eLetters to:

ELECTRONIC ARTICLE: Beena Sood, Virginia Delaney-Black, Chandice Covington, Beth Nordstrom-Klee, Joel Ager, Thomas Templin, James Janisse, Susan Martier, and Robert J. Sokol Prenatal Alcohol Exposure and Childhood Behavior at Age 6 to 7 Years: I. Dose-Response Effect Pediatrics 2001; 108: e34 [Abstract] [Full text] [PDF]

eLetters: Submit a response to this article

eLetters published:

Analyses inadequate to justify conclusions

R Curtis Ellison   (31 August 2001)

Response to Dr Ellison's letter

B Sood, "V Delaney-Black, C Covington, B Nordstrom-Klee, J Ager, T Templin, J Janisse, RJ Sokol"   (19 September 2001)

Analyses inadequate to justify conclusions 31 August 2001
R Curtis Ellison, Professor of Medicine & Public Health Boston University School of Medicine Send letter to journal: Re: Analyses inadequate to justify conclusions ellison{at}bu.edu R Curtis Ellison	My colleagues and I are concerned that the analyses described in this article have not adequately tested the hypothesis that "low" prenatal alcohol consumption is related to the behavioral outcomes described. All of the results of categorical comparisons (displayed in Tables 2-4 and Figure 1) are apparently based on univariate results only, and do not take into consideration potential confounding by the marked baseline differences among categories shown in Table 1. Thus, the results attributed to prenatal alcohol exposure could well be due to confounding by birth weight, maternal age, percent married, cigarette use, cocaine use, custody changes, SES, etc. What are needed are multivariable analyses to see if alcohol consumption in the "low" category relates to outcomes while adjusting for these factors. The step-wise regression model used to generate data for Table 5 has problems as well: (1) the use of a continuous variable for alcohol exposure does not allow for determination of the effects of low levels of intake; (2) prenatal alcohol exposure was transformed to a log scale to normalize the distribution of the data; it is unclear how the authors performed such transformation for subjects with zero alcohol intake (one cannot log-transform a value of 0); (3) use of standardized coefficients precludes interpretation of alcohol effects at various levels;1 (4) there is no assessment of possible collinearity among the many highly correlated variables included in the stepwise model; (5) the fact that prenatal alcohol remained as a "significant predictor" of certain outcomes may well be the result of confounding and, in any case, these latter analyses were not constructed in a manner permitting a comparison with the results from the univariate analyses. In sum, there is no meaningful adjusted epidemiologic measure of effect presented in this paper, so the conclusions drawn are far beyond the scope of the data. My colleagues (Kenneth J. Rothman, DrPH, Lynn L. Moore, ScD, and Yuqing Zhang, ScD) and I have worked in research on alcohol for many years and are anxious to have valid information on this very important topic. While the source data for this article are from a very atypical population for the US, in that there were a very large number of women drinking (and using drugs) during pregnancy, it is a cohort that could be used to test hypotheses regarding prenatal alcohol consumption. If the authors do not wish to do so, our group would be very happy to carry out multivariable analyses of the categorical data they have collected. Only when such results are available will it be possible to judge from this study the degree to which prenatal drinking affects behavioral outcomes in the child. 1 Greenland S, Maclure M, Schlesselman JJ, Poole C, Morgenstern H. Standardized regression coefficients: a further critique and review of some alternatives. Epidemiology 1991;2:387-392.
Response to Dr Ellison's letter 19 September 2001
B Sood, Physicians and Statisticians Wayne State University, "V Delaney-Black, C Covington, B Nordstrom-Klee, J Ager, T Templin, J Janisse, RJ Sokol" Send letter to journal: Re: Response to Dr Ellison's letter bsood{at}wayne.edu B Sood, et al.	Dr Ellison and his colleagues are concerned that the analyses described in this article have not adequately tested the hypothesis that “low” prenatal alcohol consumption is related to the behavioral outcomes described. It seems that they would like to see an analysis of covariance or multiple regression using categorical prenatal alcohol exposure. These analyses had been done and the results were not significantly different from what we presented. When a 3-level prenatal alcohol exposure variable was used (no, low and moderate/heavy), and birth weight, maternal age, marital status, cigarette use, cocaine use, custody changes and SES were included as covariates, there were still significant associations between categorical prenatal alcohol exposure and, for example, Externalizing (p=0.041) and Delinquent (p=0.043). Regression analyses were also performed using the 3-level exposure variable and the control variables listed in the manuscript. Again the results are not significantly different from what we presented. Categorical prenatal alcohol exposure was a significant predictor for Externalizing (p=0.002), Aggression (p=0.006), Delinquent (p=0.014), Internalizing (p=0.029) and Total Score (p=0.008), with categorical prenatal alcohol exposure accounting for 0.5 to 1.0% of unique variance in behavior. To further validate the conclusion that “low” levels of prenatal alcohol consumption are associated with adverse behavioral outcomes, we had also performed all the analyses using only two levels of prenatal alcohol exposure – no and low – which were defined in the original manuscript. These similar analyses also revealed that “low” levels of prenatal alcohol exposure were significantly and substantially associated with various behavior problems. Using ANCOVA the results were Externalizing (p= 0.014), Aggression (p= 0.018), Delinquent (p= 0.028), and Total Score (p= 0.037). We are confident we did the appropriate log transformation of the alcohol variable (adding a minimal constant to all scores prior to the transformation), and the reviewers were satisfied. And yes, collinearity was checked. There is no mystery as to why all these results were not presented. The tension between what will fit in the few printed pages and what is important is often difficult. It is not possible to indicate every little thing that was done within any research article. Even though some readers apparently want to see everything, the reviewers and the publisher do not. The authors’ decide which analyses best convey the results. It was our choice to make and the results are not very different. Dr. Ellison’s other comments suggest that he is unsure of our statistical abilities. Be reassured we have multiple degrees and many years of research expertise. Our unaltered conclusion, based on appropriate statistical methods and controls is that even low-level prenatal alcohol exposure is a significant predictor of parent-reported child behavior. Our well-lettered research group has over two decades of research experience in the area of prenatal alcohol exposure and includes three faculty members who teach statistics at the University level. While we were amused at the writers’ offer to re-analyze our data, we decline and are certain that the authors of the letter to which we respond are not as arrogant as they sound. Beena Sood, MD Virginia Delaney-Black, MD, MPH Chandice Covington, PhD, RN Beth Nordstrom-Klee, MA Joel Ager, PhD Thomas Templin, PhD James Janisse MA Robert J. Sokol, MD