negative effect on FLFP. This implies that in countries with higher life expectancy, which indicates a better healthcare system, water quality improvements lower FLFP. Even if this effect is very small, it contradicts the fourth hypothesis. Additionally, no coefficient aside from the interaction term and MLFP is statistically significant. This might indicate underlying endogeneity, which will be elaborated further in Section 4.3. To summarize the above, the fixed effects regressions shed doubt on most of the hypotheses underlying this thesis. The next section thus evaluates whether dividing the sample by income groups provides different insights.
4.2 Income Group Subsamples
As illustrated in Figure 1, the data illustrates different realities concerning water infrastructure between the income groupings. To evaluate whether the restriction to a smaller and potentially more uniform sample changes the results, the Access Model will be re-estimated separately for the three income groupings, namely LICs, LMICs, and UMICs. These results are displayed in Table 4. Due to the decreasing number of observations in subsample estimations, time fixed effects cannot be included as this would result in an insufficient rank for the use of robust standard errors and no F-test could be conducted.24 The exclusion of year fixed effects could
introduce a bias in the estimates given that unobserved, country-invariant shocks are not controlled for. This sheds doubt on the causal implications of the results, as a panel estimation that does not control for time fixed effects is likely picking up the influence of aggregate trends. Therefore, the results of the subsample estimations must be treated with care.
Table 4: Access Model Estimations of the Income Subsamples
(1) (2) (3)
LIC LMIC UMIC
WaterAccess -0.195 -0.525** -0.071 (0.193) (0.210) (0.486) WaterAccess2 0.006*** 0.001 0.004* (0.002) (0.001) (0.002) WaterAccess*ChildEducation -0.003 0.004* -0.006 (0.002) (0.002) (0.004) R-squared 0.534 0.3681 0.4938 Observations 202 427 537
Notes: Robust standard errors in parentheses; *p<0.1, **p<0.05, ***p<0.01. Fixed effects regression. Control variables used: GDPpc, ChildEducation, PopRural, MLFP, LifeExpectancy, Manufacturing, Unemployment,
Institution. Extensive results displayed in Appendix VI. Year fixed effects not included due to data limitations.
24 The F-test for clustered data is distributed as F(k,d-k+1). This implies that the number of independent variables
19
The water access coefficients display a negative sign in all subsample regressions; however, they only show statistical significance in the estimation of LMICs. For LMICs, the interaction term of water access and primary education is also statistically significant and shows a positive but small coefficient. In the subsample of LICs and UMICs, the quadratic water access term is statistically significant with a positive sign, while the other water-related coefficients show no statistical significance. Overall, the results are in line with previous results but show a weaker statistical significance. No striking differences between the income groupings were identified. Due to the data limitations, subsample regressions are not feasible for the Quality Model. Therefore, an income group dummy variable is introduced and interacted with the water quality indicator to evaluate how the effect of water quality on FLFP changes across different income settings. The dummy takes the value of zero for LICs. Given that data limitations do not allow for the estimation of subsamples for the Quality Model, the dummy approach provides a feasible alternative. Nevertheless, estimating subsamples is preferred as it also restricts the other variables to the respective sample. The results are displayed in Table 5. Concerning the direction of the water quality-FLFP nexus, they provide a similar insight as the entire sample. The effect of water quality on FLFP is positive in all regions, which is in line with the second hypothesis. No evidence for a non-linear water quality-FLFP relationship was identified and the interaction with the life expectancy indicator as well produced statistically insignificant results. Nevertheless, the results show that the strength of the effect differs between regions.
Table 5: Quality Model Estimations with Income Dummies
FLFP Coef. Robust Std.Err. Sig
WaterQuality 0.822 0.376 **
WaterQuality2 0.001 0.002
WaterQuality*LifeExpectancy -0.003 0.005
LMIC*WaterQuality -0.756 0.270 ***
UMIC*WaterQuality -0.482 0.271 *
Notes: *** p<0.01, ** p<0.05, * p<0.1; Observations = 536; R-squared = 0.356; Income dummy variable takes value zero for LIC. Fixed effects estimation. Year fixed effects included but not displayed. Control variables used: GDPpc, ChildEducation, PopRural, MLFP, LifeExpectancy,
Manufacturing, Unemployment, Institution. Extensive results displayed in Appendix VII.
