Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2021 Sep 1.
Published in final edited form as: Drug Alcohol Depend. 2020 Jul 2;214:108154. doi: 10.1016/j.drugalcdep.2020.108154

Stability in effects of different smoking-related polygenic risk scores over age and smoking phenotypes

Arielle R Deutsch 1,2, Arielle S Selya 1,2
PMCID: PMC7423706  NIHMSID: NIHMS1608927  PMID: 32645681

Abstract

Purpose:

Polygenic risk scores (PRSs) for smoking behavior largely fail to consider the demonstrated developmental change in genetic influence over age and stage of smoking behaviors. Additionally, few studies have examined how stage-specific smoking PRSs (e.g. for initiation vs. smoking heaviness) generalize to other stages of risk. The current study examines the stability of PRS effects over age, and how specifically calibrated PRSs associate with other smoking phenotypes.

Methods:

7228 participants were from the National Longitudinal Study of Adolescent to Adult Health, who had calculated PRSs for two smoking phenotypes, Centers for Disease Control and Prevention (CDC) smoking initiation status, and cigarettes per day (CPD). Four time-varying effects models estimated associations between both PRSs and four smoking phenotypes (CDC status, cigarettes/day on smoking days, any past-30 day smoking, and past-30 day daily smoking) over adolescence and young adulthood.

Findings:

The time-varying effects models demonstrated that both PRSs significantly associated with all four phenotypes age. PRS effects were similar, in both odds ratios and the overlap of 95% confidence intervals. There were increases in PRS associations with quantity of smoking over age, and a decrease in PRS effects over age for the CDC smoking status phenotype over early to late adolescence.

Conclusions:

Smoking PRSs can be robust predictors of smoking behavior over age. However, the lack of differentiation between specific PRSs and multiple smoking phenotypes, as well as the added contribution of both PRSs to explaining genetic variance, indicates a need to reconceptualize phenotypic measurement used to calibrate smoking PRSs.

Keywords: tobacco, genetics, polygenic risk, time varying models

1. Introduction

There is substantial interest in finding genetic profiles for smoking behavior (Kendler et al. 2012; Li et al. 2003). Recent trends include increased use of polygenic risk scores (PRSs), estimates of genetic risk for individual phenotypes, which are further used to explore polygenic liability or environmental moderation in complex behaviors (Bogdan et al. 2018). Quality control and high-powered samples are critical for score creation, reproducibility, and reliability, as smoking PRSs account for small proportions of variance (Vrieze et al. 2012). Researchers suggest enhanced genome-wide association study (GWAS) approaches and new statistical methods to improve PRS accuracy. Another possibility is a need to account for the dynamic change in genetic influence for smoking behavior over age and stage of development.

“Substance-general” and “tobacco-specific” genetic liabilities contribute to smoking risk (Palmer et al. 2009), and hundreds of variants associate with smoking behaviors (Liu et al. 2019). Common variants include the CHRNA5-CHRNA3-CHRNB4 cluster (Thorgeirsson et al. 2010); sequence variants within a cluster of nicotinic acetylcholine receptor genes, and CYP2A6, (Verde et al. 2011) which associates with nicotine metabolism. Interacting effects between CYP2A6 and CHRNA5 demonstrate a strong polygenic risk (Wassenar et al. 2011). Variants related to risk and reward neural pathways (DRD2/ANKK1) also associate with smoking (Perkins et al. 2008).

These genetic influences are not static over age. Shared environmental influences (e.g., familial environments, neighborhood/school environments) are stronger in early adolescence and mid-adolescence (White et al. 2003). Genetic influence increases over later adolescence and young adulthood (Kendler et al. 2008). Specific genetic variants relate to specific stages of use, such as initiation, age of initiation, and smoking quantity (Caporoso et al. 2009; Erzurumluoglu et al. 2019; TAG, 2010). However, multiple stages of smoking can be explained in part by a shared genetic liability (Kendler et al. 1999; Madden et al.,1999), and genetic liabilities for each stage can vary with age (Bares et al. 2015; Do et al. 2015). The association between specific genes and smoking also changes over time; some genes are more influential at earlier ages, and some are consistent over age (Chenoweth et al. 2016; Ducci et al. 2011). Patterns and trajectories of smoking behaviors, such as speed from initiation to heavier smoking, or age of onset, also influence how PRS or individual genetic variants relate to later phenotypes such as nicotine dependence (Belsky et al. 2013; Gruzca et al. 2010).

GWAS procedures for PRS creation control for age (David et al. 2012; Liu et al. 2019), rather than account for potential developmental sensitivity. Few studies examine developmental changes in PRS effects (e.g., Musci et al. 2015). Vrieze et al. (2012) demonstrated that a cigarettes-per-day (CPD) PRS related to adult, but not adolescent CPD. However, it is unknown if less “intense” smoking phenotypes (ever smoked, ever smoked in the past 30 days) have a similar developmental pattern. How (specifically calibrated) PRSs can relate to multiple phenotypes is also underexplored; for example, both CPD and initiation PRS relate to later smoking dependence (Chang et al. 2019a). As specific genes appear to be important for smoking at different ages and stages, examining the impact of different smoking PRSs on multiple smoking phenotypes over age can elucidate developmental nuances of PRS effects.

