|Home | About | Journals | Submit | Contact Us | Français|
RESEARCH IS the tool that scientists use to further their knowledge of the world. Scientists ask questions about phenomena and design research studies to answer those questions. In many instances, scientists are interested in what causes certain outcomes to occur. The most powerful and direct way to establish cause and effect is to design an experiment. In order to be considered an experiment, the study must have a control (or comparison) group and at least one experimental (or treatment) group; subjects must be assigned to groups randomly, not by choice of the subject or the investigator; and the researcher must manipulate the independent variable so that the experimental group gets the treatment and the control group does not (Polit & Hungler, 1991).
When random assignment to group is not feasible, a quasi-experimental design that requires a comparison group and manipulation of the independent variable may be used; however, in some circumstances, manipulation of the independent variable is not possible, forcing the researcher to use naturally occurring groups. For example, suppose the researcher is interested in the effects of a mother’s employment on her child’s development. One way to study this would be to compare children of employed mothers (“treatment group”) with children of nonemployed mothers (“comparison group”). Although there are two groups, random assignment of mothers to employment or nonemployment would raise both practical and ethical concerns. Manipulating either the mother’s employment status or the number of hours the mother is employed (the independent variable) also raises these concerns. Thus, the researcher must use naturally occurring groups, a design that provides very weak evidence of causality. Causal modeling, a methodological and analytical technique, may provide stronger evidence about the direction and strength of cause-and-effect relationships hypothesized by the researcher. The purpose of this two-part series is to discuss the conceptual basis for causal modeling and to provide the consumer of nursing research with guidelines for evaluating published reports of research that used causal modeling techniques.
Nursing phenomena are often complex with many variables in the causal chain. One shortcoming of the experimental design study is that it fails to provide information about the effects of other variables on the independent or dependent variables because random assignment of subjects to groups breaks any relationships between these other variables and the independent variable. Removal of these relationships strengthens the evidence for causality, but it does not represent the complexity of the process by which an outcome occurs in the real world. Some statistical techniques allow the researcher to analyze the effect of independent variables on a dependent variable while controlling for the effects of other variables on the process. These techniques tell us how much influence a variable exerts directly on the dependent variable (direct effect). However, sometimes a variable’s influence is transmitted through another variable (indirect effect). In that case, the total amount of influence the variable has on the dependent variable is underestimated by these techniques because the indirect effect is not considered. Causal modeling can be used to represent very complex relationships among a set of variables, including direct and indirect effects, which increases the model’s correspondence with reality.
Causal modeling requires the researcher to construct a model to explain the relationships among concepts related to a specific phenomenon (Asher, 1983). A causal model is a diagram of the relationships between independent, control, and dependent variables. A path model is a less complex type of causal model. Causal models are also known as “structural equation models,” “covariance structure models,” or “LISREL models.” Each of these terms has a slightly different meaning based on the specific analysis technique used but are often used interchangeably in the literature.
Construction of a causal model can be accomplished in two basic ways. The preferred method is to build the model before the study begins based on theoretical expectations. In this method, the researcher considers the relationship between the independent and dependent variables of interest. Then variables that could affect either the independent or the dependent variable or the relationship between them are added to the model, and the direction of their effects is specified. Each linkage must be supported by expectations from other theories (theoretical evidence), by results of research (empirical evidence), or by experiential knowledge. An example from the author’s research is provided to illustrate how support for the proposed linkages is described.
The causal model in Figure 1 is a simplified version* of the expected system of relationships between maternal employment status and child development that is currently being used to guide a study of this phenomenon with preschool children, some born prematurely and some born at term, and their families. Each of the linkages can be supported theoretically or empirically. Research suggests that employment has a negative effect on mother-child interaction, at least in infancy (Barglow, Vaughn, & Molitor, 1987; Belsky & Rovine, 1988), although other studies find no effect or a positive effect (Schubert, Bradley-Johnson, & Nuttal, 1980). Because of these conflicting findings, we are positing a relationship between employment and mother-child interaction, but we are not specifying whether this will be a positive or a negative relationship. Consistency between employment attitudes and employment status repeatedly has been found to have a positive effect on mother-child interaction (Benn, 1986; Bronfenbrenner, Alvarez, & Henderson, 1984; Hock, 1980).
