to be presented by James S. Koopman MD MPH
of the University of Michigan to the Colombian Association of
Epidemiology on October 8, 1997.
The focus of epidemiology on populations is said to distinguish
it from individual oriented medical care. Most epidemiological
methods, however, are designed to assess the effects of risk factors
in individuals, not populations. In its search for contol causes
of disease, epidemiology uses methods which fix its view on the
individual and causes it to ignore the population. The uses to
which standard epidemiological parameters are put assume that
causes of disease act ultimately within or upon individuals. This
holds for risk differences, odds ratios, regression coefficients,
and all other measures of association between exposure and disease.
Epidemiologic methods estimating these parameters assume that
populations are nothing more than collections of individuals and
that patterns of disease are not affected by the nature or arrangement
of interactions between individuals. That assumption is so deeply
ingrained that many epidemiologists are unaware that their analyses
depend upon it. When that assumption does not hold, the theoretical
basis for standard epidemiological methods crumbles.
Despite the theoretical weakness of of individual based risk factor
assessment, it continues to prove effective in discovering controllable
causes of disease. For that reason, some epidemiologists argue
that inconsistencies in the theoretical base of standard risk
factor assessment is no reason to abandon it for more theoretically
based methods whose foundations depend upon skills and traditions
with which most epidemiologists are unfamiliar. As more controversy
develops about specific conclusions of traditional methods, however,
the need for methods based on more comprehensive theory becomes
evident. When the benefit of identifying a risk factor like smoking
or unprotected anal sex is clear, there is no reason to be concerned
about the theoretical underpinnings for the identification. When
great costs are associated with taking either a positive or a
negative stand on a risk factor whose action is supported by some
studies and not supported by others, a situation which is ever
more frequent, the need for more comprehensive theoretical underpinnings
becomes more evident.
The resolution of controversies generated by traditional methods
is not, however, the major reason why more theoretically sound
methods are needed. The biggest need is to expand the horizons
of epidemiological inquiry. Epidemiology needs methods to get
it beyond the reductionist assumption that a population can be
understood by characterizing its individuals. New methods are
needed to discover and evaluate the controllable causes of disease
that lie precisely in the areas which traditional epidemiological
methods must ignore to preserve their validity - namely the nature
and arrangement of interactions between individuals. To illuminate
the path for these new developments, Mervyn Susser suggests a
metaphor of multilevel system interactions which he calls "Chinese
Boxes". Various social epidemiologists have suggested complementary
metaphors. I will discuss here the specifics of four approaches
toward these new horizons: 1) ecological studies, 2) discrete
individual systems analysis, 3) compartmental systems analysis,
and 4) social network analysis.
A new view of "ecological" studies can expand the horizons
of epidemiolgy. Instead of viewing ecological studies as fallacy
ridden compromises to be used for the detection of individual
risk factor effects, they can be viewed as approaches to capturing
effects that are missed by individual level analyses. These include
the arrangement of interactions between individuals in the population.
This approach is consistent with the observational inference traditions
in epidemiology. Like traditional epidemiology methods, it does
not require the elaboration of complex theories about how things
works. It mainly requires the comparison of diverse populations
in which the interactions between individuals are assessed. This
approach is illustrated by a study of malnutrition in Cali, Colombia
which demonstrated that conformation of neighborhoods is a much
stronger determinant of malnutrion than food availability at the
family level. It is also illustrated by a study in Chicago, USA
where the nature of interactions between individuals within neighborhoods
was found to be a stronger determinant of the level of violence
than the average social status, education, or wealth of individuals
in the neighborhoods.
Just like observations on individuals, the interpretation of observations
on ecological units will require a strong theoretical base regarding
how risk factors, human interactions, and population processes
generate patterns of disease in populations. Such theory can and
should be developed independently of observational methods so
that it drives the development of such methods rather than being
constrained by methods. If this is to happen, epidemiology needs
to develop new traditions of theory development. I propose that
"systems analysis" and "social network analysis"
provide valuable tools for that development. Most epidemiologists
are unfamiliar with these tools and their mathematical and theoretical
underpinnings. New software and new texts, however, are beginning
to make these more accessible to epidemiologists.
