A Twin Study on the Association Between Psychotic Experiences and Tobacco Use During Adolescence

Objective Psychotic experiences (PE) are dimensional phenomena in the general population that resemble psychotic symptoms, such as paranoia and hallucinations. This is the first twin study to explore the degree to which tobacco use and PE share genetic or environmental influences. Previous studies on the association between adolescent tobacco use and PE have not considered PE dimensionally, included negative symptoms, or accounted for confounding by sleep disturbance and stressful life events. Method An unselected adolescent twin sample (N = 3,787 pairs; mean age = 16.16 years) reported on PE (paranoia, hallucinations, cognitive disorganization, grandiosity, and anhedonia) and regularity of tobacco use. Parents rated the twins’ negative symptoms. Regression analyses were conducted while adjusted for sociodemographic characteristics, prenatal maternal smoking, cannabis use, sleep disturbance, and stressful life events. Bivariate twin modeling was used to estimate the degree of genetic and common and unique environmental influences shared between tobacco use and PE. Results Regular smokers were significantly more likely to experience paranoia, hallucinations, cognitive disorganization, and negative symptoms (β = 0.17−0.34), but not grandiosity or anhedonia, than nonsmokers, after adjustment for confounders. Paranoia, hallucinations, and cognitive disorganization correlated ≥0.15 with tobacco use (r = 0.15−0.21, all p < .001). Significant genetic correlations (rA=0.37−0.45) were found. Genetic influences accounted for most of the association between tobacco use and paranoia (84%) and cognitive disorganization (81%). Familial influences accounted for 80% of the association between tobacco use and hallucinations. Conclusion Tobacco use and PE during adolescence were associated after adjustment for confounders. They appear to co-occur largely because of shared genetic influences.

tobacco use while accounting for these factors, some of which were not considered previously.
Although tobacco might be considered an environmental exposure, twin studies estimate significant heritability of tobacco use during adolescence (36%À60%). [27][28][29][30][31] It is therefore important to consider both genetic and environmental influences when investigating the association between adolescent PE and tobacco use. Twin studies on PE indicate genetic and unique environmental influences as important; common environmental influences are small or negligible. 32 We aimed to investigate whether associations between tobacco use and PE exist in adolescence and whether they remain after adjusting for sex, age, ethnicity, socioeconomic status, cannabis use, stressful life events, sleep disturbance, and prenatal maternal smoking. Our second aim was to test whether shared genetic and environmental influences underlie adolescent PE and tobacco smoking. We consider a range of PE including paranoia, hallucinations, cognitive disorganization, grandiosity, anhedonia, and parent-rated negative symptoms (such as flattened affect and low motivation), assessed as dimensional traits.

Sample
Research participants came from the Twins Early Development Study (TEDS), a UK-based community sample of twins born between 1994 and 1996 33 who, along with their parents, participated in the Longitudinal Experiences and Perceptions Study (LEAP) 34 at age 16 years. Of the 8,534 families invited to report on PE and substance use, 3,941 (46.2%) participated. Table S1, available online, compares participating and nonparticipating families.
For regression analyses, one twin per family was chosen randomly to select a sample of unrelated individuals. After exclusions (missing zygosity data, no consent information, severe medical conditions, or perinatal complications), the sample included 3,787 individuals.
For twin analyses, zygosity was determined using a parentrated measure and confirmed using DNA testing in ambivalent cases. The sample, after exclusions, consisted of 1,342 monozygotic (MZ; 43.8% male) and 1,219 same-sex dizygotic (DZ) pairs (44.5% male). Opposite-sex DZ pairs (1,208) were not included because of limited power in bivariate twin models to test for qualitative sex differences in categorical data.
The Institute of Psychiatry Ethics Committee, King's College London, granted ethical approval for TEDS. All research participants granted informed consent. The Birkbeck Department of Psychological Sciences' Ethics Committee and the core TEDS team approved this study.

