Could exposure assessment problems give us wrong answers to nutrition and cancer questions?

Whether fruits and vegetables provide protection from cancer is of considerable public health importance—but it remains an open question. In this issue of the Journal, Hung et al. report findings from two cohort studies: fruit and vegetable intake is associated with a “modest” reduction in risk of noncommunicable (“chronic”) disease (1). This reduction is confined to cardiovascular disease. The association for cancer is null. Most evidence for a protective role of fruits and vegetables against cancer has comes from case–control studies. The authors rightly note the possibility of recall bias in such studies. Indeed, prospective cohort studies have tended to demonstrate weak or no associations between fruit and vegetable intake and malignant disease, including large-bowel cancer (2). So the question arises: does total fruit and vegetable intake really confer little or no cancer protection? (In this editorial we address only the relation of total fruit and vegetable intake to all cancer. However, similar issues can be raised for associations between individual fruits or vegetables and cancers at specific sites.) Hung et al. consider—and largely dismiss—the possibility that problems with exposure assessment have caused their study to give the wrong answer to this question. However, more consideration of this possibility is in order. The exposure assessment tool—food-frequency questionnaire (FFQ)—that is used to measure diet, including fruit and vegetable intake, is subject to substantial error, both random and systematic (3,4). Is this error sufficient to obscure an existing fruit and vegetable–cancer association? That is, could a true relative risk (RR), comparable to that seen by Hung et al. for cardiovascular disease (0.88 per increment of five daily servings of fruits and vegetables) be attenuated to the null by this error? Hung et al. argue against this possibility on three counts: First, they say that, by taking the cumulative average of several FFQs, they have reduced intraindividual random variation. This may be true, but such cumulative averaging doesn’t eliminate systematic error. Second, they assert that the FFQ has been validated against multiple weighted 1-week dietary records. We have argued elsewhere that such records are also subject to measurement error, and this error may well be correlated with that in the FFQ (4,5). Therefore the use of records as a reference instrument likely overestimates the accuracy of the food frequency questionnaire. Third, the authors assert that the observed association for cardiovascular disease means that the instrument they used is accurate enough to pick up any real association between fruit and vegetable intake and cancer. But suppose the true relative risk for cardiovascular disease is 0.75; attenuating 0.75 to 0.88 (for fruits and vegetables versus cardiovascular disease) is roughly comparable to attenuating 0.88 to 1.0 (for fruits and vegetables versus cancer). And this is considering only the univariate problem, relating disease to a single variable (fruit and vegetable intake) measured with error. The reality of nutritional epidemiologic research is multivariate: a dietary exposure such as total fruit and vegetable intake is generally examined in conjunction with energy intake and a variety of other covariates. The Hung et al. multivariate models include total energy intake, alcohol, smoking, vitamin supplement use, physical activity—13 variables in all for men and 16 for women. Measurement error affects determination of not only fruit and vegetable intake but also other dietary variables, such as alcohol and energy intake, and other covariates, such as physical activity. Moreover, the errors in measurement of these variables are likely to be correlated. The actual effect of this multivariate measurement error on true relative risk is complex. A true null relative risk could appear as an increased or decreased risk; a true protective association could be attenuated to the null (6). It would not take very much error in a few covariates (with perhaps a little intercorrelation of those errors) to attenuate a true protective association (RR 0.88, for example) to the null. Multivariate modeling in the presence of measurement error is not just an arcane statistical matter but a real concern in this field—especially if a serious attempt is being made to detect modest relative risks. Hung et al. observed a statistically significant relative risk of 0.92 for fruits and vegetables in a model including only age as a covariate. This relative risk was estimated as 0.97 (becoming statistically nonsignificant) when smoking was added to the model. They claim that the nearly null results in this simple model means that overadjustment by confounders is an unlikely explanation for the null findings in their multivariate model. However, even assuming (unrealistically) that measures of total fruit and vegetable intake contain only random error and that smoking is measured perfectly, adding smoking to the model would necessarily increase the attenuation due to random error in measurement of fruit and vegetable intake (7). This fact alone could potentially explain the additional attenuation of relative risk from 0.92 to 0.97. With more complex but more realistic measurement error, the situation could be even less predictable—and certainly consistent with attenuation of an important protective association.