Old Age Pensions and Fertility:
A
test of old age security hypothesis
in South Africa
Master Thesis
Supervised by dr. Pauline Rossi
Z. Ece Kafali
11734140
University of Amsterdam
15 August 2018
Abstract
Do the introduction of formal old age insurance reduces fertility as the old age security hypothesis predicts? This thesis uses women’s reproductive histories in Demographic Health Surveys to analyse the impact of the South African old age pension program on black women’s fertility during the period of 1993-1998. South Africa achieved racial equality in pension coverage and benefit levels in 1993, while Lesotho had not introduced an old age pension scheme until the early 2000s. Using a duration model of birth intervals, I implement a differences-in-differences framework where the change in birth spacing of South African mothers before and after 1993 is compared to that of Basotho mothers. Results do not indicate a significant impact of the old age pension program on fertility. Magnitudes of coefficients suggest two mechanisms through which the pension program affected fertility: a negative impact through the mechanism of old age security hypothesis, particularly in urban areas, and a positive impact through an income mechanism among women living with age-qualified household members in rural areas.
Statement of Originality
This document is written by Student Zeynep Ece Kafali who declares to take full responsibility for the contents of this document.
I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.
The faculty of Economics and Business is responsible solely for the supervision of completion the work, not for the contents.
Contents
1 Introduction ... 3
2 Literature Review ... 5
3 South African Old Age Pension Program ... 8
4 Data ... 9
5 Methodology ... 10
5.1
Graphical analysis ... 11
5.2
Econometric Specification... 14
5.2.1 Differences-in-Differences and a duration model of birth intervals ... 14
5.2.2 Proportional hazard model ... 16
6 Results ... 17
7 Discussion ... 19
7.1
Income Mechanism ... 19
7.2
Limitations ... 22
8 Conclusion ... 24
References ... 25
1 Introduction
It is widely accepted that the processes of economic and social evolution are accompanied by a reduction in fertility, commonly referred to as the demographic transition. One of the driving forces of this transition has been hypothesized to be the establishment of formal old age pension programs. Based on the so-called old age security hypothesis, the interaction between the old age pensions and decline in fertility is explained as follows: in the absence of well-functioning capital markets and social safety nets, children are perceived as providers of old age insurance through intergenerational transfers or different types of assistance (e.g. cohabitation). The expectation of being taken care by children during retirement creates an important motive for childbearing. Accordingly, reducing uncertainty about old age income is in return predicted to reduce the desired family size, contributing to the demographic transition.
South Africa is one of the few countries in Sub-Saharan Africa that has experienced a sharp decline in fertility over the last 50 years1. This thesis seeks to test whether the expansion
of old age pension program to the black2 population in the early 1990s has played a role in
South African fertility transition. Near the end of the apartheid system, South Africa committed to achieve racial equality in social pension coverage and benefits. Lesotho, an enclaved country within South Africa, established an old age pension scheme in 20043. I exploit
similar fertility trends among black South African and Basotho women and treat it as a natural experiment. I implement a differences-in-differences approach to evaluate the impact of the old age pension program on fertility during the ‘post-pension’ period of 1993-1998. I infer the validity of old-age security hypothesis in the case that black South African mothers lengthened their birth intervals relative to Basotho mothers. The underlying conceptual framework of this strategy is that, in a given reproductive lifespan, intervals between successive births of a mother are negatively related to the eventual number of births she will give. Demographic literature recognizes the birth spacing behaviour as a conscious fertility limitation technique
1 South African fertility dropped from 6.04 in 1960 to 2.45 children per woman in 2016 while the same
rate for Sub-Saharan Africa moved from 6.64 to 4.85 (World Development Indicators).
2 The survey data used in this thesis classifies South African population into four different racial groups:
black/African, coloured, white, Asian/Indian
among transitional populations. There is evidence of an adverse relationship between the length of birth intervals completed family size (Anderton and Bean, 1985; Casterline and Odden, 2016). In the econometric specification, I use semi parametric estimations of a duration model of birth intervals4. Duration analysis suits the quantification of the dependent variable
(number of months between consecutive births) and the time-to-event nature of the dataset. Moreover, I discuss an income mechanism that could have affected fertility. I investigate such mechanism by comparing the birth spacing of women living in extended families with and without age-qualified household members.
Investigating the interaction between social security programs and fertility in developing countries, particularly in Africa, has important implications. Poor regions of underdeveloped countries have the fastest population growth in the world (Haub and Gribble, 2011). United Nations projects the world population to grow from 7.6 billion today to 9.8 billion in 2050, while more than half of this growth is expected to take place in Africa. High fertility brings about risks in child and maternal health, lower human capital and shortage of natural resources (Chowdhury, 2010). Traditions, lack of contraceptive knowledge and child mortality are generally attributed to be the driving forces of high fertility in poor areas. Nevertheless, if desired family size was actually related to some private incentives, such as old age security motive, then spreading family planning programs and access to child health services alone would not be effective in achieving the necessary reduction in fertility. Hence, studying this link is vital to understanding demographic effects of social policies as well as determining necessary actions to target high fertility.
The rest of this paper is structured as follows. Section 2 presents a literature review. Section 3 provides a brief history of South African old age pension system. Section 4 describes the data. Section 5 introduces a graphical analysis of fertility and the econometric specification used for the analysis. Section 6 reports the results. Section 7 first investigates an income mechanism, then discusses the limitations of the empirical design. Section 8 concludes.
4
Duration model of birth intervals is commonly used in different analyses of fertility. Some examples
are: Rossi and Rouanet (2015) testing gender preference in Africa, Ghilagaber and Gyimah (2004) evaluating the impact of socio-economic and cultural factors on birth intervals, Heckman and Walker (1990) examining the relationship between male income/female wages and the timing of birth.
