Class notes

Understanding Health:
Theoretical challenges and
possible approaches
September 25, 2006
Evolving perspectives on povertyhealth link
• C19. Miasma-style: multiple interacting factors
but no clear mode of action
• c. 1920-30. Agent-host-environment triad; poor
environments constrain host resistance & limit
behaviours (nutrition, hygiene, etc.)
• c.1950-1960. Patterns of causes; interacting
chains of events (Morris, 1964)
• c.1960-1985. Risk factor approach (e.g., MRFIT)
focused interventions for specific diseases:
reverse engineering etiology
Critiques (1)
• Epidemiology has produced a “Hotch-potch of
multivariate associations between diseases and
lifestyle risk factors” (Tannahill, 1992)
• There are almost no necessary (or sufficient)
causes.
• Chains of events a simplification; multiple,
interacting sequences occur together. Field or
systems theory may be helpful (Morris, 1964).
• Susser (1973) “agent and host are in continuing
interaction with an enveloping environment”
• “The multiple cause black box paradigm of the
current risk factor era in epidemiology is growing
less serviceable” (Susser, 1973)
Critiques (2)
• Pearce (1996): “Epidemiology has become a set
of generic methods for measuring associations
of exposure and disease, rather than functioning
as part of a multidisciplinary approach to
understanding the causation of disease in
populations. We seem to be using more and
more advanced technology to study more and
more trivial issues, while the major population
causes of disease are ignored.”
• Inherent vagueness of the risk factor concept.
Critiques (3)
• Hennekens & Buring (1987): “… the use of
multivariate analysis can appear like a
‘black box’ strategy in which all of the
variables are entered (…) and the net
result is a single value representing the
magnitude of the association between the
exposure and the disease after the effects
of all confounders have been taken into
account.”
Evolving perspectives (2)
• 1990s. Bringing the context back in: ‘Chinese box
epidemiology’ (Susser & Susser, 1996).
Concentric circle models. Multilevel, but
interacting processes; analytical approach not
clear.
• 1995 onwards: lifestyles
lifecourse. Brings
time dimension back in.
• 2000 onwards. Multilevel analyses; hierarchical
modeling. Confounding factors studied in their
own right. Critique of reductionism.
• Opening up the black box: molecular & genetic
epi.
Critiques: Weiss & Buchanan
• Statistical methods unsuited to detecting many-to-many
relationships, each with small effects
• Individual cases often multifactorial (or multiple paths
from single cause to disease)
• Diseases given same name may be distinct
• Many alleles can cause single disease; selection acts on
phenotypes, not genotypes.
• Scientific method can be fallible: false falsifications can
reject acceptable hypotheses. For example, when a
disease comes to be defined by its cause, the causal
hypothesis is no longer falsifiable
• True probabilistic causation vitiates replicability &
falsifiability
e.g., ‘Dissecting complex disease’. Int J Epidemiol 2006;35:562
The many-to-many relation, with
common pathway
Critiques (4)
• Multilevel analyses retain the basic linear
regression models and mechanistic notions of
causation
• It moves beyond focus on adding up figures on
individual risks, but has not re-thought
explanation; has not accommodated complexity
• Relationships between variables are not
necessarily static but evolve through experience
and over time
• Non-linear interactions not covered well
• Not clear whether equivalent analyses should be
applied at individual and collective levels
Possible directions
• Reconsider the meaning of chance &
“random error” in regressions;
• Structured chance (Bagatelle metaphor)
• Bring the individual back in: formally
include susceptibility. Models include
– Epigenetic landscapes (Beattie, 2005, from
Waddington, 1940). Models concurrent
interacting influences of genes & environment
– Or probabilistic neural networks (PNNs)
Structured
randomness
(Bagatelle)
Random, but with
environmental
influences, and
different probabilities
of high scores
What may a complexity approach look like?
(1) Waddington’s Epigenetic Landscape
(1957)
The ball rolls downward, but may take
many different routes, each of which then
sub-divides again. While features of the
landscape will influence which route it
takes, the landscape itself changes over
time, with erosion and as a result of the
balls rolling down.
Waddington also drew the underside
of the diagram, representing the
surface of the hill as evolving, pulled by
numerous strings, each attached to
a gene, so the landscape in which we
interact is influenced by nature and
by nurture.
http://www.usc.edu/hsc/dental/odg/jaskoll01.htm
Complexity perspective (2)
Probabilistic Neural Network
Inputs are processed through
multiple, hidden (cf. black box)
nodes that have multiple links.
The prediction of the outcome
derives mainly from the pattern
of interconnections between
nodes, not from the complexity
of each. The effect of each
‘variable’ can change according
to the status of others in the
system (which was what we saw
with smoking and occupation in
the Whitehall study)
Complexity perspective (3)
Branch track diagrams
(Further decision
nodes)
Decides to join
fitness club
Grudgingly starts
walking program
Overweight patient
Chooses not to join
(Personality
determines shift to
different set of
response options)
Distressed by perceived
implication of being fat;
resentment reinforces
sedentary lifestyle.
Triumph of idleness