Previous - Next
This is the old United Nations University website. Visit the new site at http://unu.edu
Structural models of family social health theory
Conceptualization of the family system model
Justification
of the model
An application of structural modelling
Javanese
family models
Nigerian
family models
Notes
Appendix: Variable description and composite
index construction
References
Conceptualization of the family system model
In this study we have used data from Indonesia and Nigeria to develop an exploratory empirical model of family-level characteristics that determine child welfare and development. Through this model we hope to improve our understanding of the ways in which family resources influence family management, beliefs, and caring behaviours, including emotional climate and child-care quality within the home, and through which family management, beliefs, and caring behaviours influence the development of the children. Because child development is measured multidimensionally, this model integrates several family components that potentially influence each dimension of child development. Although we have conducted this analysis with respect to child outcomes, we believe that the well-being of other family members is similarly determined. We used structural equation modelling with latent variables (Bollen 1989), also known as LISREL (Jöreskog and Sörbom 1989).
Myers (1992) distinguished between the definitions of child growth and child development. He defined growth as "a change in size," whereas development was defined as "changes in complexity and function." For practical purposes, we use the term "child development" to refer to both child growth (as measured by anthropometric indicators) and child mental development (characterized by the mental development index or IQ, the physical development index for younger children, and the social development index for older children). The terms are separated for measurement and analysis purposes, however.
The general family functioning model is presented in figure 8.1. In this general conceptual form, the model is applicable to both Nigerian (the Yoruba) and Indonesian (the Javanese) cultures. The overall set of variables and the indicators designed to measure each factor may vary between the Yoruba and the Javanese (see Appendix to this chapter) because the same type of data were not always available for both cultures, and some factors that influence family management, beliefs, and caring behaviours are culturally specific.
The first level of this model, "Family Resources," includes both material (no. 1) and social (no. 2) resources. Many factors at this level of the model are measures of socio-economic status, such as housing, food budget, literacy and educational level, and media use. These factors feed into the next level of the model, "Family Management, Beliefs, and Caring Behaviours." These encompass measures of hygiene (no. 3) and feeding practices (no. 4), parents' caregiving behaviours (affection and attention, no. 5), and academic stimulation (no. 6). These factors, in turn, determine the child's growth and development (no. 7).
Figure 8.2 presents a more detailed version of the model. Additional factors not listed in figure 8.2 are community endowment and support from within the family. These two factors are conceptually related to sociocultural support and are tested only in the Javanese models (fig. 8.3-8.7). We hypothesize that sociocultural support influences family management and caring behaviours. Sociocultural support includes support from the family and other kin (the father's involvement in child care and the mother's satisfaction from support received from family and relatives) and community support/endowment (the number of health facilities and community health indicators). We do not have information on maternal satisfaction and community endowment in the Nigerian data set; therefore, the role of these factors is tested in the Javanese data set only. Our literature review on Javanese families indicates that fathers also are involved in child-care activities, especially after the child has begun to walk; therefore, it is culturally appropriate to test this aspect in the Javanese setting with samples of older children. The father's involvement in child-rearing in Yoruba families is less clear and, therefore, was not tested.
Fig. 8.1 Simplified conceptual family functioning model, based on Nigerian data
Fig. 8.2 Detailed conceptual model, based on Nigerian data. See Appendix for full explanation of abreviated terms
Two other factors not listed in figure 8.2 are modernizing lifestyle and child personality; these are included in figures 8.8-8.10 (Nigerian models). Modernizing lifestyle is expected to influence parental care behaviours. This includes "modern" child-rearing behaviours, such as parents eating together with the child, positive affection rather than negative disciplinary behaviour, and a delay in requiring the child to perform housework. In the polygamous culture of Nigeria, it includes the level of the father's involvement in the household. Since no data on these factors were available for the Indonesian sample, and it is uncommon for the Javanese fathers to eat or sleep separately from a wife and children, modernization within the family was tested only in Yoruba families.
The child's characteristics are hypothesized to influence parental care behaviours. These factors were examined only in the Nigerian model because no data were available for the Javanese model.
Family resources
Professionals in child development increasingly acknowledge the influence of the context in which the individual child grows. Bradley (1989) emphasized the importance of understanding parents' responses to other family factors, such as income and social status, although there is little research in this area by child development specialists.
