Recent Changes in Micro-Level Determinants of Fertility in India: Evidence from National Family Health Survey Data

Abstract

This paper investigates the determinants of fertility and their changes in recent years, empirically, drawing upon large household data-sets in India, namely National Family Health Survey data over the period 1992–2006. It is found that there is a negative and significant association between the number of children and levels of parental education when we apply OLS, ordered logit and pseudo panel models, while in the case of IV model only mother's literacy is negatively associated with the number of children. The results of OLS and ordered logit models imply that households belonging to Scheduled Castes (SCs) tend to have more children than other social groups. Our results suggest that policies of national and state governments to support social infrastructure, such as the provision of education at various levels, and its promotion for both males and females, particularly for households belonging to SCs, would make a significant contribution to efforts to reduce fertility and decelerate population growth.

Notes

¹ See Vallely (2008) for the debate on this issue.

² Poverty head-count ratio based on the national poverty line in 2004/2005 is 28.7% (Himanshu, 2007).

³ The theory of demographic transition explains the common pattern of transition in population history. While the first stage of transition before economic modernization sees stable population due to high birth and death rates, the population grows rapidly in the second stage, where death rates decline more rapidly than birth rates, for example, through better educational systems and medical and health-care facilities only available in modernized society. The population becomes stable again in the third stage, when further modernization and better education cause fertility to go down.

⁴ For example, the average annual growth rate of GDP per capita of India in 1992–2006 is 4.7%, while the primary and secondary school enrolment increased from 70.1% in 1992 to 90.2% in 2006 (calculated by World Development Indicator 2010).

⁵ Browning & Gørtz (2012) found that in Denmark women's relative expenditures (over men's) are positively influenced by their higher relative wages, but women's relative leisure is negatively affected by their relative wages. As women with a higher bargaining power who tend to have less leisure are likely to have fewer children to avoid their leisure time being further squeezed, Browning and Gørtz's results seem consistent with a negative effect of women's bargaining on fertility. The bargaining–fertility relationship is thus complex, but we consider only the general effect of education on fertility, due to the data constraints.

⁶ It is assumed here that the bargaining coefficient, γ, is exogenously determined by, e.g. female education or cultural factors: in other words, γ affects the household decision on the number of children, but not the other way around. However, the bargaining coefficient, γ, can be endogenous in reality: that is, the household decision on the number of children could in turn affect γ, as modelled by Basu (2006), who assumes the endogeneity of γ in the collective–bargaining model. This endogeneity is not taken into account in the present analysis.

⁷ While we assume that only parental education affects the bargaining power in the model, due to the data constraints, other factors, such as the difference in ages or labour participation of parents, could affect the relative bargaining power (e.g. Imai et al., 2012). It is possible that the preferences A_i may reflect some factors influencing the bargaining, but we have not modelled the interactions explicitly. As we primarily focus in the present study on the effects of parental education, rather than bargaining, on fertility, we do not use any proxy for the intra-household bargaining in our econometric models.

⁸ As the NFHS data do not have income or consumption data, the relation between fertility and income cannot be examined. We have instead estimated the effect of owned land on fertility and have found a negative and significant coefficient estimate.

⁹ See http://www.nfhsindia.org/index.html for a detailed description of NFHS.

¹⁰ TFR is the average number of children that would be born to a woman over her lifetime if she were to experience the exact current age-specific fertility rates through her lifetime, and she were to survive from birth through to the end of her reproductive life.

¹¹ Given the presence of heteroskedasticity in our large sample, we have opted for robust OLS model by using the White–Huber sandwich estimator, not robust Tobit model, because the latter is likely to be inconsistent with heteroskedasticity (Greene, 2003). We owe this useful comment to one of the referees.

¹² There is a high correlation between neo-natal mortality and fertility, because a mother who has lost her baby is more likely to have another baby, analytically and empirically shown by Bhalotra & van Soest (2008) in the Indian context.

¹³ The problem is partly alleviated by applying pseudo panel where age cohorts are introduced in the model.

¹⁴ We have also used “wealth index”, which is based on the data of different household assets, characteristics and access to infrastructure, and have obtained similar results.

¹⁵ See, for example, Wooldridge (2006, Chap. 15) for the details of ordered logit model. The use of ordered probit model or Poisson regression model gives broadly similar results, and thus we report only those of ordered logit model.

¹⁶ It is not possible to calculate the share of teenagers who were married or have given birth in the total number of teenagers, due to data constraints. However, the number of married couples in which a wife was a teenager (15–19 years) is much smaller than that of any other age groups.

¹⁷ In our earlier version (Imai & Sato, 2010), we estimated the same model using three rounds of National Sample Survey data in 1993, 1999 and 2005 and obtained broadly similar results by using a proxy for the fertility rate based on the number of children under 15 years old in a household. However, given the possible measurement error in this proxy, we have decided to present only the results based on NFHS.

¹⁸ We expect that grandmothers' education at the village level has a positive effect on mother's literacy in the first stage of IV, but the coefficient estimate is negative and significant. This may be because in the village (or district/state) where education was more widely available a generation ago, government has spent a larger budget on literacy projects, and thus some reversal may have occurred.

¹⁹ If the positive coefficient estimate of father's literacy in Case (E) is valid, we would conjecture that educated men tend to have a greater bargaining power and impose the traditional value of large family on women. This is contradictory to the results of Case (A) or (C).

²⁰ Note that we excluded Muslim households with more than one female spouse of a male household head.

²¹ As the pseudo-panel model involves averaging mother's and father's education at state level, the average education standards of mothers and that of fathers are highly correlated. Inclusion of both mother's and father's education in the model has led to the results which are counterintuitive and incoherent, due to multicollinearity, and it was not possible to insert both at the same time.

²² The results are not reported but will be furnished on request.

The authors have benefited from valuable comments at various stages from Barrientos Armando, Sonia Bhalotra, Pranab Bardhan, Per Eklund, Raghav Gaiha, Raghbendra Jha, Kunal Sen, Shoji Nishijima, Yoshifumi Usami, Koji Yamazaki and participants in seminars at Harvard, Manchester, Bristol, Kobe, Osaka City and Doshisha Universities and the international conference on “New Directions in Welfare” at St Catherine's College, Oxford University, 29 June–1 July 2009 and the joint Kobe–Hanyang Workshop at RIEB, Kobe University on 28 May 2010. Valuable comments from two referees are gratefully acknowledged. The views expressed are, however, those of the authors and do not necessarily represent those of the organizations to which they are affiliated.

The authors are grateful for the financial support of Grant-in-Aid for Scientific Research (S) (#21221010), the Australian Research Council-AusAID Linkage grant LP0775444 in Australia, and a small grant from DFID and the Chronic Poverty Research Centre at the University of Manchester in the UK. Research support from RIEB, Kobe University, for the first author is gratefully acknowledged.