A Genetic Instrumental Variables Analysis of the Effects of Prenatal Smoking on Birth Weight: Evidence from Two Samples

Abstract

There is a large literature showing the detrimental effects of prenatal smoking on birth and childhood health outcomes. It is somewhat unclear, though, whether these effects are causal or reflect other characteristics and choices by mothers who choose to smoke that may also affect child health outcomes or biased reporting of smoking. In this paper, we use genetic markers that predict smoking behaviors as instruments to address the endogeneity of smoking choices in the production of birth and childhood health outcomes. Our results indicate that prenatal smoking produces more dramatic declines in birth weight than estimates that ignore the endogeneity of prenatal smoking, which is consistent with previous studies with non-genetic instruments. We use data from two distinct samples from Norway and the United States with different measured instruments and find nearly identical results. The study provides a novel application that can be extended to study several behavioral impacts on health and social and economic outcomes.

Acknowledgments

This research uses data from the Norway Facial Cleft Study (NCL) and Add Health. The research using the NCL sample was supported by NIDCR 1 R03 DE018394 and 1 R01 DE020895-01 and in part by the Intramural Research Program of the NIH, National Institute of Environmental Health Sciences. Add Health is a program project designed by J. Richard Udry, Peter S. Bearman, and Kathleen Mullan Harris and funded by grant P01-HD31921 from the National Institute of Child Health and Human Development, with cooperative funding from 17 other agencies. Special acknowledgment is due Ronald R. Rindfuss and Barbara Entwisle for assistance in the original design. We thank Jeremy Green for helpful comments and research assistance. The authors do not have any conflicts of interest in this work.

Notes

*George L. Wehby and Jason M. Fletcher contributed equally to this paper. For questions on the Norway sample analysis, contact George L. Wehby([email protected]). For questions on the Add Health analysis, contact Jason M. Fletcher([email protected]).

¹The use of genetic instruments is sometimes referred to as “Mendelian Randomization” in epidemiology, but the approach is a standard instrumental variable application with genetic instruments (Wehby, Ohsfeldt, and Murray 2008).

²See study Web site http://www.niehs.nih.gov/research/atniehs/labs/epi/studies/ncl/publications.cfm

³The genotyped sample did not include samples of 171 control children due to inadequate and/or low quality DNA samples. The analytical sample excludes 85 cases from the genotyping sample due to missing data on the study variables, including genotypic data. The data loss is not correlated with any characteristics that are related to smoking and birth weight as described below, and is therefore thought to be random and not systematic.

⁴The first trimester is a critical period for fetal development and maternal exposures during this period are likely to have large effects on fetal growth and birth weight. The majority (about 74 per cent) of mothers who smoked at pregnancy continued to do so during the first trimester.

⁵In a sensitivity analysis, we estimated the impact of daily smoking (i.e., a minimum of 1 cigarette per day) during the first trimester on birth weight.

⁶The smoking rates in the Norway sample reported at the first prenatal visit at about 10.3 weeks of gestation on average (available through the Medical Birth Registry) were lower than first-trimester smoking rates reported after delivery in the NCL survey, suggesting that smoking status later in pregnancy may not accurately reflect first trimester smoking due to quitting during pregnancy (Lie et al. 2008).

⁷We identify first births by examining whether the reported age of the mother at the time the pregnancy ended was the lowest age of all observations for that mother.

⁸The Add Health sample had no data on number of cigarettes smoked—only number of packs smoked.

⁹These expectations are based on maternal perceptions of the biologic and environmental risk factors that contribute to maternal and child's health (health, endowments). However, maternal risk perceptions and health endowments are inadequately measured in typically available datasets for birth outcomes studies, including the samples used in this study. The net bias in estimating the effects of smoking on birth weight by estimating Equation 1 via OLS is a function of the average positive and negative biases in an available sample and cannot be signed a priori. For example, a potential mother may choose to smoke prior to pregnancy in part due to her perceptions of her health risks and how these risks might be affected by smoking. During pregnancy, the mother will decide to continue to smoke or stop in part due to her perceptions of her health risks during pregnancy, of fetal health risks, and of the effect of smoking on these risks. Thus, a woman may decide to smoke during pregnancy in part because she considers herself to be healthy and considers that continuing to smoke will have no adverse effects on her and her child's health. In this case, unobservable indicators of health endowments that are positively correlated with both smoking and child health (such as no history of low birth weight in the family or in previous pregnancies, history of smoking without health problems in the family, and others) may result in an underestimation (positive bias) of the negative effects of smoking on birth weight. Conversely, mothers who smoke are likely to have, on average, stronger preferences for current versus future consumption and are therefore more likely to engage in other unhealthy behaviors besides smoking, such as poor nutrition, lack of exercise, drug use, overall risk taking, and others. These factors likely have negative effects on birth weight. If some of the health behaviors correlated with smoking and birth weight are unobserved, as is typically the case in available datasets, the effects of smoking on birth weight may be overestimated (negative bias). The net bias in estimating the effects of smoking on birth weight is a function of the average positive and negative biases in an available sample and cannot be signed a priori.