Family resources alone are not enough to promote positive child development, however. Much depends on how the family translates the available resources into positive child-caring behaviours. Many researchers believe that interventions or actions to improve the caring behaviours within the family can be implemented with little or no improvement in family income (Engle 1992). Parental education, especially maternal education, is consistently related to improved child nutritional status and lower infant mortality (Heller and Drake 1979; Behrman and Wolfe 1984), better child feeding and health practices (Zeitlin et al. 1988), a higher level of child stimulation (Zeitlin, Ghassemi, and Mansour 1990), greater use of health care services (Grant 1984), lower fertility, and more child-centred care behaviours (Ware 1984). Some researchers argue that the positive association between mother's schooling and child health is not due primarily to the effect of schooling per se; instead, schooling serves largely as a proxy for unobserved characteristics related to her childhood background. A study in Nicaragua that controlled for the heterogeneity in unobserved maternal endowment through the shared childhood experience of adult sisters found that the impact of the mother's education declines substantially. Behrman and Wolfe (1987) also found that maternal schooling effects evaporate when the maternal endowment (abilities, habits, and health status related to childhood family background) is controlled for, thus raising doubts about standard estimates of the impact of maternal schooling on child welfare.
A study of infant development in South India confirms the argument that maternal childhood experience is related to positive caring behaviour. Landers (1989) found adequate growth and cognitive performance among three-monthold babies despite the poor environment and high biological risks. In this case, cultural patterns of child care, such as maternal proximity and a high degree of psychosocial commitment of the mothers, supported positive child development during infancy.
In our Javanese model we used measures of the social network to represent family social resources. The social network provides families with a sense of integration in the community and an opportunity to obtain interpersonal gratification and material resources. In the industrialized world, social isolation has been associated with child abuse (Belsky 1984). Women of higher social class in Java were found to interact less often with neighbours and, in Nigeria, to exchange child care less often with neighbours (Olusanya 1981). In the United States, low-income families in Detroit viewed neighbourhood friends as more intimate than non-neighbourhood friends, whereas high-income families described their neighbourhood friends as less intimate than non-local friends (Fischer 1977); non-local friends can be part of the social network too, however.
Family management, beliefs, and caring behaviours
Our measures of "Family Management, Beliefs, and Caring Behaviours" are similar to "family climate" (Schneewind 1989) and "parental personality" (Belsky 1984) and include feeding and hygiene behaviours along with other factors in child caregiving noted by Engle (1992) in her definition of "care." In our own case, these measures rely on the mother's report of family practices and on observations of the mother-child dyed and the home environment. Family climate or parental personality imply a human locus of control or executive function with values operated by the family as a unit. The term "Family Management and Caring Behaviour" in this study is defined as (1) the extent to which decisions to allocate family resources are made and monitored by a central executive function for the achievement of a rational optimization strategy, (2) health and hygiene behaviours, as determined by cultural norms, and (3) parental personality.
Family climate, or "inner-family socialization activity" (Schneewind 1989) or "parental personality" (Belsky 1984), is a transactional relationship between parents and children. It determines to what extent the family's potential resources can be transformed for the benefit of family members. Family climate includes maternal affection and environmental stimulation.
Belsky (1984) asserts that parental personality is the most influential factor in determining child outcome. Optimal child development is promoted by parents whose caregiving is attentive, physically demonstrative, stimulating, responsive, and non-restrictive. Based on a literature review, Belsky, Lerner, and Spanier (1984) found that a high level of overall maternal attentiveness at five months (as measured by the frequency with which the mother looked at, touched, held, or spoke to the baby) predicted high levels of infant exploratory activity. Attentiveness may be a proxy for other maternal qualities and behaviours that facilitate the infant's intellectual growth because other positive caregiving personality traits are also able to predict the intellectual and emotional development of infants (Belsky, Lerner, and Spanier 1984). A review of the positive deviance literature found that caregiver-child interactions associated with adequate growth and development included a high degree of physical interaction (holding and hugging), speaking and responding to a child's vocalizations, establishing frequent eye contact, responding to the child's needs, creating a stimulating physical environment for the child, and enhancing the child's initiative and creativity (Zeitlin, Ghassemi, and Mansour 1990).
Studies of family ecology (e.g. the amount and availability of toys and rooms) did not support the assumption that the physical environment was important for development of cognitive skills (Hwang, Lamb, and Broberg 1989; Kreppner 1989). When the quality of mother-child interactions was added to environment variables, a greater association was found (Wachs 1984). Family interactions, such as the involvement of fathers in child care, seemed more significant for child development than the physical environment (Hwang, Lamb, and Broberg 1989).