1.1. Current Study

Change in smoking-related genetic risks over age and stage may be necessary for developmentally-informed PRS creation. The current study explored developmental change in the association between two smoking PRSs and four smoking phenotypes representing a continuum of smoking stages by intensity. Phenotypes included two related to PRS calibration (Centers for Disease Control and Prevention/CDC-defined smoking status and CPD), which served as the “endpoints” of the continuum, and two “intermediary” phenotypes (current smoking status and current regular (daily) smoking status). We expected both PRSs to associate with all smoking phenotypes, but to also differ in strength such that they would have stronger associations with their respective calibrated phenotype. We also expected PRS effects on phenotypes to be stronger at later ages, and that associations between each PRS and phenotypes would start to overlap at later ages.

2. Methods

2.1. Participants

Participants were from Waves 1–4 of the National Longitudinal Study of Adolescent to Adult Health (Add Health; Harris et al. 2009). A nationally-representative United States sample was first assessed at grades 7–12 in 1994–95 (Wave 1), and then followed up in 1996 (Wave 2), 2000–2001 (Wave 3) and 2004–2005 (Wave 4). Individuals were genotyped at Wave 4 (N=9129), and restricted to four genetic ancestry groups (European, African, Hispanic and East Asian) to account for population stratification. Only individuals who participated in at least Waves 1, 3 and 4 (N=7228) were included, ensuring all participants had adequate adolescent and young adult coverage.

2.2. Measures

2.2.1. Centers for Disease Control and Prevention (CDC) Smoker Status.

CDC smoker status (smoking at least 100 cigarettes; CDC, 2008) was not directly measured in Add Health and therefore was approximated as a quantity (number of days smoked) × frequency (cigarettes per smoking day) measure over the past 30 days. We added this score cumulatively over the four waves (e.g., Wave 1 CDC status used only the Wave 1 quantity × frequency score, wave 2 CDC status added Wave 1 and 2 quantity × frequency score) to approximate a “lifetime” cigarette quantity. At each wave, those who reported smoking at least 100 cumulative cigarettes were coded as “1,” while those who reported smoking less than 100 cumulative cigarettes, or those who reported never smoking cigarettes, were coded as “0”.

2.2.2. Smoking Quantity.

At all waves, smoking quantity was a count variable assessing the number of cigarettes smoked on each smoking day in the past 30 days. For Waves 1, 2 and 4, individuals who reported ever smoking a full cigarette, but who did not smoke in the past 30 days were coded with a “0”. For Wave 3, smoking quantity was only assessed for individuals who reported smoking every day in the past 30 days, and smokers who did not report daily smoking in the past 30 days were coded with a “0”. This variable was converted to a six-point continuous variable that increased in half-pack/10 cigarette increments, ranging from 0 (no smoking in the past 30 days) to 5 (smoking more than 2 packs/40 cigarettes per smoking day), as our analytic method does not allow count distributions.

2.2.3. Current Smoker.

At each wave, individuals who reported smoking at least once in the past 30 days were coded as current smokers (1), while those who did not were coded as non current smokers (0).

2.2.4. Current Regular Smoker

At each wave, individuals who reported smoking daily for the past 30 days were coded as current regular smokers (1), while those who did not were coded as non current regular smokers (0).

2.3. Polygenic Risk Scores

2.3.1. PRS Calculation.

Genetic collection, genotyping, and PRS calculation was conducted by the Add Health research team. Available data for interested researchers includes the participant PRS and the ancestry principal components scores. Genotyping and quality control information is extensively detailed elsewhere (Highland et al. 2018). PRS calculation was done using a modified version of the PRSice wrapper for R within PLINK software (Chang et al. 2015). Each smoking PRSs was calculated using the procedure outlined in Dudbrige (2013), as the sum across single-nucleotide polymorphisms of the following product: the single-nucleotide polymorphisms’s allele frequency in the Add Health sample multiplied by the single-nucleotide polymorphism’s regression coefficient with the phenotype in an independent discovery sample. Raw PRS were then standardized (mean=0, standard deviation=1) within each ancestry group to control for population stratification and limit between-group bias. All details regarding Add Health PRS calculation, as well as raw PGS summary statistics and distributions by ancestry are found in the Add Health PRS user’s guide (Braudt and Harris, 2018).

The discovery sample was based on a 2010 meta-analysis GWAS genome wide analysis study conducted by the Tobacco and Genetics Consortium (TAG), and validation of the 15 most significant regions utilized from two other GWAS consortiums (totaling N=143,023 across the three consortia; studies that belonged to multiple consortia were only included once). The Add Health dataset was not part of these consortia. Phenotypes included 1) current/ever smoker, as defined by CDC smoking status definition (≥100 cigarettes/lifetime) and 2) CPD, using the average number of cigarettes smoked per day, or the maximum number smoked in a single day. Each GWAS accounted for the number of copies of the minor allele and, when the case-control status of participants when applicable. All meta-analysis files are publically available at https://www.med.unc.edu/pgc/download-results/. (Tobacco and Genetics Consortium, 2010).

