Elsevier

Social Science & Medicine

Volume 56, Issue 8, April 2003, Pages 1677-1691
Social Science & Medicine

Undernutrition in Benin—an analysis based on graphical models

https://doi.org/10.1016/S0277-9536(02)00162-4Get rights and content

Abstract

Undernutrition is one of the most important health problems in developing countries. Examining its determinants implies the investigation of a complex association structure including a large number of potential influence variables and different types of influences. A recently developed statistical technique to cope with such situations are graphical chain models. In this paper, this approach is used to investigate the determinants of undernutrition in Benin (West Africa). Since this method also reveals indirect influences, interesting insight is gained into the association structure of all variables incorporated. The analysis identifies mother's education, socioeconomic status, and religion as three variables with particularly strong direct and indirect linkages to undernutrition.

Introduction

Undernutrition among children is one of the most important health problems afflicting developing countries. It is an important cause of infant and child mortality, compromises the social and intellectual development of children, and reduces labor productivity (UNICEF, 1998; Sen, 1999).

Modeling the determinants of undernutrition at the micro level is a complex undertaking. While it is clear that undernutrition is fundamentally related to inadequate dietary intake and the incidence, severity, and duration of disease, these factors themselves are related to a large number of intermediate, underlying, and basic causes operating at the household, community, or national level (UNICEF, 1998). Among the most important factors are probably the education, wealth, and income situation of the parents, household size, birth order, religion, and sex of the child, the availability of clean water, adequate sanitation, immunization, and primary health care services, and the level of disease prevalence in the surrounding community.

These many causal factors generate two different problems for analysts that seek to identify the determinants of undernutrition. First, the different determinants of undernutrition occupy a different position in a dependence chain and cannot adequately be modeled by including them all in a linear regression. For example, the inadequacy of dietary intake is an immediate cause of undernutrition, while the education of the parents plays a more indirect role as it can affect the ability of the household to secure adequate nutrition, may affect the health knowledge and care practices, and may affect the availability of clean water, sanitation, and health care. In fact, UNICEF has made a useful distinction between immediate, intermediate, and underlying causes of undernutrition. Smith and Haddad (1999) use a similar method of classifying the different determinants along a dependence chain. Any empirical strategy that attempts to identify the determinants of undernutrition must recognize the existence of such a dependence chain.

Second, there are likely to be close correlations between many of the factors affecting undernutrition. For example, the education of the parents may be closely correlated with the income situation of the household, which in turn may be closely related to the availability of water, sanitation, and health services. Any empirical investigation of this topic should be able to explicitly model these associations and identify their strength.

Graphical chain models are able to deal with these two issues. They are a tool for modeling and illustrating complex association structures that are placed along a dependence chain. They are therefore ideally suited to model the complex dependence chain involved in undernutrition of children.

In this paper, a graphical chain model for undernutrition is estimated for Benin in West Africa using data from the 1996 Demographic and Health Survey. The model considers some thirty immediate, intermediate, and underlying factors associated with undernutrition and models the association structure between them. With this approach, we are able to identify important pathways in the dependence chain for undernutrition. In particular, we identify many pathways through which mother's (and father's) education, household socioeconomic status, and religion have an effect on undernutrition. At the same time, this approach allows us to identify areas for further investigation of the complex dependence chain leading to undernutrition in developing countries.

This paper is organized as follows. The next section addresses the problem of how to measure undernutrition. We introduce the basic idea of graphical chain models in the following section and describe the data set thereafter where we also provide some descriptive statistics. The model selection strategy for fitting such a model to a multivariate data set is presented in the following section after which we provide a detailed discussion of the results, followed by a short conclusion.

Section snippets

Defining and measuring undernutrition

Undernutrition among children is usually defined in relation to their anthropometric status. Analysts typically distinguish between chronic undernutrition (stunting) that is defined as insufficient height for age, acute undernutrition (wasting) that is defined as insufficient weight for height, and a combination of the two (underweight) that is defined as insufficient weight for age (UNICEF, 1998). Wasting and stunting are in a dynamic interaction. For example, if children stop growing and thus

Graphical models

Multivariate data sets typically imply complex association structures, not only between the response variables of interest and their potential explanatories, but also among the explanatories themselves. To learn about such a structure, sophisticated statistical analysis tools are required. In this respect, graphical models are a very promising, recently developed approach (cf. Lauritzen & Wermuth, 1989; Wermuth & Lauritzen, 1990; Cox & Wermuth, 1993) which allows for identifying associations

The data set

The data are drawn from the 1996 Demographic and Health Survey in Benin in West Africa. The DHS draws a representative sample of women of reproductive age and then administers a questionnaire and does some anthropometric measurements of them and their children that were born in the previous 3 years. The questionnaire covers a wide range of questions regarding household socioeconomic status, health access and behavior, fertility behavior, reproductive health, and HIV/AIDS. The questionnaire also

Fitting the chain graph

Having formulated the rough dependence chain the task of a statistical analysis is now to identify direct, indirect, symmetrical as well as asymmetrical dependencies among all variables, and of course the corresponding conditional independencies. This task calls for an appropriate selection strategy which should provide an optimal solution to the trade-off of extracting only a low number of associations to get a simple structure without leading to an oversimplification. For fitting graphical

Analysis

The chain graph presented in Fig. 4 shows the direct and indirect associations of each variable with stunting and wasting. Although this figure may at first glance appear rather complicated, it is fairly easy to identify different pathways through which the various variables directly or indirectly influence the response variables. Estimated regression parameters of the direct associations between the immediate, intermediate, and underlying factors with stunting and wasting are given in Table 4.

