How valid are measures of beliefs about the causes of illness? The example of myocardial infarction

Abstract

In Multi-Trait Multi-Method (MTMM) studies of causal attributions for laboratory events, there is little evidence of convergent and discriminant validity for attribution measures. We report the first MTMM study to investigate the validity of two methods of eliciting causal beliefs for an illness, specifically, myocardial infarction. Adult respondents (N = 107) listed causes of MI, then completed questionnaire rating scales for causal beliefs for MI. Measures were compared using both Campbell and Fiske's approach to MTMM analyses, and a Confirmatory Factor Analysis approach. Neither single item measures causal beliefs, nor scales of causal beliefs derived using exploratory factor analysis provided much evidence of convergent and discriminant validity. Confirmatory factor analysis showed that a model containing only causal belief factors provided a moderately good fit to the data. Adding a questionnaire method factor significantly improved the fit of the model, as well as substantially changing the pattern of factor loadings: loadings of questionnaire items on causal belief factors were markedly reduced. These results highlight major problems with the measurement of causal beliefs, and in particular question the validity of factor analysis of questionnaire measures of causal beliefs. They also suggest that at least some of the MI causal belief factors reported in the literature are artefacts of common questionnaire method variance.

Keywords:

• Causal attribution
• validity
• myocardial infarction
• multi-trait multi-method

Acknowledgements

David French was supported by a Wellcome Trust Prize Studentship (reference Number 047585/Z/96/Z) while this work was conducted. Theresa Marteau was funded by The Wellcome Trust.

Notes

In these analyses, the term “single item” measure denotes both a single questionnaire item, and also the total score for an individual causal belief, as shown in Table I.

An alternative method of estimating association was also calculated, involving dichotomising the frequency count data for the eight categories on the listing task, to reflect “mentioned” versus “not mentioned”. Biserial correlation coefficients were then calculated between these eight dichotomised scores and the ratings on the eight structured questionnaire items. This alternative method produced estimates of association which were highly similar to those reported here. For example, the validity coefficients (in bold in Table III) using respectively, (a) the method employed, (b) biserial correlations are: cholesterol (0.107, 0.130), lack of exercise (0.328, 0.295), eating fatty foods (0.128, 0.105), heredity (0.344, 0.331), high blood pressure (0.196, 0.196), smoking (0.309, 0.309), stress (0.222, 0.244), type of work (0.292, 0.264).