2.3.2. Ancestry-Specific Principal Components.

Ancestry-specific principal components identified respondents in the four restricted ancestry groups. Including these scores is a common method in PRS research to minimize population stratification bias within a diverse sample (Martin et al. 2019), however there is no recommended number of scores to include (at least 5 is recommended by Braudt and Harris, 2018). We included 10 out of the 20 calculated principal components scores.

3. Statistical Analysis

We estimated four time-varying effect models (TVEM), using the weighted TVEM SAS macro (User’s Guide; Dziak et al.2017) to account for complex survey data. Weighted TVEM allows for binary and linear regression distributions. TVEM is a semiparametric regression-based method that estimates coefficients over a dynamic, function of a continuous variable such as age (Shikyo et al. 2012). Intercepts and slopes are not constrained to a specified parametric shape (e.g. linear) but are estimated empirically. The Weighted TVEM macro only uses unpenalized b-splines to estimate curves. Robust standard errors are constructed using the sandwich formula. Clustering was accounted for at the highest level (e.g., within regions compared to within persons; see TVEM User’s Guide for rationale).

Age at interview was used as the time variable (range = 12–32 years, to ensure that the number of individuals at each age was sufficient, e.g., above 40). All dependent variables (the 4 phenotypes) were re-coded from a wave-based to an age-based format (e.g., score at age 12, age 13). Scores were based on the age of the interview; all other age scores were coded as missing for each participant. For the CDC model, both lifetime reported smokers and nonsmokers were included. For the smoking quantity, current smoking, and regular smoking models, only those who ever reported smoking a full cigarette were included.

Both PRSs were included in each model as time-varying predictors. Ancestry-specific principal components and gender were included as time-invariant covariates. Outputs for TVEM include a graphical plot of estimates over time, as well as output of the estimate and 95% confidence intervals (CI) CIs for 100 points estimated between the minimum and maximum points of time. Statistical significance is determined by the 95% CIs. There is no formal way to test whether effects (or groups) significantly differ from each other beyond examining overlap in CIs. Non-overlapping CIs indicate that effects significantly differ, and overlapping CIs suggest (but not definitively; Cumming and Finch, 2001) no significant difference between effects. We also ran four additional European ancestry-only models (total N=4601; smoker-only N=3656) to account for any potential within-group bias that could not be accounted for by the ancestry principal components (Supplemental Figures 1–4). We did not run models for the other three ancestry groups due to considerable sample size differences (African total N=1489; Hispanic total N=766; East Asian total N=372), which would make for largely underpowered TVEM models (TAG, 2010).

4. Results

4.1. Descriptive Results

Table 1 displays the means, standard deviations and ranges, or percentages for all wave-level smoking phenotypes and demographics. Correlations between wave-level phenotypes and both PRSs are available in Supplemental Table 1. Notably, the correlation between the PRSs was small (r = 0.03, p>.01).

Table 1.

Full and smoking only sample size by wave, and summary statistics for wave-based and non wave-based variables

Mean (SD) [Range] / Frequency N
Total Gender
 Male 45.89% 3317
 Female 54.11% 3911
Total Ancestry
 European 63.66% 4601
 African 20.60% 1489
 Hispanic 10.60% 766
 East Asian 5.51% 372
Total Ever Smoked
Cigarette
 Yes 74.52% 5386
 No 25.48% 1842
Age
 Wave 1 15.55(1.72) 7228
[12–21]
 Wave 2 16.17(1.63) 5553
[13–21]
 Wave 3 22.42(1.73) 7228
[18–28]
 Wave 4 28.92(1.72) 7228
[25–32]
CDC Status (full sample)
 Wavel Y: 11.31% N: 88.69% 7228
 Wave 2 Y: 16.48% N: 83.52% 5553
 Wave 3 Y: 30.88% N: 69.12% 7228
 Wave 4 Y: 36.68% N: 63.32% 7228
Smoking Quantity (smoking only sample)
 Wave 1 0.44 (0.69) 5334
 Wave 2 0.56 (0.74) 4099
 Wave 3 0.72 (0.90) 5317
 Wave 4 0.73 (0.89) 5345
Current Smoking Status (smoking only sample)
 Wavel Y: 36.38% N: 63.62% 5334
 Wave 2 Y: 45.97% N: 54.03% 4099
 Wave 3 Y: 47.31% N: 52.69% 5317
 Wave 4 Y: 49.93% N: 50.07% 5345
Current Regular Smoking Status (smoking only sample)
 Wave 1 Y: 27.75% N: 72.24% 5334
 Wave 2 Y: 29.31% N: 70.69% 4099
 Wave 3 Y: 31.69% N: 68.31% 5317
 Wave 4 Y: 31.27% N: 68.73% 5345
Open in a new tab

4.2. Time-Varying Effect Models

4.2.1. CDC Smoking Status

Figure 1 displays the TVEM results for the association between CDC smoking status and both PRS over age. Dotted portions of the effect (black) lines indicate points of time in which the associations were significant as indicated by the 95% CIs. Both CDC PRS and CPD PRS significantly associated with CDC smoking status over adolescence and young adulthood. Although the effects of both PRSs slightly decreased over age, effects were relatively stable. A higher CDC PRS associated with higher odds of CDC smoker status from age 12 (OR=1.79, 95% CI: 1.12–2.87), to age 30 (OR=1.15, 95% CI: 1.01–1.26). Higher CPD PRS also associated with higher odds of CDC smoker status starting at age 13 (OR=1.41, 95% CI: 1.00–1.99), and ending at age 30 (OR=1.14, 95% CI: 1.05–1.22). PRS effects were similar to each other, as suggested by the 95% CIs and comparable ORs. There was almost complete overlap in the two PRS effects starting at age 25 (CDC PRS OR=1.19, 95% CI: 1.05–1.29; CPD PRS OR=1.19, 95% CI: 1.10–1.29).

