Advances in Consumer Research Volume 19, 1992 Pages 688-695
A FACTOR ANALYTIC INVESTIGATION OF BELK'S STRUCTURE OF THE MATERIALISM CONSTRUCT
Seth R. Ellis, University of San Diego
Materialism has been identified as a potentially important dimension to further the understanding of consumer behavior. A series of articles, principally by Belk, has traced the development of a scale to measure aspects of materialism: possessiveness, nongenerosity, and envy. Belk implies that materialism may be a second-order factor composed of these traits, and possibly other unspecified traits. This research extends the study of the reliability and validity of the materialism construct through the use of confirmatory factor analysis on both the implied first-order and second-order congeneric factor models as well as on a congeneric composite model. All of the models tested did not meet the criteria for most generally accepted goodness of fit indicators. It is suggested that continued work is necessary to develop a set of unidimensional measures for the subscales.
Belk (1984) believes materialism to be a critical but neglected issue in consumer behavior. The materialism construct may be an important segmentation criteria during the 1990's, especially if there is a shift away from a monetary orientation and self-centeredness (Van Gorder 1990). Advertising has been condemned as a force promoting materialism as the means to happiness, status seeking, social stereotypes, shortsightedness, selfishness, and preoccupation with sexuality and conformity. But measures for materialism have not been available to assess empirically any of these charges (Belk & Pollay 1985; Pollay 1986; Lantos 1987).
In addition to the assessment of a possible causal link with a variety of social ills facing the modern world, measures for materialism may also be useful marketing tools in the quest for profitable segmentation strategies. Response functions for segments based on ranges within the materialism dimension may differ for various product types, product quality, and service quality.
Belk hypothesizes that materialism is composed of several traits including, but not necessarily limited to, possessiveness, nongenerosity, and envy. Possessiveness is defined as "the inclination and tendency to retain control or ownership of one's possessions (Belk 1983, p. 267)." However, anecdotal evidence suggests that the role of objects as cultural icons to emphasize social ties or accomplishments may not be related to the possessiveness component of materialism (Wallendorf and Arnold 1988). Nongenerosity is defined "an unwillingness to give possessions to or share possessions with others (Belk 1984)." Belk (1985, p. 268) reports Schoeck's definition of envy as "displeasure and ill will at the superiority of another person in happiness, success, reputation, or the possession of anything desirable."
After traditional scale development for the three measures of materialism traits, Belk found moderate reliabilities for the scales based on the inter-item and item-to-total correlations. His findings showed coefficient alphas of .68, .72, and .80 for the scales measuring possessiveness, nongenerosity, and envy, respectively.
Belk (1985) also offers evidence that the three subscales are oblique factors of the materialism construct by finding non-zero correlations between the sums of the items for each subscale. This furthers speculation that materialism may be a second-order factor responsible for the correlation among these first-order factors. However, based on this correlation matrix and other evidence Belk (1985) proposes the use of the aggregation of all three scales as a measure of materialism, and believes it to be adequate for the exploration of some of the macro issues facing consumer behavior with respect to materialism.
The nature of the materialism construct, with respect to these subscales and their measures, can be further tested with hypotheses by applying the confirmatory factor analytic techniques of Joreskog (1969). Recent developments have emphasized the theory-testing abilities of this technique (Bentler & Bonett 1980; James, Mulaik and Brett 1982). Competing hypotheses regarding the factor structure can be tested with various chi square statistics. When an a priori model has been specified, the goodness of fit of the model to a set of data can be determined (Joreskog & Sorbom 1986).
The purpose of this study was to use the confirmatory factor-analytic methodology to provide a further test of Belk's scale items for the first-order constructs of possessiveness, nongenerosity, and envy. If these scales are satisfactory measures of the first-order constructs, then this methodology also can be used to test a hypothesis that materialism is a second-order factor composed, in part, of some combination of these constructs.