Measures
Psychotic Experiences. The Specific Psychotic Experiences Questionnaire (SPEQ) 34 included three subscales measuring "positive" PE (paranoia, hallucinations, and grandiosity), two measuring "negative" PE (parent-rated negative symptoms and anhedonia), and one cognitive disorganization subscale. Subscales consisted of eight to 15 items and asked about frequency or severity of recent PE. Subscales showed good-to-excellent internal consistency (Cronbach's a ¼ 0.77-0.93) and test-retest reliability over 9 months (r ¼ 0.65-0.74). Negative symptoms were parent-rated; other subscales were selfrated.
Untransformed SPEQ scores were used in descriptive statistics. Subscales with a positive skew (paranoia, hallucinations, grandiosity, and parent-rated negative symptoms) were square root transformed. PE were standardized to have a mean of 0 and a standard deviation (SD) of 1.
Tobacco Use. A tobacco use variable was created with three levels. Nonsmokers were those who answered "no" to the question "Have you ever smoked a cigarette (including rollups)?" Occasional and regular smokers were defined based on the item "How many cigarettes have you smoked, in total, in your lifetime?" Occasional smokers were those who had smoked <50 cigarettes and regular smokers !50. Those who had not smoked in the past year were considered nonsmokers. The threshold of 50 cigarettes (rather than a higher threshold) was chosen because adolescents have had less access to resources and opportunity to smoke than adults. For smoking frequency by group, see Table S2, available online.
Confounding Variables. Variables adjusted for during multiple regression included sex, age, ethnicity, socioeconomic status (SES), prenatal maternal smoking, and selfrated measures (at age 16 years) of cannabis use, sleep disturbance measured by the Pittsburgh Sleep Quality Index (PSQI), 35 and stressful life events (SLE) measured by an abbreviated Coddington Life Events Record. 36 Cannabis use was determined from "yes" responses to "Have you ever tried cannabis?" Those who indicated no use during the past year were considered nonÀcannabisusers. SES was a standardized score derived from parent qualifications and employment and mother's age at birth of her first child (ascertained at first contact when the twins were aged 18 months). Prenatal maternal smoking was determined at first contact from "yes" responses to "Did you smoke cigarettes while pregnant?" SLE scores were square root transformed to reduce skewness. Variables were standardized for regression analyses.

Data Analyses
Differences among nonsmokers, occasional smokers, and regular smokers on SPEQ scores and covariates were tested using one-way analyses of variance (for continuous variables) and c 2 tests (categorical variables). Pairwise comparisons were performed using Bonferroni-corrected p values.
Regression Analyses. Regression analyses were conducted using the lm() function in R. 37 PE were the outcome variables and tobacco use the exposure. Missing data were omitted on a pairwise basis. Linear regression between PE and tobacco use was run to estimate unadjusted models. In adjusted models, predictor variables were entered simultaneously. The presence of multicollinearity was assessed by computing variance inflation factors (VIF), where VIF > 4 indicates multicollinearity. Significance thresholds were set at p < .05 and adjusted R 2 values reported to account for the number of predictors in models.

Structural Equation Twin
Modeling. The twin design enables individual differences to be decomposed into genetic and environmental components. If a trait is heritable, MZ within-pair similarity is higher than DZ similarity. Shared environmental influences make children growing up in the same family similar as indexed by DZ within-pair similarity being greater than half MZ within-pair similarity. Nonshared environmental influences make children growing up in the same family different, present where MZ within-pair similarity is less than unity.
Structural equation twin modeling was conducted in OpenMx 2.0 38 for R 37 . The effects of sex and age were regressed out of SPEQ scales after normalization. Twin models were fitted if the correlation between PE and tobacco use was >0.15 to allow for enough covariance to be decomposed into genetic and environmental influences. Liability-threshold models, which assume an underlying normal distribution to ordinal data, were fitted for univariate tobacco use models. For bivariate analyses, joint ordinalÀcontinuous twin models were fitted that modeled liability to tobacco use and variation in PE.
Twin correlations were calculated using intraclass correlations. Saturated models, constraining means, thresholds, and phenotypic correlations across twin order, were run between PE and tobacco use to calculate cross-twin cross-trait (CTCT) and phenotypic polyserial correlations. Saturated models provide a full description of the data prior to decomposing variance/covariance into genetic (A), common environmental (C), and unique environmental (E) influences. ACE models were compared to saturated models. Subsequently, statistical significance of the variance components was tested by systematically fixing each to zero in submodels (while retaining E because it contains residual error) and comparing to ACE models. The most parsimonious model is preferred and was identified if a likelihood-ratio test at p < .05 indicated a not significantly worse fit compared to full models and based on the lowest Akaike Information Criterion (AIC).
The extent to which the same genes or environments influence PE and tobacco use was estimated from the genetic correlations (r A ), shared environmental correlations (r C ), and unique environmental correlations (r E ). The bivariate heritability (biva 2 ) and equivalent bivariate values for shared (bivc 2 ) and unique environments (bive 2 ) provided estimates of the degree of covariation between tobacco use and PE explained by A, C, or E. These values were divided by the phenotypic correlations to calculate the proportion of the covariance explained by genetic or environmental influences.

