Published online June 1, 2007
PEDIATRICS
Vol. 119
No. 6
June 2007, pp.
1252-1253
(doi:10.1542/peds.2007-0725)
Neonatal Brain Volumetric Studies: Regression Analysis and Interpretation: In Reply
Nehal A. Parikh, DO
Department of Pediatrics
Robert E. Lasky, PhD
Kathleen A. Kennedy, MD, MPH
Jon E. Tyson, MD, MPH
Center for Clinical Research and Evidence-Based Medicine
University of Texas Medical School
Houston, TX 77030
We thank Dr Vasileiadis for his letter highlighting some key aspects of our article and observational data sets in general.1 We agree with his comments about the challenges of assessment and adjustment for confounding factors to uncover any true associations between risk factors of interest and cerebral volumes. His letter describes one of several statistically acceptable approaches for variable selection and model building.2–4 As outlined in the statistical-analysis section of our article, we used prespecified criteria for inclusion of covariates to avoid bias in selecting the regression model. The goal of variable selection is to achieve a balance between simplicity (with as few independent variables as possible) and fit (with as many independent variables as needed). Although the inclusion of additional covariates to improve model fit seems desirable, adjustments made by models that include too many covariates, even when statistically significant, may be spurious. Especially for preliminary smaller studies such as ours, inclusion of too many variables in the final model will produce numerically unstable models because of overfitting.5 Therefore, a simpler model may actually estimate the true relationship more accurately than a model with many independent variables.
It is evident from the work of Hüppi et al,6 and the strong association we reported, that postmenstrual age at MRI scan has the greatest impact on brain volumes. As such, we adjusted all of our analyses for this important covariate. Including birth weight in the final model addressed group differences in size at baseline. Adjusting brain volumes for differences in body size at MRI scan, as suggested by Dr Vasileiadis, could understate the true effect of steroids on regional and total brain volumes if the neurologic effects of dexamethasone are mediated through adverse effects on growth/size because of the reduction in growth resulting from dexamethasone. A more common approach in assessing regional volumes is to adjust for total intracranial volume as a covariate in the regression equation. Because dexamethasone may reduce overall brain growth and intracranial volume, this approach could also understate regional volume differences. In any case, adjusting for total intracranial volume resulted in little or no change in the P values in our 5 regions of interest (P values ranged from .07 to .001).
With respect to including uncomplicated germinal matrix-intraventricular hemorrhage as a covariate, we did not find an association between cerebral volumes and this potential determinant or, as we tested in our study, white matter injury, a more powerful determinant of brain volumes.7 As stated in our discussion, our significant associations may have resulted, in part, from incomplete adjustment of measured and/or unmeasured confounders. Such a problem, of course, is not unique to this area of investigation; any observational study may suffer from residual confounding. Confounding is less likely to occur in randomized trials, which is the best study design for resolving the effects of postnatal steroids on brain development in high-risk preterm infants. We are currently in the process of completing such a trial8 and hope to address these lingering questions.
REFERENCES
- Parikh NA, Lasky RE, Kennedy KA, et al. Postnatal dexamethasone therapy and cerebral tissue volumes in extremely low birth weight infants.
Pediatrics. 2007;119
:265
–272[Abstract/Free Full Text]
- Harrell FE Jr, Lee KL, Mark DB. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors.
Stat Med. 1996;15
:361
–387[CrossRef][ISI][Medline]
- Harrell FE Jr.
Regression Modeling Strategies. New York, NY: Springer; 2001
- Katz MH.
Mutivariable Analysis: A Practical Guide for Clinicians. 2nd ed. Cambridge, United Kingdom: Cambridge University Press; 2002
- Hosmer D, Lemeshow S.
Applied Logistic Regression. 2nd ed. New York, NY: John Wiley & Sons Inc; 2000
- Hüppi PS, Warfield S, Kikinis R, et al. Quantitative magnetic resonance imaging of brain development in premature and mature newborns.
Ann Neurol. 1998;43
:224
–235[CrossRef][ISI][Medline]
- Thompson DK, Warfield SK, Carlin JB, et al. Perinatal risk factors altering regional brain structure in the preterm infant.
Brain. 2006;130
:667
–677[CrossRef][ISI][Medline]
- A randomized trial of hydrocortisone in very preterm infants at high risk for neurologic and pulmonary impairments. Available at: www.clinicaltrials.gov/ct/show/NCT00167544. Accessed April 9, 2007
PEDIATRICS (ISSN 1098-4275). ©2007 by the American Academy of Pediatrics