2 Literature Review
A large section of theoretical work emphasizes that the need for economic support in old age and the magnitude of inter-generational transfers from children to parents results in higher demand for children (e.g., Leibenstein, 1975; Caldwell, 1978; Willis, 1980). For instance, a model of fertility choice, developed by Boldrin and Jones (2002), predicts that more children will result in more support for parents if children cooperate in providing old age support. An important implication of their model is that modifications in social and economic conditions, such as improving access to financial markets as well as social security pensions, will decrease fertility.
A number of cross-section analyses support the theory by providing a negative correlation between the Social Security system, in particular government-provided pensions, and fertility. Boldrin et al. (2015) use data from 104 countries in 1997, with the majority of their sample consisting of developed countries. Investigating the link between the expenditures on the Social Security system as a fraction of GDP and the total fertility rate, the authors control for per capita GDP, infant mortality rate and the share of population aged 65 and older. Their results reveal that a 10% increase in the size of the Social Security is associated with a 0.7-1.6 child decrease in the total fertility rate. Moreover, they calculate that roughly 50% of the decrease in fertility between 1950 and 2000 in U.S. and European countries is explained by the expansion of the pension systems. Nevertheless, it is very likely that there are omitted factors (e.g., human capital) affecting both the explanatory variables and the fertility rate.Nugent and Gillaspy (1983) study this link in a developing country. They use a cross-section of 34 counties in Mexico and analyse the changes in child-woman ratios between 1960 and 1970. Utilizing the extension of social security to sugarcane sector in 1963, they use sugar cropping as a proxy for the influence of participation in the old age pension system. They conclude that being covered by the old age pensions reduces the child-women ratio by more than one-tenth, however, one could also suspect a link between the fertility preferences and occupation.
A part of the literature emphasizes that the sons are perceived as the main providers of old age insurance, resulting in sex preference. Ebenstein and Leung (2010) investigate whether the old age pension motive was one of the reasons behind the strong son preference in China after the One Child Policy. In 1991, China initiated an old age pension scheme at chosen villages where people between 20-60 years old could choose to contribute premiums. Benefiting from a household survey conducted in 2002, the authors exploit information about program availability and participation at village level, and the number and sex of participant’s children. In different specifications, they regress a dummy variable about the participation, on failure to have a son and the number of children separately for daughters and sons. Results indicate that the families without sons and those with fewer children were more likely to participate in the pension program. Again, these results do not prevail causality because people who do not have a son preference or prefer having fewer children could also have a more modern view toward the formal pension scheme, and villages that have enforced the One Child Policy strictly could also have been more likely to implement the program
.
Even though studies presented above are prone to endogeneity issues, they are highly suggestive of the predictions of the old age security hypothesis.To the best of my knowledge, there is no empirical study that assesses the causal impact of old age pension program on fertility. Using quasi-experiments, some studies provide evidence regarding the validity of insurance motive for fertility and also that more children result in more monetary transfer and assistance to parents. Banerjee et al. (2014) show that parents perceive their children as old age insurance by investigating a reverse relationship: how do fertility restrictions affect parent’s saving decisions? The authors use China’s One Child Policy as an instrument for fertility and the urban sample was collected by them in 2008 in the form of a household survey. They find that an exogenous decrease in fertility results in higher savings. Lambert and Rossi (2016) investigate whether sons are perceived as widowhood insurance for inheritance by Senegalese women, demonstrating that widowhood insurance motive indeed affects women’s fertility choices. They use the information on completed birth history in a household survey conducted in 2006-2007. Using a duration model of birth intervals, they find that women at most risk of not having access to inheritance (whose
husband has children from ex-wives) through a son intensify their fertility until they have one, compared to women at less risk. Oliveira (2016) investigates another relevant question: do children actually take care of their parents when parents’ productivity declines? She exploits the incidence of first-born twins as an instrument for number of children and uses a survey of elderly individuals conducted in 2011 in China. She finds that having more children is associated with higher financial transfers from children to parents and higher likelihood of co-residing with an adult child. Studies presented above show that the old age motive for fertility is legitimate and a decrease in fertility as a result of secured future income is very likely to be practical.
Despite the fact that old age pensions essentially target elderly population, benefits could reach younger individuals as well, particularly in developing countries where extended families and sharing of income is fairly common. Moreover, if the benefit level was an important amount, one could expect an income effect on fertility as the cost of a child to the household would decrease significantly. The link between income and fertility generated an extensive debate. Initially, theoretical literature viewed children as normal goods and argued that a decrease in the cost of a child will have a positive impact on fertility and some empirical studies provided evidence supporting the theory. However, many others have found an opposite or ambiguous link, generating further theoretical considerations. A study focusing on the contemporary developing world, Chatterjee and Vogl (2018), makes a distinction between long-term and short-long-term impacts of income growth on fertility. They argue that during fluctuations fertility is procyclical, contrary to long-term growth. They explain this difference by the lack of alteration in returns (e.g. child investment), prices (e.g. women’s wages) and the presence of intertemporal substitution effects in the short-term. Some researchers are interested in examining this link in the context of government policies implicitly or explicitly influencing the fertility behaviour. Whittington et al. (1990) investigates the demographic consequences of personal tax exemption for dependents in United States for the period of 1913-1984. The exemption policy was essentially conducted to relief the tax burden on the low-income households. The authors conclude that increases in the tax value of the exemption had a significant and positive impact on aggregate family birth decisions. Stecklov et al. (2006)
investigate the demographic externalities of poverty programs. They find that the Family Assistance Program in Honduras, which was essentially implemented to increase human capital by providing conditional cash transfers, raised the birth rate by between 2-4 percentage points in 2002. Though, it is argued that this result was due to a flaw in the program design, incentivising marriage. Overall, the impact of higher income on fertility has not yet brought to a conclusion. It is also of interest to investigate a possible income effect of the old age pension program on fertility, because it may abate as well as intensify the predicted negative impact by the old age security hypothesis.
