Reexamining the effects of gestational age, fetal growth, and maternal smoking on neonatal mortality
© Ananth and Platt. 2004
Received: 23 March 2004
Accepted: 01 December 2004
Published: 01 December 2004
Skip to main content
© Ananth and Platt. 2004
Received: 23 March 2004
Accepted: 01 December 2004
Published: 01 December 2004
Low birth weight (<2,500 g) is a strong predictor of infant mortality. Yet low birth weight, in isolation, is uninformative since it is comprised of two intertwined components: preterm delivery and reduced fetal growth. Through nonparametric logistic regression models, we examine the effects of gestational age, fetal growth, and maternal smoking on neonatal mortality.
We derived data on over 10 million singleton live births delivered at ≥ 24 weeks from the 1998–2000 U.S. natality data files. Nonparametric multivariable logistic regression based on generalized additive models was used to examine neonatal mortality (deaths within the first 28 days) in relation to fetal growth (gestational age-specific standardized birth weight), gestational age, and number of cigarettes smoked per day. All analyses were further adjusted for the confounding effects due to maternal age and gravidity.
The relationship between standardized birth weight and neonatal mortality is nonlinear; mortality is high at low z-score birth weights, drops precipitously with increasing z-score birth weight, and begins to flatten for heavier infants. Gestational age is also strongly associated with mortality, with patterns similar to those of z-score birth weight. Although the direct effect of smoking on neonatal mortality is weak, its effects (on mortality) appear to be largely mediated through reduced fetal growth and, to a lesser extent, through shortened gestation. In fact, the association between smoking and reduced fetal growth gets stronger as pregnancies approach term.
Our study provides important insights regarding the combined effects of fetal growth, gestational age, and smoking on neonatal mortality. The findings suggest that the effect of maternal smoking on neonatal mortality is largely mediated through reduced fetal growth.
Birth weight is arguably one of the strongest predictors of infant survival, yet its role as a causal predictor of mortality is poorly understood . This is at least partly because low birth weight (<2,500 g) is a construct of two intricately intertwined components: preterm delivery and reduced fetal growth, or both. Our lack of understanding of the complex relationship among birth weight, gestational age and perinatal mortality stems from mixing etiologically distinct pathways to mortality, namely effects chiefly due to fetal maturity (i.e., gestational age) versus those related to fetal growth.
Disentangling the intricate pathways of gestational age and fetal growth to neonatal mortality gets even more complicated by the consideration of a third factor – maternal smoking during pregnancy. Smoking has been clearly associated with poor reproductive outcomes, including increased risk of preterm birth, stillbirth, and a range of other outcomes [2–6]. Recent studies suggest a more direct and stronger association between maternal smoking and "fetal growth" (birth weight-for-gestational age) than with preterm delivery , suggesting that the effect of smoking on mortality may be largely mediated through restricted fetal growth rather than preterm delivery.
To better understand the relationship among these indices of "fetal wellbeing", we examined neonatal mortality in relation to standardized birth weight (i.e., z-score birth weight), gestational age, and smoking during pregnancy. We applied nonparametric logistic regression based on generalized additive models to examine neonatal mortality in relation to 3 factors.
Data for this study were derived from the 1998–2000 United States vital statistics data files (live births linked to infant deaths), assembled by the National Center for Health Statistics of the Centers for Disease Control and Prevention . The analysis was restricted to singleton live births, with neonatal mortality defined as deaths within the first 28 days. Gestational age assignment in these data are predominantly based on self-reported last menstrual period, with a small fraction (<5%) based on the clinical estimate . Further, the National Center for Health Statistics imputed missing gestational ages in these data files prior to release of the data .
Information on smoking during pregnancy was available in two forms on the vital statistics data: one as an indicator variable (yes or no), and the other as a continuous variable denoting the number of cigarettes smoked per day during pregnancy. Both of these smoking measures were based on maternal self-report. Information on smoking patterns across different trimesters in pregnancy was not available on the vital records.
Fetal growth was defined as birth weight-for-gestational age, and was expressed as gestational age-specific birth weight z-score. This z-score construct is interpreted as units of standard deviations from the population-specific mean birth weight at each gestational age. The z-score or standardized birth weight follows a Gaussian distribution with mean 0 and variance 1.
