Evidence for a Causal Relationship Between Academic Achievement and Cigarette Smoking

Abstract

Introduction

Academic achievement (AA) is associated with smoking rates. Can we determine the degree to which this relationship is likely a causal one?

Methods

We predict smoking in male conscripts (mean age 18.2) assessed from 1984 to 1991 (N = 233 248) and pregnant females (mean age 27.7) receiving prenatal care 1972–1990 (N = 494 995) from AA assessed in all students at 16. Instrumental variable (IV) analyses used the instrument month-of-birth as in each school year, older children have high AA. Co-relative analyses used AA-smoking associations in the population, cousins and siblings to predict the AA-smoking relationship in MZ twins, thereby controlling for familial confounding.

Results

In males, higher AA was associated with a substantial decrease in risk for smoking (odds ratio [OR] [95% confidence intervals [CIs]] per standard deviation [SD] = 0.41 [0.40–0.41]) while the parallel figures obtain from our IV and co-relative analyses were 0.47 (0.39–0.57) and 0.51 (0.43–0.60), respectively. In females, these figures for pre-pregnancy smoking were, respectively, 0.39 (0.39–0.39), 0.50 (0.46–0.54) and 0.54 (0.51–0.58). Results for heavy versus light smoking suggested a causal effect but were inconsistent across methods. However, among females smoking prior to pregnancy, AA predicted a reduced risk for continued smoking with ORs for uncontrolled, IV, and co-relative analyses equaling, respectively, were 0.54 (0.53–0.55) 0.68 (0.56–0.82) and 0.78 (0.66–0.91), respectively.

Conclusions

Two different methods produced consistent evidence that higher AA has a causal effect on reducing smoking rates and increasing cessation rates in smoking pregnant females. Improving AA may result in meaningful gains in population health through reduced smoking.

Implications

This study provides consistent evidence across two different methods that high AA is causally related to reduced rates of smoking and increasing rates of smoking cessation among pregnant women. Our results suggest that interventions that improve educational achievement in adolescence would reduce tobacco consumption, thereby improving public health.

Introduction

Smoking is responsible for 11.5% of all global deaths and is the strongest risk factor for all-cause, all-age attributable disability-adjusted life-years in the United States and most other developed countries.¹ Smoking during pregnancy is associated with elevated rates of adverse outcomes, including low birth weight, birth defects, and growth retardation.² Identifying causal mechanisms linked to smoking can inform preventive interventions targeting those mechanisms to reduce the public health burdens of smoking.

Poor levels of academic achievement (AA) in the middle- and high-school have been associated, in observational studies, with elevated rates of smoking,^3,4 and lowered rates of quitting.⁵ Individual and school-based interventions for children and adolescents have shown that improving AA results in declines in drug use and smoking.^6,7 However, with rare exceptions, follow-ups have been too short to determine whether such interventions produce a longer-term impact on tobacco use. Furthermore, although these intervention studies suggest a causal relationship between AA and subsequent smoking, the magnitude of this association may be upwardly biased by a range of possible confounders not always accounted for. For example, familial or community factors may increase risk for both poor AA and high smoking risk.

To examine the nature of the association between AA and smoking behavior in adulthood, we utilize two causal inference methods.⁸ The first—an instrumental variable (IV) analysis⁹—utilizes month-of-birth as an instrument. Sweden has a cutoff date for enrollment in school so that in classes, nearly all students come from the same birth year and therefore differ in age by as much as 12 months. As in several other European countries, in Sweden, within the same class, older students perform better academically than their younger classmates.^10,11

Our second causal inference method utilizes a co-relative design¹² that examines the association between AA and smoking in the general population and then in cousins, full-siblings, and monozygotic (MZ) twin pairs. Our analyses focus on estimating, within MZ pairs, the magnitude of association between differences in AA and differences in risk for smoking, thereby controlling for genetic effects and the impact of rearing in the same household and community and exposures to similar schools and peer groups.

