Presentation on theme: "Presentation, data and programs at:"— Presentation transcript:
1 Presentation, data and programs at: Simple Causal GraphsSimple Casual GraphsHein StigumPresentation, data and programs at:Apr-17H.S.
2 Causal graphs Simple causal graphs Directed Acyclic Graphs (DAGs) Proper analysis (adjust or not)Direction of biasDirected Acyclic Graphs (DAGs)Formal toolInventory of variablesCausal inferenceApr-17H.S.
3 Exposure-Disease influenced by C C can be:ConfounderIntermediate (in 2. Path)ColliderEffect modifierUse graphsDetermine C-typeChoose analysisCEDApr-17H.S.
4 Example Exposure Disease Covariates Pysical Activity: PA Diabetes type 2: D2CovariatesSmoking: SHealth Conscious: HCOverweight: OvBlood Pressure: BPApr-17H.S.
5 Linear models Best model? Model choice can not be based on data only. Best model?Likelihood ratio tests or Akaike criteria mod 4All changes in PA effect considered important mod 4Claim mod 2.Model choice can not be based on data only.Need extra info or assumptions.All est significant (except PA=0)20% D2 overall. PA from 0-5. Max PA is -10 pp.Apr-17H.S.
7 Confounder: Smoking Should adjust for Smoking Stratify Regression S Negative bias+-biasedtruePAD2-3-2Should adjust for SmokingStratifyRegressionApr-17H.S.
8 Confounder 2 Adjust for Smoking or for Health Consciousness +-HCSNegative biasbiasedtruePAD2-3-2Adjust for Smokingorfor Health ConsciousnessAssume all following models are adjusted for smokingProblem: if effect from HC to D2 (via diet )If we include both, the simple pos/neg bias calc will not work unless both pathways have the same signCould adjust for HC if measured, or for both Smoking and Diet.DAGs will not give this info, a important shortcoming.Apr-17H.S.
9 Intermediate (in 2. path): Overweight +-Alt 1: Ignore OverweightTotalPAD2Alt 2: Two models:Direct cIndirect b1*cTotal c2+ b1*c1 -2.0OvOvc1Assume all these model are adjusted for smokingb1=-0.25, c1=2, b1*c1=-0.5b1c2PAPAD2Simply adjusting for Overweight is not OK!Apr-17H.S.
10 Select limping subjects Collider ideaTwo causes for limpingHip arthritisLimpKnee injurySelect limping subjects+Limp++Hip arthritisKnee injury-Conditioning on a collider induces an association between the causesCondition = (restrict, stratify, adjust)Bias direction?Hip arthritis and knee injury are not associatedHip arthritis and knee injury are rare eventsApr-17H.S.
11 Collider: Blood Pressure BPPositive bias if we adjust+-truebiasedPAD2Should not adjust for Blood PressureProblem if selection is connected to BPStudy of subjects with high BP, decide to analyze PA-D2 associationOr, based on invitation letter advertising BP-measurement, people with high BP select into studyApr-17H.S.
12 Best model (so far) Model 2 is best Used extra info in graphs to decideAll est significant (except PA=0)20% D2 overall. PA from 0-5. Max PA is -10 pp.Apr-17H.S.
13 Effect modifier: Sex Alt 3 : Ignore Sex Alt 1 : Two models PAD2Alt 1 : Two modelsEasyNo test for interactionInefficient (12 estimates)p=number of covariatesEstimates=2(p+1) versus p+2Alt 2: Model with interactionTechnicalTest for interactionEfficient (7 estimates)Apr-17H.S.
14 Effect modifier: Sex Model with interaction term Linear modelTest for interactionWald test on b3=0If significant interactionSex is coded 0 for Males and 1 for FemalesThe effect of PA (1 unit increase)-2.5-1.5Apr-17H.S.
16 Smoking and LRTI The truth is out there? LRTI=Lower Resperatory Tract InfectionsWant: effect of smoking in pregnancy on LRTI in childrenHave: 40% response, high education is overrepresentedBest causal estimate:Crude smoke-LRTI under 100% response?Crude smoke-LRTI under 40% response?LRTISmokeEduc-SSmoking in pregnancyLower Respiratory Tract InfectionsEducation a confounder?Strong selection on educ in data, effect?Education is a confounderSelection representspartial adjustmentApr-17H.S.
17 Smoking and LRTI, ex 2 LRTI Smoke Educ S Education is a not a confounderCrude smoke-LRTI in population is unbiasedCrude smoke-LRTI in sample is biased, S is a colliderAdjusted smoke-LRTI in sample is unbiasedApr-17H.S.
18 Ethnicity and lung function Exposure Ethnic groupOutcome Lung functionCovariates Hemoglobin, heightDraw DAGSuggest analyzes/modelsModel with all covariates meaningful?Lung funcHemoHeightEthnicApr-17H.S.
19 Models Model 1 Lung func Ethnic Model 2 Lung func Hemo Height Hart rateModel 3Model 4Lung funcHemoHeightEthnicModel 1: total effect of ethnic group on lung functionModel 2: adjusted effect of hemoglobin on lung functionModel 3: Add all covariates, if effect of ethnic group, then some physio variable is missingModel 4: ethnic group as a effect modifier of hemoglobin on lung functionApr-17H.S.
20 Summing upIn a study of 2 variables, a 3. variable may have 4 effects:Confounder, Intermediate, Collider, Effect modifierNot distinguished from data, need extra infoCasual graphs help use the extra infoThe 3 first situations can not be gleaned from the dataApr-17H.S.