Presentations in this series Introduction Self-matching Proxies Intermediates Instruments

Presentations in this series1. Introduction2. Self-matching3. Proxies4. Intermediates5. Instruments6. Equipoise

Avoiding Bias Due toUnmeasured Covariates

Alec Walker

T D

X

T D

XRandomization

T D

XRandomizationSelf-matching

T D

XSelf-matchingProxies Proxies

Randomization

T D

XSelf-matchingProxies Proxies

Randomization

IntermediatesIntermediates

7

A Cautionary Example:Vaccination for influenza is associated with a reduction in the apparent risk of almost every subsequent serious health event.

8

Group Health Cooperative of Puget Sound, 1995-2002

“We evaluated a cohort of 72527 persons 65 years of age and older followed during an 8 year period and assessed the risk of death from any cause, or hospitalization for pneumonia or influenza, in relation to influenza vaccination, in periods before, during, and after influenza seasons. Secondary models adjusted for covariates defined primarily by diagnosis codes assigned to medical encounters.”

Jackson LA, Jackson ML, Nelson JC, Neuzil KM, Weiss NS. Evidence of bias in estimates of influenza vaccine effectiveness in seniors. International Journal of Epidemiology 2006;35:337–344

9

....

...

.

...

..

...

.

Vaccinees had lower risk than nonvaccinees of hospitalization for pneumonia or influenza during and after the influenza season, as expected.

10

Vaccinees had lower risk than nonvaccinees of hospitalization for pneumonia or influenza during and after the influenza season, as expected.

....

...

Vaccinees also had lower risk than non-vaccinees before the influenza season began.

11

Vaccinees had lower risk than nonvaccinees of hospitalization for pneumonia or influenza during and after the influenza season, as expected.

....

...

Vaccinees also had lower risk than non-vaccinees before the influenza season began.

This pre-season reduction cannot have been a causal effect of vaccination.

12

Vaccinees had lower all-cause mortality than non-vaccinees before, during and after flu season.

13

Adjusting for many baseline factorsatrial fibrillation, heart disease, lung disease, diabetes mellitus, dementia, renal disease, cancer, vasculitis/rheumatologic disease, hypertension, lipid disorders, pneumonia hospitalization in previous year, and 12+ outpatient visits

slightly magnified the bias.Some of the controlled factors may have been instruments, resulting in Z-bias with respect to unmeasured confounders.

Influenza Vaccine

(Low) Mortality

Routine preventive care

General good

health

Baseline risk

factors

15

Confounder with a mediated effect on disease

E D

C

I

16

Post-treatment intermediates

Vaccine Death

Anticipation of future treatments

Immunosuppressive treatments

Cancer

time

17

Control for Post-treatment intermediates

Vaccine Death

Anticipation of future treatments

Immunosuppressive treatments

Cancer

time

18

Control for Post-treatment intermediates

Vaccine Death

Anticipation of future treatments

Immunosuppressive treatments

Cancer

time

19

Control for Post-treatment intermediates

Vaccine Death

Anticipation of future treatments

Immunosuppressive treatments

Cancer

time

Ignorance of the indications for therapy may justify controlling for a “downstream” time-varying covariate.

20

Intermediate Variables for Predictors of Treatment

21

• Propensity models are typically constructed in such a way as to capture proxies for predictors of treatment (previous lecture)

22

• Propensity models are typically constructed in such a way as to capture proxies for predictors of treatment (previous lecture)

• They also captureintermediates

U

I

T D

23

• Propensity models are typically constructed in such a way as to capture proxies for predictors of treatment (previous lecture)

• They also captureintermediates

U

I

T D

24

Intermediate Variables – Summary • Blocking the a variable on a unique causal path from a

confounder to either– Outcome or– Treatmentis sufficient to block the confounding effect

• In medicine we almost never know what a doctor is thinking about a patient, but we do often know his or her actions. These are intermediate variables on the pathways that tie – Diagnosis– Prognosis, and – Treatment

• Events that follow after treatment are not necessarily intermediates, and should be controlled if they are intermediates for unmeasured confounders.

Presentations in this series1. Overview

and Randomization2. Self-matching3. Proxies4. Intermediates5. Instruments

Avoiding Bias Due toUnmeasured Covariates

Alec Walker

25