3 South African Old Age Pension Program
The South African government established the social pension program in 1928 for relatively poor white workers that were not covered by the existing occupational pension scheme. In 1944, the pre-apartheid state extended the coverage to other ethnicities; Blacks, Coloureds and Indians (Burns et al., 2005). Even then, benefits were determined by race and were highly discriminatory, particularly against the black population. In practice, compared to rural Africans, payments in 1944 were 12, 6 and 5 times higher for Whites, Coloureds and Indians respectively (Devereux, 2001). During the apartheid era, discriminatory practices widened. Even though there are no reliable estimates of the coverage, it is mentioned that the differential delivery systems and intentional discriminative practices left the majority of legally eligible black South Africans uncovered. Furthermore, those who received the pensions among Africans were entitled to an extremely lower amount of benefits and a much stricter means-test compared to other ethnicities (Duflo, 2003).
By the end of the 1980s, the anti-apartheid movement gained ground and bilateral negotiations for ending segregation started. In 1989, the government promised to achieve racial equality in the pension program and expansion efforts started two years later (Duflo, 2003). By the beginning of 1993 the promise was achieved, and social pensions were effectively operating countrywide including remote rural areas (Case and Deaton, 1998). The pension system is universal, non-contributory and subject to a means-test. It covers women over the
age 60 and men over the age 65, with monthly “means” equal or less than R370 (in 1994) (Case and Deaton, 1998). In 1993, benefits for Africans had almost doubled relative to 1980 and reached a level that was roughly twice the median per capita income of African households. In the same year, 80% of the age-qualified African women and 77% of the age-qualified African men reported having received the pension, with vast majority receiving the maximum amount (Case and Deaton, 1998). Take-up rates were significantly higher for Africans compared to other ethnicities, indicating the efficiency of the means test.
As mentioned earlier, the expansion efforts started in 1991, however, it is not possible to estimate pension coverage before 1993, mainly because earlier surveys excluded the homelands where most of the African population lived. Moreover, there is no information about an earlier announcement of the expansion efforts to the African population. Thus, in my analysis I consider January 1993 as the turning point of the pension program for the African citizens. It is conceivable to assume that the program was widely known at that time and that reproductive people were expecting to receive such pensions during retirement, in contrast with the previous years. Moreover, there is no information about the existence of a social pension scheme earlier than 2004 in Lesotho, leaving Basotho women “untreated” by the old age pensions during the period analysed in this thesis.
4 Data
The analysis is based on Demographic and Health Surveys (DHS) collected in South Africa in 1998 and in Lesotho in 2004. DHS interviews 15 to 49-year-old women and provide nationally representative data on health, population and nutrition in developing countries. Most importantly, it provides a detailed reproductive history of interviewed mothers. The survey contains separate data files for different units of analysis. More specifically, individual women’s data, which include characteristics and reproductive history of all interviewed women, are used in the computation of age-specific fertility rates. The births dataset includes the same information as individual women’s dataset but the unit of analysis in this data is the children born to eligible women, hence it excludes women with no reproductive history. I use this
dataset in the duration analysis. It allows me to observe the year and month of birth of all children and to measure the intervals between consecutive births in months. It also allows me to calculate the mothers’ age at each birth. The successive birth intervals and the mothers’ age at birth are the main components of my duration analysis. I also use the household dataset to identify the mothers that live with or without an age-qualified household member.
Surveys are similar across South Africa and Lesotho, both containing a large number of observations, however, the South African DHS was conducted 6 years earlier than that of Lesotho. This requires restricting the sample to the women whose birth years overlap in both surveys. As a result, women’s year of birth range between 1955, the earliest birth year in Lesotho DHS, and 1983, the latest birth year in South Africa DHS. Furthermore, I subtract 80 months from the mean century month code (cmc) of the date of interview in Lesotho DHS. This allows me to equalize the mean interview dates in both surveys, both corresponding to March 1998. Then, I eliminate all children that were born after the new interview date in Lesotho. I treat the youngest children that remain in the dataset, for whom the successive birth interval is unknown, as a right-censored observation. I do this in order to consider the fertility during similar periods in both countries. Finally, since my analysis is based on the fertility of black South African women, I exclude all other ethnicities from the South African survey.
There are three waves of DHS conducted in South Africa until today: DHS 1998, DHS 2003 and DHS 2016. Ideally, three surveys could be utilized as a repeated cross-sectional data in order to carry out a more comprehensive analysis of the fertility trend in South Africa after the pension expansions. This would permit observing the fertility behaviour of a larger amount of younger generations that were treated with the old-age pensions since the beginning of their reproductive lifespan. Nevertheless, DHS 2016 is on-going, and DHS 2003 is not publicly available and, despite contacting various institutions I could not gain access to the latter.
This section consists of two parts: first part introduces a graphical analysis of fertility over cohorts; second part presents the econometric specification used to estimate the impact of old age pensions on fertility.
5.1 Graphical analysis
This part presents a graphical analysis of the age specific fertility-rates (ASFRs). It is the basis of the common trends assumption that will be introduced in the following part. This analysis is conducted in order to create a time dimension in fertility using only one cross-sectional survey. Women are classified into five-year cohorts based on their year of birth5. Then, fertility
rates are computed by five-year age groups separately for each cohort6. The analysis takes into
account every observed birth in the DHS7.