These methods are applied to two Swedish national samples with smoking data: 18-year-old male military recruits and females coming for standard pre-natal care at their first pregnancy. In both groups, we attempt to predict from levels of AA rates of current smoking and among smokers, quantity smoked. In pregnant females smoking prior to pregnancy, we predict from AA levels those who quit versus persist.

Methods

We analyzed information on individuals from Swedish population-based registers with national coverage. These registers were linked using each person’s unique identification number replaced by a serial number to preserve confidentiality. Ethical approval was provided by the appropriate Regional Ethical Review Board. We used data from The National School Registry that contains AA for all students in grade nine (usually age 16). Education is mandatory for all children in Sweden between age 7 and 16. From 1988 to 1997, scores were assessed by a peer-referencing system and had had minimal inflation over time and were normally distributed. From 1998 onwards, the score was expressed on a quantitative scale, which we standardized by year and gender.

We studied smoking status in two different registers. For males, we used the Swedish Military Conscription Register, which includes results of conscription examinations for ~97% of Swedish males usually assessed at age 18. Self-reported smoking was recorded in 7 different categories (None, 1–10, 11–20, >20 cigarettes/day, 1, 1–2, >2 packages of tobacco a week). In our analyses, we defined smokers as individuals with any type of smoking, and heavy smokers as >10 cigarettes/day or >2 packages a week. For females, we used the standardized prenatal exams from the population-based nationwide Swedish Medical Birth register covering 97%–99% of all births. These exams categorized maternal smoking into nonsmoker, smoking 1–9, or 10 or more cigarettes per day at two time points, 3 months prior to pregnancy, and at first trimester. We defined smokers as individuals with any type of smoking behavior, and heavy smokers as >10 cigarettes/day 3 months prior to pregnancy.

For males, we had information on both smoking behavior and school achievement for individuals born 1984–1991 while for pregnant females we had information on individuals born 1972–1990. We used information from the first pregnancy among females with more than one pregnancy. The school grade was normally made-up of individuals who turned 16 between January 1 and December 31. But for approximately 3% of individuals, AA was instead registered the year they turned 15 or 17. The month-of-birth of these individuals was not randomly distributed; more individuals born in January were registered with AA the year they turned 15. More individuals born in December were registered with AA the year they turned 17. As the association between month-of-birth and AA was central to our further analyses, we modeled this non-randomness. Using all individuals regardless of age at AA, we fitted a regression model predicting AA based on month-of-birth and age at registration. In the below analyses we used, for individuals registered the year they turned 15 or 17, their predicted value, and for individuals with AA the year they turned 16, their true value.

In the first model, we used logistic regression to investigate the risk of smoking as a function of high AA. The odds ratio (OR) represents the decreased risk for smoking for 1 standard deviation (SD) increase in AA. This is the AA-smoking crude association used for comparison. We then used an IV approach to control for unmeasured confounding using month-of-birth as an instrument.¹³ Below, we empirically evaluate the suitability of month-of-birth as an instrument. We used a two-stage-regression model. The first step was a linear regression of month-of-birth on AA. The predicted values were thereafter used in the logistic regression models as exposure variable (instead of the actual AA value). Given that the key assumptions of IV analysis are fulfilled, the unobserved and observed confounders should be equally distributed among the predicted values of AA so that resulting ORs for AA will be controlled for unmeasured and measured confounding. To obtain the 95% confidence intervals (CIs), we used nonparametric bootstrap with 1000 replications.