Figure 1.

Figure 1.

Open in a new tab

Weighted TVEM model estimating PRS odds ratios for attaining CDC smoker status over age. Gray lines represent 95% confidence intervals. Dotted regions represent statistical significance (i.e. 95% confidence intervals exclude an odds ratio of 1).

4.2.2. Smoking Quantity

As seen in Figure 2, both PRSs significantly associated with the quantity of smoking over adolescence and young adulthood. The CDC PRS significantly associated with quantity of smoking starting at age 13 (b=0.03, 95% CI: 0.01–0.06) and continuing until age 30 (b=0.05, 95% CI = 0.02–0.10). CPD PRS significantly associated with quantity of smoking starting at age 17 (b=0.02, 95% CI: 0.01–0.05), increasing in effect and peaking at age 22 (b=0.08, 95% CI: 0.04–0.11), and ending at age 30 (b=0.04, 95% CI: 0.0 –0.07). Although the pattern of association over age differed for the two PRS, similar beta coefficients and overlapping CIs suggested that the effects were similar in strength.

Figure 2.

Figure 2.

Open in a new tab

Weighted TVEM model estimating PRS effects for the quantity of smoking per day over age. Gray lines represent 95% confidence intervals. Dotted regions represent statistical significance (i.e. 95% confidence intervals exclude a beta coefficient of 0).

4.2.3. Current Smoking / Current Regular Smoking Status

Figures 3a and 3b show associations between PRSs and current smoking and current regular smoking, respectively. CDC PRS significantly associated with current smoking status over a longer period of age (ages 13–31) compared to CPD PRS (ages 18–29). The pattern of associations for CDC PRS slightly decreased, and then increased, such that the effect was strongest at later ages (age 31, OR=1.16, 95% CI: 1.03–1.31). The pattern of associations for CPD PRS was the opposite; with an initial increase, and then decrease in effect, such that the strongest effects were at age 23 (OR=1.20, 95% CI: 1.10–1.30). PRS odds ratios were similar, and there was overlap between CIs, suggesting that the effects did not significantly differ in effect from each other.

Figure 3a–b.

Figure 3a–b.

Open in a new tab

Weighted TVEM model estimating PRS effects for current smoking (A) and current regular smoking (B). Gray lines represent 95% Confidence intervals. Dotted regions represent statistical significance (i.e. 95% confidence intervals exclude an odds ratio of 1).

For current regular smoking, there was minor divergence in the PRS associations. CDC PRS was significant over ages 13–16, during which there was a decrease in effect (13 years: OR=1.85; 17 years: OR=1.12). After age 17, the CDC PRS effect remained stable and significant between the ages of 21 (OR=1.11, 95% CI: 1.01–1.21) and 28 (OR=1.08, 95% CI: 1.01–1.16). The CPD PRS was significant starting at age 14 (OR=1.16, 95% CI: 1.01–1.36), remained relatively stable until age 31 (OR=1.18, 95% CI: 1.01–1.38). Similar odds ratios and overlap between CIs suggested the PRS effects were similar in strength.

4.3. Ancestry-specific models

Ancestry-specific models are presented in Supplemental Figures 1–4. As demonstrated, the European ancestry-only models share the same patterns of results for all smoking phenotypes as the full models. There were slight differences, mainly related to the extent to which the 95% CIs indicated that effect sizes were significant or not at a given age. However, the effect sizes themselves for all models were close to the effect sizes for the full-sample models.

4.4. Post Hoc Analyses

Given similarity in PRS effects, we further probed differences by examining how each PRS contributed variance to latent smoking phenotypes. Four structural equation models were estimated using Mplus Ver 8. We created four smoking phenotype latent factors with a psychometric, factor modeling approach (Deutsch et al. 2017). This approach assumed that individual smoking phenotypes at different time points were related to each other, but did not assume a shape to the intercept or slope (similar to TVEM). Both PRSs were regressed on to each latent factor. To examine contributions of each PRS to smoking factors, we compared the model fit by constraining individual PRS effects to zero, using the -2LL chi-square difference test. We also empirically tested if PRS significantly differed by constraining them to be equal to each other. As with the TVEM analyses, we also included the first 10 ancestry principal components groups, and excluded non-smokers from the CPD, current smoker, and current regular smoker models.

Overall, both PRS effects were similar for each latent phenotype, and accounted for a small amount of variance. As seen in Supplemental Table 2, constraining either PRS to zero significantly worsened model fits for all phenotypes. In addition, the PRS were not interchangeable - when one PRS was constrained after the other, there was a significant drop in model fit for all phenotypes (regardless of which PRS was constrained first). PRS effects did not significantly differ from each other in all models except for the smoking quantity model, which indicated that the CPD PRS had a significantly stronger effect on smoking quantity compared to the CDC PRS.