Let W(c, w) be a set of women who are in birth cohort c and let B(c, a, b) be a set of women in W(c, w) who have given birth while they were in age group a. w and b denote the total number of women in each set, where w ≥ b. Each woman i in B(c, a, b) has given k'(,*
number of births in age group a. Thus, the total number of births of B(c, a, b) in age-group a is:
∑ kic,a
b
i=1
Every woman in W(c, w) has spent different periods of time in age group a. Let tic,a denote the number of years the woman i spent in age group a at the time of the interview. Given ti
c,a∈ (0, 5], the total number of person-years spent by W(c, w) in age group a is:
∑ tic,a
w
i=1
I compute separately for every cohort c, the yearly number of births in each applicable age group a:
5 Five-year cohorts are: 1955-57, 1958-62, 1963-67, 1968-72, 1973-77 and 1978-82. The oldest cohort
consists of 3 birth years instead of 5. This is because the sample is restricted to the earliest birth year of 1955 in Lesotho.
6 Five-year age groups are: 15-19, 20-24, 25-29, 30-34, 35-39 and 40-45.
7 As mentioned earlier, this analysis takes into account only the children born before the new survey date in Lesotho.
ASFR(c, a) =∑ ki
c,a b i=1
∑wi=1tic,a
Figure 1 presents how ASFRs for each age group evolve from oldest to the youngest cohort in South Africa and in Lesotho. Each line represents a different age group specified by a colour. An example of the interpretation of the graph is as follows: The South African cohort of 1968-72 had 0.1 child per year when they were between 15-19 years old, the same cohort had roughly 0.16 child per year between the ages 20-24, and their fertility rate decreased to about 0.14 child per year in the age group 25-29. An ASFR could be calculated if a given cohort had been exposed to that age group at the time of the survey. For example, ASFR for 15-19 age group could be calculated for all cohorts because all observations were 15 or older in 1998, whereas ASFR for 40-44 age group could only be calculated for the 1955-57 birth cohort because it is the only cohort that had been exposed to the 40-44 age group at the time of the survey.
This representation enables to observe the trend of ASFRs over time, owing to the fact that each cohort was in a given 5-year age group during different periods of time. For example, the 1978-82 birth cohort of women were between 15 and 19 years old during the period 1993-988. Consequently, the last point on 15-19 ASFR line includes only the children that were born
between these years. The oldest cohort of 1955-57 was 15-19 years old during the period 1970-76. Thus, the blue line represents how the yearly number of births in age group 15-19 evolved between the years of 1970-98. The last point on each line strictly corresponds to the post-pension period of 1993-1998. The second last point includes births in both pre- and post- pension periods (1988-1996), and the remaining points correspond to the pre-pension period (1970-1991).
Figure 1 shows a decreasing trend in ASFRs over cohorts in both countries. Lesotho’s ASFRs are higher than that of South Africa. A clear difference in fertility trends between the two countries appears in the orange and grey lines: South African cohorts of 1955-57 and 1958-62 had almost the same number of children per year in age groups 20-24 and 25-29, while
8 This cohort was actually in the 15-19 age group between 1993-2001 but I can only observe the births
Basotho cohort of 1958-62 had considerably less children per year in the same age groups compared to the older cohort. Otherwise, changes in ASFRs between other cohorts are more or less similar in the pre-pension period, however, it is not clear how the South African fertility differed from that of Lesotho following the pension program.
Figure 1: Age-specific fertility rates by 5-year age groups
0 0,05 0,1 0,15 0,2 0,25 0,3 1955-57 1958-62 1963-67 1968-72 1973-77 1978-82 A ge S pe ci fic F er ti lt iy R at e Cohort
South Africa
15-19 20-24 25-29 30-34 35-39 40-44 0 0,05 0,1 0,15 0,2 0,25 0,3 1955-57 1958-62 1963-67 1968-72 1973-77 1978-82 Ag e Sp ec ifi c Fe rt ilt iy R at e CohortLesotho
15-19 20-24 25-29 30-34 35-39 40-44Source: Own calculations using South Africa DHS 1998, Lesotho DHS 2004. South African sample is restricted to black women. Number of births and
5.2 Econometric Specification
5.2.1
Differences-in-Differences and a duration model of birth intervals
As described in the introduction, the underlying reasoning of my empirical strategy is that the number of births a woman gives in her fertile lifespan increases (decreases) as she shortens (lengthens) the duration between her births. The pension program is predicted to reduce fertility; consequently, I expect South African women to have lengthened their birth intervals in the post-pension period.
To evaluate the causal impact of the pension program on birth spacing of South African mothers, I use a differences-in-differences (DiD) framework. DiD assesses the impact of the intervention by comparing the treated group to the control group before and after the intervention. In this specific analysis, I compare the pre- and post-pension change in the birth intervals of South African women relative to Basotho women. The dependent variable is thus the intervals between successive births quantified in months. This framework relies on the following identifying assumption: in the absence of old-age pensions black South African women’s fertility would have followed the same trend as Basotho women (Figure 1)9. Under
this assumption, the coefficient of interest, or the DiD term, is denoted as following: 𝐷𝑖𝐷 = (𝔼[T|SA = 1, Post = 1] − 𝔼[T|SA = 1, Post = 0])
− (𝔼[T|SA = 0, Post = 1] − 𝔼[T|SA = 0, Post = 0])
The unit of observation is the children of interviewed mothers. T denotes duration between births 𝑛 and (𝑛 + 1) measured in months of a mother, where 𝑛 ≥ 1; SA equals 1 when
9 I acknowledge that the pre- and post-pension trends are not perfectly clear on Figure 1 and trends
are not perfectly parallel in the pre-pension period. I discuss the issues of the empirical design in the limitations part.
person-years are weighted by the sample weights. Calculations are conducted using the Stata module tfr2 by Schoumaker (2013).
the mother is from South Africa and 0 when she is from Lesotho; Post equals 1 if the child was born in the post-pension period and 0 otherwise.
