Correcting for measurement error in nutritional epidemiology

Ruth Keogh

MRC Biostatistics Unit

MRC Centre for Nutritional Epidemiology in Cancer Prevention and Survival

IPH Showcase, 14 February 2011

Measuring dietary intake

• Food frequency questionnaires (FFQ)• Diet diaries/24 hour recalls• Biomarkers

There is no gold standard measurement

UK dietary cohort consortium

• EPIC-Norfolk + 5 other cohorts• Case-control studies nested within cohorts• 4-7 day diaries, FFQs• Colorectal cancer, breast cancer, prostate cancer

What is the association between long term or usual dietary intake and disease risk?

Food frequency questionnaires (FFQ)

Diet diaries

Diet diaries

Choice of instrumentFFQs• Designed to measure long term intake• Inexpensive used for large populations• Subject to substantial measurement error

Diet diaries• Measure actual intake• Very expensive to process• More highly correlated with objective biomarkers of intake• Still subject to error

UK dietary cohort consortium • One of only a small number of studies using diaries/24 hr recalls as

the main instrument• Interested in correcting for error in dietary diary measurements

Effects of measurement error1. Biased associations – usually attenuated

2. Loss of power to detect associations

3. Can hide nonlinear association shapes

Effects of measurement error1. Biased associations – usually attenuated

2. Loss of power to detect associations

3. Can hide nonlinear association shapes

Effects of measurement error1. Biased associations – usually attenuated

2. Loss of power to detect associations

3. Can hide nonlinear association shapes

Correcting for measurement error

i

i

T = True average daily intake

R =Diet diary measurement

True diet-disease association

i i ilogit {Pr(D =1|T )} =α+βT

Estimating when we can’t observe Ti

i i i ilogit {Pr(D =1|R )} =α+βE(T |R )

Linear regression calibration model

i 0 1 i iT = λ + λ R +e

Fitting the model

We need to understand the structure of the error in diet diary measurements Ri

i 0 1 i iT = λ + λ R +e

Random error in diary measurements:

i i iR = T +Suppose we have another diary measurement for some people

i1 i i1

i2 i i2

R = T +

R = T +In fact we can show that

i2 0 1 i1 iR = λ + λ R +e

Fibre intake and colorectal cancer

OR per 6g/day increase

95% CI

Using FFQ 0.90 0.75-1.08

Using diet diary 0.84 0.71-1.00

Corrected for measurement error in diary

0.77 0.62-0.97

Dahm CC, Keogh RH et al. Dietary Fiber and Colorectal Cancer Risk: A Nested Case–Control Study Using Food Diaries. JNCI 2010

However…We do not believe that diet diaries are subject only to random error

i i iRandom error R = T +

Biomarker studies suggest

• error in diaries depends on true intake• errors in replicate diaries are not independent

i i iSystematic error R = α+βT +

We cannot estimate the extra parameters using repeat measurements

• to correct for error we need unbiased measurements, e.g. biomarker• biomarkers not available for most nutrients

Different assumptions about the type of error in Ri can give very different corrected estimates

Challenges

• Measurement error can have severe effects on observed diet-disease associations

• What we assume about the form of the error can have strong effects on ‘corrected’ estimated associations

• Biomarkers can help us to understand the structure of the measurement error

Also…

• We often want to adjust for other dietary variables, such as total energy intake

• All dietary variables in the model are measured with error…