3 Causal Models Part I: Sufficient Causes Matthew Fox Advanced Epidemiology 1 2 Review of This Morning “Modern” epidemiology Goal of etiologic research – Why not everything is as we were taught – Statistics Review of study designs, measure of effect – – 3 Valid and precise estimate of the effect of exposure on disease Odds ratios and case-control sampling Hopefully changed a few minds (and set the tone) This Session Why are we in this business? – Things may have missed in intro epi – How causal models and causal inference helps clarify what we do and how we do it Sufficient Causes Model – 4 What is the goal of epi investigations? Rothman’s Model (Sufficient Cases Model) I turn on a light switch and the light doesn’t go on. The globe (bulb?) is burned out. What prevented the light from going on? 5 I turn on a light switch and the light does go on. What caused the light to go on? 6 Is there any disease for which all cases are attributable to one and only one cause? 7 Smoking causes lung cancer. So why doesn’t every smoker get lung cancer? 8 Is the occurrence of disease deterministic or random within what we know? Is a coin flip random? 9 What is a “strong” risk factor? What determines strength? 10 Why study causal models? Helps understand why we do what we do – Gives meaning to our measures of effect – 11 What do we mean when we say that smoking causes lung cancer? Theoretical and practical Helps clarify important epidemiologic concepts A definition of a cause An antecedent event, characteristic, or condition that was necessary for the occurrence of disease at the time it occurred all other things being fixed – – – – 12 Antecedent Necessary At the time it occurred Other things fixed The Sufficient Cause Model Ken Rothman 13 History Goes back to John Leslie Mackie (1964) Necessary causes: – – Sufficient causes: – – If x is a sufficient cause of y, then x necessarily implies y Since other things can cause y, y does not imply x Typical use of cause refers to: – 14 If x is a necessary cause of y, then y necessarily implies x x does not imply y will occur Insufficient and non-redundant parts of unnecessary but sufficient causes (INUS) The Sufficient Cause Model Lots of ways to get a disease – – Mechanisms exist independent of us – 15 Think of each way as a pie Called a sufficient cause But we’re susceptible to them if we acquire the components Go through life picking up exposures and filling in pies The Sufficient Cause Model Person is susceptible to multiple diseases – – – – – 16 Diseases have multiple sufficient causes Each sufficient cause has multiple component causes Each component cause has attributes Shared components between sufficient causes It is theoretically possible every case of outcome has a unique pie Take home message 1: All disease causation is multifactoral and mechanisms are complex 17 Sufficient Causes Minimally sufficient – Necessary cause – – Component cause appearing in all sufficient causes for a disease Poole proposes “universally necessary” Complementary component causes – 18 Each sufficient cause has a unique set of components and none is extraneous Set of component causes required to complete a sufficient cause, aside from one (exposure) How much do we know? U is often largest piece of a sufficient cause – But if we could specify the mechanisms perfectly, could we predict all disease? – – 19 We understand poorly disease causation Deterministic Might be infinite combinations In risk factor epidemiology, we focus on one component and ignore the complement So what might HIV infection look like? One SCM (pie) might be: – Exposed to HIV through sex Unvaccinated (OK for now), no natural immunity No condom use – Why doesn’t everyone exposed through sex get HIV? U – – Is each component necessary? – – 20 Circumcision? STDs? Genetic factors? No use of microbicide? If so, take away one and you prevent disease Must be more causes than just the HIV virus Other SCMs exist – Mother to child, transfusion, needle stick, etc. The Sufficient Cause Model Components can be positive or negative – Component causes should be specific – Can be identical except for timing Don’t need to understand entire pie to prevent – 21 Lack of vaccination Removing one piece renders the pie incomplete Disease and Causation The sufficient cause acts when all of the component causes have been gathered Disease occurs at completion of temporally last component cause – Each disease occurrence has a latency – 22 Model is deterministic Time between its occurrence and its detection Component Causes (Exposures) Component causes (i.e. exposures we might want to study) have attributes – – – 23 Dose Duration Induction period (NOT LATENCY) Specifying the attributes improves the resolving power of the study Does smoking cause lung cancer? Smoking is too imprecise – – Infinite combinations – Each might have a different risk Ignoring the attributes means we are lump all exposures (from 1 lifetime cigarette to a 10 packs a day) together as exposed – 24 How much for how long? What type? Starting at what age? This biases towards no effect! Dose attributes Time weighted average dose – Maximum dose – Grams of alcohol per kilogram body weight Cumulative dose – 25 Highest adult body weight Body weight or surface area scaled – Grams of fat per day Pack years of cigarettes Duration attributes Total time of exposure – Biologically relevant time of exposure – Years of driving after age 25 Time of exposure after gathering another component cause – 26 Smoking before first pregnancy Time of exposure beyond a minimum – years employed HIV infection after HPV infection Take home message 2: To study effects of exposures, the exposures must be precisely defined 27 Induction period attributes Induction period is the time between completion of a COMPONENT cause (i.e. the exposure of interest) and completion of the SUFFICIENT cause (i.e. disease occurrence) – – – – 28 Induction period doesn’t characterize disease Characterizes component