In the second approach, the co-relative design, we examined if the regression results (ie, the crude association between AA and smoking) reflect confounding by familial risk factors. From the Swedish Multi-Generation and Twin Registers, we identified all MZ twin, full-sibling, and cousin pairs. Using a stratified logistic regression model, with a separate stratum for each relative pair, we refitted the analysis (ie, the risk of smoking as a function of AA). The OR is then adjusted for a range of unmeasured genetic and environmental factors shared within the relative pair. MZ twins share 100% of their genes and a large part of environmental factors suggesting that the OR for MZ twins is controlled for all possible confounding by genes and shared environment. Full-siblings and cousins share, respectively, on average 50% and 12.5% of their genes identical by descent. Finally, we combined all four samples (ie, population, twin, full-siblings, and cousins) into one dataset in which we performed two analyses. The first allowed all parameters for each sample to be independent (ie, similar to four separate analyses). In the second, we modeled the association between AA and smoking with two parameters: one main effect and one as a linear function of the genetic resemblance; that is, 0 for the population, 0.125 for the cousin, 0.5 for the sibling, and 1 for the MZ twins. The OR for the second parameter gives an indication of the size of the familial confounding. If the second model fitted the data well, as indexed by the Akaike’s Information Criteria (AIC), we also obtained an improved estimation of the association among all relatives, but especially MZ twins, where the data was sparse.

We replicated all our analyses among individuals categorized as smokers with heavy smoking used as the outcome variables. For females, we also used information on the continuation of smoking during pregnancy in females who smoked pre-pregnancy. All statistical analyses were performed using SAS 9.4.¹⁴

Results

Males

IV Analyses

Current smoking information at conscription was available on 233 248 males at a mean (SD) of age of 18.2 (0.4) (Table 1). At the assessment, 12.9% reported smoking. Mean AA was 0.75 SDs lower in smokers than nonsmokers. Among smokers, the mean AA of heavy smokers was 0.36 SDs lower than light smokers.

Table 1.

Descriptive Information on the Samples Studies

	N	Year of birth	Academic achievement (SD)	Pairs of relatives		N	Year of birth	Academic achievement
Males					Males
No current smoking	203 270 (87.2%)	1987 (2.2)	+0.26 (0.8)	Monozygotic twins	No current smoking	457	1985 (0.9)	+0.32 (0.9)
Smoking	29 978 (12.9%)	1987 (2.2)	−0.44 (1.0)		Smoking	59	1985 (0.7)	−0.22 (0.8)
1–10 CPD	22 476 (75.0%)	1987 (2.2)	−0.39 (1.0)	Full siblings	No current smoking	45 089	1987 (2.1)	+0.32 (0.8)
>10 CPD	7502 (25.0%)	1986 (2.0)	−0.75 (1.0)		Smoking	5407	1987 (2.1)	−0.37 (0.9)
Females				Cousins	No current smoking	53 753	1987 (2.1)	+0.27 (0.8)
No smoking pre-pregnancy	391 040 (79.0%)	1980 (4.9)	+0.17 (0.9)		Smoking	7047	1987 (2.1)	−0.46 (1.0)
Smoking	103 955 (21.0%)	1981 (5.0)	−0.68 (1.0)
					Females
1–10 CPD	51 383 (49.4%)	1981 (5.0)	−0.51 (1.0)	Monozygotic twins	No smoking pre-pregnancy	1387	1979 (4.1)	+0.31 (0.9)
>10 CPD	52 572 (50.7%)	1981 (5.0)	−0.90 (1.0)		Smoking	291	1980 (3.8)	−0.72 (1.0)
Smoking pre-pregnancy only	68 547 (66.3%)	1980 (5.0)	−0.52 (0.9)	Full siblings	No smoking pre-pregnancy	121 346	1980 (4.5)	+0.19 (0.9)
Smoking pre-pregnancy and first trimester	34 815 (33.7%)	1981 (5.0)	−1.08 (1.0)		Smoking	29 978	1981 (4.6)	−0.66 (1.0)
				Cousins	No smoking pre-pregnancy	198 048	1980 (4.8)	+0.18 (0.9)
					Smoking	50 188	1981 (4.7)	−0.70 (1.0)

CPD = cigarettes per day; SD = standard deviation.

We first examine the current smoking status. The ability of an IV to provide information about causal influences requires that it should meaningfully predict the independent variable (low AA) and show minimal association with the dependent variable (smoking) except that mediated through the independent variable. Aside from an anomaly in January births, we see a clear monotonic association between our adjusted estimate of month-of-birth and AA (Supplementary Appendix Figure 1). Modeled as a linear effect, the regression coefficient (±95% CIs) was estimated at −0.01843 (−0.01947; −0.01739). Being 1 month younger reduced AA by 1.8% of 1 Std.

