| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Letter |
Cancer Research Center of Hawaii, University of Hawaii, Honolulu, Hawaii
In the recent report from the Honolulu-Asia Aging Study, the authors analyzed the association of midlife tofu consumption with brain function and structural changes in late life by using linear and logistic multivariate regression models [1]. We are concerned that the incorrect imputation of missing exposure values has led to an overestimate of the odds ratios.
Questions concerning consumption of tofu were asked at the 1965–1968 and 1971–1974 interviews among 3734 participants tested with the Cognitive Abilities Screening Instrument (CASI) at the 1991–1993 examination. Although the authors performed initial analyses based on the dietary exposure at the two time points separately, most analyses are presented for a 4-level composite index based on both tofu variables and categorized as low-low, low, high and high-high. There are 596 individuals who were non-respondents to the 1971–1974 interview and hence have no information for the second measurement of tofu consumption. The authors have implicitly assigned the missing values for tofu consumption for all 596 individuals; they have treated the missing individuals in the same way as those who reported one serving of tofu per week in the second interview. As a result, subjects with missing exposure values were put into either the "low" or "high" categories, but could not be placed into the "low-low" or "high-high" categories. However, the individuals who were missing on the second measurement should have had the opportunity to belong to any of the four categories. This implied imputation can result in biased estimation of associations and variances due to misclassification of some subjects. In particular, the report found that the "low" group had an elevated risk of a lower CASI score, compared to the reference "low-low" group (odds ratio=1.42, 95% confidence interval=(1.06–1.90)). However, 30% (466 out of 1574) of the "low" group had a missing tofu value at the second interview. Therefore, it seems plausible that the missings in this group (n=466) are at higher risk than the individuals in the "low-low" group. If some of the 466 subjects had been assigned to the reference group using probability-based imputation, the baseline risk in the reference group would be higher, and the relative risk in the "low" group would be reduced. Therefore, the assignment of missings only to the "low" and "high" groups probably led to an overestimate of the odds ratios in this report. Alternatively, the analyses could have been done omitting the non-responses to the second dietary assessment if there was evidence to demonstrate that the data are "missing at random" [2].
There is now a sizable biostatistical literature on the treatment of missing data and on the effect of misclassification of exposure on estimation of risk [2,3]. There are many possible ways to handle the problem of missing data. Probability-based imputation is a general and flexible method for dealing with missing data problems. In these data, the missing second values could have been imputed based on regression models of tofu consumption in the second interview on other variables, such as the tofu consumption in the first interview, age, birth in Japan and years of childhood spent in Japan. Multiple imputation allows for the variability introduced by the imputation to be taken into account in the estimation of the variance of risk estimates. In addition, the results of several approaches to handling the missing values (e.g., probability-based imputation, omitting those subjects) could have been compared in a sensitivity analysis. A thorough discussion of the possible effects of misclassification is needed in the interpretation of analyses based on mismeasured or missing data.
In conclusion, we believe that it is premature to issue dietary recommendations about the risk of eating tofu. The problem outlined here, as well as other methodological issues in observational studies of diet and disease, suggest that the results should be interpreted more cautiously.
References
Pacific Health Research Institute, Honolulu (L.W., H.P.), Hawaii
University of Hawaii at Manoa, Honolulu (L.W., H.P., G.W.R., K.M., J.H.), Hawaii
Department of Veterans Affairs, Honolulu (L.W., G.W.R.), Hawaii
Kuakini Medical Center, Honolulu (H.P., K.M.), Hawaii
Louisiana State University (J.N.), Baton Rouge, Louisiana
University of Kentucky (D.D.,W.M.), Lexington, Kentucky
Guo and co-workers express concern relating to the possible impact of missing data and imputation methods on our finding of an association of midlife tofu consumption with functional and structural evidence of accelerated brain aging in members of the Honolulu-Asia Aging Study cohort (JACN, April 2000). In longitudinal studies, missing data due to attrition or partial non-response are often predictive of morbidity, including dementia or cognitive decline. This is certainly relevant to their concerns.
In the HAAS we are most fortunate to have dietary data obtained twice in midlife (1965–1967 and 1971–1974) for the prediction of neuropsychological endpoints assessed later (1991–1993). As Guo and colleagues noted, our initial analyses utilized data from these two dietary assessments separately. We reported poor cognitive function in late life to be associated with midlife tofu intake in independent analyses using diet data collected at the 1965–1967 and 1971–1974 interviews separately. Neither missing data nor imputations were involved in either of these analyses.
Their questions concern the four-stratum composite index we created by simultaneously considering tofu-consumption answers provided by the same participant at the two interviews. The "low-low" and "high-high" strata included only persons with no missing data and who were reported consistently low or consistently high in tofu consumption at both interviews. Our intent was to focus the analysis on participants whose usual dietary patterns during midlife were ascertained with greatest certainty. The "low" and "high" strata were defined on the basis of intermediate or inconsistent consumption frequencies or by using only the 1965–1967 interview when information was unavailable from the 1971–1974 interview.
To address the concerns of Guo and colleagues as specifically as possible, we conducted supplemental analyses identical with those used to generate Table 2 in our paper, except that the "low" and "high" tofu intake strata were split into subsets depending upon whether the 1971–1974 intake data were present or missing. Associations of the resulting six strata of midlife tofu intake with late life cognitive impairment were estimated for the full cohort (3081 normal, 555 cognitively impaired) or after exclusion of participants with a history of stroke (2818 normal, 420 cognitively impaired). Odds ratios and 95% confidence limits associated with the 11 control factors were essentially unchanged. Results for the tofu strata are shown in the Table here.
As is apparent, limiting the analysis to persons with no missing data caused the odds ratio for cognitive impairment associated with "low" tofu intake to fall to 1.26 from 1.42 (for the full "low" group, as noted in our paper), becoming statistically non-significant, but remaining midway between the "low-low" and "high" groups. Otherwise, excluding persons without 1971–1974 intake information had minimal influence on our basic findings.
Results for Tofu Strata
| |||||||||||||||||||||||||||
The higher odds ratios observed among persons with missing data suggest that non-response at the 1971–1974 examination was by itself an independent predictor of subsequent cognitive impairment. While non-response is widely recognized as an early indicator of morbidity, the actual mechanism underlying this association cannot be discerned in our study. In the men with missing information from the 1971–1974 interview, the odds ratio for cognitive impairment was somewhat greater among the "high" tofu consumers (2.25) than among the "low" consumers (1.87). The numbers of subjects in these strata were small, and the difference was not statistically significant.
As regards the final paragraph of their letter, it appears that Guo and co-workers have imputed a recommendation that did not appear in our paper. We did not propose "...to issue dietary recommendations about the risk of eating tofu." The profound importance of these issues requires that the response to our observations not be premature or hasty. Recommendations on the basis of our findings alone would be premature. Nonetheless, should our findings be supported by other research, the public health implications will be enormous. Similarly, assurances that soyfoods and their pharmacologically active constituents can be consumed in any amount without risk should be based on solid, adequate research observations rather than on the absence of such information.
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