20
4.3 The Endogeneity Problem and Solution
The testable model of this thesis likely suffers from endogeneity concerns, which render the identification of a truly causal effect problematic.25 This section outlines possible sources of
this endogeneity and makes use of an instrumental variable approach to assess these concerns. The first possibly endogenous variables are the water-related ones. In line with Koolwal and van de Walle (2013), I argue that the placement of water infrastructure could be endogenous. More economically active women could be able to finance easier water access or better water treatment facilities, for example a connection of their property to a general water network or a filter for their water source. Additionally, regions with low FLFP might be preferred by water provision incentives (Ilahi, 2001). Further reasons for the endogenous placement of water infrastructure could be related to the aim of having a say in decision making, a frequently cited consequence of women’s economic activity (Klasen et al., 2019). Economically active women might be more concerned to reduce their water collection time burden and the health implications resulting from contaminated water, and therefore may lobby local governments to provide water infrastructure or push to construct such infrastructure on their own. Besides, more economically active women might consciously locate closer to water sources due to a higher time saving preference. Therefore, work decisions might be simultaneously determined with having water on the premises or the quality of the main water source. This would suggest that more economically active women are more likely to have water on their premises or to consume higher quality water. Consequently, the WaterAccess and WaterQuality coefficients might be biased. Additionally, the literature review in chapter two suggested that the relationship between FLFP and GDP could suffer from endogeneity. In particular, the possible occurrence of reverse causality must be taken into account. In the underlying models, we assumed causality from GDP to FLFP and not vice-versa. However, having highlighted the growth implications of FLFP, it became clear that a country’s GDP could also benefit from having a large female workforce.26
Facing endogeneity, the OLS estimates are inconsistent.27 In econometric estimations, an
instrumental variable (IV) approach is commonly used to deal with such endogeneity concerns. Finding relevant and exogenous instruments can, however, prove difficult. Endogeneity concerns related to the water-FLFP nexus feature prominently in the literature. However, given that most studies utilize micro-level data, they mainly use geographical variables as instruments, such as municipal population density (Grogan & Sadanand, 2013) or annual municipal rainfall (Nauges, 2017). Due to large within-country heterogeneity and data constraints, such IVs are not suitable for estimations at the macro-level. A possible solution is
25 For the causes and implications of endogeneity, please refer to Hill et al. (2012), Chapter 10.
26 It could also be argued that reverse causality arises from other variables such as MLFP. Men’s decisions on
whether or not to be economically active could depend on women’s economic activity. This thought will, however, be neglected.
27 At the household and individual level, endogeneity in FLFP regressions could result from unobservable,
21
to use the lag of the endogenous variable as an instrument (Nauges, 2017; Bun & Harrison, 2019). This is economically feasible as water access, water quality, and GDP per capita in year t are expected to depend to some extent on the values in the previous year, t-1. I therefore make use of a Two-Stage Least Squares (2SLS) estimator applying the one-year lag of the endogenous variables as instruments. Given that the access and quality models include quadratic terms and interactions of the endogenous variables, the respective coefficients must also be treated as endogenous. This bears the danger of the so-called forbidden regression.28 To avoid this, I follow Wooldridge (2002)’s recommendation to use the quadratic term of the fitted values from the first stage regression as an instrument for the quadratic endogenous variable. Simultaneously, the interaction of the fitted values with the primary education and life expectancy variables, respectively, are used as instruments for the interaction term.Regressing the residuals against the IVs and the exogenous variables confirms the exogeneity of the instruments.29 The first stage regressions show that the instruments are valid (Appendix VIII). Standardized tests displayed by the statistical software confirm the validity and significance of the instruments and reject the possibility of weak instruments (Stock & Yogo, 2005). The model is found to be identified and can thus be consistently estimated (Hill et al., 2012).
The Access Model is estimated in a 2SLS procedure with country and year fixed effects. Standard errors are robust to heteroskedasticity. The results are reported in Table 6, Column 1. The postestimation test for endogeneity confirms the concerns and makes the instrumental variable approach the preferred and consistent specification for the Access Model.30 The results confirm the insight gained from the fixed effects estimation and thus still contradict the first hypothesis to some extent. The previously identified negative linear effect of water access on FLFP is confirmed. The positive and significant coefficient of the quadratic water access variable shows that this negative effect decreases with increases in the water access share. Even if the coefficient of the quadratic variable is small, it has the potential to turn the direction of the water access-FLFP relationship. At sufficiently high water access shares, the relationship turns positive, so that the previously identified U-shape is confirmed. Besides, the positive and significant coefficient of the interaction term confirms the previous results. This supports the third hypothesis that water access jointly with primary education exercises a positive effect on FLFP. The marginal effect of water access on FLFP therefore depends on the level of water access as well as the educational enrollment. Section 4.4 will provide further insight into the magnitude of the water access-FLFP relationship.
28 Wooldridge (2002) describes this as the attempt to mimic a 2SLS approach by replacing the non-linear function
of an endogenous variable with the respective non-linear function of the fitted values from a first stage estimation. For a discussion please refer to Wooldridge (2002), especially Chapter 9.5.2.
29 Given that the model is exactly identified, meaning that the number of endogenous variables equals the number
of instruments, an overidentification test is not possible and this approach is thus a feasible alternative.
30 For a discussion of consistent specifications see Hausman and Taylor (1981), or the related summary by Hayashi