The raw association between month-of-birth and smoking was modest but statistically significant with later months (younger age) predicting smoking: OR (±95% CIs) = 1.014 (1.010; 1.018). However, when AA was added to the model, the association between month-of-birth and smoking disappeared: hazard ratio (HR) = 1.000 (0.997; 1.004).

Higher AA strongly predicted reduced smoking rates: OR (per SD) = 0.41 (0.40–0.41) (Table 2). Our IV analysis produced a slightly weaker association OR = 0.47 (0.39–0.57). Among current smokers, the observed ORs for high AA and heavy versus light smoking was 0.68 (0.66–0.70). Results with our IV analyses were comparable but imprecisely known: 0.44 (0.22–0.88).

Table 2.

Association (Odds Ratios [OR] With 95% Confidence Intervals) Between Low Academic Achievement and Current Smoking, Heavy vs. Light Smoking Among Smokers, and Smoking During Pregnancy Among Those Smoking Prior to Pregnancy

Model	Sex/predicted contrast	Crude model OR	Instrumental variable OR	Co-relative model (monozygotic twins extrapolated) OR
	Males
A	Current smoking vs. not smoking	0.41 (0.40; 0.41)	0.47 (0.39; 0.57)	0.51 (0.43; 0.60)
B	Among smokers, heavy vs. light smoking	0.68 (0.66; 0.70)	0.44 (0.22; 0.88)	0.82 (0.52; 1.30)
	Females
C	Pre-pregnancy current smoking vs. not smoking	0.39 (0.39; 0.39)	0.50 (0.46; 0.54)	0.54 (0.51; 0.58)
D	Among smokers, heavy vs. light smoking	0.66 (0.65; 0.67)	0.82 (0.69; 0.98)	0.99 (0.86; 1.15)
E	Among pre-pregnancy smokers, smoking vs. not smoking during pregnancy	0.54 (0.53; 0.55)	0.68 (0.56; 0.82)	0.78 (0.66; 0.91)

Co-Relative Analyses

Samples sizes of subjects are seen in Table 1. Noteworthy is the small sample of MZ twins. Table 3, model A, provides the observed HRs between low AA and smoking in the entire male sample and then in first-cousins, full-siblings, and monozygotic twin pairs discordant for AA level. Aside from the low HR in cousins, the HRs are progressively higher (closer to unity) for more closely related relative pairs. However, the observed association within monozygotic twin pairs is estimated very imprecisely. On the right side of this table, we apply our genetic model to these estimates. The predicted HRs are close to those observed except that now the key estimates for the MZ twins are known much more precisely. As measured by AIC, the genetic model fits slightly worse than the observed estimates. The key estimates for the AA-Smoking association in discordant MZ pairs is 0.51 (0.43–0.60).

Table 3.

Results of the Co-Relative Analyses (Odds Ratios [ORs] ± 95% Confidence Intervals [CIs])

