American Journal of Epidemiology Advance Access originally published online on May 8, 2007
American Journal of Epidemiology 2007 166(1):17-18; doi:10.1093/aje/kwm095
American Journal of Epidemiology Copyright © 2007 by the Johns Hopkins Bloomberg School of Public Health All rights reserved; printed in U.S.A.
Response to Invited Commentary
|
Terry et al. Respond to "Antecedents of Obesity"
Mary Beth Terry1,
Ying Wei2 and
Denise Esserman2,3,4
1 Department of Epidemiology, Mailman School of Public Health, Columbia University, New York, NY
2 Department of Biostatistics, Mailman School of Public Health, Columbia University, New York, NY
3 Department of Biostatistics, School of Public Health, University of North Carolina, Chapel Hill, NC
4 School of Medicine, University of North Carolina, Chapel Hill, NC
Correspondence to Dr. Mary Beth Terry, Department of Epidemiology, Mailman School of Public Health, Columbia University, 722 West 168th Street, Room 724 A, New York, NY 10032 (e-mail: mt146{at}columbia.edu).
Received for publication February 5, 2007.
Accepted for publication February 21, 2007.
 |
INTRODUCTION
|
|---|
The commentary by Drs. Gillman and Kleinman (
1) on our paper
examining pre- and postnatal predictors of adult body mass index
(BMI) (
2) highlights the analytic and design challenges of conducting
these types of investigations.
|
Choice of model
|
|---|
We agree with Gillman and Kleinman that employing a variety
of analytic approaches illuminates the sensitivity of findings
to model specification. We compared standard analytic approaches
(ordinary linear and logistic regression) with quantile regression.
Logistic regression and its generalization, multinomial (polytomous)
regression, estimates the covariate effect on the risk of exceeding
a prespecified threshold determined by the marginal distribution
of the outcome (
3). These techniques are useful for applications
in which a natural cutoff value exists (
4,
5). In our case,
sparse cell counts at suggested cutoff points (>30 kg/m
2,
<18.5 kg/m
2) limited meaningful polytomous modeling. Quantile
regression does not rely on prespecified thresholds; rather,
it examines how the conditional distribution of the outcome
varies with the covariates (
6). While all three statistical
models in our study supported associations between maternal,
infant, and early-life factors and adult BMI, quantile regression
allowed for additional inferences regarding the relative magnitudes
of the associations for smaller, average, and larger women.
|
Exposure construct
|
|---|
Although we considered alternative exposure constructs, we selected
percentile changes because our interest was in the impact of
percentile rank changes relative to constant percentile rank
(growth trajectory) on adult BMI. We found similar associations,
however, when modeling
z score differences. In models based
on absolute weight measures, rather than rates, weight at age
7 years was the only postnatal measure associated with adult
BMI. This was probably due to the strong positive correlation
in absolute measures over time as opposed to the weaker and
negative correlation in growth rates. The negative correlation
between birth weight and postnatal growth rates means that few
high birth weight babies increase in percentile rank; however,
those that do have a higher BMI in adulthood, as our final models
suggest.
|
Model-building
|
|---|
Although we used standard progressive modeling (table 2) to
assess potential mediation, we recognize that these methods
may be limited in the presence of confounding and interaction
(
7). We also used path modeling (
8) to assess mediation but
did not fully explore its multivariate regression extension—structural
equation modeling (
9)—because it relies on modeling the
mean level. Both quantile regression and path analyses confirmed
that all of the variables included in our final model were independent
predictors.
|
Design challenges
|
|---|
We were fortunate to build upon data from the New York City
site of the National Collaborative Perinatal Project, which
prospectively collected pre- and postnatal data on offspring
up to the age of 7 years (
10). We agree that clinical evaluations
of anthropometric factors are necessary to explore plausible
biologic explanations for these associations; however, given
the geographic diversity of the adults in our study, we were
limited to questionnaire data on body size. Empirical evaluation
of human life-course questions will always rely on piecing together
information from multiple lines of evidence and multiple time
periods. Rejuvenating the National Collaborative Perinatal Project
cohort and similar birth cohorts may be an efficient first step
(
11).
|
Data inference
|
|---|
Gillman and Kleinman (
1) argue that targeting interventions
to the more narrow pregnancy window may ultimately prove more
efficient for obesity prevention. Although our models suggested
that maternal weight gain may be one important feature, other
variables such as childhood weight gain and maternal BMI remained
associated with BMI at age 40 years. The overall fit of the
model was based on the collection of variables and not any single
variable. Limiting pregnancy weight gain may result in many
adverse outcomes, including lower birth weight among babies,
who then may be at risk of obesity from rapid postnatal growth.
Continued focus on maintaining a healthy BMI throughout life,
however, will improve a woman's own health and may ultimately
prove to influence her offspring's.
|
ACKNOWLEDGMENTS
|
|---|
Conflict of interest: none declared.
|
References
|
|---|
- Gillman MW, Kleinman K. Invited commentary: antecedents of obesity—the analysis, interpretation, and use of longitudinal data. Am J Epidemiol (2007) 166:14–16.[Abstract/Free Full Text]
- Terry MB, Wei Y, Esserman D. Maternal, birth, and early-life influences on adult body size in women. Am J Epidemiol (2007) 166:5–13.[Abstract/Free Full Text]
- Hosmer DW, Lemeshow S. Applied logistic regression. 1st ed. (1989) New York, NY: John Wiley and Sons, Inc.
- Terry MB, Gammon MD, Schoenberg JB, et al. Oral contraceptive use and cyclin D1 overexpression in breast cancer among young women. Cancer Epidemiol Biomarkers Prev (2002) 11:1100–3.[Abstract/Free Full Text]
- Terry MB, Neugut AI, Bostick RM, et al. Risk factors for advanced colorectal adenomas: a pooled analysis. Cancer Epidemiol Biomarkers Prev (2002) 11:622–9.[Abstract/Free Full Text]
- Koenker R, Hallock K. Quantile regression. J Econ Perspect (2001) 51:143–56.
- Kaufman JS, Maclehose RF, Kaufman S. A further critique of the analytic strategy of adjusting for covariates to identify biologic mediation. Epidemiol Perspect Innov (2004) 1:4.[CrossRef][Medline]
- Alwin DF, Hauser RM. The decomposition of effects in path analysis. Am Sociol Rev (1975) 40:37–47.[CrossRef][ISI]
- Bollen KA. Structural equations with latent variables. (1989) New York, NY: John Wiley and Sons, Inc.
- Broman S. The Collaborative Perinatal Project: an overview. In: Handbook of longitudinal research—Mednick SA, Harway M, Finello KM, eds. (1984) Vol 1. New York, NY: Praeger Publishers. 185–227.
- Susser E, Terry MB. A conception-to-death cohort. Lancet (2003) 361:797–8.[CrossRef][ISI][Medline]
CiteULike 
Connotea 
Del.icio.us What's this?
Related articles in Am. J. Epidemiol.:
- Maternal, Birth, and Early-Life Influences on Adult Body Size in Women
- Mary Beth Terry, Ying Wei, and Denise Esserman
Am. J. Epidemiol. 2007 166: 5-13.
[Abstract]
[FREE Full Text]
TOP
INTRODUCTION
Choice of model