ABSTRACT

Background: Researchers often use census-derived measures of socioeconomic status (SES) when personal information is not available. Theory predicts that the resulting misclassification will blunt associations between outcomes and SES and that control for confounding by SES will be less effective. The purpose of this paper was to examine the magnitude of this problem using data from the National Population Health Survey (NPHS).

Methods: Subjects were 4,037 respondents to the NPHS who were linked to the Ontario Health Insurance Plan. An ecologic measure of income was obtained by linkage of subjects' postal codes to the Census.

Results: The relationships between the ecologic-level measure and health outcomes or health services utilization were attenuated in comparison to the relationships relative to the direct measure of household income. The ecologic measure also produced poorer control for confounding by income in the analysis of other health relationships.

Conclusions: Many interesting public health and health services questions can be addressed only with the use of ecologic level socioeconomic information. While most of the results were qualitatively similar when the direct and ecologic measures were compared, researchers and users of research findings should be aware that attenuated or potentially misleading findings may result from the use of these methods.

Personal or household incomes are important correlates of health status and the utilization of health services in Canada,1 the United States,2 and Europe.3 Information on personal measures of income and socio-economic status (SES) is occasionally available to researchers, particularly when analyzing national surveys such as the National Population Health Survey (NPHS). In many cases, however, personal information is not available, and researchers are obliged to infer personal SES from ecologic measures obtained, for example, from census data.4

Canadian researchers make use of administrative data to study questions pertinent to health care and public health policy.5,6 There has been worry about the possibility of misclassification of SES in the use of these data. Demissie7 found substantial discrepancy between area-based SES measures and SES assessed at the individual level in Montreal. Sin8 studied the validity of using postal codes to identify groups of people with low socio-economic status. he reported that the use of Forward Sortation Area (FSA) postal codes as the only marker to identify people with low SES may result in substantial misclassification of personal poverty. Glazier9 reported that, although address inaccuracy was found in Ontario's health care registry, serious socio-economic misclassification occurred at a low rate and did not introduce significant bias in the calculation of hospitalization rates by socio-economic group. Southern et al.10 assessed the agreement between FSA and Enumeration Area (EA) derived income levels in Alberta. They found that the variability in EA-derived income quintiles was large for any given FSA-derived income quintile and recommended that EA-derived measures should be used when individual data are not available.

Measurement theory predicts that a group level measure will contain more measurement error than the direct household measure. Accordingly, the relationship between the group level measure and a health outcome will be attenuated in comparison to the true relationship relative to the direct measure of socio-economic status.11 Mustard and colleagues12 examined the validity of using ecologie measures of SES as proxies for individual-level measures by using a 5% sample of households in Manitoba and linking the records of the Manitoba Health Services Insurance Plan and the Census of 1986. They concluded that the hypothesis that risk estimates derived from ecologie income measures will be attenuated relative to estimates obtained from household income was not supported. The purpose of this paper was to make use of a unique data set to re-examine the effect on measures of health status, and health care utilization, of using ecologie rather than personal measures of income.

METHODS

The subjects were the Ontario respondents to the first wave, 1994/95, of the NPHS.13 Respondents were asked about their household income, and Statistics Canada derived an index of income adequacy on the basis of household income and household size. In 1994 dollars, the four categories of the index were as follows: lowest income, less than $10,000; lower middle income, $10,000 to $22,499; upper middle income, $22,500 to $59,999; and highest income, $60,000 or more for 1 or 2 people. For households with more than 2 people, the income ranges were scaled accordingly.

The Ontario Ministry of Health linked the NPHS file to the Ontario Health Insurance Plan (OHIP) providers' database, using the respondent's health insurance number as the identifier. The Ministry generated a data file containing a record of each service provided to NPHS respondents by physicians in the year before their interviews. Postal codes from the OHIP file were used to link subjects to Enumeration Areas from the 1996 Census, and median household incomes in each EA were extracted.

Several measures of self-reported health status, and one of health services utilization, were selected for correlation with direct and proxy measures of household income. Subjects who reported their health as fair or poor were compared with those who reported their health to be good, very good, or excellent. Self-reports of the diagnosis of arthritis, back problems, heart disease, diabetes and the use of diabetes medications, asthma and the use of medication for asthma were also examined. The sum of all Assessment Fees (A or K codes) submitted to OHIP for each subject by General Practitioners in the year prior to the NPHS interview was used as a measure of health services utilization.

The prevalence of dichotomous conditions was analyzed with logistic regression and the cost of health services by multiple linear regression. The variables included in the models are presented in the tables of results. Survey weights were used in all calculations and a robust variance estimator14 was computed to allow for the complex survey design. The Akaike Information Criterion (AIC)15 was used to compare non-nested logistic regression models. The smallest value of AIC indicates the best fit to the data. The F-score was used to compare non-nested linear regression models. The larger value of F indicates a better fit to the data.

RESULTS

Study subjects and income groupings

The Ontario component of the 1994/95 NPHS included 4,804 Ontario residents aged 12 years or more, representing some 9.4 million residents of the province. Individual-level household income estimates were available for 4,637 (96.5%) subjects. Linkage of the OHIP-derived postal codes to the 1996 census provided estimates of EA-level median household income for 4,188 (87%) of subjects. Measures of both self-reported and EA income were available for 4,037 (84%) respondents, and these are the subjects of this analysis.

In order to compare income groups of similar size, the EA-derived median incomes were ranked, and subjects were placed into EA-groups of (approximately) the same size as the self-report income categories. The distribution of the subjects by income category is shown in Table I.

Relations between self-reported and enumeration area household incomes

Figure 1 shows the distribution of selfreported incomes within each of 4 categories of Enumeration Area household incomes. Each EA category is clearly not homogeneous with respect to self-reported household income, so that the potential for misclassification exists when using ecologie data for analysis.

Relations between self-reported health status and personal and ecologie measures of income

Table II shows the results of logistic regression models explaining self-reported health status as fair or poor (N=555) versus good or better (N=3,482). These models are adjusted for age and gender and high school graduation status is included as an example of a variable that might be of substantive interest for some analyses, and for which we wish to control for confounding by income. One model includes selfreported income and the other the EAderived ecologie measure. The Model AICs, reported in the bottom line, indicate that the model containing self-reported income provides a substantially better fit to the data than the model including the ecologic measure. This improved fit is reflected in larger z-scores for most of the explanatory variables. For the income categories, the inverse trend in reporting ill health is sharper for the self-report than for the ecologie measure.