Model		Observed ORs	Predicted ORs
A	Current smoking (M)
	Population	0.41 (0.40; 0.42)	0.40 (0.39; 0.41)
	Cousin	0.38 (0.36; 0.41)	0.41 (0.40; 0.42)
	Sibling	0.47 (0.43; 0.51)	0.45 (0.42; 0.49)
	MZ twins	0.54 (0.15; 1.92)	0.51 (0.43; 0.60)
	AIC	46752.673	46755.895
B	>10 CPD vs. 1–10 CPD (M)
	Population	0.68 (0.66; 0.70)	0.68 (0.66; 0.70)
	Cousin	0.68 (0.54; 0.84)	0.70 (0.66; 0.74)
	Sibling	0.75 (0.60; 0.95)	0.75 (0.60; 0.94)
	MZ twins	-	0.82 (0.52; 1.30)
	AIC	33593.887	33591.952
C	Current Smoking Pre-Pregnancy (F)
	Population	0.39 (0.39; 0.39)	0.38 (0.37; 0.39)
	Cousin	0.40 (0.39; 0.41)	0.40 (0.39; 0.41)
	Sibling	0.46 (0.44; 0.47)	0.46 (0.44; 0.47)
	MZ twins	0.45 (0.17; 1.20)	0.54 (0.51; 0.58)
	AIC	108451.74	108448.58
D	>10 CPD vs. 1–10 CPD pre-pregnancy (F)
	Population	0.66 (0.65; 0.67)	0.66 (0.65; 0.67)
	Cousin	0.68 (0.64; 0.71)	0.69 (0.68; 0.71)
	Sibling	0.81 (0.76; 0.88)	0.81 (0.75; 0.87)
	MZ twins	1.56 (0.34; 7.14)	0.99 (0.86; 1.15)
	AIC	147679.44	147676.68
E	Smoking pre-pregnancy and first trimester vs. smoking only pre-pregnancy (F)
	Population	0.54 (0.53; 0.55)	0.54 (0.53; 0.55)
	Cousin	0.55 (0.52; 0.59)	0.56 (0.55; 0.58)
	Sibling	0.65 (0.60; 0.71)	0.65 (0.60; 0.70)
	MZ twins	0.20 (0.01; 4.17)	0.78 (0.66; 0.91)
	AIC	130983.60	130981.08

AIC = Akaike’s Information Criteria; CPD = cigarettes per day; MZ = monozygotic.

Model B in Table 2 presents the co-relative analyses for quantity smoked. We have no informative MZ pairs and again, the observed ORs for cousins are higher than expected. However, the predicted estimates fit better by AIC than those observed. The critical estimate within MZ pairs is compatible with the population estimate but so imprecisely known as to be uninformative (0.82 [0.52–1.30]).

Females

IV Analyses

Information about smoking prior to pregnancy was available on 494 995 females at a mean (SD) of age at the assessment of 27.7 (4.3) (Table 1). At the assessment, 21.0% of the females reported smoking 3 months prior to pregnancy. The mean AA was 0.85 SDs lower in the smokers than the nonsmokers. Among those smoking prior to pregnancy, heavy smokers had a mean AA 0.39 SDs lower than light smokers. Among those who smoked pre-pregnancy, those who persisted in smoking while pregnant had a mean AA 0.56 SDs lower than those who quit (Table 1).

We first apply our causal inference methods to the pre-pregnancy smoking status. Aside from an anomaly in January births, we see a clear monotonic association between our adjusted estimate of month-of-birth and AA (Supplementary Appendix Figure 2). Modeled as a linear effect, the regression coefficient (±95% CIs) was estimated at −0.02446 (−0.02526; −0.02366). Being 1-month older in the school class at age 16 was associated with a reduction in AA by 2.4% of a standard deviation.

The raw association between month-of-birth and smoking was modest but statistically significant: OR (±95% CIs) = 1.015 (1.013; 1.017). However, when AA was added to the model, the association between month-of-birth and smoking risk disappeared: OR= 0.997 (0.994; 0.999).

In our entire female sample, high AA had a strong negative association with smoking: OR (per SD) =0.39 (0.39-0.39) (Table 2, model C). Our IV analysis produced a modestly weaker association OR = 0.50 (0.46–0.54). Amongst those smoking pre-pregnancy, high AA predicted a lower risk for heavy (vs. light) smoking: OR = 0.66 (0.65–0.67). Our IV analysis produced a weaker but significant association (0.82 [0.69–0.98]; Table 2, model D). Among the women smoking pre-pregnancy, AA predicted cessation of smoking during pregnancy. For this comparison, the overall OR was 0.54 (0.53–0.55), while the results for IV analyses was 0.68 (0.56–0.82) (Table 2, model E).

Co-Relative Analyses

Samples sizes of subjects are seen in Table 1, the sample size of sample informative MZ twins again being small. Table 3, model C, provides, on the left side, the observed HRs between low AA and smoking in the female sample and then in first-cousins, full-siblings, and monozygotic twin pairs discordant for their level of AA.