We do not have direct measures on family cohesion. We postulate that families with a high degree of paternal involvement are more cohesive. The pattern of fathering also reflects the quality of the marital relationship. Belsky, Lerner, and Spanier (1984) pointed out that marital quality can be a powerful predictor of fathering. We assume that fathers who are involved in child care are not in conflict with their spouses or are at least available to care for their children. When considering family variables that affect children's mental health, enduring family conflict is seen to have the most negative impact (Colletta 1991).
Support and community endowment
Sociocultural support determines the quality of care the mother is able to give her infant. "Intimate" support is provided by the immediate family, and "community" support comes from neighbourhood and workplace associations (Zeitlin, Ghassemi, and Mansour 1990).
Bronfenbrenner, Moen, and Garbarino (1984) reviewed the significance of health services in the community in the context of family well-being and concluded that the use of health services is influenced by access to services, their availability and quality, and the orientations of service providers to the clients. They also noted that the provision of health services to individuals rather than to the family is a major source of inefficiency and inequities. The disorganization of health services in the United States, for example, in which different services are provided in different locations, creates logistical difficulties for families seeking multiple services. This has led to efforts to develop a new model of comprehensive health care along neighbourhood lines.
The idea of posyandu services (integrated health care services, family planning, nutrition education, immunization, prenatal care, supplementations, and oral rehydration therapy) in Indonesia provides one-stop multiple services for the family rather than for individuals. There is ample evidence that this kind of service has positive effects on child welfare in Indonesia.
Grandmothers in both Nigeria and Indonesia provide child care and emotional support (See chs. 6 and 7). Rohner's cross-cultural study (Rohner 1975) found that mothers who are unable to get away from the burden of child care from time to time are more likely to withdraw warmth and affection and to reject their children than mothers who have another adult to help them with child care. When grandparents are present, children are more likely to receive adequate warmth. Similar findings also were observed in some African cultures (Minturn and Lambert 1964) and in developed countries (Bengston and Robertson 1985; Hwang, Lamb, and Broberg 1989).
Alternate caregivers, especially grandparents, not only assume some of the burden of child care, but also serve as additional teachers of social skills and models for adaptive behaviour. Moreover, they can improve the quality of parental behaviours by providing emotional support and advice.
In Nigeria, fathers enjoy playing with their young children and may be alternative caretakers, but only if they regularly live with the child. In our sample, most fathers (74 per cent) always live in the same house as their wives and children but might occupy a separate bedroom. This pattern also has been documented in Kenya (Whiting and Whiting 1975). The intimacy of the father's role in child care is also determined by the degree of modernization (Caldwell and Caldwell 1977).
In Java (ch. 6), fathers can be secondary caregivers even before the children reach the age of five. Usually, husbands live in the same house as their wives and children. This conforms to the Whitings' findings that when a father shares a room with his wife, he is more likely to become involved in child care.
Community endowment is defined structurally. The community is
... a territorially bounded social system or set of interlocking or integrated functional subsystems (economic, political, religious ... ) serving a resident population, plus the material culture of physical plant through which the subsystems operate. The community concept does not include such characteristics as harmony, love, or intimacy ... but it does include a minimum of consensus. (International Encyclopedia of Social Sciences 1968, 3:163)
Modernizing lifestyle
Kohn's work (Kohn 1969) showed that working-class fathers, whose jobs require compliance and obedience, tend to hold values that stress obedience in their children, whereas middle-class fathers, whose jobs require effective intellectual functioning and self-direction, value intellectual development and independence in their children (Kohn 1969). Our literature review on Javanese families shows that there is little class difference with respect to child obedience and compliance: child obedience is still preferred by the majority of Javanese parents, even among the professional parents (about 50 per cent, according to Hoffman's study [Hoffman 1988]). However, LeVine, Klein, and Owen (1967), Lloyd (1970), and our own study found in Nigeria that élite and modernizing parents appeared to be less restrictive of their children's aggressive behaviour, and to value self-reliance and responsibility more than traditional parents.
In the case of Nigeria, modernization within the family can be partially characterized by the living arrangements. The norm for the traditional family is usually polygamy, when a man can afford more than one wife. This means that the same family may be more "modern" in its early phases and more traditional in later phases when affluence permits the acquisition of more than one wife. In monogamous families, the father lives in the same house all the time.
Child's characteristics
In Nigeria, the child's characteristics of individuality are assumed to be related to parental care. Belsky's literature review (Belsky 1984) cited several studies that linked the child's temperament with parenting and concluded that a difficult temperament, especially in infancy, can undermine parental functioning. The relationship between a child's personality and parental care behaviours can be reciprocal, however (Morris 1969). Parental personality, such as responsiveness and the ability of the mother to form a strong bond with her child, also can shape the child's personality.