5. Discussion

This study examined the developmental change of smoking PRS effects. Overall, both PRSs were stable over age and had similar patterns of effects on each phenotype. Although post hoc analyses indicated that each PRS had unique contributions to all smoking phenotypes (e.g., they were not necessarily interchangeable with each other), overlap of the effects in the TVEM models indicated that PRS as representatives for the genetic risk of specific stages or constructs of smoking should be carefully considered. Below, we discuss the results with the understanding that they may only apply to individuals of specific ancestry groups (European).

We hypothesized that PRS effects would be stronger in adulthood compared to adolescence (Vrieze et al. 2012; Kendler et al. 2008). The quantity model supported this hypothesis, demonstrating modest increases for both PRS effects over age. There was also a slight increase in the effect of the CPD PRS on regular smoking status over age. This is consistent with results indicating that polygenic risk scores including genes related to the CHRNA5-CHRNA3-CHRNA4 cluster have stronger effects on CPD and related phenotypes in early adulthood, but not adolescence (Verize et al. 2012).

We found the opposite effect for the CDC-smoking status model, A higher association at earlier ages may relate to the overall negative relationship between genetic liability to smoke/initiate smoking and age at which one starts smoking (Chang et al. 2019b). It is also possible that higher PRS effects at earlier ages in the CDC model reflects age-related normative smoking trends within the United States at the time of the survey (approximately 12.7 for Add Health participants, Kaestle 2015). Individuals who start smoking at atypically higher ages may have relatively low genetic influence (and may be motivated more by environment). Finally, there was minor change for PRS effects on current smoking or current regular smoking models over age. That they are significantly associated with the phenotypes across almost all of adolescence and young adulthood indicates that both PRSs are robust indicators of “broad” smoking behavior.

The most surprising finding was the similarity between CDC smoking and CPD PRS effects, and almost phenotype, contrary to our expectations. Post-hoc tests indicated that both PRSs uniquely contributed a small amount of variance to the phenotypes. There was one exception, the stronger association between CPD PRS and the smoking quantity latent score (i.e., the PRS and its most closely-related phenotype). Overall, there was a high overlap in the pattern of association between each PRS and the varying phenotypes, and overlap in CIs suggests these effects were similar to each other. However, significant differences between PRS may be clearer in studies with larger samples. It is also possible that phenotypic similarity also explains these results. CPD and overall smoking initiation (smoking 100+ cigarettes over the lifetime) may be too similar to each other (as one is nested within the other). Although the two PRSs share few genetic variants between them (see TAG 2010, supplemental tables 4 and 5), and the PRS were only minimally correlated, both sets of genes may share similar associations with “active” smoking behaviors, as studies indicate partial shared genetic influence between multiple smoking stages (Kendler et al. 1999; Madden et al.,1999), and have a moderate genetic correlation (r=0.37; Nivard et al. 2016) It is possible that these PRSs both contribute to genetic liability for a broader smoking risk phenotype. High-powered studies and enhanced genomic methods may build more comprehensive genetic profiles of risk that can distinguish between different, but related, smoking stages. As this is one of the first studies comparing effects for PRS calibrated using different phenotypes for an individual substance, replication is required.

5.1. Limitations and Future Directions

The PRS were based on results from the 2010 TAG study. Replications of this study are needed, using GWAS results from newer studies larger samples that have uncovered more genetic variants related to smoking phenotypes (Erzurumluoglu et al. 2019; Liu et al. 2019). The current PRS was based on a GWAS with a European ancestry-only sample, and results may not apply to majority or exclusively non-European samples. PRS developed in the TAG 2010 study have been demonstrated to not generalize to some non-European samples (e.g., Otto et al., 2016), and therefore, generalizing current results to non-European ancestry groups should be done with caution. As a relatively new statistical method, TVEM has no formal power tests. A sample of at least 100 participants is sufficient given 10–25 observations per participant (i.e., 1000–2500 data points; Tan et al. 2012), and power increases as the number of time points increases. In the current study, all participants had either three or four observations over 20 time points. Conservatively, sample size requirements can be estimated based on the sample size at each time point; we estimate a sample size of N≈1000 within each age point to be necessary to detect characteristically small effect sizes of PRSs. Thus, overall our study is sufficiently powered, particularly for certain age ranges (14–18, 22–24, 28–31); however there is potentially limited statistical power in the intermediate and surrounding ages (e.g., approximately 40). Power decreases in TVEM binary models as the number of zeroes compared to ones increase (Yang et al. 2017). Thus, rarer endorsement of phenotypes at specific ages may have reduced model power at these points. Although the CPD PRS was calibrated using a count variable, we utilized a continuous variable, as Weighted TVEM does not estimate count models. Although results may have changed with a count variable, we provided a substantial range (from 10 cigarettes/half a pack or less to more than 40 cigarettes/2 packs) per day), capturing 97% of our sample.