Aside from the anomalously low OR in cousins, the ORs are progressively lower for more closely related relative pairs. However, the observed association within monozygotic twin pairs is estimated very imprecisely. The right side of this table presents results from our genetic model, which fits the data very well. The key estimates for the AA-Smoking association in discordant MZ pairs are 0.54 (0.51–0.58).

Table 3, model D, presents co-relative results for the association between AA and quantity smoked. The genetic model fits well and estimates a null OR in discordant MZ twins: 0.99 (0.86–1.15). The comparison between AA levels among those who smoked pre-pregnancy and did versus did not continue to smoke while pregnant is seen Table 3, model E. Here the very small number of observed discordant MZ twin pairs produced non-sensible estimates with very large CIs. However, the genetic model fitted well and estimated the OR between high AA and persistence of smoking while pregnant at 0.78 (0.66–0.91).

Discussion

Consistent with prior studies,^1,3,4 in both men and women, we found a substantial inverse association between AA as our predictor variable and subsequent smoking as our dependent variable, with an approximately 60% reduction for every SD increase in AA. Four estimates of the potential causal effects of high AA on smoking were available from IV and co-relative analyses in males and females. These results were reassuringly similar in predicting an approximately 50% reduction in smoking per SD increase in AA. These results provide consistent and substantial support for the hypothesis that the observed AA-smoking association largely reflects causal effects of AA on smoking, with moderate contributions made by background confounders. Of note, for both males and females, our IV—month-of-birth—met IV requirements as, in the presence of the predictor variable, it had no significant association with the outcome variable.

In both male and female smokers, a 1 SD increase in AA was associated with an approximately 33% decrease in the odds of heavy smoking. However, results from our causal analysis methods were variable and confidence intervals quite broad. Of the four estimates of causal effects, only two were statistically significant. Overall, our analyses provide at most modest support for the hypothesis that, among smokers, AA has a casual effect on quantity smoked.

In female subjects who reported smoking prior to pregnancy, each SD of higher AA was associated with a reduction in the odds of continuing smoking during pregnancy of 46%. Both our causal inference methods produced significant results but the magnitude of the associations, while not precisely known, were clearly smaller than the observed association. Our results suggest that the observed association between high AA and smoking cessation during pregnancy is likely, in part, causal but almost surely also results from the impact of confounding factors.

This report contributes to a large literature documenting widespread robust associations between education and health outcomes^15–17 and addresses one of the critical questions in this literature—the degree to which the association is causal.^15,18 While our study cohorts with data on AA are too small and young to obtain stable risk estimates for key smoking-related diseases, in the entire Swedish sample with current smoking data (n = 1 543 890), ORs for the current smokers for subsequent diagnoses of carcinoma of the lung, myocardial infarction, and Chronic Obstructive Pulmonary Disease are, respectively, 4.81 (4.35–5.32), 4.26 (3.98; 4.55), and 11.63 (10.85–12.46). Furthermore, among the full sample of women who reported smoking prior to pregnancy and then continued to smoke versus quit (N = 277 590), the ORs for preterm birth and small for gestational age infants were, respectively, 1.33 (1.20–1.37) and 2.19 (2.10–2.29). While we cannot directly test the association between AA and smoking-related diseases and adverse pregnancy outcomes, indirect evidence supports the hypothesis that high AA would result in a meaningful reduction in risks for these and other adverse smoking-related outcomes.

Through what mechanisms could month of birth impact on AA? Dixon et al.¹⁹ propose a model of “relative age effect” in which older children in each school year develop a spiral of accumulated advantage as a result of increased maturity and opportunity which leads to a positive self-concept and increased motivation with the effect working in the opposite direction for the younger children. Key to this hypothesis is the consistent evidence that within each school year, birthdate impacts appreciably on academic performance.