In order to assess the change in the birth spacing between pre- and post-pension periods, I benefit from the panel structure of DHS. The survey allows me to observe every birth of mothers until the survey date, some occurring before 1993 and some from January 1993 onwards. I compare the birth spacing behaviour of “young” mothers in the post-pension period to that of the “old” ones in the pre-pension period. More specifically, imagine two mothers who have given birth to their 𝑛𝑡ℎ child at the same age in the pre-pension period,
where 𝑛 ≥ 1. This implies that in the absence of the pension program, the two mothers had similar fertility preferences in terms of timing. Moreover, both mothers had at least one more birth thereafter. The older mother’s subsequent birth interval(s) occurred in the pre-pension period whereas those of the younger mother occurred in the post-pension period. The older mother’s birth interval following her 𝑛𝑡ℎ birth serves as a counterfactual for that of the
younger mother. This is the key assumption in the assessment of the pre and post-pension difference.
The DiD framework is implemented by using a duration model of birth intervals. In duration analysis
(
also known as survival analysis) there is a determined event, the so-called failure, occurring after a period of time, the so-called duration or survival time. It is used to analyse the expected duration until the failure occurs. In fertility context, failure is determined as childbearing. This analysis is more appropriate regarding the quantification of the dependent variable compared to conventional linear models (e.g. OLS). Most importantly it handles the following issue: DHS does not provide a complete fertility history of the mothers, especially of the young ones, because the mothers might have more children after the interview date. Consequently, the duration between the latest birth in the survey (right-censored observation) and the potential successive birth of mothers is unknown. Nevertheless, it is known that the potential birth interval lasted at least for the length of time between the date of the latest observed birth and the interview date. Duration analysis gives the advantage to identify right-censored observations and exploit the information on the right-censored durations. An assumption made in this analysis is that the procedure causing the censoring isindependent of the duration. As the interview date is not related to the birth intervals of a mother, this assumption is satisfied.
5.2.2 Proportional hazard model
To estimate the impact of old age pensions on birth intervals, I use a semi parametric Cox proportional hazard model (Cox, 1972). As mentioned before, my outcome variable T is the duration between successive births measured in months. In duration analysis there are two key concepts: hazard (or failure) function and survival function. Hazard function summarizes the instantaneous intensity of transition from duration to the failure (Jenkins, 2005). In this analysis, it gives the instantaneous probability to give another birth at time 𝑡 and takes the following form:
𝜆(𝑡) = 𝜆𝑜(𝑡) × exp (𝑋𝑛′β)
𝜆𝑜(𝑡) is the baseline hazard function that is assumed to be common to all individuals.
𝑋𝑛′ is a vector of covariates associated with the mother at each birth. Baseline hazard function
describes how the hazard of birth per time unit varies over time when all covariates take the value zero. An advantage of Cox model is that it does not require making any assumption about the form of the baseline hazard function, hence it is unspecified. exp (𝑋𝑛′β) is a
person-specific function of covariates that summarizes the effects of personal covariates on the baseline hazard function. Then, survival function that gives the probability of not giving birth until t is modelled as following:
𝑆(𝑡) = Pr(𝑇 > 𝑡) = exp (− ∫ 𝜆𝑜(𝑡)du × exp(𝑋𝑛′β) 𝑡
0
)
In my analysis, 𝑋𝑛 includes 𝑷𝒐𝒔𝒕, a dummy that takes the value 1 if the birth occurred
in the post-pension period and 0 if it occurred in the pre-pension period; 𝑷𝒐𝒔𝒕 × 𝑺𝑨, an interaction term between Post and a dummy 𝑆𝐴 that equals 1 if the mother is from South Africa and to 0 if the mother is from Lesotho; 𝑴𝒐𝒕𝒉𝒆𝒓_𝑨𝒈𝒆, the age of mother at birth and
𝑩𝒊𝒓𝒕𝒉_𝑶𝒓𝒅𝒆𝒓, the order number a child is born to the mother. Standard errors are clustered at mother level to allow for correlation between different birth intervals of a mother. I include mother fixed effects assuming that the covariates associated with mothers influencing fertility remain constant over time.
In proportional hazard models, the estimated coefficients are hazard ratios. Two individuals that have the covariate vectors 𝑋𝑖 and 𝑋𝑗 will have the hazard functions as
followed: ℎ(𝑡|𝑋𝑖) ℎ(𝑡∣𝑋𝑗) =𝜆𝑜(𝑡) × exp(𝑋𝑖 ′β) 𝜆𝑜(𝑡) × exp(𝑋𝑗′β) = exp(𝑋𝑖− 𝑋𝑗)′𝛽 𝐼𝑓 𝑋𝑖,𝑘− 𝑋𝑗,𝑘= 1 𝑡ℎ𝑒𝑛, ℎ(𝑡|𝑋𝑖) ℎ(𝑡∣𝑋𝑗) = exp(𝛽𝑘)
In this analysis, hazard ratio 𝑒𝛽k indicates the relative hazard of giving birth between two
mothers that differ only by one unit of the 𝑘𝑡ℎ covariate. 𝑒𝛽1 and 𝑒𝛽2 are hazard ratios on
𝑷𝒐𝒔𝒕 and 𝑷𝒐𝒔𝒕 × 𝑺𝑨, respectively. 𝑒23 is the hazard ratio between the births that occurred
after and before the intervention in Lesotho. 𝑒𝛽1+𝛽2 measures the same ratio among black
South African women. My coefficient of interest 𝑒𝛽2, gives how the hazard of birth in South
Africa differed from Lesotho in the post-pension period. Hazard ratios are interpreted by comparing to 1: if 𝑒𝛽2< 1, it means that the hazard of failure (or birth) in the post-pension
period is lower than that in the pre-pension period in South Africa compared to Lesotho. South African women lengthened their birth spacing and fertility decreased following the pension expansion, and vice versa. If old-age security hypothesis is valid, I expect 𝑒𝛽2 < 1.
