Presentation on theme: "Where is Epidemiology going? Jan P Vandenbroucke Bern, STROBE meeting August 2010 Part II."— Presentation transcript:
Where is Epidemiology going? Jan P Vandenbroucke Bern, STROBE meeting August 2010 Part II
Four topics The ‘surge’ of Comparative Effectiveness Research New statistical techniques (or old ones that are suddenly popular) New methodologic insights (confounding, selection bias, interaction, mediation..) The call for registration of observational research
New methodologic insights Largely based on causal graph theory = Direct Acyclic Graphs (DAGs), counterfactual theory & sufficient component causes theory Names associated: J Pearl, P Spirtes (see Google books)
Sequence Confounding Selection bias Effect Modification and Interaction Mediation
Confounding (1) Anecdote: 30 years ago at Harvard: “up to now everybody was wrong about confounding” – sentence literally repeated by new generation at Harvard School of Public Health
Confounding (2): classic ‘triangle’ Exposure CONFOUNDER Outcome Image copied from net – Univ of Nottingham Health Services Research Course
C (3) in words: Confounder is Determinant / risk factor of outcome Somehow associated with the exposure Should not be an intermediate between exposure and outcome That is what we have in STROBE (Box)
C (4) New theory Dissatisfaction with loose definition of ‘determinant of’ and ‘somehow associated’ All possibilities spelled out in causal graphs where arrows are no longer ‘descriptive’ but denote causes
C (5) Several possibilities Figures from internet book by Hernan & Robins
C (6) “Confounder” is anything that can remove confounding if stratified upon Here, U is unknown, L is known – if U strong cause of L, suffices to stratify for L – Fine point: underlies new ideas about propensity score
C (7) What’s the big deal? The new view can more easily solve situations that are difficult to conceptualize with the classic definition In particular variables that are affected by exposure or by outcome – they often seem to satisfy definition of ‘confounder’ (are associated with exposure and outcome, are not intermediates). Yet, controlling for them introduces confounding
C (8) Example: the M-bias - L is associated with A via U2 and L is associated with Y via U1. U1 and U2 are unkown - There is no confounding because there is no direct arrow from L to either A or Y - If you adjust for L, you make U1 and U2 associated via L, and within its levels L becomes a confounder, as it will be a cause of A via U2 and a cause of Y via U1
C (9) There are other instances… However, in most situations the classic triangle suffices, provided: –You do not adjust variables that are affected by exposure or by outcome Except for M-bias, which will usually be weak (goes over too many arrows), the extra condition can also be seen logically without DAGs: other consequences of an exposure or events that follow the outcome should not enter the analysis (they give selection bias) – you’d never do that in RCTs!
C (10) Solution in “Modern Epidemiology” 3rd ed Rothman, Greenland, Lash Teach both: –in Chapt 9 classic, with warnings –In Chapt 11 DAG-based Mind: New theory is a ‘must’ in case of time- dependent exposures that affect therapy time- dependently & vice versa; main example is estimation of treatment effect in observational follow-up studies of HIV+ persons (ample literature by Hernan, Robins – experience in Bern)
C (11) STROBE (my proposal + Charlie Poole Stick to ‘classic’ with additional warning about variables affected by exposure or by outcome & mention advanced theory in available books Add that problems may become intricate and necessitate DAG-based theory in case of time- dependent exposures and treatments that follow from the exposure level but also affect the exposure level Show the elementary DAGs?? Charlie Poole: if people base important part of their reasoning on DAGs: show the DAGs
Selection Bias (1) In STROBE, Box, predominantly explained in context of case-control studies, as ‘wrong’ choice of cases (associated with exposure) or of controls (associated with exposure). Good explanation in follow-up studies and demarcation with confounding was always difficult & led to many an epi quarrel
SB (2) Hernan et al. Epidemiology 2004 Selection bias = collider bias in DAG theory: –Either because of selection in study –Or because of unwarranted adjustment in analysis (see previous theory about confounding) Permits common explanation of selection bias in case-control and cohort studies and reasonably clear demarcation from confounding
SB (3) Figures from internet book Hernan & Robins Case-control: exposure affects disease, but also inclusion in study as either case or control Follow-up study: differential loss to follow-up. Exposure affects loss to follow-up differently in exposed and nonexposed A very different DAG from the confounder DAG
SB (4) Others…
SB (5) New terminological confusion… Proposal by Greenland to call all selection biases “Berksonian bias”, whether or not due to design or analysis Proposal by Hernan/Robins to call –Confounding: common causes of exposure and outcome –Selection bias: common effects of exposure and outcome Proposal by DAG aficionados to omit all names, and only reason in terms of DAGs – DAGs will give the picture of the diverse biases that might operate, and they need no names.