The Measurement Model
The three dimensions of materialism are identified in Figure 1. Items P1 through P9 are Belk's measures for the possessiveness construct, items N1 through N7 are measures of nongenerosity, and items E1 through E8 measure envy. The arrows from each construct to the items are factor loadings and the lone arrows represent measurement error. Uncorrelated measurement errors are assumed. The curved lines linking the factors represent correlations among the factors. The specific statements for the scale items are found in Table 1.
The data for the tests were gathered as part of a larger survey conducted in a large, southwestern metropolitan area. Personal interviews were conducted and the interviewers read the statements to the respondents. Each respondent was given a coding card with a seven-point scale from one equals strongly agree to seven equals strongly disagree and asked to rate each statement. There were 148 usable responses gathered. Two-thirds of the respondents were female.
THE FIRST ORDER FACTOR MODEL OF BELK'S MATERIALISM CONSTRUCT
The construct validity of the factors was evaluated by Joreskog's confirmatory factor analysis procedure using LISREL VII (Joreskog and Sorbom 1989). A one-factor and an oblique, three-factor model
BELK'S MATERIALISM SUBSCALE ITEMS
were specified and compared to a null model. A reference metric was provided in each model by setting all of the diagonals in the factor correlation matrix (PHI) to one.
Confirmatory factor analytic techniques set a priori definitions of the theoretical factor structure. Factor loadings can be constrained to zero thus allowing the estimation of pure factors in the sense that the test will load only on the targeted factors. In all of the models tested here, the diagonal elements of the factor correlation matrices (PHI) were set at one to fix the scales of measurement so that the common factors would have unit variance. This puts all of the parameter estimates in a standardized metric. Each model was estimated with the correlation matrix in order to find appropriately comparable values of the root mean square residual (RMSR).
The confirmatory factor analysis literature has yet to settle on one particular criterion for model acceptability. Commonly used criteria are included in Table 3 for each model tested. LISREL 7 output provides the chi square statistic (CHISQ), the degrees of freedom (df), a goodness of fit index (GFI), an adjusted goodness of fit index (AGFI), and the root mean square residual (RMSR). Smaller values for the chi square statistic, the chi square statistic divided by the degrees of freedom (CHISQ/df), and the root mean square residual indicate better fitting models. Better model fits are indicated as the goodness of fit indicator and the adjusted goodness of fit indicator approach 1.0.
Another useful index reflects the improvement in fit of one model with respect to another model, typically the null model. The normed fit index (Bentler and Bonett 1980) reflects the proportion of chi square eliminated by the competing model. This index ranges from zero to one; Bentler and Bonett suggest a minimum value of .90 when the competing model is compared to a fully-restricted null model. For confirmatory factor analysis the fully-restricted null model specifies an individual factor for each measurement item with factor variances set equal to one and no covariance is allowed among the factors. Also, all of the factor loadings are set equal to one and there is no correlation specified among the errors in the measurement variables. Such a model should result in a large chi square value that many theoretically grounded alternative models will surpass with respect to various goodness of fit indices based on estimations from a reasonable data set.
The literature suggests several potential models. In one study, Belk (1985) aggregated the items in the three subscales to form a one-factor solution. In fact, he advocates its use as an overall factor for materialism (see Model 1 in Table 2). In discussing the subscales, Belk alludes to an alternative possibility that materialism is a second-order factor composed, in part, of the three first-order subscales of possessiveness, nongenerosity, and envy. If this is true, substantial positive correlation should be evident between the first-order constructs. Another interpretation of the situation is that materialism is a composite construct composed, in part, of the three subscales. Such an interpretation does not require positive correlations among the component constructs.