6 Results
Cox estimation results are presented on Table 1. Colum (1) reports the hazard ratios for the whole sample. In magnitude, both 𝑒𝛽1 and 𝑒𝛽1+𝛽2 are smaller than 1, suggesting a decrease in
fertility in both countries. Hazard ratio 𝑒𝛽1, which is 0.90, states that the hazard of birth
the pension program decreased the hazard of birth in South Africa by 10%. Nevertheless, both ratios are not statistically significant.
I further investigate whether the pensions had heterogenous effects in urban and rural areas. Nugent (1985) states that the conditions determining the importance of old age motive for fertility, such as lack of profitable means of accumulating assets, are more likely to be found in rural compared to urban areas. Hence, the impact of pensions is expected to be stronger among rural women. Column (2) and (3) present the estimation results for rural and urban areas respectively. DiD terms are not significantly different than 1, however, the magnitude of the ratios indicates the opposite of the hypothesis: in rural areas pensions seem to have had no impact or a very small positive impact on fertility whereas in urban areas it had a stronger negative effect than in the whole sample.
Table 1
Cox estimation.
Hazard ratios = eβk
Whole
Sample Rural Urban
(1) (2) (3) Post 0.90 0.80** 0.82 (.064) (.060) (.185) Post x SA 0.90 1.05 0.88 (.084) (.106) (.228) Mother_Age 1.09*** 1.13*** 1.11*** (.008) (.011) (.015) Birth_Order 0.55*** 0.54*** 0.44*** (.015) (.018) (.028) Log-likelihood -18533 -12370 -6015 Observations 28000 18398 9602 Number of failures 18479 12357 6122 Clusters 9527 5726 3801
Stratified at woman level. Robust standard errors are in parentheses (clustered at woman level). *** Significant at 1% level, ** Significant at 5% level * Significant at 10% level. Breslow method for ties among non-censored durations. Column (2): sample restricted to rural areas, Column (3): sample restricted to urban areas. Post is a dummy equal 1 if the observation was born in 1993 and onwards and 0 otherwise.
7 Discussion
The results of the analysis suggest that the old age pension program did not have a significant impact on fertility in South Africa. In this section, I first examine the income mechanism that could have possibly offset the negative impact of the pension program and could explain the insignificant results. In the second part, I discuss the limitations of the empirical design.
7.1 Income Mechanism
In developing countries, a prominent share of people lives in large extended families. Shared accommodation may bring about the sharing of further resources, especially money. If so, social policies might have unpredictable outcomes by affecting the preferences and behaviours of different demographic groups than the targeted one. Such a mechanism might have taken place among black South Africans in 1990s. As benefit amounts were considerably large, black population was quite poor and the share of extended families was significant. In the South African sample, roughly 21% of African mothers lived with one or more age-qualified household member. In urban areas the rate was 16% while in rural it reached 28% in 199810. In this part,
I investigate an income mechanism from pensions to fertility by examining separately the fertility behaviour of women living with and without an age-qualified household member. It is important to note that in DHS there is no information about the pension receipt or the income, therefore, I assume that those who are age-qualified receive the old age pension. This is not a very strong assumption because means-test excluded very few Africans, particularly in rural areas, whose income was quite low compared to the test.
To investigate the income effect, I restrict the South African sample into two extended family groups according to the age of the elderly household member(s): The “eligible” group includes mothers living with one or more household member who has been age-qualified for pension since 199311. As the survey was conducted in 1998, the age range is defined as 65-70
10 Own calculations using South African DHS 1998
11
In DHS there is no information about how long the members had been living in the household,
for female and 70-75 for male members. In the “non-eligible” group, the sample is limited to women whose household does not include elderly members who were age-qualified during the period 1993-1998. The oldest age range possible for this group is 55-60 for female and 60-65 for male members. Women that have members from both age groups are placed in the eligible group. Some descriptive statistics of the two groups are presented in Table 2. Mothers in the eligible group are significantly older than those in the non-eligible group, however, age at first birth and first marriage are not significantly different between the two groups. Both groups mostly reside in rural areas. Even though the means are quite close, Column (3) indicates that women in the eligible group are significantly less educated and live in bigger households. For the control group, the Lesotho sample is restricted to mothers living with a household member who is aged between the ranges of 55-70 for female and 60-75 for male members12.
The cox estimation presented earlier is applied separately to two groups using the new Lesotho sample. Table 3 reports the results of this analysis. Coefficients of interest are not significantly different than one. Looking at the magnitudes in Columns (1) and (2), mothers in the non-eligible group have lengthened their birth spacing following the expansion of pension program whereas mothers in the eligible group did not change their birth spacing. In rural areas, the magnitude of the hazard ratio for the eligible mothers is striking; it indicates that the hazard of fertility increased by 49% as a result of the pension expansions. This analysis could not be conducted for urban areas due to insufficient number of observations. Overall, the coefficients are consistent with a positive income mechanism offsetting the predicted negative impact of the pension expansion.
years at the time of the survey. Nonetheless, sharing of income does not necessarily require sharing the same house.
12 Ideally, this age range should have been the same as the South African eligible group (65-70 for
female and 70-75 for male). But in that case the number of observations in the Lesotho control group was less than the half of the number of observations in the other groups. This obliged me to choose a wider age range for the control group.