SB (6) For STROBE (My proposal + Charlie Poole) Mention paper Hernan, new definition: selection bias = collider bias Give practical examples of case-control and follow-up Expand explanation of selection bias due to design of study by DAGs? (To show difference with confounding DAG??) Charlie Poole: if people base important part of their reasoning on DAGs: show the DAGs
Effect Modification & Interaction (1) New insights in several publications by VanderWiele/Robins, based on sufficient- component causes & counterfactual theory (also explained in internet book Hernan/Robins) VanderWiele, Epidemiology 2007, 2008, 2009, 2009 In STROBE: proposal of fourfold table (for two dichotomous effects), so that interaction can be assessed on a risk difference scale as well as on a relative risk scale.
EM & I (2) New conceptual proposal: Difference between effect modification and interaction (VanderWeele, On the distinction…, Epidemiology 2009) Effect modification: when the interest is in one causal effect that may differ in two situations, e.g. in RCT: one treatment effect that may differ over the strata of say, old vs. young. Comparison of two strata, each with their own RD or RR.
EM & I (3) New conceptual proposals Interaction: when the interest is in two causal effects that may influence each other. E.g. in RCT: two potential treatments whose effects may differ over the several combinations. Comparison of four strata with four risks, in terms of RD or RR.
EM & I (4) In observational research Effect modification: there is only interest in the confounders of the one cause that is investigated. No interest in the effect of the variables that define the strata (old, young). Like in RCT with single therapy. Interaction; there is interest in the confounders of both causes, and both need proper adjustment or control.
EM & I (5) New analysis proposals Basic measure of effect modification and interaction remains risk difference, based on counterfactual models. Handy rules for derivation whether data analyzed by a RR model (logistic, etc) will show interaction on a RD scale: –If effect monotonic (no cause has preventive effect in some individuals); suffices that RR of interaction >=1 –If not, interaction RR must be >=2
EM & I (6) for STROBE (My proposals) Introduce difference between Effect Modification & Interaction Refer to refinements of calculation rules to estimate RD interaction on a RR scale. Note: VanderWeele has paper in press in which he will propose an amendment on STROBE to treat effect modification differently from interaction (different lay-out of tables etc).
Mediation (1) Classic theory “When an intervening variable is controlled in an analysis, the initial association between the independent and the dependent study variables disappears or is markedly reduced” (Susser, Causal reasoning in the health sciences, 1973, page 122) Inversely, if adjusting does not remove effect, there is a ‘direct effect’ not mediated by this intermediary variable
M (2) In STROBE we wrote: item 16 Inappropriate decisions may introduce bias, for example, by including variables that are in the causal pathway between exposure and disease (unless the aim is to assess how much of the effect is carried by the intermediary variable) We got early letter saying that we were totally wrong, quoting Cole & Hernan IJE 2002
M (3) Cole & Hernan 2002 You think it is… but it is…… Unmeasured confounder between intermediary and outcome
M (4) Immediate critique: Also in real world? Blakely IJE 2002 Confounding needs to be very strong Needs to go in opposite directions, which is often counterintuitive
M (5) All possibilities for residual effect after adjustment for intermediary. Le Cessie et al, unpublished
M (6) For STROBE Box about mediation?? – mainly stating that classic idea too simple, many reasons why a ‘pseudo-direct effect’ remains One reason being unmeasured confounding of relation between intermediary and outcome (which has as yet to be proven as a real life possibility)