class: center, middle, inverse, title-slide # Measurement bias ## What If: Chapter 9 ### Elena Dudukina ### 2021-03-03 --- # 9.1 Measurement bias .pull-left[  ] .pull-right[ - Independent errors  ] --- # 9.2 The structure of measurement error: dependent misclassification .pull-left[ - No single structure (unlike for confounding or selection bias) - Independence and nondifferentiality ] .pull-right[ - Dependent errors  ] --- # 9.2 The structure of measurement error: differential misclassification .pull-left[ - Differential misclassification - Outcome ➵ how exposure was measured (recall bias) - Exposure ➵ how outcome is measured (detection bias) ] .pull-right[ - Independent errors - Differential misclassification   ] --- # 9.2 The structure of measurement error: dependent & differential misclassification .pull-left[ - Dependent - Outcome ➵ how exposure was measured (recall bias) - Exposure ➵ how outcome is measured (detection bias) ] .pull-right[ - Independent errors - Differential misclassification   ] --- # Fine point 9.1 - Measurement error will result in bias - Except if A and Y are unassociated and the measurement error is independent and nondifferential - "The magnitude of the measurement bias depends on the magnitude of the measurement error" - "Causal diagrams do not encode quantitative information, and therefore they cannot be used to describe the magnitude of the bias" --- # 9.3 Mismeasured confounders .pull-left[ - A: drug use - Y: liver disease - L: hepatitis history via questionnaire ] .pull-right[ - the backdoor A ⬅️ L ➡️ Y cannot be blocked by conditionning on L*  ] --- # 9.3 Mismeasured confounders .pull-left[ - A: drug use - Y: liver disease - L: hepatitis history via questionnaire ] .pull-right[ - the backdoor A ⬅️ L ⬅️U ➡️Y cannot be blocked by conditionning on L*  ] --- # 9.3 Mismeasured confounders - Mismeasurement of confounders may also lead to appearance of effect modification (EMM) - If L=0 and L=1 strata differently report `\(L*\)`, stratification by `\(L*\)` would produce appearance of EMM --- # 9.3 Conditioning on mismeasured collider - Selection bias  --- # 9.4 Intention-to-treat effect: the effect of a misclassified treatment .pull-left[ - Z: randomization - A: treatment - Y: outcome - U: unmeasured ] .pull-right[ - Z ➡️ Y arrow is present when there is unblinding or allocation concealment failure - exclusion restriction (assumption: no arrow from Z to Y) - Effect of Z is intention-to-treat effect (ITT)  ] --- # 9.5 Per-protocol effect - Causal effect of the treatment itself and not of randomization to treatment level - Lack of exchangeability between A=1 and A=0 - Back to observational epidemiology realm --- # Fine point 9.2 As-treated vs per protocol .pull-left[ As-treated - Y=1 in those A=1 vs Y=1 in those A=0 regardless of Z - Confounded (9.11 & 9.12) - Feasible and estimable given L (9.13) Per protocol - Only those who have adhered to the study protocol (A=Z) - Y=1 in those Z=1 vs Y=1 in those Z=0 - ITT in per-protocol population - Selection bias ] .pull-right[   ] --- # ITT ITT is a lower bound for per-protocol effect - ITT is closer to null (conservative) - Not in safety studies - ITT is null if there is no effect - ITT assumes monotonicity (same direction of the effect in all individuals) --- # Per protocol - Time-varying - Report both ITT and per-protocol effects - trade-off between bias due to potential unmeasured confounding vs misclassification of exposure - IV approach - Adjustment approach --- # References Hernán MA, Robins JM (2020). Causal Inference: What If. Boca Raton: Chapman & Hall/CRC (v. 31jan21)