Family functioning is defined as the quality of relationships among family members. Again, research findings are conflicting but support the importance of investigating the effect of employment on family functioning. Some researchers find no adverse effects on marital satisfaction (Locksley, 1980), and others find negative effects (Hardesty & Betz, 1980). In a series of studies, Youngblut, Loveland-Cherry, and Horan (1991, 1993, in press) found that parents in nonemployed-mother families were more satisfied with their family’s relationships than parents in employed-mother families when the family’s preterm infant was 18 months old (in press), but not at 3 months (1991) or at 9 and 12 months (1993). The path from consistency to family functioning is supported by a positive relationship between degree of choice in the employment decision and parents’ perceptions of family relationships and family cohesion (Youngblut et al., in press).
The link between mother-child relationship and child development is strongly supported by the work of many researchers (see Sameroff & Chandler, 1975, for review of the full-term literature; or Meisels & Plunkett, 1988, for review of the preterm literature). Because research on the effects of family functioning on preterm infant development is sparse, this link is supported with theoretical literature. Family systems theory (Friedman, 1986) suggests that quality of family relationships will affect the individual members of the family. Therefore, if the mother-child relationship has an effect on child development, it is likely that family relationships will also have an effect on the child’s development. In addition, both Bronfenbrenner (1985) and Hoffman (1989) have advocated broadening inquiry beyond the mother-child relationship to include the family. Finally, the effect of mother’s education level on child outcomes is included as a control variable with empirical support by many researchers (Heynes, 1982; Hoffman, 1989). When constructed theoretically, the causal model will serve as a guide for study design and implementation and for data analysis.
Causal models can also be built empirically. In this case, construction of the model is based on the correlations among the variables that were actually obtained in the study. Although this may be more appealing and perhaps simpler, there are several issues that need to be considered. First is the issue of relying on simple (bivariate) correlations to indicate relationships. A correlation between two variables indicates the total amount of overlap between the two variables. This overlap represents the combination of direct effects, indirect effects, and other effects. Direct effects are those that are due to one variable when the effects of other variables are controlled. Indirect effects are transmitted through intervening, or mediating, variables. Other effects could be due to sampling or measurement error or to the effect of another variable on both of the variables being correlated (Pedhazur, 1982).
For example, if the correlation between employment and child development is high, one might consider putting a direct path between the two variables (Figure 2, solid line, top diagram). However, it may be that the quality of the mother-child relationship mediates the effect of employment on child development (Figure 2, broken line, top diagram). If so, controlling for quality of mother-child interaction would show that there is really no correlation between employment and development. Or it may be that the correlation between employment and development (Figure 2, curved broken line, bottom diagram) is due to the effect of neonatal health status on both employment and development (Figure 2, solid lines, bottom diagram); in other words, the correlation is a coincidence (known as a spurious correlation) that is due to the impact of a third variable on the other two. Thus, relying on the bivariate relationships would be misleading and would obscure the “true” path of the effect.
Importantly, using the correlations obtained with the study’s data changes the analysis from hypothesis testing to a fishing expedition. Causal models represent a system of hypotheses about the interrelationships of variables when the model is constructed theoretically before the data are collected. Testing the model then represents testing each of the implied hypotheses. However, if the model is constructed from the data, it can no longer be considered to represent a system of hypotheses and the analysis becomes post hoc exploration of the data. This practice of data snooping often yields results that are not theoretically defensible, do not make sense conceptually, and are less likely to be replicated in future research.
In summary, causal modeling is an important tool for knowledge development in nursing, and its use is increasing. Therefore, consumers of nursing research must have some familiarity with the conceptual underpinnings of causal modeling and an appreciation of how the technique may be used. In addition, the readers’ ability to critique published reports of model testing is crucial so that they can evaluate the results of these studies. The first part of this two-part series was aimed at increasing the reader’s level of sophistication regarding causal modeling techniques. The second part will provide guidelines for evaluating published reports of research.
Supported in part by a grant from the National Institute for Nursing Research (No. ROI-NR02707) and by an administrative supplement from the Office of Research on Women’s Health through the National Institute for Nursing Research.
Address reprint requests to JoAnne M. Youngblut. PhD. RN. Frances Payne Bolton School of Nursing. Case Western Reserve University, 10900 Euclid Ave, Cleveland, OH 44106-4904.