| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
US Food and Drug Administration, Center for Biologics Evaluation and Research, 1401 Rockville Pike, Rockville MD 20852
American Beverage Association, 1101 16th Street, NW, Washington, DC 20036, E-mail: mstorey{at}ameribev.org
Dear Sir:
We appreciate the opportunity to respond to the letter from Malik et al regarding our recently published meta-analysis (1) of the relation between sugar-sweetened beverage consumption and body mass index (BMI; in kg/m2) in children and adolescents. Malik et al are correct that the studies published by Blum et al (2) and Newby et al (3) should have been scaled by a factor of 12, because they estimated their models using 1 oz as the unit of analysis for beverage consumption. Both of these studies estimated a small negative association between sugar-sweetened beverage consumption and BMI that was not statistically significant. Scaling these 2 studies by a factor of 12 results in an overall estimate of 0.03 (95% CI: –0.01, 0.07) BMI unit change per serving compared with the original estimate of 0.02 (95% CI: –0.01, 0.04). Although we regret the scaling error, this does not affect any of the substantive conclusions of the article. An erratum, featuring corrected versions of Figures 1 and 2 and a corrected version of the online-only data set (see "Supplemental Data" in the current online issue), also appears in this issue of the Journal (4).
We respectfully disagree with the other points raised in the letter. First, we used the correct weights in our analysis and reported both the fixed-effects and random-effects estimates and the test for heterogeneity. Anyone can verify this by checking the code we provided [see the option "second(random)" under Supplemental Data in the current online issue]. The forest plot we used simultaneously displayed both the fixed- and random-effects estimates so that readers could compare the 2 results. Second, Phillips et al (5) presented the most challenges for scaling, because it used a unique combination of methods and units. We contacted the corresponding author to request additional information, but we did not receive any more information on the study. We considered excluding the study from the quantitative analysis and discussing it qualitatively, but we were concerned that excluding a statistically significant study would be inappropriate. As part of our overall sensitivity tests for influential studies, we estimated a model excluding the Phillips et al study, and the overall random-effects estimate changed from 0.02 to 0.01. In the corrected data, the random-effects estimate changed from 0.03 to 0.02 when Phillips et al's study was excluded.
After careful consideration of the options that were available and assessment of the impact they would have on our results, we used the coefficient and SE (calculated from the P value using p2ci) for those females who consumed the largest percentage of calories from soda in the BMI z-score model (Table 3). We also explored an alternative scaling that converted the BMI z-score for that difference to BMI units using the LMS method (6) and then divided by the estimated number of servings based on the percentage of calories from soda data (Table 1). This alternative scaling was very close to the original—0.121 compared with the original 0.178—and did not affect the estimate of the overall effect size. We were uncomfortable with the number of assumptions used in the scaling, particularly because the percentage-of-calories-from-soda variable could be affected by either the numerator (energy from soft drinks) or the denominator (total energy). We believe that we exhausted all options to put the Phillips et al study in a common scale, selected the most defensible option, and demonstrated through extensive sensitivity tests that no other plausible option would lead to a substantively significant difference in our findings.
Third, Malik et al argue that statistical models that do not adjust for total energy are more appropriate and would lead to a different conclusion. This argument hinges on an empirical and a theoretical question. The empirical question is whether the results would be different if all the studies had reported unadjusted results. If the results differ, the theoretical question is whether adjusted or unadjusted models are most appropriate.
The empirical question could be answered if all of the study authors provided adjusted and unadjusted results. Without the data, it is impossible to know whether the alternative results would lead to different conclusions. In the small number of studies that reported both adjusted and unadjusted figures, the differences were generally small and not in a consistent direction. For example, Ludwig et al (7) reported a 0.20 unadjusted and a 0.24 adjusted coefficient. Berkey et al (8) reported a 0.028 unadjusted and a 0.015 adjusted coefficient for males and a 0.021 unadjusted and a 0.019 adjusted coefficient for females. If these differences are typical, there will be little difference between meta-analyses based on adjusted or unadjusted coefficients.
There is no consensus that unadjusted models are the most appropriate. Otherwise, all of the studies would have reported unadjusted results. One of the primary concerns with unadjusted models is omitted variable bias. If the excluded energy is correlated with both soft drink consumption and BMI, the coefficient for the soft drink consumption will be biased. Even Willett (9) emphasized the importance of controlling for energy. He made the point: "[b]efore attributing an effect to a specific nutrient, however, the burden is on the epidemiologist to demonstrate that the association of this nutrient with disease is independent of caloric intake" (p 284). We were following this advice and common practice in the field by focusing on the adjusted values.
Fourth, the alternative, stratified analysis presented by Malik et al cannot show that the results of the adjusted studies would have been different if they had not adjusted for energy. That question can only be answered by discovering what the unadjusted results would be. We are also concerned that the analysis misclassified as unadjusted the Phillips et al study. Phillips et al used the percentage of calories from soda (calories from soda divided by total energy) as the independent variable. Such a variable is highly related to caloric intake (7, p 285) and is not comparable with unadjusted studies. Removing Phillips et al from the unadjusted group reduces the effect size to 0.06 (95% CI, 0.01, 0.10), and the difference between adjusted and unadjusted models is no longer statistically significant.
Fifth, ignoring the 4 studies that collectively showed a –0.03 association (95% CI: –0.11, 0.04)—as proposed by Malik et al—and placing all of the weight on the 5 studies that collectively showed a 0.08 association (95% CI: 0.03, 0.13) does not present an accurate picture of the overall body of evidence.
Finally, the authors of the letter say that our focus on adjusted values suggests "the presence of bias." Readers may judge for themselves whether we have adequately corrected our minor error, explained our approach, and responded to the other arguments. We are confident that the study meets the highest scientific standards for transparency and replicability. An independent scientist critiqued our methodology, reviewed the manuscript, and provided comments on our study design and methods at various stages of the project. We have made all of our data and analysis files publicly available for extended peer review. If using adjusted values is evidence of bias, we are in good company, which includes the authors of all the articles that used only adjusted values.
ACKNOWLEDGMENTS
The original research was conducted while all of the authors were affiliated with the University of Maryland and was supported with funding from the American Beverage Association (ABA). RAF is currently a Senior Risk Assessment Expert with the US Food and Drug Administration (FDA), Center for Biologics Evaluation and Research. MLS is currently a Senior Vice President for the ABA. PAA is a consultant for the ABA. The views expressed in this article are those of the authors and may not represent those of the University of Maryland, the ABA, the FDA, or any other organization with which any of the authors may have been affiliated in the past. RAF declared no conflicts of interest other than the original source of funding for the project. MLS is employed by the ABA but had no other conflicts of interest. PAA is a consultant for the ABA but had no other conflicts of interest.
REFERENCES
This article has been cited by other articles:
|
|
 |
|
  M. S Vanselow, M. A Pereira, D. Neumark-Sztainer, and S. K Raatz Adolescent beverage habits and changes in weight over time: findings from Project EAT Am. J. Clinical Nutrition, December 1, 2009; 90(6): 1489 - 1495. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
