Skip Navigation



American Journal of Epidemiology Advance Access published online on May 13, 2008
American Journal of Epidemiology, doi:10.1093/aje/kwn117
This Article
Right arrow Abstract Freely available
Right arrow FREE Full Text (PDF) Freely available
Right arrow All Versions of this Article:
168/1/9    most recent
kwn117v1
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Add to My Personal Archive
Right arrow Download to citation manager
Right arrowRequest Permissions
Right arrow Disclaimer
Google Scholar
Right arrow Articles by Oakes, J. M.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Oakes, J. M.
Social Bookmarking
Add to CiteULike   Add to Connotea   Add to Del.icio.us  
What's this?

American Journal of Epidemiology © The Author 2008. Published by the Johns Hopkins Bloomberg School of Public Health. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org.

Invited Commentary: Rescuing Robinson Crusoe

J. Michael Oakes

From the Division of Epidemiology and Community Health, University of Minnesota, Minneapolis, MN

Correspondence to Dr. J. Michael Oakes, Division of Epidemiology and Community Health, University of Minnesota, 1300 South Second Street, Suite 300, Minneapolis, MN 55454 (e-mail: oakes007{at}umn.edu).

Received for publication March 15, 2008. Accepted for publication March 19, 2008.


   ABSTRACT
TOP
 ABSTRACT
INTRODUCTION
 WHAT ARE AGENT-BASED MODELS?
 WHY AGENT-BASED MODELS?
 CONCLUSION
 References
 
Estimating the independent effect of "place" on health outcomes has proven quite difficult. In this issue of the Journal, Auchincloss and Diez Roux contribute a lucid introduction to agent-based simulation models and argue that they may be a fruitful alternative to current approaches to the problem. Insofar as conceptual understanding must precede empirical investigation, this author agrees. Given the obvious shortcoming of pure simulations, the key benefit of agent-based models lies in their ability to alter our thinking and/or theory. Among other things, the approach permits analysts to model (i.e., conceptualize) system dynamics, more realistic social treatment effects, endogenous contexts, and a more congenial image of human behavior.

computer simulation; environment and public health; epidemiologic methods; health behavior; models, theoretical; residence characteristics; systems theory


   INTRODUCTION
TOP
 ABSTRACT
 INTRODUCTION
WHAT ARE AGENT-BASED MODELS?
 WHY AGENT-BASED MODELS?
 CONCLUSION
 References
 
In this issue of the Journal, Auchinloss and Diez Roux (1) contribute a lucid introduction to agent-based simulation models and argue that they may be a fruitful alternative to current approaches aiming to tease out the effects of neighborhood contexts on health outcomes. For the reasons outlined below, I wholeheartedly agree. This commentary aims to amplify a few of the important points made in the review article and to locate them in the broader context of epidemiologic inquiry, especially as concerns social epidemiologic theory.


   WHAT ARE AGENT-BASED MODELS?
TOP
 ABSTRACT
 INTRODUCTION
 WHAT ARE AGENT-BASED MODELS?
WHY AGENT-BASED MODELS?
 CONCLUSION
 References
 
To reiterate, agent-based models are computer simulations that researchers design and use to model how agents (e.g., hypothetical human beings) interact to produce group-level (i.e., social or system-level) phenomena. An agent-based modeler's computer becomes her social laboratory. Accordingly, researchers hypothesize how a change in inputs and/or rules of interaction will affect outcomes, run the virtual experiment, and then compare the observed with the hypothesized results. Theories, or more precisely, predictions, may thus be supported, refuted, or advanced by examining unanticipated group- or system-level outcomes. The canonical agent-based experiment is described by Epstein as follows:

Situate an initial population of autonomous heterogeneous agents in relevant spatial environments; allow them to interact according to simple rules, and thereby generate—or grow—the macroscopic regularity from the bottom up (2, p. 7).

