Multidimensional Assessment of Spirituality/Religion in Patients with HIV: Conceptual Framework and Empirical Refinement

Abstract

A decade ago, an expert panel developed a framework for measuring spirituality/religion in health research (Brief Multidimensional Measure of Religiousness/Spirituality), but empirical testing of this framework has been limited. The purpose of this study was to determine whether responses to items across multiple measures assessing spirituality/religion by 450 patients with HIV replicate this model. We hypothesized a six-factor model underlying a collective of 56 items, but results of confirmatory factor analyses suggested eight dimensions: Meaning/Peace, Tangible Connection to the Divine, Positive Religious Coping, Love/Appreciation, Negative Religious Coping, Positive Congregational Support, Negative Congregational Support, and Cultural Practices. This study corroborates parts of the factor structure underlying the Brief Multidimensional Measure of Religiousness/Spirituality and some recent refinements of the original framework.

Technical Appendix

Factor Analytic Techniques, Measurement Models, and Hypothesis Testing Criteria

Factor Analysis Techniques

Two major distinctions are important in factor analysis procedures. First, exploratory factor analysis (EFA) differs from confirmatory factor analysis (CFA). Furthermore, there are two major types of EFA: principal components analysis (PCA) and exploratory-common factor analysis (E-CFA). EFA is data-driven: items are inputted with a view to reduce a larger number of items to a smaller number of factors. In contrast, CFA is theory-driven: a specific configuration of factor structure is posited, and CFA provides a test of whether that structure fits the data. Although PCA is often used, it has two limitations. First, it yields factors that are not necessarily unidimensional, even when each item “loads” only on one factor, because it is not specific to an underlying measurement model. Second, PCA provides the best fit of a set of items, but not necessarily a robust fit; PCA results often are not corroborated when CFA “tests” the PCA-suggested factor structure. This latter limitation is true of E-CFA as well.

Evaluation of CFA Models. Because CFA is theory-driven, the idea of “good model fit” has two aspects that are best separated analytically. First, one must determine whether one can trust the parameter estimates (e.g., the loadings and correlations among factors). And second, one must determine whether the parameter estimates corroborate one’s measurement ideas. CFA determines a “best solution possible” by iteratively deriving parameter estimates for the hypothesized model and comparing the inputted covariance matrix with the covariance matrix derived from the parameters. This process of successive approximation results in a solution that minimizes the differences between the two matrices. However, this “best solution possible” may not be a good fit per se. The foremost measure of goodness-of-fit is a chi-square analysis that attempts to ascertain whether the two matrices differ only by an amount falling within sampling variability. A relatively low chi-square value with a correspondingly high P value (given the degrees of freedom) indicates that one can trust the parameter estimates. However, chi-square is confounded by sample size because chi-square increases with sample size. Thus, a number of indicators are generally examined to determine whether the parameters estimates are dependable. Nevertheless, being able to trust the parameter estimates is a necessary but not sufficient condition for the hypothesized model’s corroboration. Of course, untrustworthy parameter estimates indicate a poor model fit and would lead one to reject the theorized model. However, if the indicators of fit suggest that one can trust parameter estimates, one must examine these estimates to determine whether they conform to the hypotheses, i.e., that they are statistically significant and sufficiently large versus statistically insignificant.

Measurement Models

One typology distinguishes between reflective and formative measurement models (Technical Appendix Fig. 1). A reflective model is one in which correlations among items are due to their common cause, the latent variable. A unidimensional factor or a group of unidimensional factors will take the form of a reflective measurement model(s). In contrast, a formative measurement model is one in which component parts (items) make up the whole (latent variable), i.e., they form the latent variable. The component parts (items) can be, but do not have to be, correlated; here, item correlations are not due to the latent variable. Although reflective models are typically validated by using psychometric techniques, such as factor analysis, such techniques should not be applied to validate formative models. Validation of formative models occurs initially through establishing content validity—that is, making sure that all key elements of the target/latent variable are incorporated into the measure. Inter-item correlations provide some indication for the type of measurement model that fits a particular scale, but do not constitute conclusive evidence.

Fig. 1

Reflective versus formative measurement models and first and second (higher)-order measurement models

Among reflective measurement models, different “ordered” factors are possible (see Technical Appendix Fig. 1). Items may parse into subsets with each subset loading on one factor. The factors are first-order factors. The set of factors may logically correlate because they share a common cause, a higher-ordered factor (a second-order factor).

Hypothesis Testing Criteria

The analysis had two objectives: to determine whether the parameter estimates (e.g., the loadings and correlations among factors) could be trusted and to examine whether the parameter estimates corroborated the hypothesized measurement ideas. The study proceeded by first running a model with a sample size of 200 so that the chi-square statistic reflected the difference in the inputted covariance matrix and the matrix derived from the best fit loadings, with minimum distortion due to sample size. When examination of the indicators of fit suggested that the parameter estimates were trustworthy, the model was rerun with the actual sample size of 450 patients to test the statistical significance of individual parameters. Indicators of fit considered to be noteworthy and trustworthy were as follows: A model that fits well would have a small chi-square value with a correspondingly larger P value (exceeding 0.05), indicating that the inputted observed variances and covariances do not differ from the variances and covariances based on the model’s solution, to a greater extent than expected due to sampling variability. Other indicators of a good fit were high goodness-of-fit index (GFI) and adjusted goodness-of-fit index (AGFI) values—the closer these values were to 1.0, the better the fit. With a large number of variables, values greater than or equal to 0.8 were desirable. Fewer observed variables would increase the acceptable thresholds. The standardized root mean square residual (SRMR) is a type of standard error reflecting the difference in corresponding elements of a correlation matrix for the variables based on the estimated loadings versus those of the inputted covariance matrix. The tested model should be interpreted as being “drawn” from a population where about two-thirds of the differences in correlations are between plus or minus the SRMR value; if the two correlation matrices are drawn from the same population, one would expect the SRMR to be small. Furthermore, the Critical N indicates the sample size at which the P value for the chi-square statistic equals 0.01. In a well-fitting model, the Critical N value is 200 or greater. (Note that many would accept a chi-square statistic with an associated P > 0.01 as indicating an acceptable model fit, particularly when the number of observed variables is large.) The Critical N presented was the “tipping point” for P = 0.01 rather than P = 0.05. Finally, the study followed the convention that an item was not acceptable if it loaded on more than a single factor (items must have had a factor complexity equal to 1).