¹⁰A few studies (Evans and Ringel 1999; Rosenzweig 1983, among others) have used an IV strategy to account for self-selection into prenatal smoking. As we discussed earlier, the main contribution is to exploit a new source of variation generated by genetic instruments that vary at the individual level as opposed to group or state level instruments of tax rates.

¹¹ http://zork.wustl.edu/nida/Results/data1.html.

¹²These were GSTM1, UGT1A7, NAT1, NAT2, CYP2E1, CHRNA4, GSTT1, CYP2D6, GABBR2, GABRB3, DDC, GAD1, and KCNJ2. We also consider SNPs in the gene ACTN1 which has been identified in a recent GWA study of smoking and that have been genotyped in this sample (Caporaso et al. 2009).

¹³rs1041983 (NAT2), rs1930139 (GABBR2), rs1432007 (GABRB3), and rs4906908 (GABRB3) are correlated with smoking participation and are used as instruments. rs1041983 (NAT2), rs721398 (NAT2), rs1432007 (GABRB3), rs2059574 (GABRB3), rs5758589 (CYP2D6), rs2268973 (ACTN1) are correlated with the number of cigarettes and used as instruments. NAT2 and CYP2D6 are genes of detoxification pathways and have been implicated in several types of cancer including breast, prostate, bladder cancer, and others (Abdel-Rahman et al. 2000; Sanderson, Salanti, and Higgins, 2007) especially when combined with smoking and alcohol, though results are generally inconsistent across studies. Given that the study is limited to women of childbearing age and is focused on birth outcomes, it is unlikely that these variants affected the studied outcomes through their effect on cancer risks.GABBR2 and GABRB3 are genes that code receptors for the neurotransmitter GABA, which are involved in neurological inhibition. GABBR2 has been implicated in smoking behaviors. Previous studies of smoking genetics that included GABRB3 did not report significant results in coding for cytoskeletal proteins and has overall no well-documented disease associations and functions, but a SNP in ACTN1 has recently been found in a GWA study to be significantly related to a threshold indicator of number of cigarettes per day (Caporaso et al. 2009).

¹⁴All of these genes except for DRD4 are considered to be candidate genes for smoking by the NICSNP Nicotine Project. However, DRD4 have been linked in previous studies to smoking behaviors (Laucht et al. 2008; Hutchison et al. 2002). Comings et al. (1996) find an association between DRD2 and smoking behavior, whereas Jin et al. (2006) find an association between MAOA and smoking behaviors, and Gerra et al. (2005) find an association between 5-HTT and smoking behaviors.

¹⁵There is no consistent evidence in the literature for interactive effects between the genetic variants used as instruments in the analysis of the Norway sample. Therefore, we do not include interaction terms between these variants in the Norway sample analysis. Using binary indicators as instruments for the main effects of these genetic variants as done in the Norway data model is suboptimal as these indicators have insignificant effects, which weakens the first stage.

¹⁶Four new SNPs have recently been added to the Add Health data, but they were weaker predictors of smoking behaviors of pregnant women in our sample than those we use in this paper.

¹⁷This may occur due to the correlations between alleles that are tightly linked within a certain genomic area on a certain chromosome (referred to as linkage-disequilibrium).

¹⁸See Fletcher and Lehrer (2009); Ding et al. (2006, 2009); and Norton and Han (2008) for other applications that use genetic markers as instruments. These studies use a similar approach to evaluate the instrument validity (Ding 2006; Ding et al. 2009; Fletcher and Lehrer 2009; Norton and Han 2008).

¹⁹Appendix C reports the full regression results along with the tests of the IV assumption, and Appendix D reports the coefficients of the first stages of the 2SLS models.

²⁰The exogeneity of smoking participation is not rejected based on a Hausman test (0.189).

²¹The exogeneity of cigarettes is not rejected based on a Hausman test (p = 0.107).

²²Appendix F reports the full OLS and 2SLS regression results.

²³The Add Health Wave IV data have recently been released. At a reviewer's request, we attempted to increase our sample by including the births that occurred between the two waves of data. Though we were able to double our sample, the instruments were somewhat weaker in the larger sample, (F-stat < 2). The point estimates were nearly identical to those in the tables (−570 in the large sample versus −586 in this paper). The reason for the weaker instruments in the large sample are unknown but could be related to the composition of the new births, which are from older, more advantaged mothers. It appears that these genetic variants play a smaller role in the smoking decisions of these mothers.

²⁴First-stage results are available in Appendix G.

²⁵In the Add Health models that adjust for twin-birth status, the smoking coefficient is −144 (marginally significant), −551 and −571 under OLS, 2SLS, and LIML, respectively. The instrument effects on smoking are also virtually unaffected (F-statistic = 4.22). Further detailed results are available upon request.

²⁶The rates of smoking reported after delivery in the Norway sample were higher than those reported during the first prenatal visit. This may be due to some mothers' stopping smoking before their first prenatal visit (around 10 weeks of pregnancy on average in this sample) or due to underreporting of smoking during prenatal visits; Lie et al. 2008). Therefore, biased reporting of smoking cannot be completely ruled out as a potential contributor to underestimation of smoking effects by OLS in the Norway sample.