Sholom Wacholder additional slides

advertisement

SOME ADDITIONAL POINTS ON

MEASUREMENT ERROR IN

EPIDEMIOLOGY

Sholom

May 28, 2011

Supplement to Prof. Carroll’s talk II

Measurement error (ME) in epidemiology

• ME may be more important than confounding

• Examples from my work

• Best solution based on my experience:

– Avoid or minimize ME

Examples (1)

• Molecular epidemiology and biomarkers

– Reduce Coefficient of Variation (CV) by reducing lab error

• Population variation remains

• Genetics

– GWAS: power loss depends on LD between markers and tagging SNPs

• Best characterized by

R

2

• No theorems on D’ or recombination fraction

– Kin cohort analysis

• Known Mendelian rules allow inference of genotype from relative’s

Examples (2)

• Misclassification of disease

– E.g.: Screening leads to diagnosis

•  Misclassification of disease, differential by screening

•  Screening studies use mortality, not diagnosis, as endpoint

Differential misclassification

• Standard research practices minimize differential misclassification

– Occupational epidemiology: Industrial hygienists blinded to disease status

– Molecular epidemiology: Blinding in the lab

– Randomized controlled trial:

• Double blind treatment assignment

• Blinded (masked) ascertainment of disease

Differential misclassification:

Case-control

• Qx: “recall bias” is always a potential

• Direct evidence of recall bias is weak

• Molecular studies: biomarkers can be affected by disease progression

• Pre-diagnostic biospecimens needed

• Or perhaps from before disease initiation

• Cohort studies needed

• Genetics/genomics

– Genotype calling, QC

– DNA quality of cases and controls may be different

• Obtained from different sources

• Controls and cases from different studies

– As in studies of rare cancers using shared controls with previous genotypes

– Recall genotype from optical density?

Randomized Controlled Trials (1)

Differential misclassification of tx:

• Intention to treat (ITT) analysis

– Effect of dose assigned

– Controls confounding

– Estimate of effect often biased

• Effect of dose actually received

– May be more interesting

– Can be subject to confounding

• Solutions?

– ITT + understanding error model for dose received

– ITT + instrumental variable approach

Randomized Controlled Trials (2):

Differential misclassification

• HPV vaccine to prevent CIN2 due to HPV16/18

– CIN2 is cervical precancer

• Early trials used HPV assays only for HPV16/18

• Example of misclassification

– CIN2 caused by HPV type not affected by vaccine

– HPV16/18 found in lesion only in placebo arm

– Contributes to apparent benefit from vaccination

• NCI and other recent studies test for all oncogenic HPV types

Alloyed gold standard

• Use of alloyed gold standard in validation studies can lead to overcorrection for bias in regression calibration models

– Wacholder et al., 1993, PMID: 8322765

• Average of multiple 24h recalls can be distorted

Missing data (MD) vs. ME

• MD and ME: absence of true value of the variable

• ME: proxy variable available;

• MD: No proxy

• MD and ME: Statistical approaches available when you understand underlying mechanisms

Summary

1. Measurement error is pervasive

– Even in trials, not just epidemiology

2. Understand the causes of measurement error

– minimize bias;

– increase power, efficiency

– Understand mechanism generating observations via

– Pilot studies

– Validation studies

– Replication studies

– Inter-observer studies

1. Statistical insights can help at design and analysis stages

Download