While they share some characteristics, agent-based model simulations are not the same as the now conventional Markov Chain Monte Carlo simulations (Bayesian) epidemiologists are using. Simply put, Markov Chain Monte Carlo methods aim to simulate data or a parameter space for a more or less conventional epidemiologic analysis, such as a sensitivity analysis. By contrast, the outcome of an agent-based model is not data, per se, but the process of how and why group phenomena emerge in the first place. No subsequent regression or life-table analysis typically follows.

A key point is that agent-based models incorporate responsive and/or adaptive aspects of human behavior. Agents are modeled as heterogeneous purposive actors who observe their virtual surroundings, learn from experience, and act accordingly. Therefore, unlike common regression models and related linear theories of disease causation, agent-based models explicitly permit feedback loops and other nonlinear dynamic phenomena. There is some precedent for this kind of thinking (3, 4), but unlike equation-based efforts, agent-based models exploit the iterative power of computers to "grow" phenomena irreducible to equations.


   WHY AGENT-BASED MODELS?
TOP
 ABSTRACT
 INTRODUCTION
 WHAT ARE AGENT-BASED MODELS?
 WHY AGENT-BASED MODELS?
CONCLUSION
 References
 
Thanks in large part to the work of Diez Roux (5, 6) and Diez Roux et al. (7), the idea that multilevel (social) contexts matter in shaping both exposure patterns and disease outcomes is now firmly established. Although clearly a good thing, two major issues merit attention.

First, it does not take an epidemiologist to appreciate that residing in a disadvantaged area is unhealthy. Untrained observers have long noted the relation and, at least as far back as 1830 (approximately 20 years before Snow's cholera research), Benoiston de Chateauneuf produced a modern-looking table of mortality ratios by age and poverty status (8).

Second, and simply put, the ultimate goal of neighborhood effects research is to estimate the effect of moving a poor person to a better neighborhood (or vice versa) so as to inform policy makers about the health impacts of impoverished areas as opposed to impoverished individuals (9). In other words, the goal is to estimate the independent effect of an impoverished area on a person's health regardless of his or her own poverty status. This is a laudable goal, but a large set of methodological challenges must be addressed before we should have confidence in results. Among these are the related issues of 1) selection bias and exchangeability and 2) within- and between-group dynamics. Even setting aside the cloudy results of relevant field experiments (10), it now seems patently obvious that observational designs and conventional epidemiologic methods, including the multilevel regression model, cannot yield credible neighborhood effect estimates. Indeed, my recent own empirical work has shown that 1) one cannot identify the effects of neighborhood poverty on American Indian infant mortality (11) and 2) when a more defensible comparison group is used, the effect of neighborhood density and street connectivity on total physical activity is negligible (12). Given commonsense intuitions about the effects of context on health, neither result is intellectually satisfying. Should we give up or change our approach?

I believe it is time to change our approach. For the following interrelated reasons, at least, agent-based models hold some promise for advancing neighborhood effects research, if not social epidemiology more generally.

First, agent-based models permit analysts to model system dynamics. A deep criticism of neighborhood effects research is that it assumes away posttreatment responses (9). This means that, for example, once a poor person is relocated to a better neighborhood, the residents of the target neighborhood are assumed to not respond by protesting, moving away, or other such things. Similarly, current approaches imply that, when economic development improves a given location, the expected increased demand for it does not increase property values, increase taxes, or otherwise induce gentrification (13). By their very nature, agent-based models permit an analysis of longer-term (perhaps "equilibrium" or "net") impacts of a proposed intervention.

Second, agent-based models permit analysts to more realistically model social treatment effects. A deep criticism of neighborhood effects research is the assumption that what happens to one resident has no impact on any other resident (9, 14). The so-called "stable unit treatment value assumption" is a subtle idea that essentially says for a given effect to be identifiable, the treatment given to a particular subject cannot spill over or interfere with the treatment to another subject within the same group. The reason for this is that, if there is spillover, the treatment given may not be the treatment received, and there is thus no obvious way to define a treatment much less estimate its effect (15). It follows that the obvious impacts of charismatic leaders, peer pressure, social fads, electoral momentum, and so forth cannot be incorporated into a conventional statistical analysis of neighborhood effects. Agent-based models are not so constrained; they permit analysts to model the subtle effects of leaders, momentum, social norms, and so forth on health outcomes.