Conclusion

Using graphical chain models for investigating malnutrition in Benin has enabled to identify the determinants and their structure of associations in the system. A great advantage is that the displayed graph allows an easy recognition of direct and indirect pathways between variables. Through that, factors which appear to have no impact since they do not directly influence the response, could still be identified as important indirect influences.

As an example, mother's education influences

Acknowledgements

Financial support by the SFB 386, Deutsche Forschungsgemeinschaft is gratefully acknowledged. We thank an anonymous referee as well as workshop participants in Munich and Rostock for helpful comments and discussion.

References (31)

  • S.T. Andersson et al.

    Alternative Markov properties for chain graphs

    Scandinavian Journal of Statistics

    (2001)
  • A. Blauth et al.

    Interactive analysis of high-dimensional association structures with graphical models

    Metrika

    (2000)
  • A. Caputo et al.

    A graphical chain model derived from a selection strategy for the sociologists graduates study

    Biometrical Journal

    (1999)
  • R.G. Cowell et al.

    Probabilistic networks and expert systems

    (1999)
  • D.R. Cox et al.

    Linear dependencies represented by chain graphs (with discussion)

    Statistical Science

    (1993)
  • D.R. Cox et al.

    Tests of linearity, multivariate normality and the adequacy of linear scores

    Applied Statistics

    (1994)
  • D.R. Cox et al.

    Multivariate dependencies—models, analysis and interpretation

    (1996)
  • M. Das Gupta

    Selective discrimination against female children in rural Punjab

    Population and Development Review

    (1987)
  • D. Edwards et al.

    A fast procedure for model search in multidimensional contingency tables

    Biometrika

    (1985)
  • D. Edwards et al.

    A fast model selection procedure for large families of models

    Journal of the American Statistical Association

    (1987)
  • M. Frydenberg

    The chain graph Markov property

    Scandinavian Journal of Statistics

    (1990)
  • P. Giudici et al.

    Decomposable graphical Gaussian model determination

    Biometrika

    (1999)
  • S. Højsgaard et al.

    BIFROST—block recursive models induced from relevant knowledge, observations, and statistical techniques

    Computational Statistics and Data Analysis

    (1995)
  • S. Klasen

    Nutrition, health and mortality in Sub-Saharan AfricaIs there a gender bias?

    Journal of Development Studies

    (1996)
  • S. Klasen

    Marriage, bargaining, and intrahousehold resource allocationExcess female mortality among adults during early German development

    Journal of Economic History

    (1998)
  • Cited by (24)

    • Reducing child malnutrition in Nigeria: Combined effects of income growth and provision of information about mothers' access to health care services

      2010, Social Science and Medicine
      Citation Excerpt :

      These estimates indicate that possession of a vaccination card is associated with about a 1.4-point improvement in child z-scores. Thus, as in Caputo, Foraita, Klasen, and Pigeot (2003), if child stunting provides a reliable overall measure of long-term effects of malnourishment, Table 3 estimates imply child vaccination programs play a major role in Nigeria’s overall strategy to address long-term child health effects such as stunting, as well as short- and long-term cumulative undernourishment as measured by low weight-for-age. The coefficients for family size and its square are jointly significant at the 5 percent level in Table 3 regressions 1–4.

    • Using graphical chain models to analyze differences in structural correlates of undernutrition in Benin and Bangladesh

      2008, Economics and Human Biology
      Citation Excerpt :

      Related to such a production function approach and of particular relevance for our purposes is the framework proposed by UNICEF (1998) which divides the determinants of undernutrition into immediate, intermediate, and underlying determinants which already informed the work on Benin. Similarly to Caputo et al. (2003) we postulate the dependence chain as follows (Fig. 1), from left to right: In the first block we put our pure response variables wasting (Wa) and stunting (St).4 The next block includes variables that have an immediate influence on these Z-scores such as those variables relating to nutritional intake and illness episodes.

    • A SINful approach to Gaussian graphical model selection

      2008, Journal of Statistical Planning and Inference
    • Analyzing infant mortality with geoadditive categorical regression models: A case study for Nigeria

      2004, Economics and Human Biology
      Citation Excerpt :

      Many variables turned out to be insignificant so that we chose a reduced set of covariates reported in Tables 4 and 5.5 Furthermore, some of the socio-economic factors are highly correlated (Caputo et al., 2003), hence we systematically split the highly correlated socio-economic factors and then included them in separate models. When interpreting these covariates, one should therefore note that they capture more broadly socio-economic conditions that the particular indicator included (e.g. the electricity variable, which is highly correlated with other indicators of socio-economic status, is more an indicator of general socio-economic status than of narrowly access and use of electricity).

    View all citing articles on Scopus
    View full text