In addition to the full analysis, we also examined in a sub-analysis the impact of implausible birth weight/gestational age combinations on overall results. These implausible birth weight/gestational ages were identified if infants' birth weights were outside the gestational age-specific birth weight cutoffs . This was done to examine the impact of largely apparent gestational age errors (e.g., infant delivered at 26 weeks with a birth weight of 4,000 g) on neonatal mortality.
There were 11,677,103 singleton live births from which we excluded infants with missing birth weight or gestational age (n = 237,433), and birth weight <500 g or gestational age <24 weeks (n = 28,732). Since smoking data was not reported on vital statistics in California, Indiana, New York state, and South Dakota , births from these states were also excluded (n = 1,326,841). After all exclusions, 6,117,808 singleton live births remained for analysis.
We examined the distributions of z-score birth weight, gestational age, and number of cigarettes smoked per day, and compared these distributions between the two groups of neonatal mortality. Neonatal mortality was then modeled using nonparametric logistic regression based on generalized additive models . GAM is one modeling approach that makes no assumptions about the functional form of the exposure-disease relationship except for smoothness, i.e., continuity of the dose-response function and its low-order derivatives . When combined with more traditional modeling approaches, GAMs are powerful graphical tools that can provide interesting insights about complex relationships. While polynomial models  could be used to the same end as GAM-based approaches, such models result in restricted shapes, especially at the tail of the distribution, and may not be as statistically efficient as nonparametric models. Therefore, these models were not considered.
All regression models were adjusted for the confounding effects due to maternal age and gravidity (i.e., number of pregnancies). We examined the associations between neonatal mortality and each of the 3 factors z-score birth weight, gestational age, and number of cigarettes smoked per day separately. We then fit a full model for mortality after forcing all 3 predictors (in addition to the confounders) as described in the Appendix [see additional file 1]. The independent effect of each of these 3 factors on neonatal mortality was assessed by comparing the residual deviances  between nested models (i.e., comparing the residual deviances from a full model to a model without the predictor). Under the large sample assumption, the deviance has an approximate chi-square distribution, with degrees-of-freedom for the test being the difference in the degrees of freedom between the nested models being compared. We also examined the distribution of partial residuals  from fitting the model to assess departures from adequate fit.
In addition, we tested for all possible two-factor interactions between the predictors. Although all interactions were statistically significant (owing to the large study size), none provided any additional insights that were different from a model that contained no interaction terms. Therefore, we did not consider assessing two-way interactions in the analysis.
All statistical analyses were performed in S-Plus (Insightful Corporation, Seattle WA) version 6.2 on the UNIX (Sun Microsystems, Inc: Palo Alto, CA). Nonparametric logistic regression models were fit using the gam( ) function based on the it loess scatterplot smoother , using the default span of 50%. Given the large size of the study, small changes in the span resulted in statistically significant improvement in the fit, while offering very little clinical insight. Thus, we resorted to the default span.
Distributions of birth weight, gestational age, and maternal smoking in relation to neonatal survival status
Maternal age (years)†
Birth weight (grams)†
Birth weight <2,500 grams
Birth weight <1,500 grams
z-score birth weight†
Gestational age (weeks)†
Delivered <37 weeks
Delivered <34 weeks
Delivered <32 weeks
Smoking during pregnancy
For decades, several researchers have focused on trying to understand the complex biological relationship among pregnancy duration, infant size, and neonatal mortality. Not only are gestational age and birth weight highly correlated, but both are powerful predictors of neonatal mortality [14–16]. The chief findings from our study include (i) z-score birth weight and preterm delivery (independent of birth weight) exert strong influences on neonatal mortality; (ii) the effect of maternal smoking is mediated largely through reduced fetal growth and, to a smaller extent, through shortened gestation; and (iii) mortality among babies born to smoking mothers is virtually higher at every z-score birth weight (independent of gestational age) than those born to nonsmoking mothers.
The inverted "J"-shaped relationship between birth weight and mortality essentially holds for analyses relating to gestational age and mortality. While birth weight is considered a marker for fetal size, gestational age is thought of as an indicator of fetal maturity. Almost 3 decades ago, Susser and colleagues  proposed that gestational age is causally precedent to birth weight (implying that birth weight is in the causal pathway of the gestational age-mortality relationship). Wilcox and Skjaerven  examined close to 400,000 singleton births from Norway in an effort to separate the influences of birth weight and gestational age on neonatal mortality. They showed that, comparisons using the "relative birth weight" scale, there were two strong and separable factors related to mortality: gestational age independent of birth weight, and relative birth weight at any given gestational age.