Most of the literature relating education and health behaviors, including smoking examine years of education, not AA.^15,20 We suggest two plausible mechanisms through which AA might predict risk for smoking. First, students who succeed academically typically develop positive attachments to school, which facilitates their commitment to a prosocial lifestyle, including reduced risk for substance use, including tobacco.²¹ Those who perform poorly at school are more likely to reject prosocial attitudes in favor of more deviant behaviors, including smoking.²² Second, AA is likely to be causally related to a range of measures of self-efficacy,^23,24 and those who are high in such traits are less likely to initiate smoking and, if initiated, are less likely to become dependent and more likely to quit.^25,26

Policy changes like increasing taxes on tobacco and banning smoking in public places have been successful in reducing the prevalence of smoking.²⁷ However, because they do not address the key internal causal factors related to smoking, they may have limited impact, especially among youth. The current finding that AA is causally related to smoking suggests that investment in improving educational quality could provide a useful way to improve population health.²⁸ Such interventions include supplementary academic services (eg, tutoring) to students with low AA and school-wide efforts to create more supportive learning environments. A few such programs have been shown to not only increase AA, but also prevent smoking and, in some cases, reduce alcohol and other drug use, thereby addressing multiple public health problems.^6,7 However, evaluations of many AA programs have not assessed impact on smoking in the short- or long-term, and the current results suggest that such analyses should occur. Other studies indicate that school-based smoking prevention programs are effective,^27,29 and that the impact of these services is enhanced when offered in conjunction with environmental strategies.³⁰ Thus, comprehensive approaches that target causal factors and involve regulatory and economic policy changes may have the greatest potential to reduce population rates of smoking.³¹

We recently examined the potentially causal impact of AA on drug abuse in Sweden using the same methods we have employed here to examine smoking.³² The results were very similar. Using both month of birth as an IV and our co-relative analyses suggest that a large proportion of the association between low AA and subsequent risk for drug abuse was causal in nature.³²

Although causal inference from observational data should be considered tentative, confidence can be increased substantially through the use of multiple inferential methods³³ recently termed “triangulation”.³⁴ The larger the difference in methods, the stronger the resulting inferences can be given similar findings. Our IV and co-relative analyses are divergent in implementation and theoretical assumptions suggesting that some confidence can be placed in our findings of the causal impact of high AA on reduced rates of current smoking and increased quit rates when pregnant.

Limitations

These results should be interpreted in the context of five potentially significant methodological limitations. First, our measures of smoking in both of our samples were relatively crude and in the males, confusingly included both measures of cigarettes per day or packages per week, which we did our best to harmonize. Second, the instrument for our IV analyses—month-of-birth—was not without limitations. Three percent of the sample were not tested at age 16. We statistically corrected this anomaly and presented the results when those cases were excluded in the Supplementary Appendix Table 1, obtaining quite similar results. Furthermore, methodological concerns have been raised with of month-of-birth as an instrument because of potential systematic differences between children born at different times of the year. Indeed, in our sample, month-of-birth was weakly predicted by parental education. However, results did not appreciably differ when we included this variable as a covariate (Supplementary Appendix Table 2). While imperfect, our IV analyses with month-of-birth are unlikely to be seriously biased.

Third, our co-relative design does not control for environmental confounders specific to individuals that could impact both on AA and risk for smoking. Fourth, reverse causation is theoretically possible as smoking in school is associated with poor current and future AA.^4,35 However, such effects are unlikely to be causal as nicotine acutely improves human attention and memory³⁶ and IV analyses specifically protect against the impact of reverse causation.³⁷

Finally, while for simplicity, our descriptions of the causal relations between AA and smoking have assumed a simple linear relationship, in reality, the causal links are almost certainly more complex. For example, poor AA is likely to lead to increased socialization with other students with minimal interest in school and high rates of deviant behaviors, including smoking, which, via well-documented peer effects,^38,39 will result in increased smoking rates.

Supplementary Material

A Contributorship Form detailing each author’s specific involvement with this content, as well as any supplementary data, are available online at https://academic.oup.com/ntr.

Funding

This project was supported by grant R01DA030005 from the National Institutes of Health, grants from the Swedish Research Council (2016-01176), as well as Avtal om Läkarutbildning och Forskning (ALF) funding from Region Skåne.

Declaration of Interests

None declared.