Belk's research suggests that an appropriate first-order model may be the oblique, three-factor, congeneric model (shown as Model 2 in Table 2). This model permits correlations among the three factors but maintains unidimensional measures. Indeed, a successful second-order model would be expected to have positive and significant correlations among the first-order factors. Perusal of the correlation matrix for the three factors and their respective standard errors indicates that only the correlation between possessiveness and envy is not significant. Model 3 in Table 2 is structured in a similar fashion but reflects the results of discarding those items with large modification indices (>5) on factors for which they are not theoretically prescribed. These large indices suggest that the items are not congeneric and, therefore, they were discarded in the interest of maintaining unidimensonal measures. Scale items that are not unidimensional in nature are usually considered unacceptable because they violate an assumption of test theory (see Lord and Novick 1968). However, from a substantive stance, discarding those items to produce a more parsimonious model may change the nature of what the scales purportedly measure.
Table 3 summarizes the fit indices for the theoretical models and the empirically derived reduced models as well as two null models. The first null model (MA) was developed for the 24 item models and the second null model (MB) was developed for the 16 item, reduced models. As expected neither provides a particularly good fit on any of the indices.
Although Belk argues for the aggregation of the three factors into one general factor called materialism, this model (M1) is not particularly successful. Although there is some improvement over the null model, the normed fit index of .29 is not near the recommended value of .90 for a good fitting model. The other indicators also suggest a poor fit with the data.
Models 2 and 3 (M2, M3) are both oblique, congeneric three-factor models with the reduced version (M3) showing the results of the removal of those items that may not be unidimensional. This reduced version compares very favorably with the full 24 item model. Nonetheless, the fit indices are not conclusive: the adjusted goodness of fit indicator almost reached .90, the root mean square residual did slip below .10, but the normed fit index of .67 did not approach the suggested minimum for the fully restricted null model of .90. Also, this model, and all of the empirically derived reduced models, may capitalize on the individual nature of this particular data set.
The literature suggests that materialism may be a higher-order construct composed, in part, of the scales just investigated. The implied second-order factor model is shown in Figure 2. Although not all three subscales are positively and significantly correlated with each other, it might be argued there are enough positively correlated factors that might be capturing some aspects of this higher-order materialism construct.
The goodness of fit indicators associated with the second-order factor model (M4 in Table 2) suggest this model is essentially equivalent in quality to the three-factor first-order models (M2) but not to the reduced version of the three-factor first-order model (M3). In other words, none of the indicators suggest the second-order model is a good representation of the data, but it fares no worse than the relevant competing models tested here.
Although a second-order factor structure is theoretically defensible from the previous research that has been done, the structure of materialism could be specified as a composite construct, that is, as an aggregation of constructs, much like socioeconomic status in that it is considered a composite of social class, education and occupation and not a causal factor. This view of materialism does not require positively correlated constructs, but simply defines the materialism construct as an aggregate of subscales. Also, as Belk points out, there may be other, as yet unidentified, constructs that are additional constituents of materialism. This composite view of the materialism construct is shown in Figure 3 and for this particular model equal weights are specified for the three subscales by fixing ones in the diagonal of the gamma matrix. Additionally, the possibility of other constructs influencing the composite materialism construct is modeled by freely estimating the size of the uniqueness associated with the composite. The goodness of fit indices resulting from testing this model are listed as models 5 and 6 (M5, M6) in Tables 2 and 3. The full composite model appears to be essentially equivalent to both the full, second-order model and the full, three-factor, first-order model. Analogous results are apparent for the reduced version of the composite model, thus the models fare no better and no worse than their respective first and second-order counterparts.
FACTOR LOADINGS AND FACTOR CORRELATIONS
GOODNESS OF FIT MEASURES
MATERIALISM AS A SECOND ORDER FACTOR
MATERIALISM AS A COMPOSITE CONSTRUCT
SUMMARY AND CONCLUSION
Materialism has been identified as a potentially important dimension to further the understanding of consumer behavior. A series of articles, principally by Belk, has traced the development of a scale to measure aspects of materialism: possessiveness, nongenerosity, and envy (Belk 1982, 1983, 1984a, 1984b; Belk and Pollay 1985). Belk implies that materialism may be a second-order factor composed of these traits, and possibly other unspecified traits. This research extends the study of the reliability and validity of the materialism construct through the use of confirmatory factor analysis on the implied first-order and second-order factor models as well as a composite model alternative. None of the tested models were acceptable from the point of view of generally accepted goodness of fit indicators, however, the versions using a reduced set of scale items performed markedly better than the full set of scale items. The reduced set was empirically derived from the data and the models tested were not specified as a priori tests so their validity is questionable.
