Our study findings suggest a significant disagreement between individual-level and area-level SES measures. The poor agreement between individual-level and area-level SES measures was found equally in both lower and higher SES groups. Overall, area-level SES measures had different patterns of association with health outcomes from individual-level SES measures.
In our study setting there is a relatively heterogeneous distribution of people in terms of race, employment opportunities, etc within geographic boundaries. Thus, area-level measures of SES may not be suitable proxy measures for individual
SES, which can be quite heterogeneous within these boundaries. The weighted Cohen’s κ indices, indeed, suggest a poor agreement between area-level versus individual-level SES measures, ranging from 0.14 to 0.20 for education and 0.16 to 0.19 for income. As a result, significant proportions of individuals were misclassified (10–35% in education and 27–31% in income) as shown in and . One also may attribute poor agreement of individual SES in our study to the unit of measure at aggregate levels. The census block-group level used in our study is the smallest geographic unit for which income and other SES data are published. SES or other attributes in the community including census block-level SES data is not publicly available. Typically, a block group contains between 600 and 3000 people with an optimum size of 1500 people33
as compared with census tracts, which typically have between 1500 and 8000 people, with an average size of about 4000 people.34
The census tract is the next higher aggregate level, and is commonly used for a measure of neighbourhood environment.35
Thus, given the census block group as the smallest unit of census SES data that is publicly available, census block group is likely to minimise within-group heterogeneity. In rural parts of Olmsted County, however, block groups can be over 80 square miles in area, introducing wide heterogeneity in what are conventionally considered ‘neighbourhoods’, resulting in a significant misclassification bias.
We postulate that poor agreement between area-level versus individual-level SES measures may result in different patterns of associations of individual-level and area-level SES measures with health outcomes. Indeed, while individual-level income was strongly associated with risk of low birth weight, area-level income measures were not. Compared to individual-level SES measures, area-level measures showed weaker associations with health outcomes. Geronimus et al
reported a weaker association between aggregate measures and health outcomes and suggested a careful application of aggregate measures as a proxy measure for individual SES.36
Similarly, a study comparing the effects of SES at individual-tract and census-tract or block-group levels on high blood pressure, height, smoking and number of full-term pregnancies showed that area-based socioeconomic indicators greatly underestimated the impact of SES on the outcomes.14
Also, the geographic unit used for area-level SES measures may not be a mere census-defined boundary sharing similar SES characteristics, but rather the unit may be a socially defined neighbourhood affecting the health of individuals along with individual SES.37,38
Thus, even if area-level SES measures were associated with risks of being low birth weight, overweight or smoking exposure in household, the associations could be attributed to an influence of neighbourhood environment, which could be difficult to disentangle from that of individual-level SES measures. For instance, O’Campo et al
reported that the census tract-level variables modified the associations between the individual-level factors and the risk of low birth weight.38–40
Similarly, Black and Macinko41
demonstrated that both individual and neighbourhood factors were independently associated with reduced obesity. Thus, neighbourhood or area-level factors themself can be an important contextual factor influencing health outcomes and should be considered in health studies independent of socioeconomic measures.
Previous studies have reported a substantial lack of agreement between individual-level and area-level information.15–18,20,25,26
Similarly, Marra et al17
found poor agreement between individual-level and aggregate-level measures of income and education, and the weighted Cohen’s κ index ranged from 0.13 to 0.27. A recent study showed that individual-level and area-level SES measures have an average concordance rate 64–69%.17,15–19,25,26
Also, few studies assessed the agreement between individual-level and area-level SES measures in the context of health outcomes among children. Thus, our study is the first study based on a population-based sampling frame that examined the agreement between individual-level and area-level SES measures and their relationships with childhood health outcomes in a mixed urban–rural setting.
Our study results have implications for research and public health. If the purpose of a study was to assess the role of SES in etiological pathways at an individual level, that is, generalising the findings to individuals, using area-level SES as a mere surrogate measure for individual SES may not be suitable without considering the study setting.
In mixed urban–rural settings, census areas or neighbourhoods can be relatively homogeneous in terms of the distribution of race, ethnicity, employment, etc across census areas or neighbourhoods, but have varying SES of individuals within neighbourhoods or small residential areas. In such instances, SES measured on the entire census area may considerably misrepresent the SES of individuals in a study. A clear example of such misrepresentation could occur within a large and relatively sparsely settled census area in which residents of differing SES are widely scattered.27,28
Area-level proxies have been used as control variables in studies analysing the influence of different risk factors on specific health outcomes. Recognising the importance of area-level SES measures, we make a few suggestions. First, area-level proxies need to be carefully utilised in health research depending on study settings. Second, given the potentially different, yet important, constructs for individual-level versus area-level SES, multilevel approaches taking into account both individual-level and area-level SES measures should be considered when formulating public health strategies or policy. Third, for those studies using area-level SES measures in health research (eg, secondary dataset), we suggest stratified analysis for the main results by study settings if dataset permits: (1) large MSA, (2) small MSA, (3) small urban non-MSA (urban–rural mixed area) and (4) the remaining rural areas to assess whether the results are consistent across different study settings.
Our study has a few limitations. Individual income, education and health outcomes were self-reported measures. Also, the survey respondents were primarily female caregivers, incurring the possibility of a non-differential misclassification bias with regard to the study results. Since our study was based on a telephone survey that interviewed ‘adult who is most knowledgeable about the health of the children in the home’, the majority of respondents were female caregivers who is unlikely to affect the relationship of individual SES with area-level SES and health outcomes. Another possible limitation is a discrepancy in the timing of measurement as area-level SES measure were derived from the 2000 US census, whereas an individual-level SES telephone survey was conducted in 2006. Although the response rate was only 62%, this is considered as an adequate rate for a telephone-administered survey of more than 5 min.42
In addition, the characteristics of non-respondent caregivers were likely to be randomly distributed.29
Despite limitations, our study has some strengths. Our study was a cross-sectional study using a population-based sampling frame, representative of families with children in the community. Olmsted County is an excellent setting in which to conduct some epidemiological studies because medical care is virtually self-contained within the community under the auspices of the Rochester Epidemiology Project.43
Lastly, we collected multiple measures of SES and health outcomes for analysis.
In conclusion, given the poor agreement between individual-level and area-level SES measures, different patterns of their associations with health outcomes and the potential influence of neighbourhood environment on health, area-level SES measurement may not be a suitable proxy measure of individual SES. Our results call for development of individual-level SES measures overcoming the absence of SES measures in commonly used datasets.