Scaling Methods for Categorical Self-Assessed Health Measures

Presenter: Patricia Cubí-Mollá, University of Alicante

Abstract

Rationale:
The lack of a continuous health valuation is a major drawback in health analyzes over broad populations. On the contrary, general surveys usually include self-assessed health questions, where the respondents must choose among different health levels or categories. The use of these categorical responses to approximate a continuous health variable is a usual procedure in health studies. The most common approaches (ordered probit/logit model and interval regression model) assume that health is an unobservable, latent variable that is normally distributed. However, this is a rough assumption, since many studies have reported skewness in the distribution of health (that is, a great majority of the population reporting good health).

Objectives:
In this work we propose methods for scaling SAH measures, considering that the latent health variable is log-normally distributed. These new methods are compared to the existing scaling procedures. The purpose is to examine to what extent the actual health valuations can be approximated by the predicted values of the scaling approaches.

Data:
Data from the Catalonia Health Survey 2006 are used to provide the results (a sample of 15,648 individuals). In this survey every respondent reports a categorical SAH evaluation. Also, utility values can be derived by means of the EQ-5D descriptive system provided by the survey. Thus, we can take advantage of having actual continuous health values in this study. The validity of the scaling approaches is assessed by measuring to what extent the health values derived from SAH (conditioning on a set of socioeconomic factors) suit the actual health values, available in the survey.

Results:
In general, models under log-normality outperform the other approaches. In particular, the ordered probit model (considering health as normally distributed) is clearly surpassed by the others. The interval regression model under normality approximates the actual health tariffs in a similar way to the same model under log-normality; however, the latter seems to match better the lower values. Surprisingly, the ordered probit model under log-normality procedure is the one that better approximates the distribution of health, especially at working with the VAS tariff. It is also the regression model that is closer to the ordinary least-squares predictions.

Conclusions:
If defining health-related quality of life measures with cardinal properties from continuous variables is a demanding challenge, then, obtaining them from ordinal variables is even a more difficult task. Assigning a numerical valuation for a category only masks the ordinal relationship between categories. If the main goal of an analysis is obtaining quality weights from health states, regression methods are, therefore, a powerful tool as scaling procedures. The results obtained in this paper can provide a new benchmark for the proper cardinalization of health measures.