cause-disease pair Every disease has a component cause with zero induction time Failure to exclude induction time from person time biases towards the null Take home message 3: Diseases don’t have induction times 29 Induction Time Example Diethylstilbestrol, adenocarcinoma of the vagina – – – Other processes assumed to occur in the interim – – 30 A synthetic non-steroidal estrogen, given to pregnant women to prevent miscarriage (’40s-’70s) Exposure is known to have occurred during gestation Cancer occurs in the offspring between 15-30 years of age Other components in the causal pie still occur Adolescent hormonal activity may be one If outcome can’t occur before 10 years, don’t include 1st 10 years of person-time – Similar to immortal person-time What about promoters or catalysts? Catalyst of diseases – Are they causes of disease? – – 31 Anything that speeds up (or slows down) the occurrence of a disease that would occur anyway Remember the “at the time it occurred” part of the definition of a cause Does it matter to you? Applications of the sufficient cause model The effect of the index condition, relative to the reference condition: – – Interaction between component causes – 32 The number of completed sufficient causes among those with index condition Minus number of completed sufficient causes among those with reference condition Arises when one or more sufficient causes contains both component causes Take home message 4: Interaction is ubiquitous 33 Strength is Determined by Complements Strength of a risk factor – Is determined by the relative prevalence in the population of the causal complements – 34 Typically measured on the relative scale Also affected by the competing risks of other sufficient causes for the same disease Imagine a gene-environment interaction U U G E Phenylketonuria (PKU) - a genetic disorder characterized by a deficiency in the enzyme to metabolize the amino acid phenylalanine. Untreated, it can cause problems with brain development. However, PKU is a rare genetic diseases that can be controlled by diet, one low in phenylalanine. 35 Imagine a gene-environment interaction Imagine a population where: – – – 36 10% get disease through U no matter what 60% of the population has U and G completed If we randomly assigned the exposure (diet), would the relative risk be high or low? Imagine a gene-environment interaction Randomized to get E Randomized to not get E U G U E G U E G U E G U E G U E G U E U Exposed 70% get disease (10% U + 60% U/G/E) G U G U U G U RR = 70%/10% = 7 37 U U G G G Unexposed 10% get disease (10% U) Imagine a gene-environment interaction Imagine a population where: – 10% get disease through U no matter what – – 38 Same as 1st example 10% of the population has U and G completed If we randomly assigned the exposure, would the relative risk be high or low? Imagine a gene-environment interaction Randomized to get E Randomized to not get E U U Exposed 20% get disease (10% U + 10% U/G/E) Unexposed 10% get disease (10% U) G U E G U RR = 20%/10% = 2.0 39 Imagine a gene-environment interaction Imagine a population where: – – 40% get disease through U no matter what 10% of the population has U and G completed – 40 Same as last If we randomly assigned the exposure, would the relative risk be high or low? Imagine a gene-environment interaction Randomized to get E U U U U Exposed 50% get disease (10% U + 40% U/G/E) Randomized to not get E U U U U G U E G U RR = 50%/40% = 1.25 41 Unexposed 40% get disease (40% U) Another example Vaccines are usually extremely protective – But what if we tested our vaccine in a population where everyone had natural immunity? – – – 42 What we often call a strong protective effect Lack of natural immunity is a piece in the pie (causal complement) When it is common, the effect appears strong When it is rare, the effect appears weak Take home message 5: “Strength” of an exposure’s effect is a function of: 1) prevalence of causal complements (what we usually ignore) and 2) the % of all disease that goes through mechanisms without the exposure (U for short) 43 % with spina bifida Illustrates arbitrariness of effects: Effect of folic acid on: Spina bifida 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.85 0.65 0.55 Folic Acid No FA 0.35 0.25 0.05 RD = 0.2 RR=0.2 44 Assume same sample size (N=100) A RR=0.63 B RR=0.75 C % with neural tube closure Illustrates arbitrariness of effects: Effect of FA on: Neural tube closure 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.95 0.75 0.65 0.45 0.35 0.15 RD = 0.2 RR=1.3 45 Folic Acid No FA Same A RR=1.4 RR=2.3 B C Take home message 6: We are often best studying rare occurrences when possible 46 Proportion of disease caused by… What % of all cases are caused by genes? What % of are caused by environment (dietary phenalynine)? G U E 47 Phenketunuria - 100% of all cases Take home message 7: The % of all cases of a disease attributable to different causes can (and will) sum to > 100% 48 Advantages of the SCM Applies to disease mechanisms in individuals – Illustrates: – – – – – 49 Tells us about how disease occurs, not just what caused it How component causes act together Attributes of component causes How strength is function of complements The ubiquity of interaction Process of causation, although not complete Disadvantages of the SCM Cannot easily apply to populations – Deterministic in nature – – Assumes no randomness to disease causation Doesn’t easily deal with continuous variables Obscures importance of reference group definition – 50 Though our examples show can be used this way Comparison is between those with component cause and those without the component cause under study Disadvantages of the SCM Difficult to assess validity of measures of association – No distinction between mutable, immutable variables – 51 Can sex can be a component cause? Does not include temporal sequence – No easy definition of bias Could be revised to do so Mechanisms cannot be fully articulated