Many genetic variants relate to smoking, and many associate with multiple behaviors (Evans et al. 2020). Some smoking phenotypes are nested within each other (e.g., you must initiate smoking to be a regular smoker). There is evidence that smoking initiation and cessation have unique genetic influences (Chang et al. 2019a). Less is known about qualitative stages or trajectories of “active” smoking (Belsky et al. 2013). Researchers may benefit from reconceptualizing substance use measurement for calibrating scores to improve PRS predictability. Possibilities include using developmentally-informative phenotypes such as trajectories (e.g., Brook et al, 2008; Mathew et al. 2018) which can encapsulate both age and stage-specific change, or using multiple common phenotypes for PRS calibration into latent classes or profiles that can provide qualitative distinctions between smoking groups (e.g., non-smokers, non-regular smokers, light regular smokers).

5.2. Conclusion

Although PRSs explain a small portion of variance for many phenotypes, they can be useful tools to explore how phenotypes relate to each other and the environment (Bogdan et al. 2018). Researchers continually improve methods to examine genetic architectures underlying complex behaviors, championing large sample sizes (So and Scham, 2017) and sophisticated genetic analyses (Vilhjalmsson et al. 2015). Advancing genetics research might also require continued discussions about the methodology of PRS creation, given that associated phenotypes can be dynamic in nature. Future work on developing PRSs should focus on distinguishing between different behavioral phenotypes and accounting for age-dependency of genetic effects.

Supplementary Material

1
2

Highlights:

  • Different smoking PGS have similar effects on multiple smoking behaviors

  • Association between smoking PGSs and smoking quantity increase over age

  • Association between smoking PGSs and ever smoking ≥ 100 cigarettes decline over age

Acknowledgements:

This study was funded by NIH grants P20GM121341-03 and P50DA039839-5. This research uses data from Add Health, a program project directed by Kathleen Mullan Harris and designed by J. Richard Udry, Peter S. Bearman, and Kathleen Mullan Harris at the University of North Carolina at Chapel Hill, and funded by grant P01-HD31921 from the Eunice Kennedy Shriver National Institute of Child Health and Human Development, with cooperative funding from 23 other federal agencies and foundations. Information on how to obtain the Add Health data files is available on the Add Health website (http://www.cpc.unc.edu/addhealth). No direct support was received from grant P01-HD31921 for this analysis

Role of Funding Source:

Funding for this study was provided by NIGMS grant P20GM121341-03. NIGMS had no role in in the study design, collection, analysis or interpretation of the data, writing the manuscript, or the decision to submit the paper for publication.

Footnotes

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Conflict of Interest:

None

References

  1. Barnes CB, Kendler KS & Maes HH 2015. Developmental changes in genetic and shared environmental contributions to smoking initiation and subsequent smoking quantity in adolescence and young adulthood. Twin Res Hum Genet, 18, 497 – 506. doi: 10.1017/thg.2015.48 [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Belsky DW, Moffitt TE, Baker TB, Biddle AK, Efans JP, Harrington H, …& Caspi A 2013. Polygenic risk and the developmental progression to heavy, persistent smoking and nicotine dependence. JAMA Psychiatry, 70, 534 – 542. doi: 10.1001/jamapsychiatry.2013.736. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Bogdan R, Barranger DAA & Agrawal A 2018. Polygenic risk scores in clinical psychology: Bridging genomic risk to individual differences. Ann Rev Clin Psychol, 14, 119 – 157. doi: 10.1146/annurev-adolescence-050817-0804847 [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Braudt DB, & Harris KM 2018. Polygenic scores PRS in the National Longitudinal Study of Adolescent to Adult Health Add Health - Release 1. Retrieved from https://www.cpc.unc.edu/projects/addhealth/documentation/guides/PRS_AH1_UserGuide.pdf on 2/4/2020
  5. Brook DW, Brook JS, Zhang C Whiteman M, Cohen P& Finch SJ 2008. Developmental trajectories of cigarette smoking from adolescence to the early thirties: Personality and behavioral risk factors. Nicotine Tob Res, 10, 1283 – 1291. doi: 10.1080/14622200802238933. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Caporaso N, Gu F, Chatterjee N, Sheng-Chih J, Yu K, Yeager M, …& Bergen AW 2009. Genome-wide and candidate gene association study of cigarette smoking behaviors. PLoS ONE 4e4653. doi: 10.1371/journal.pone.0004653 [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Centers for Disease Control and Prevention 2008. Cigarette smoking among adults-United States, 2007. MMWR Morbidity and Mortality Weekly Report, 57, 1221–1226. [PubMed] [Google Scholar]
  8. Chang L, Witfield JB, Liu M, Medland SE, Hickie IB, Martin NG …& GSCAN Consortium. 2019a. Associations between polygenic risk for tobacco and alcohol use and liability to tobacco and alcohol use, and psychiatric disorders in an independent sample of 13,999 Australian adults. Drug Alcohol Depend, 205, 107704. [DOI] [PubMed] [Google Scholar]
  9. Chang L, Couvy-Duchesne B, Liu M, Medland SE, Verhulst B, Bentosch EG, …& GSCAN Consortium. 2019b. Association between polygenic risk for tobacco or alcohol consumption to licit and illicit substance use in young Australian Adults. Drug Alcohol Depend, 197, 271 – 279. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Chenoweth MJ, Sylvestre M, Contreas G, Novalen M, O’Loughlin J, & Tyndale RF 2016. Variation in CYP2A6 and tobacco dependence throughout adolescence and in young adult smokers. Drug Alcohol Depend, 158, 139 – 146. doi: 10.1016/j.drugalcdep.2015.11.017 [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Cumming G & Finch S 2001. A primer on the understanding, use, and calculation of confidence intervals that are based on central and noncentral distributions. Educ Psychol Meas, 61, 532–574. Doi: 10.1177/0013164401614002. [DOI] [Google Scholar]
  12. David SP, Hamidovic A, Chen GK, Bergen AW, Wessel J, Kasberger JL,…& Furberg H 2012. Genome-wide meta analyses of smoking behaviors in African Americans. Transl Psychiatry, 2, e119 doi: 10.1038/tp.2012.41 [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Deutsch AR, Wood PK & Slutske WS 2017. Developmental etiologies of alcohol use and their relations to parent and peer influences over adolescence and young adulthood: A genetically informed approach. Alcoholism Clin Exp Res 41, 2151 – 2162. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Do EK, Prom-Wormley EC, Eaves LJ, Silberg JL, Miles DR & Maes HH 2015. Genetic and environmental influences on smoking behavior across adolescence and young adulthood in the Virginia Twin Study of Adolescent Behavioral Development and the transitions to substance abuse follow-up. Twin Res Hum Genetics, 18, 43 – 51. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Ducci F, Kaakinen M, Pouta A, Hartikainen A, Veijola J, Isohanni M,…& Jarvelin M 2011. TTC12-ANKK1-DRD2 and CHRNA5-CHRNA3-CHRNB4 influence different pathways leading to smoking behavior from adolescence to mid-adulthood. Biol Psychiat, 69, 650 – 660. doi: 10.1016/j.biopsych.2010.09.055. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Duncan L Shen H, Meijesen J, REssler K, Feldman M Peterson R & Dominguue B 2019. Analysis of polygenic risks score usage and performance in diverse human populations. Nat Commun, 10, 3328. doi: 10.1038/s4167-019-1112-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Dziak JJ, Li R, & Wagner AT 2017. Weighted TVEM SAS macro users’ guide Version 26. Retrieved from methodology.psu.edu. University Park PA: The Methodology Center Penn State; on 3/4/2020 [Google Scholar]
  18. Evans LM, Johnson EC, Melroy-Grief WE, Hewitt JK, Hoeffer CA, Keller MC,…& Ehringer MA 2020. The role of a priori-identified addiction and smoking gene sets in smoking behaviors. Nicotine Tob Res, ntaa006. doi: 10/1093/ntr/ntaa006 [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Erzurumluoglu AM, Liu M, Jackson VE, Barnes DR, Datta G, Melbourne CA, …& Howson JMM 2019. Meta-analysis of up to 622,409 individuals identifies 40 novel smoking behavior associated with genetic loci. Mol Psychiatr, doi: 10.1038/s41380-018-0313-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Gruzca RA, Johnson EO, Krueger RF, Breslau N, Saccone NL, Chen L … & Bierut LJ 2010. Incorporating age of onset of smoking into genetic models for nicotine dependence: Evidence for interaction with multiple genes. Addict Bio, 15, 346 – 357. doi: 10.1111/j.1369-1600.2010.00220.x [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Harris KM, Halpern CT, Whitesell E, Hussey J, Tabor J, Entzel P, and Udry JR. 2009. The National Longitudinal Study of Adolescent to Adult Health: Research Design [WWW document]. Retrieved at http://www.cpc.unc.edu/projects/addhealth/design on 5/10/2020
  22. Highland HM, Avery CL, Duran Q, Li Y & Harris KM (2018). Quality control analysis of Add Health GWAS data. Retrieved at https://www.cpc.unc.edu/projects/addhealth/documentation/guides/AH_GWAS_QC.pdf. On 5/1/2020