Systems analysis in epidemiology can use discrete individual simulation
approaches as well as continuous population segment approaches
such as compartmental models. These two approaches have complementary
abilities and deficiencies. Tools for the discrete individual
approach have been developed by the "National Micropopulation
Simulation Resource" at the University of Minnesota and are
available on the web at http://www.nmsr.labmed.umn.edu/nmsr/NMSR.html.
We are using these tools extensively in our work but unfortunately
it takes an experienced C++ programmer to apply them. The advantage
of these tools is that they can address many things which compartmental
models cannot and that they can interdigitate with social network
analysis approaches. The disadvantage is that the basis for using
these tools to advance fundamental theory is not well developed
outside of operations engineering and the publications in the
operations engineering literature are quite often not comprehensible
to epidemiologists. We are using these tools to make more evident
where the deficiencies of risk factor analyses lie and to explore
approaches to overcome those deficiencies. For example we have
shown that the erroneous assumption of no interaction effects
in standard epidemiological analysis is very likely causing a
serious misperception of the effects of oral sex on HIV infection
levels. We have constructed discrete individual simulations where
30% of all HIV transmission occurs during oral sex but where the
standard methods show oral sex to have no risk at all or to be
protective. We have also used discrete individual simulations
to explore designs for HIV vaccine trials which are based on couples
or populations rather than individuals. The system dynamics of
HIV infection suggests that vaccine induced immune responses will
not protect individuals from infection but they will dramatically
reduce the contagiousness of individuals during the early stages
of HIV infection. Our analysis suggests that such effects could
quite readily bring HIV transmission to a halt. Based on discrete
individual simulations we have developed a vaccine trial design
which can efficiently detect these effects which would be completely
missed by a standard trial.
The compartmental model approach to systems analysis is likely
to be more accessible to epidemiologists. This approach has been
a cornerstone for theory development in infectious disease which
pursues the patterns of interactions between individuals which
determine levels of infection in populations. A major text illustrating
some basic advances in this area is "Infectious Diseases
of Humans" by Roy Anderson and Robert May. This approach
should be just as valuable to address social determinants of chronic
diseases and I think it will be essential for the development
of an epidemiology of violence. User friendly software for this
approach has been developed which mainly requires pointing and
clicking and filling in a few formulas. I have developed a course
and am writing a text on how to pursue this approach. The text
is being made available on the web at http://www.sph.umich.edu/~jkoopman/802Web/Course.htm.
It outlines how to use Stella to construct compartmental models
that serve to help develop population level theory in epidemiology.
Stella is available from High Performance Systems at http://www.hps-inc.com/.
We have recently used this approach to demonstrate that HIV treatment
programs are likely to have significant effects upon HIV transmission.
Our analysis shows, however, that those effects will depend upon
how individuals are interacting and not just on the risk factors
and responses to treatment of individuals or the level of infection
in the population.
Social Network Analysis also provides a cornerstone for the pursuit of new epidemiological theory and methods. An excelent introduction to this discipline is provided by the text "Social Network Analysis" written by Stanley Wasserman and Katherine Faust. This approache is particularly valuable because it provides a vision of how epidemiological data sets should be organized to address interactions between individuals rather than assumming them away as the standard data structures do. The individual risk factor approaches of epidemiology organize data on individuals in different rows with data on variables relevant to those individuals arranged in columns. The arrangement of individuals in the rows is assummed not to affect the analytic results. That means that how different classes of individuals relate to each other is assummed not to affect the analysis. Social network analysis, in contrast, completely focuses on the interactions between individuals. Each individual or class of individuals is represented in both the rows and columns of a social network analysis data structure. Individual level variables must be specified as to how they are affected by contact with different types of individuals. While the development of analytic tools for epidemiology based upon such data structures is in its infancy, the vision provided by the data structure provides a guiding light for the development of new theory and methods.