On these similar lines, Herman and Hastie  examined neonatal mortality in relation to (absolute) birth weight and gestational age. They initially speculated that among preterm (<37 weeks) babies, maturity would serve as a strong predictor of mortality, while among term babies, the increased mortality was probably due to growth restriction. However, their analysis showed that mortality was associated only with birth weight and not with gestational age. Their approach to analysis may have suffered from collinearity (between birth weight and gestational age), perhaps leading to the attenuated gestational age-mortality relationship . Coory  analyzed neonatal mortality in relation to birth weight and gestational age. He concluded that both birth weight and gestational age have independent effects on mortality, and that both are fundamental risk-adjusting variables. However, he was cautious in not interpreting the effects of gestational age, but focused his interpretations almost entirely on birth weight. Our construction of standardized birth weight z-score was developed conditional on gestational age. Thus, this birth weight z-score (independent of gestational age) enabled us to assess the effects of shortened gestation and fetal growth restriction on mortality.
It is widely acknowledged that smoking mothers give birth to infants that are lighter compared with those born to nonsmoking mothers. This reduction in birth weight is thought mainly to result in fetal growth restriction, as well as to shortened gestation [19, 20]. Although the precise mechanism by which smoking during pregnancy affects the fetus is unclear, two possible pathways have been proposed. Smoking results in increased capillary fragility and vasoconstriction of arterial walls, leading to reduced blood flow to the uterus and eventually to the placenta . The second is the "fetal hypoxia" hypothesis, whereby smoking leads to a villous shrinkage due to an alteration in the thickness of the villous membrane, thereby reducing oxygen transfer to the fetus . Both mechanisms are likely to increase the risk of uteroplacental bleeding in pregnancy , which, in turn, increases the risk of not only neonatal deaths [20, 24], but also preterm delivery and growth restriction . Our study provides circumstantial evidence that after the general effects of (shortened) gestational age and (reduced) fetal growth are accounted for, smoking has little direct impact on neonatal mortality.
Our study has some limitations. First, errors in the estimation of gestational age [25, 26] are likely to affect our results to some extent. Our study was based on gestational age largely determined from the date of last menstrual period as opposed to one based on early ultrasound. Sonographically estimated gestational age is likely to shift the overall gestational age distribution to lower gestational ages  sometimes by as much as a full menstrual cycle , possibly due to delayed ovulation or amenorrhea. Second, the impact of congenital malformations and chromosomal abnormalities on the risk of neonatal death could have been partly responsible for the findings noted here. Although data on malformations are contained on the vital statistics files, they are recorded poorly. Third, although we adjusted all the analysis for maternal age and gravidity, the study does not take into account other known or suspected risk factors for neonatal mortality. These risk factors may account for a part of the associations noted here, but is unlikely that these factors could explain the powerful effects of fetal growth restriction and preterm delivery on neonatal mortality. Finally, non-differential misclassification of smoking data on vital records is likely  and may have attenuated the smoking-mortality association to some extent.
Application of generalized additive regression models to examine neonatal mortality appears useful towards understanding the complex biological relationship amongst the predictors. However, we make no claim that GAMs serve as adjuncts to other modeling approaches; on the contrary, we believe that GAMs can provide the first step toward modeling complex exposure-disease relationships.
Our study provides important insights about the combined effects of gestational age, fetal growth, and smoking during pregnancy on neonatal mortality. Both standardized z-score birth weight and preterm delivery are strongly associated with neonatal mortality, and the effect of maternal smoking appears largely mediated largely through reduced fetal growth and, to a smaller extent, through shortened gestation.
We thank Drs. KS Joseph, Michael Kramer, John Preisser, David Savitz, John Smulian, Anthony Vintzileos, and Michelle Williams, for their helpful discussions; for reviewing an earlier draft of the manuscript and providing critical comments and insights. We also thank the referees for their thoughtful comments that improved the manuscript considerably.
This paper was presented at the 15th annual meeting of the Society for Pediatric and Perinatal Epidemiologic Research, held in Palm Desert, CA, June 2002.
Dr Ananth is supported, in part, through a grant (R01-HD038902) awarded to him from the National Institutes of Health, USA. At the time this research was conducted, Dr Platt was a career scientist of the Canadian Institutes for Health Research, Canada.
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.