Third, agent-based models permit the analysis of endogenous contexts (i.e., exposures) on health outcomes. As far as I can tell, all neighborhood effects research presumes that neighborhood contexts are exogenous or not produced, even in part, by residents. This presumption is implicit in the socioecologic model of health (16), which essentially posits that environments (e.g., state policy) are upstream causal factors to downstream health outcomes (e.g., lung cancer), and the causal arrow goes from the macro to the micro only. Few appear to appreciate the importance of the other direction and thus the endogeneity of contexts. For example, few seem to appreciate that state policies are a function of voter preferences, lobbyists, and legislative interests; that is, few acknowledge that state policy is endogenous. The conceptual distinction between exogenous and endogenous environments is especially important in neighborhood studies, because it means that subjects of a study actually create the context in which they reside. Where possible, parents keenly interested in safe and supportive environments for their children will help to build a playground and participate in neighborhood watch associations. Although slippery, the fact is that not only do environments affect people (i.e., macro-to-micro transitions), but people also affect environments (i.e., micro-to-macro transitions) (17). Thus, simple analytical efforts aiming to disentangle an environment from its inhabitants seem doomed to fail. By treating environments as possibly endogenous, agent-based models offer an advance over current approaches.

Fourth, compared with conventional epidemiologic theory, agent-based models permit a more congenial image of humankind. The elemental subject of most epidemiologic investigations is a passive host (perhaps even a cell or organ system) subject to the forces of nature. Purposive human activity and/or interaction is rarely part of an epidemiologic explanation. This is not all bad as great progress has been made and, in some cases, an analysis of cellular activity is entirely appropriate. Nonetheless, it is an unfortunate irony that so much research in epidemiology, the science of population health, incorporates so little about human preferences, choice, socialization, exploitation, or adaptation. One need only consider the equivalence of methodology between the epidemiology of livestock and humans to appreciate that some change is necessary (18). Agent-based models may help to propel it by forcing analysts to model human interaction.

Fifth, because they approach problems from a systems process perspective, agent-based models offer insight into some limitations of current epidemiologic explanation. For the most part, epidemiology's explanatory paradigm is one of statistical difference between groups: the exposed and unexposed. The mechanisms through which exposures emerge and population change occurs are rarely considered. The reason for this may lie with the ongoing confusion about the difference between statistical inference and scientific explanation (19). Importantly, agent-based models are focused solely on the mechanisms of change. Epstein makes the point with the following statement: "If you didn't grow it, you didn't explain it" (2, p. xxi). It follows that agent-based models may help to focus epidemiology's attention toward the social processes that allocate noxious exposures, susceptibilities, and treatments to some and not others (20).


   CONCLUSION
TOP
 ABSTRACT
 INTRODUCTION
 WHAT ARE AGENT-BASED MODELS?
 WHY AGENT-BASED MODELS?
 CONCLUSION
References
 
Except for raising the awareness that context matters—which is no small thing—I think that epidemiologic efforts to estimate neighborhood effects have not been all that scientifically fruitful. Looking back over the last 5–10 years' research, it seems to me that the key insight is that our simple models (e.g., multilevel regressions) and theories (e.g., socioecology) are unable to incorporate the important aspects of the phenomena under investigation. Perhaps this should not be surprising. Many great minds have struggled with the complexity of social life. Consider the 1969 Nobel prize winner for physics, Murray Gell-Mann, who upon being asked about the relatively slow pace of social science quipped, "Physics would be a lot harder if particles could think" (21, p. 5).