Table 2
Descriptive Statistics, South Africa
Eligible Non-eligible
Mean Mean Difference
(1) (2) (3)
Age at interview date 30.72 28.93 1.79***
(.372) (.305) (.557)
Age at first birth 19.88 19.65 0.24
(.186) (.146) (.273)
Age at first marriage 21.96 21.25 0.71
(.471) (.360) (.722)
Type of place of residence
(urban=1) 0.40 0.46 -0.06 (.023) (.020) (.035) Years of schooling 8.03 8.77 -0.75** (.186) (.149) (.277) Household size 7.72 7.30 0.42** (.155) (.110) (.200) Observations 465 633
Robust standard errors are in parentheses. *** Significant at 1% level, ** Significant at 5% level * Significant at 10% level. In Column (1), South African sample is restricted to mothers living with one or more household members aged between 60-65 (female), 65-70 (male). In Column (2) same restriction is done for members aged between 55-60 (female), 60-65 (male). Column (3) presents the difference between Column (1) and (2). Data: South African 1998 DHS
Table 3
Cox estimation: Income mechanism
Hazard ratios = 𝑒25 Whole Sample Rural
Eligible Non-eligible Eligible Non-eligible (1) (2) (3) (4) Post 0.88 0.97 0.81 .95 (.207) (.229) (.199) (.229) Post x SA 1.01 0.80 1.49 1.02 (.344) (.242) (.555) (.242) Mother_Age 1.11*** 1.17*** 1.15*** 1.20*** (.008) (.039) (.049) (.039)
Birth_Order 0 .53*** 0.432*** .51*** .41*** (.063) (.052) (.070) (.052) Log-likelihood -1149 -1226 -964 -926 Observations 2,173 2,416 1,671 1,736 Failures 1,255 1,373 1,017 1,028 Clusters 873 1040 641 715
Stratified at woman level. Robust standard errors are in parentheses (clustered at woman level). *** Significant at 1% level, ** Significant at 5% level * Significant at 10% level. Breslow method for ties among non-censored durations. Eligible group consists of mothers living with one or more household member aged between 60-65 (female), 65-70 (male). For the non-eligible group same restriction is done for members aged between 55-60 (female), 60-65 (male). In columns (3) and (4) sample restricted to rural areas.
7.2 Limitations
The empirical design of this analysis is vulnerable to problems that are preventing me from inferring causality. Consequently, I can neither rule out the predictions of old age security hypothesis, nor argue the presence of a positive impact of higher income on fertility. This part explains the issues as well as some remedies.
One of the best ways to look at the fertility trend over time would be to calculate total fertility rates in two time points, before and after the pension program. Unfortunately, this approach could not be pursued with the available data. As a result, I looked at the age-specific fertility rates over cohorts in order to understand the fertility trends over time (Figure 1). This is not a perfect representation of the time trend as the pre- and post-pension periods cannot be distinguished perfectly. The above limitation arose from the fact that the ideal data were not available.
Important concerns result from the utilization of Lesotho as a counterfactual fertility trend for South Africa. First concern is that, as mentioned in the econometric specification section, the parts that correspond to the pre-pension period are not perfectly parallel (Figure 1), possibly violating the common trend assumption. Another reason that Lesotho may not have been the ideal control group is the differences in fertility levels between two countries. DiD framework allows for pre-intervention differences in levels and adjusts for these differences
by subtracting the pre-intervention difference between two groups from the post-intervention difference. Nonetheless, in the fertility concept this assumption is questionable because the pace of decline is likely to depend on the level of fertility13. Hence, one could argue that Lesotho
will show a similar fertility trend as South Africa during the further phases of the fertility transition even in the absence of an old age pension scheme. Despite the limitations that Lesotho produced, it appeared as the best control candidate among the rest of the available countries.
It would have been optimal to exploit a variation in pension take-up among black South African population. Earliest data available that allows to examine a convincing coverage of the program is the World Bank’s Integrated Household Survey that was conducted in 1993. I used this data to estimate the social pension coverage over provinces. The take-up rates were already very high and considering the income levels, the variation across provinces was indicating that it was due to the means-test.
One way to deal with the limitations of the control group could be to use a method called propensity score matching. Propensity score matching uses observed characteristics of individuals in the treatment and control groups to construct a treatment group that has similar pre-treatment characteristics as the control group. On average, these matched sets will have similar balanced baseline covariates, therefore, they will be more likely to show a similarity in the outcome variable than the unmatched sample (Rosenbaum and Rubin, 1983). A good matching requires determining the baseline covariates that are prognostic of the outcome, but that were not affected by the treatment. In DHS, some of the available covariates that could be used in this method are the type of place of residence, years of schooling, literacy, access to electricity and a wealth index constructed using some variables (roof/wall materials, possession of goods such as car and refrigerator). Despite using various combinations of variables and different matching algorithms, I was unable to equalize the fertility levels (ASFRs) in the pre-pension period. Moreover, the literature is neither clear about how to accommodate sample
13 United Nations suggests that the pace of decline is positively related with the level of fertility
(United Nations Population Division (2002) “Fertility Levels and Trends in Countries with Intermediate Levels of fertility”).
weights in the matching context, nor how to use matching weights in a Cox regression. Hence, I could not pursue this strategy.
Another issue in the design could be that in the early 1990s, access to old age pensions was not the only thing that changed in favour of the black population in South Africa. The system of institutionalised racial discrimination that lasted more than 50 years was coming to an end. During this process of change, some other factors are likely to have affected the fertility behaviour. For instance, an improved access to health, education and financial services or simply a better sense of security could have simultaneously played a role with the pension program in stimulating the fertility transition. Hence, even in the case of identical pre-pension levels and trends across South Africa and Lesotho, it would have been difficult to argue an isolated impact of the social pension program on fertility.
8 Conclusion
The South African government achieved racial equality in social pension coverage, means-test and benefit levels in 1993. Old age pensions represented a large cash transfer to elderly households, benefit levels reaching about twice the median income per capita of rural African households. This thesis tested the old-age security hypothesis that predicts a decrease in fertility levels as the uncertainty about future income diminishes. It is very likely that the old-age pension program was widely known in the beginning of 1993 among the black South African population, and the benefit levels were significant enough to make a difference in the expectations about the future income.