The synergistic relationship between child growth and child development
Child growth and child mental development are mutually related. The synergistic relationship between these factors has been noted by several authors (Zeitlin, Ghassemi, and Mansour 1990; Myers 1992). Zeitlin and coworkers (Zeitlin, Ghassemi, and Mansour 1990) reviewed numerous studies to examine the link between psychosocial well-being, nutritional thriving, and health. They concluded that psychological stress has a detrimental effect on the use of nutrients, whereas psychological well-being enhances growth.
Myers (1992) also reviewed the literature to establish the link between growth and mental development. He offers the following summary:
... satisfying psychosocial needs can have an effect on nutritional status through its effect on metabolism linked to stress reduction, and by helping to produce changes in the care demanded and provided. At the same time, nutrition is seen to have an effect on psychosocial development, operating primarily through its impact on attention, responsiveness, independence, irritability, and affect. Nutrition is one of a complex of factors operating to influence that development and associated behavior. (Myers 1992,188)
An application of structural modelling
Quantitative estimates of the relative strength of pathways between elements in the family behaviour model are obtained by specifying and estimating several structural equation models. Structural equations with latent variables permit the analyst to join all the important theoretical relationships in a data set together in the same comprehensive mathematical model. The term "structural" refers to the assumption that the parameters both are descriptive measures of association and reveal an invariant "causal" relation (Bollen 1989). In fact, this technique does not prove causality, but rather tests whether the causal assumptions defined in the model match the associations "found" in a sample of data. The use of structural modelling may be thought of as an attempt to represent explicitly both the direct influence of one variable on another, and the indirect influence that may occur through a third variable. An advantage of structural modelling is that it allows for separate estimates of these direct and indirect effects.
"Path" diagrams, such as those in figures 8.3-8.10, are a convenient way of visualizing these effects. The overall theory is used to specify a set of pathways between a set of idealized variables. Idealized variables are usually abstract concepts that cannot be precisely measured in practice. These idealized variables are referred to as "latent variables" and are shown as ovals on the path diagram. A latent variable is understood to be an underlying cause that influences a set of measurable outcomes, or "measured variables," represented by rectangles. We infer the values of the latent variable indirectly by measuring the outcomes it causes, and by entering these measured variables into factor analysis. The factor on which they load is a mathematical representation of the latent or idealized variable. Paths from each oval latent variable link it to its rectangular measured (or "proxy") variables. A single path from one variable to another represents a single direction of causation. When two paths link a pair of variables, they represent the simultaneous influence of the variables on each other.1 For example, in the Javanese growth model (fig. 8.3), growth is a latent variable that explains some of the variation of two measured variables - weight-for-age Z-score (WAZ), and height-forage Z-score (HAZ).
Prior to estimating a structural model, it is necessary to choose a scale for the latent variables, and to make sure that all of the paths can be identified with the available information. Following common practice, we have allowed each latent variable to have the same scale as one of the measured variables it determines. Our figures present the standardized values of all path coefficients for easy comparison. Estimating a structural model involves estimating the path coefficients and the error variances. The estimates presented here are maximum likelihood estimates generated by the CALIS procedure of the SAS statistical program.2 The path coefficient, like a standardized regression coefficient, can be thought of as a measure of the strength of a relationship (or the percentage of variance in the dependent variable explained by the independent variable).
The coefficients show the relative importance of the pathways. The asterisks reflect the significance levels at which we can reject the hypothesis that the true path value equals zero (* represents p-value between .05 and .025 level, D represents p-value between .025 and .01 Ievel, and # represents <0.01 Ievel).
There is no measure for the overall fit of a system of equations that is as simple to interpret as the R2 measure used for a single regression equation. Two descriptive goodness-of-fit measures are the BentlerBonnet statistic (BB fit index) and the adjusted GFI (goodness-of-fit index): the BB fit index gives the change in the value of a test statistic as a proportion of its value in a baseline model; the adjusted GFI gives the explained proportion of variances and covariances with an adjustment for degrees of freedom.
The total effects of each independent variable of interest can be broken down into direct and indirect effects. While the path coefficients shown in the figures represent the direct effects of one variable on another, the total effects represent both the direct and indirect effects. Both effects are of potential interest for policy purposes.