As Auchincloss and Diez Roux (1) make clear, the obvious shortcoming of agent-based models is that simulated results may be a far cry from actual empirical ones. Although it is true that sooner or later fieldwork must be done—and I still think that neighborhood-randomized trials are the best path to progress (9, 22)—let us not prematurely dismiss the benefit of a conceptual advance. The fact is that counterfactual accounts of causality (23) are entirely without empirical support, and directed acyclic graphs and related theory (24, 25) are conceptual tools for guiding analysis and perhaps action. Additionally, regression adjustment, sampling weights, multiple imputation, and essentially all modern tools for statistical inference are based on convenient fictions (e.g., ceteris paribus). Moreover, although most know that Box and Draper (26) famously pointed out that we should measure our models not by how perfectly they represent reality but rather by their utility for understanding the same, few recognize that Box (27, 28) insisted that scientific progress required an iterative process between conceptual thinking and empirical investigation (i.e., think Formula investigate). It seems to me that, given agent-based models' potential to alter the paradigm through which epidemiologists think about how social systems produce the health of populations, a modest investment in them is appropriate.

Convincing the skeptic of the potential utility of agent-based models will presumably require a collective recognition that, since the advent of germ theory, epidemiology has relied on a model of disease in which hosts (individuals, cells) interact with agents (bacteria, ice cream) in environments (neighborhoods, jungles). Although this classic triadic model has been useful in many areas, it forecloses insight in others, such as social epidemiology. As I have stated before (18, 29), boiled down to its core, the triadic model implies Robison Crusoe (30), a metaphorical character whose health depends on his genes and behavior, along with the natural island environment. Excluded are the impacts of other people or emergent social phenomena, such as strategic interaction, norms, and economic policy—issues that are at the center of social epidemiologic inquiry. Thus, beyond their obvious utility for better understanding of neighborhood effects, agent-based models may help us to return Robinson Crusoe to society and thereby improve our collective ability to understand and enhance the health of populations.


   ACKNOWLEDGMENTS
 
Conflict of interest: none declared.


