Tests of Data Quality, Scaling Assumptions, and Reliability of the Danish SF-36

Abstract

We used general population data (n = 4084) to examine data completeness, response consistency, tests of scaling assumptions, and reliability of the Danish SF-36 Health Survey. We compared traditional multitrait scaling analyses to analyses using polychoric correlations and Spearman correlations. The frequency of missing values was low, except for elderly people and people with lower levels of education. Response consistency was high and compared well with results for the U.S. SF-36. For respondents with computable scales in all eight domains, scaling assumptions (item internal consistency, item discriminant validity, equal item–own scale correlations, and equal variances) were satisfactory in the total sample and in all subgroups. The SF-36 could discriminate between levels of health in all subgroups, but there were skewness, kurtosis, and ceiling effects in many subgroups (elderly people and people with chronic diseases excepted). Concerning correlation methods, we found interesting differences indicating advantages of using methods that do not assume a normal distribution of answers as an addition to traditional methods.

Introduction

A critical step in the development of health status measures is to evaluate their measurement properties. In the evaluation of such questionnaires, the health status measurement field has to some extent developed criteria that are unique to or especially important in the health field (e.g., response burden and responsiveness), but most criteria and techniques have been adopted from the area of psychological testing. In particular, statistical techniques from classical psychometrics [1] have been adopted to examine scaling assumptions, reliability, and validity. In principle, these techniques assume that data are continuous and normally distributed, but the techniques often work well also with categorical rank scaled data 1, 2, that is, the kind of data used in health status measurement. The use of methods from classical psychometrics in the evaluation of health status measures may, however, still be debatable, as health status data are often skewed.

The psychometric properties of health status measures are examined as part of the development process, but it is also recommended to do new psychometric analyses when a questionnaire is translated into another language [3]. Further, because the validity and reliability of a questionnaire are specific to the setting and the population [1], a check of psychometric properties may be appropriate if the questionnaire is to be used in new settings or with groups for whom it has not previously been tested.

This article concerns the psychometric properties of the Danish translation of the MOS SF-36 Health Survey. The translation and adaptation of the Danish SF-36 followed the procedures of the International Quality of Life Assessment (IQOLA) project 3, 4, 5, an international collaboration to translate and implement the SF-36 across languages and cultures. Initial studies have shown the adequacy of the Danish translation as judged from backward translations and from independent evaluation of the conceptual equivalence, use of common language, and clarity of the Danish translation [5]. Factor analytic studies have established that the factor structure is very similar to the U.S. version 6, 7. Analyses by the Rasch model of the Physical Functioning scale and contingency table analyses of differential item functioning (DIF) in all scales have found that some items show DIF but that these cases of DIF have only a small impact on the total score for each scale in comparisons of general population data 8, 9.

We present data here on the psychometric properties of the Danish SF-36, focusing on data completeness, response consistency, tests of scaling assumptions, and the reliability of the Danish SF-36. As has been done in studies of the U.S. SF-36 [10], these properties are analyzed for the total population and subgroups defined by age, gender, education, and disease status. In addition, we investigate some more general issues concerning the choice of different statistical methods in testing scaling assumptions. Given that the Danish general population sample data used in this paper include many healthy people, our data are more skewed than the data originally used to validate the SF-36 in the United States 10, 11. For this reason, we examined whether our results changed when using methods that are not dependent on the assumption of normal distribution of the observed variables.