  23. Kaestle CE 2015. Age of smoking milestones: Longitudinal inconsistencies and recanting. J Adolesc Health, 56, 382–388. doi: 10.1016/j.jadohealth.2014.12.005 [DOI] [PubMed] [Google Scholar]
  24. Kendler KS, Chen X, Dick D, Maes H, Gillespie N, Neale MC & Riley B 2012. Recent advances in the genetic epidemiology and molecular genetics of substance use disorders. Nat Neurosci, 15, 181 – 189. doi: 10.1038/nn.3018. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Kendler KS, Schmitt E, Aggen SH, & Prescott CA 2008. Genetic and environmental influences on alcohol, caffeine, cannabis, and nicotine use from early adolescence to middle adulthood. Arch Gen Psychiatry, 65, 674–682. doi: 10.1001/archpsyc.65.6.674. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Kendler KS, Neale MC, Sullivan P, Corey LA, Gardner CO & Prescott CA 1999. A population-based twin study in women of smoking initiation and nicotine dependence. Psychol Med, 29, 299–308. doi: 10.1017/S003329177008022 [DOI] [PubMed] [Google Scholar]
  27. Li MD, Cheng R, Ma JZ & Swan GE 2003. A meta-analysis of estimated and environmental effects on smoking behavior in male and female adult twins. Addiction, 98, 23–3. doi: 10.1046/j.1360-443.2003.00295.x [DOI] [PubMed] [Google Scholar]
  28. Liu M, Jiang Y, Wedow R, Li Y, Brazel DM, Chen F,…& Vrieze S 2019. Association studies of up to 1.2 million individuals yield new insights into the genetic etiology of tobacco and alcohol use. Nat Genetics, 51, 237 – 244. doi: 10.1038/s41588-018-0307-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Madden PAF, Heath AC, Pedersen NL, Kaprio J, Koskenvuo MJ & Martin NG 1999. The genetics of smoking persistence in men and women: A multicultural study. Behav Gen 29, 423 – 431 doi: 10.1023/a:1021674804714 [DOI] [PubMed] [Google Scholar]
  30. Musci RJ, Masyn KE, Uhl G, Maher B, Kellam SG & Ialongo NS 2015. Polygenic score × intervention moderation: AN application of discrete-time survival analysis to modeling the timing of first tobacco use among urban youth. Dev Psychopathol, 27, 111 – 122. doi: 10.1017/S0954579414001333 [DOI] [PubMed] [Google Scholar]
  31. Nivard MG, Verweij KJH, Mincica CC, Treur JL, International Cannavis Consortium, Vink JM & Boomsma DI (2016P. Connecting the dots, genome-wide association studies in substance use. Mol Psychiatry, 21, 733–735. doi: 10.1038/mp.2016/14 [DOI] [PubMed] [Google Scholar]
  32. Otto JM, Gizer IR, Bizon C, Wilhelmsen KC & Ehlers CL (2016). Polygenic risks scores for cigarettes smoked per day do not generalize to a Native American population. Drug Alcohol Depend, 167, 95 – 102. doi: 10.1016/j.drugalcdep.2016.07.029 [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Palmer RHC, Young SE, Hopfer CJ Corley, Stallings, Crowley TJ & Hewitt JK 2009. Developmental epidemiology of drug use and abuse in adolescence and young adulthood: Evidence of generalized risk. Drug Alcohol Depend, 102, 78 – 87. doi: 10.1016/j.drugalcdep.2009.01.012. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Perkins KA, Lerman C, Grottenhalter A, Ciccocippo MM, Milanack ME, Cynthia CA …& Benowitz NL 2008. Dopamine and opioid gene variants are associated with increased smoking reward and reinforcement due to negative mood. Behav Pharmacol, 19, 614 – 649. doi: 10.1097/FBP.0b013e32830c367. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Shikyo MP, Lanza ST, Tan X, Li R, & Shiffman S 2012. Using the time-varying effect model TVEM to examine dynamic associations between negative affect and self-confidence on smoking urges: Differences between successful quitters and relapsers. Prev Sci, 13, 288–299. doi: 10.1007/s11121-011-0264-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. So H & Sham PC 2017. Improving polygenic risk prediction from summary statistics by an empirical Bayes approach. Sci Rep, 7, 41262. doi: 10.1038/srep41262 [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Tan X, Shiyko MP, Li R, Li Y & Dierker L 2012. A time-varying effect model for intensive longitudinal data. Psychol. Methods 17, 61–77. doi: 10.1037/a0025814 [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. The Tobacco and Genetics Consortium 2010. Genome-wide meta-analyses identify multiple loci associated with smoking behavior. Nat Genetics, 42, 441–447. doi: 10.1038/ng.571 [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Thorgeirsson TE, Geller F, Sulem P, Rafnar T, Wiste A, Magnusson KP, …& Steffanson K 2010. A variant associated with nicotine dependence, lung cancer, and peripheral arterial disease. Nature, 452, 638 – 642. doi: 10.1038/nature06846. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Vrieze SI, McGue M & Iacono WG 2012. The interplay of genes and adolescent development in substance use disorders: Leveraging findings from GWAS meta-analyses to test developmental hypotheses about nicotine consumption. Hum Genet, 131, 791 – 801 doi: 10.1007/s00439-012-1167-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Vilhjalmsson BJ, Yang J, Finucane HK, Gusev A, Lindstrom S, Ripke S,…& Won H 2015. Modeling linkage disequilibrium increases accuracy of polygenic risk scores. Am J Hum Genet, 97, 576 – 592. doi: 10.1016/j.ajhg.2015.09.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. Wassenaar CA Dong Q, Wei Q, Amos CI, Splitz MR & Tyndale RF 2011. Relationship between CYPA6 and CHRNA5-CHRNA3-CHRNB4 variation and smoking behaviors and lung cancer risk. J Natl Cancer Inst, 103, 1342–1346. doi: 10.1093/jnci/djr237 [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. White VM, Hopper JL, Wearing AJ & Hill DJ 2003. The role of genes in tobacco smoking during adolescence and young adulthood: A multivariate behavior genetic investigation. Addiction, 98, 1087 – 1100. doi: 10.1046/j.1360.0443.2003.00427 [DOI] [PubMed] [Google Scholar]
  44. Yang S, Cranford JA, Li R, Zucker RA & Buu A 2017. A time-varying effect model for studying gender differences in health behavior. Stat Methods Med Res 26, 2812 – 2820. doi: 10.1177/0962280215610608 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

1
NIHMS1608927-supplement-1.docx (18.2KB, docx)
2
NIHMS1608927-supplement-2.docx (55.7KB, docx)

RESOURCES