   References
TOP
 ABSTRACT
 INTRODUCTION
 WHAT ARE AGENT-BASED MODELS?
 WHY AGENT-BASED MODELS?
 CONCLUSION
 References
 

  1. Auchincloss AH, Diez Roux AV. A new tool for epidemiology: the usefulness of dynamic-agent models in understanding place effects on health. Am J Epidemiol (2008) 168. 000–00.
  2. Epstein JM. Generative social science: studies in agent-based computational modeling. (2006) Princeton, NJ: Princeton University Press.
  3. Blume LE, Durlauf SN. Identifying social interactions: a review. In: Methods in social epidemiology—Oakes JM, Kaufman JS, eds. (2006) San Francisco, CA: Jossey-Bass/Wiley. 281–309.
  4. Homer JB, Hirsch GB. System dynamics modeling for public health: background and opportunities. Am J Public Health (2006) 96:452–8.[Abstract/Free Full Text]
  5. Diez Roux AV. Investigating neighborhood and area effects on health. Am J Public Health (2001) 91:1783–9.[Abstract/Free Full Text]
  6. Diez-Roux AV. A glossary for multilevel analysis. J Epidemiol Community Health (2002) 56:588–94.[Abstract/Free Full Text]
  7. Diez Roux AV, Merkin SS, Arnett D, et al. Neighborhood of residence and incidence of coronary heart disease. N Engl J Med (2001) 345:99–106.[Abstract/Free Full Text]
  8. Coleman W. Death is a social disease: public health and political economy in early industrial France. (1982) Madison, WI: University of Wisconsin Press.
  9. Oakes JM. The (mis)estimation of neighborhood effects: causal inference for a practicable social epidemiology. Soc Sci Med (2004) 58:1929–52.[CrossRef][Web of Science][Medline]
  10. Kling JR, Liebman JB, Katz LF. Experimental analysis of neighborhood effects. Econometrica (2007) 75:83–119.[CrossRef][Web of Science]
  11. Johnson PJ, Oakes JM, Anderton DL. Neighborhood poverty and American Indian infant death: are the effects identifiable? Ann Epidemiol. (in press).
  12. Oakes JM, Forsyth A, Schmitz KH. The effects of neighborhood density and street connectivity on walking behavior: the Twin Cities walking study. (Electronic article). Epidemiol Perspect Innov (2007) 4:16.[CrossRef][Medline]
  13. Deluca S, Rosenbaum JE. If low-income blacks are given a chance to live in white neighborhoods, will they stay? Examining mobility patterns in a quasi-experimental program with administrative data. Housing Policy Debate (2003) 14:305–46.[Web of Science]
  14. Vanderweele TJ. Ignorability and stability assumptions in neighborhood effects research. Stat Med (2007) 27:1934–43.[CrossRef]
  15. Sobel ME. What do randomized studies of housing mobility demonstrate? Causal inference in the face of interference. J Am Stat Assoc (2006) 101:1398–407.[CrossRef][Web of Science]
  16. Stokols D. Establishing and maintaining healthy environments: toward a social ecology of health promotion. Am Psychol (1992) 47:6–22.[CrossRef][Medline]
  17. Coleman JS. The foundations of social theory. (1990) Cambridge, United Kingdom: Belknap.
  18. Oakes JM, Kaufman JS. Introduction. In: Methods in social epidemiology—Oakes JM, Kaufman JS, eds. (2006) San Francisco, CA: Jossey-Bass/Wiley. 3–20.
  19. Greenland S. Summarization, smoothing, and inference in epidemiologic analysis. Scand J Soc Med (1993) 21:227–32.[Web of Science][Medline]
  20. Macy MW, Willer R. From factors to actors: computational sociology and agent-based modeling. Annu Rev Sociol (2002) 28:143–66.[CrossRef][Web of Science]
  21. Gintis H, Bowles S, Boyd R, et al. Moral sentiments and material interests: origins, evidence, and consequences. In: Moral sentiments and material interests (2005) Cambridge, MA: MIT Press.
  22. Hannan PJ. Experimental social epidemiology: controlled community trials. In: Methods in social epidemiology—Oakes JM, Kaufman JS, eds. (2006) San Francisco, CA: Jossey-Bass/Wiley. 335–64.
  23. Pearl J. Causality: models, reasoning, and inference. (2000) New York, NY: Cambridge University Press.
  24. Oakes JM, Kaufman JS. Using causal diagrams to understand common problems in social epidemiology. In: Methods in social epidemiology (2006) San Francisco, CA: Jossey-Bass/Wiley. 393–428.
  25. Greenland S, Pearl J, Robins JM. Causal diagrams for epidemiologic research. Epidemiology (1999) 10:37–48.[CrossRef][Web of Science][Medline]
  26. Box GEP, Draper NR. Empirical model-building and response surfaces. (1987) New York, NY: Wiley.
  27. Box G. Statistics and quality improvement. J R Stat Soc (A) (1994) 157:209–29.
  28. Box GEP. Science and statistics. J Am Stat Assoc (1976) 71:791–9.[CrossRef][Web of Science]
  29. Oakes JM. An analysis of American Journal of Epidemiology citations with special reference to statistics and social science. Am J Epidemiol (2005) 161:494–500.[Abstract/Free Full Text]
  30. Defoe D. Robinson Crusoe. (2007) New York, NY: Oxford University Press.

Add to CiteULike CiteULike   Add to Connotea Connotea   Add to Del.icio.us Del.icio.us    What's this?


This article has been cited by other articles:


Home page
 Int J EpidemiolHome page
J M. Oakes
Commentary: Individual, ecological and multilevel fallacies
Int. J. Epidemiol., April 1, 2009; 38(2): 361 - 368.
[Full Text] [PDF]

This Article
Right arrow Abstract Freely available
Right arrow FREE Full Text (PDF) Freely available
Right arrow All Versions of this Article:
168/1/9    most recent
kwn117v1
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Add to My Personal Archive
Right arrow Download to citation manager
Right arrowRequest Permissions
Right arrow Disclaimer
Google Scholar
Right arrow Articles by Oakes, J. M.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Oakes, J. M.
Social Bookmarking
Add to CiteULike   Add to Connotea   Add to Del.icio.us  
What's this?