Technical limitations of the current model
The structural equation modelling technique provides promise for future use because of its comprehensiveness in covering different techniques such as factor analysis, regression, and other econometric procedures. This study is an exploratory analysis to illustrate a technique that will become increasingly important in the areas of family and child development research. Given the exploratory nature of this study, there are some limitations that should be mentioned. We hope that our descriptions of model limitations will give insights for other researchers who will use similar techniques in the future. The limits of this study originated from both the modelling process and the state-of-the-art of the techniques that are not yet well developed. We developed complicated models with very limited sample sizes. We took into account all factors in the family simultaneously in order to eliminate some variable omission bias. This is like a correction for widely used simple techniques such as ANOVA, ANCOVA and other simplistic statistical tools that do not take into account other potential confounding factors simultaneously; therefore, despite the limitations mentioned further below, we consider this to be less biased. Also, the directions of the associations and the significance levels of parameters in this study are consistent with the theory of the family, and other empirical findings using different techniques.
A major problem in the application of maximum likelihood estimation to structural models is that most of the current theory and software is based on the assumption of multivariate normality. In practice, it is not unusual for measured variables to be ordinal scales, even when the latent variables are normally distributed.
Bollen (1989) summarizes some consequences of using categorical variables in estimating these models: first, non-normality can influence the values of test statistics; second, the coefficient estimates for the paths may be attenuated. Despite non-normality, if there is no restriction on the variances and covariances imposed in the exogenous variables, we still can get consistent estimates, and significant tests (K. Bollen 1992, lecture note). Little is known about the robustness of structural equation techniques when categorical variables are used.
Many of the variables in this study are ordinal indices based on responses from survey questions. Because of the uncertain impact of using ordinal data, we do not rely on the commonly used chi-square statistic to evaluate the overall goodness-of-fit of the model. We also are inclined to be cautious about accepting or rejecting paths based solely on the conventional significance levels of test statistics.
Although theoretical results on the impact of using categorical variables are incomplete, simulation studies suggest that the attenuation of coefficients is smaller when ordinal variables include more categories. Since all but two of the variables used in this study have more than five categories, we felt that the known gains from an explicit modelling of the measurement errors would outweigh the possible losses due to the categorical nature of many of the variables.
One suggestion for the treatment of correlations involving categorical data has been to model the categories as representing measures of underlying continuous variables. The correlations of these underlying continuous variables can then be estimated with polychoric or polyserial correlations.3 The usual Pearson correlation tends to understate these correlations by a substantial amount when there are few categories in the ordinal variables.
Since neither SAS nor other software at our disposal computes polychoric and polyserial correlations, we conducted a simple ad hoc test to see if our results were likely to be sensitive to the use of Pearson correlations. Each Pearson correlation value for continuous-ordinal variable pairs and ordinalordinal variable pairs was increased by 25 per cent and the model was reestimated. About 90 per cent of the t-statistics of estimated parameters became more significant with the modified correlations, but the basic conclusions did not change. Therefore, we have elected to present results based on the Pearson correlations with the understanding that the path coefficients may be underestimated.
The presence of multivariate outliers can cause the convergence problem, because of negative error variance. We experience a high degree of difficulty in getting the program to converge. One suggestion from the SAS manual (SAS Institute Inc. 1989) was to impose a boundary statement that forces the error variance to be 0. This can affect the parameter estimates of certain observed variables to be 1. At the time of this study, there is no correctional procedure established in the program to deal with this problem. However, this does not influence the other parameter estimates in the models. Future development of this technique is needed.
Convergence problems forced us to eliminate some covariance pathways between exogenous variables. We had too many parameters to be estimated, given the small sample sizes. For future use of this technique with similar models, larger sample sizes are needed to satisfy all the required assumptions.
Yet another weakness in this form of modelling is the underlying assumption that all relationships can be expressed in linear form. Although transformations of variables may be used to introduce some non-linearities, it is possible that non-linear effects are not fully captured. Progress on this problem awaits new developments in structural modelling.
Finally, since this study tries to explain the outcome indicators with a fully elaborated model of family dynamics, there are many parameters to be estimated with a relatively small number of observations. A standard check to determine whether the sample is likely to generate precise estimates of coefficients is to look at the ratio of sample size to free parameters (Bender 1985). gentler's standard cut-off point is 5:1. In this study, the total complete sample size in both data sets is less than 200 and there are more than 50 free parameters to be estimated, which makes the ratio less than the standard cut-off. While this does not make the model or the estimates invalid, it becomes more likely that the coefficient estimates will have high standard errors and low t-values.