Epi07_8

Continuation of 7b + =Placing individuals in a multileveled context= Idea: Different or even contradictory associations can be detected at different levels of aggregation (e.g., individual, region, nation), but not all influences can be assigned to properties of the individual—Membership in a larger aggregation can influence outcomes even after conditioning on the attributes of the individuals.

Initial notes on the Cases for multileveled context
From PT: Cases: Ecological and atomistic fallacies, Neighborhood effects, Childhood maltreatment, Effects of inequality, Obesity, Heart disease Readings: Freedman 2001, Diez-Roux 2002a, b, Coulton 1999, Korbin 2000, Marcelli 2005, 2006, Balfour 2002, Lawlor 2005


 * Preparation for class** (subject to revision if someone volunteers to co-lead this session) [//not so much revised as augmented by// JC]:

__Ecological & atomistic fallacies__
1. Bring one example of an ecological fallacy not in the readings and be ready to explain it to the other students. 2. Extract one example of an effect that is obscured by focusing only on one level, i.e., all individuals in a population. This example may come from the readings. Be ready to explain it to the other students (and to explain where you get confused, if applicable).

**__Neighborhood effects__**
Diez-Roux (2002a) notes in her commentary that “…many of the analytic issues that arise when examining neighborhood effects on health are present throughout the continuum from society to molecules. These analytic issues include, for example, nested data structures, variables and units of analysis at multiple levels, contextual effects, distal causes, and complex causal chains with feedback loops and reciprocal effects” (p.516). 1. Do you find that the issues referenced above have been successfully addressed in the studies by Balfour & Kaplan (2002) and Coulton, Korbin & Su (1999)? 2. Do the discussions by the authors on the limitations of their studies point to directions or changes in design that might strengthen the hypothesis that “neighborhood environments are causally related to heath”? In other words, what would you do differently if you were designing a study to 'empirically test different aspects of the specific processes thought to be involved in neighborhood effects on health'?

**__Income inequality, stature, and obesity__**
1. In general, did you find the study by Marcelli (2005) on the impact of family SES and income inequality on height persuasive? 2. Describe (and evaluate) the proposed pathway from income inequality to weight gain in the study by Marcelli (2006). 3. Evaluate the finding from the study that “intimates that these area-based socioeconomic factors partly influence BMI through the food distribution system” (p. 25). 4. Evaluate the claim that “geographically differentiated access to affordable healthy food causes stress in human bodies that leads to weight gain” (p. 26). How well does this claim meet the “Guidelines for Judging Whether an Association is Causal” presented in Gordis, pp.212-215?



Multi-Level Analysis
Class Presentation by EB October 24, 2007

What is Multi-level Analysis?
	An analytical approach for examining individual outcomes that allows simultaneous examination of group level and individual level variables on individual outcomes1 	“A tool to investigate more sophisticated and hopefully more realistic models of disease causation1”

Why Use Multi-level Analysis?
	For nested data structures/nested sources of variability. Acknowledges that individual level outcomes ecologically – i.e. subject to effects of environments in which individuals are nested, such as neighborhoods and health care practices 	Accounts for cross level effects of and interaction between higher/group level variables on lower/individual level outcomes, or the modification of the effects on lower level outcomes. 	Allows for examining both between group and within group variabilility, e.g. examining differences between neighborhoods as well as between individuals nested within the groups. 	Group level variables may provide information that is not captured by individual level data

What Analytical Methods can be Employed?
	Multi-level models: hierarchical linear models, random effects or random coefficient models, covariance components models 	Regression model 	Multilevel models allow you to separate the effects of context (group characteristics) and of composition (chars of individuals in groups)

How does MLA Differ from Other Approaches?
•	Allows for the simultaneous examination of the effects of group level and individual level predictors •	It accounts for the nonindependence of observations within groups •	Groups or contexts are treated as coming from a larger population of groups (not treated as unrelated) •	Allows for examination of both interindividual and intergroup variation.

MLA and Neighborhood Effects
	Sets out to examine between-neighborhood and within-neighborhood variability in outcomes 	Estimate associations of neighborhood characteristics with individual-level outcomes after adjustment for individual level confounders

Analytical Issues with Neighborhood Effects
	Where studies have found neighborhood effects, the percent of total variation between neighborhoods has been small. 	Question whether have large enough samples of nhbds to detect between-nhbd variability. 	Use of crude proxies based on existing data 	Need to develop theories specifying how nhbd factors influence health 	Need for alternative analytic techniques to better account for the complex set of relations, nested data structures, variables and units of analysis at mulitple levels, contextual effects, distal causes and complex causal chains with feedback loops and reciprocal effects.

Errors in Inference
	Ecological Fallacy 	Drawing inference about relationships at the individual-level based on relationships at the group-level. E.g. Durkheim’s conclusion that suicide rates were associated with Protestantism. 	Atomistic Fallacy 	Drawing inference regarding variability across groups based on individual level data

Challenges to MLA
	Need to develop testable theories that specify how group and individual level factors influence the distribution of health and disease 	Specification of relevant constructs and the levels at which they are defined and measured 	Separating out independent effects 	MLAs are very complex: sample size and power calculation are complex; need large sample size (esp for number of groups)



Annotated additions by students
Oakes, J. M. (2004). "The (mis)estimation of neighborhood effects: Causal inference for a practicable social epidemiology." Social Science & Medicine 58: 1929-1952. This is technical despite their efforts to make the tchnical accessible. But it questions whether multilevel analysis can show anything of causal significance. Oakes promotes randomized community interventions. The article is followed up by articulate responses from Diez-Rouz and by Subramanian (?). Even if one doesn't have expertise to take sides, it is important to know that there exists a deep debate.

Braveman, P.A., Cubbin, C., Egerter, S., et al. (2005). [|Socioeconomic status in health research: one size does not fit all]. //JAMA, 294// (22): 2879-2888. This article is a timely reminder of the necessity for health researchers to adequately and explicitly indicate and justify their choice of SES indicators in their studies. Braveman, et al examined current practices in the use of SES as well as conducting new analyses that showed many SES measures are not interchangeable. Specifically, they found that education and income are not interchangeable; income is not a proxy for wealth; occupational standards in the U.S. are not meaningful as SES measures; past SES may be just as relevant as current SES in some situations; neighborhood SES conditions can be made more meaningful by using census variables to describe neighborhoods; and the implications of potentially relevant aspects of SES that could not be measured should be acknowledged. The authors make five recommendations and promote an approach that is “outcome- and social group-specific and rests on considering explanatory pathways and mechanisms, measuring as much relevant socioeconomic information as possible, claiming to measure only what was measured, and systematically considering how important unmeasured socioeconomic factors may affect conclusions.” (JC)

Portnov, B., Dubnov, J., & Micha, B. (2007). On ecological fallacy, assessment errors stemming from misguided variable selection, and the effect of aggregation on the outcome of epidemiological study. Journal of exposure science & environmental epidemiology 17(1), 106-121. This article used a sample of 1492 school children to examine if ecological fallacy occurred as a result of assumptions about an individual on aggregate data for a group regarding air pollution. Data was collected from children regarding their health status, housing conditions, and test results from a pulmonary function test. The PFT results were average and compared to the average level of air pollution in the community in which the children live. The next analysis looked at individual pollution estimates compared to individual results from PFT. The conclusion of this study is based on the findings that indicate that ecological fallacy did not occur because of outcomes of data aggregation but because of the selection of variables chosen to measure health effects (JN).