It appears that a three-factor model is a better specification than an aggregated one-factor model, however, continued work is necessary on the scale development. One of the problems suggested by the tests is that a substantial number of the items are not congeneric. Indeed, the removal of a group of items with large modification indices immediately improved the fits of all the models tested. If unidimensional measures are sought, the overall suggestion from these tests is to continue modification of the subscales possessiveness, nongenerosity and envy so that each of the items constitutes a unidimensional measure of the respective constructs. The various goodness of fit indices suggest that any complex theories based in part on the present operationalization of these materialism subscales will be problematic, at least if they are specified as congeneric structures. As Belk has cautioned, the use of these particular scales continues to be recommended for exploratory research only.
Belk, Russell (1982), "Acquiring, Possessing, and Collecting: Fundamental Processes in Consumer Behavior," in Marketing Theory: Philosophy of Science Perspectives, Chicago: American Marketing Association, 185-190.
Belk, Russell (1983), "Worldly Possessions: Issues and Criticisms," in Advances in Consumer Research, 10, eds. Richard P. Bagozzi and Alice M. Tybout, Ann Arbor: Association for Consumer Research.
Belk, Russell (1984a), "Cultural and Historical Differences in Concepts of Self and Their Effects on Attitudes Toward Having and Giving," in Advances in Consumer Research, 11, ed. Thomas C. Kinnear, Provo, UT: Association for Consumer Research, 753-760.
Belk, Russell (1984b), "Three Scales to Measure Constructs Related to Materialism: Reliability, Validity, and Relationships to Measures of Happiness,"in Advances in Consumer Research, 11, ed. Thomas C. Kinnear, Provo, UT: Association for Consumer Research, 291-297.
Belk, Russell (1985), "Materialism: Trait Aspects of Living in the Material World," Journal of Consumer Research, Vol. 12, No. 3, 265-280.
Belk, Russell and Richard W. Pollay (1985), "Images of Ourselves: The Good Life in Twentieth Century Advertising, Journal of Consumer Research, 11, 887-897.
Bentler, P.M. & D.G. Bonett (1980), "Significance Tests and Goodness of Fit in the Analysis of Covariance Structures," Psychological Bulletin, 88, 588-606.
Campbell, D. T. and D. W. Fiske (1959), Convergent and Discriminant Validation by Multitrait-multimethod Matrix," Psychological Bulletin, 56, 81-105.
James, L.R., S.A. Mulaik, & J.M. Brett (1982), Causal Analysis: Assumptions, Models and Data, Beverly Hills, CA: Sage.
Joreskog, Karl G. and Dag Sorbom (1989), LISREL 7 User's Reference Guide, Mooresville, IN: Scientific Software.
La Du, Terence J. and J. S. Tanaka (1989), "Influence of Sample Size, Estimation Method, and Model Specification on Goodness-of-Fit Assessments in Structural Equation Models," Journal of Applied Psychology, 74, No. 4, 625-635.
Lantos, Geoffrey P. (1987), "Advertising: Looking Glass or Molder of the Masses?," Journal of Public Policy and Marketing, 6, 104-128.
Lord, F.M. & M.R. Novick (1968), Statistical Theories of Mental Test Scores, Reading, MA: Addison Wesley.
Marsh, H. W., J. R. Balla, and R. P. McDonald (1988), "Goodness-of-Fit Indices in Confirmatory Factor Analysis: The Effect of Sample Size," Psychological Bulletin, 103, 391-410.
Marsh, H. W. and Dennis Hocevar (1988), "A New, More Powerful Approach to Multitrait-Multimethod Analyses: Application of Second-Order Confirmatory Factor Analysis," Journal of Applied Psychology, 73, 107-117.