By implementing a differences-in-differences framework, I examined whether black South African women lengthened their birth spacing after 1993 compared to Basotho women. Results indicated a negative, but insignificant effect of the pensions on the fertility. Magnitude of coefficients indicated a stronger negative impact in urban areas, whereas a small positive impact in rural areas, where the old age security motive for fertility is hypothesized to be stronger. I further investigated whether the insignificant results were due to an income effect that counterbalanced the negative impact predicted by the old age security hypothesis. In this
investigation results were suggestive of a strong and positive income effect of the pension program on fertility among mothers residing with age-qualified household members in rural areas.
Leaving aside the limitations in the empirical design, this thesis suggests that the demographic consequences of old age pension programs may be subject to unintended externalities among extended families, preventing the predictions of the old age security hypothesis from occurring.
References
Anderton, D. L. and Bean, L. L. (1985) ‘Birth Spacing and Fertility Limitation: A Behavioral Analysis of a Nineteenth Century Frontier Population’, Demography. doi: 10.2307/2061176.
Banerjee, A., Meng, X., Porzio, T., & Qian, N. (2014) ‘Aggregate fertility and household
savings: A general equilibrium analysis using micro data’, NBER Working Paper No. w20050
Boldrin, M. and Jones, L. E. (2002) ‘Mortality, fertility, and saving in a Malthusian economy’, Review of Economic Dynamics. doi: 10.1006/redy.2002.0186.
Boldrin, M., De Nardi, M. and Larry, J. (2015) ‘Fertility and Social Security’, Journal of Demographic Economics. doi:10.1017/dem.2014.14
Burns, J., Keswell, M. and Leibbrandt, M. (2005) ‘Social assistance, gender, and the aged in South Africa’, Feminist Economics. doi: 10.1080/13545700500115944.
Caldwell, J. C. (1978) ‘A Theory of Fertility: From High Plateau to Destabilization’, Population and Development Review. doi: 10.2307/1971727.
Case, a and Deaton, A. (1998) ‘Large gash transfers to the elderly in South Africa’, The Economic Journal. doi: 10.1111/1468-0297.00345.
Casterline, J. B. and Odden, C. (2016) ‘Trends in Inter-Birth Intervals in Developing Countries 1965–2014’, Population and Development Review. doi: 10.1111/j.1728-4457.2016.00134.x.
Chatterjee, S., and Vogl, T. (2018) 'Escaping Malthus: Economic Growth and Fertility Change in the Developing World', American Economic Review, 108(6), 1440-67.
Chowdhury, S. (2010) ‘Determinants and Consequences of High Fertility : A Synopsis of the Evidence’, World Bank.
Cox, D. R. (1972) ‘Regression models and life tables’, Journal of the Royal Statistical Society. Series B:
Devereux, S. (2001) ‘Social Pensions in Namibia and South Africa’, Institute of Development
Studies.
Duflo, E. (2003) ‘Grandmothers and Granddaughters: Old-Age Pensions and Intrahousehold Allocation in South Africa’, The World Bank Economic Review. doi: 10.1093/wber/lhg013. Ebenstein, A. and Leung, S. (2010) ‘Son preference and access to social insurance: Evidence rom China's rural pension program’, Population and Development Review. doi: 10.1111/j.1728-4457.2010.00317.x.f
Ghilagaber, G. and Gyimah, S. O. (2004) ‘A Family of Flexible Parametric Duration Functions & their Applications to Modelling Child-Spacing in Sub-Saharan Africa’, PSC Discussion Papers Series.
Haub, C. and Gribble, J. (2011) ‘The world at 7 billion’, Population Bulletin. doi: 10.9745/GHSP-D-14-00012.
Heckman, J. J. and Walker, J. R. (1990) ‘The Relationship Between Wages and Income and the Timing and Spacing of Births: Evidence from Swedish Longitudinal Data’, Econometrica. doi: 10.2307/2938322.
Jenkins, S. P. (2005) 'Survival analysis', Unpublished manuscript, Institute for Social and Economic Research, University of Essex, Colchester, UK, 42, 54-56.
Lambert, S. and Rossi, P. (2016) ‘Sons as widowhood insurance: Evidence from Senegal’, Journal of Development Economics. doi: 10.1016/j.jdeveco.2016.01.004.
Leibenstein, H. (1975) ‘The economic theory of fertility decline’, The quarterly journal of economics. doi: 10.2307/1881706.
Nugent, J. B. (1985) ‘The Old-Age Security Motive for Fertility’, Population and Development Review. doi: 10.2307/1973379.
Nugent, J. B. and Gillaspy, R. T. (1983) ‘Old age pensions and fertility in rural areas of less developed countries: some evidence from Mexico’, Economic Development and Cultural Change.
Oliveira, J. (2016) ‘The value of children: Inter-generational support, fertility, and human capital’, Journal of Development Economics. doi: 10.1016/j.jdeveco.2015.12.002.
Rosenbaum, P. R. and Rubin, D. B. (1983) ‘a. “The Central Role of the Propensity Score in Observational Studies for Causal Effects”’, Biometrika. doi: 10.1093/biomet/70.1.41.
Rossi, P. and Rouanet, L. (2015) ‘Gender Preferences in Africa: A Comparative Analysis of Fertility Choices’, World Development. doi: 10.1016/j.worlddev.2015.03.010.
Schoumaker, B. (2013) ‘A stata module for computing fertility rates and TFRs from birth histories: Tfr2’, Demographic Research. doi: 10.4054/DemRes.2013.28.38.
Stecklov, G. et al. (2006) ‘Demographic Externalities from Poverty Programs in Developing Countries: Experimental evidence from Latin America’, American University.
Whittington, L. A., Alm, J. and Peters, H. E. (1990) ‘Fertility and the personal exemption: implicit pronatalist policy in the United States’, American Economic Review.
Willis, R. J. (1980) ‘The old age security hypothesis and population growth’, In: Burch TK, ed. Demographic behavior: interdisciplinary perspectives on decision making. Boulder, Colorado, Westview Press. doi: 10.3386/w0372.