Regression Models
The VIF for all predictors ranged between 1.07 and 1.51, which indicated no multicollinearity between predictor variables (see Table S3, available online, for correlations between variables). Unadjusted models (Table 2) indicated that regular smoking, compared to not smoking, significantly (p < .05) predicted higher scores on all PE subscales. Occasional smoking, compared to not smoking, significantly predicted higher scores for paranoia, hallucinations, and cognitive disorganization and lower scores for anhedonia, and did not predict grandiosity and parent-rated negative symptoms. Models accounted for 0.3% to 2.5% of variance in PE scores, most for cognitive disorganization, paranoia, and hallucinations.
Adjusted models accounted for 4.6% to 22.6% of variance in PEs, the highest being for cognitive disorganization (22.6%), paranoia (16%), and hallucinations (14.5%). Standardized coefficients indicated that paranoia increased by 0.13 and 0.17 SD in occasional and regular smokers, respectively, compared to nonsmokers. Models for hallucinations, cognitive disorganization, and parent-rated negative symptoms indicated an increase of 0.24, 0.34, and 0.23 SD, respectively, in regular smokers compared to nonsmokers; occasional smoking did not significantly predict these PE. Regular tobacco use did not significantly predict grandiosity in adjusted models. Anhedonia scores decreased on average by 0.12 SD in occasional smokers compared to nonsmokers. Compared to unadjusted models, standardized coefficients for tobacco use generally decreased in adjusted models. Sensitivity analyses using generalized estimating equation models to include both twins did not affect our conclusions (see Table S4, available online).

Genetic and Environmental Influences on Psychotic Experiences and Tobacco Use
Phenotypic correlations (Table 3) were sufficiently large (>0.15) to run bivariate models between tobacco use and paranoia, hallucinations, and cognitive disorganization.
Univariate MZ twin correlations (Table 3) were higher than DZ correlations for tobacco use, paranoia, hallucinations, and cognitive disorganization, which implied genetic influences (A). Common environmental influences (C) were indicated for hallucinations and tobacco use, and somewhat for paranoia, because DZ correlations were greater than half the MZ correlations, but not for cognitive disorganization. Unique environmental influences (E) were indicated for all measures because MZ correlations were <1.   A full ACE liability-threshold model was indicated for tobacco use (see Table S5, available online) with A ¼ 0.32 (CI ¼ 0.17À0.49), C ¼ 0.51 (0.36À0.64), and E ¼ 0.17 (0.13À0.29). A third of the variance in tobacco use was explained by genetic influences and half due to common environment. Univariate models for SPEQ subscales have been published previously from the TEDS sample 21 showing genetic (A ¼ 0.27À0.54) and nonshared environmental (E ¼ 0.12À0.50) influences explain most variation in PE.
Cross-twin cross-trait (CTCT) correlations (Table 3) were higher in MZ than in DZ pairs, indicating A on the covariance between PE and tobacco use. The DZ CTCT correlations were greater than half those of MZ correlations for paranoia and hallucinations (but not for cognitive disorganization), implicating C influences on covariation. Some E influences on the covariation between tobacco use with paranoia, hallucinations, and cognitive disorganization were suggested, as MZ CTCT correlations were lower than phenotypic correlations.
Fit statistics for bivariate models (Table 4) indicated that, compared to saturated models, ACE models did not fit significantly worse. Partial AE models, in which C parameters were dropped for PE and for covariance paths between tobacco and PE (C was retained for tobacco because univariate results indicated that C influences tobacco use), were compared to the full ACE models. Partial    AE models did not fit significantly worse than ACE models for tobacco use with both paranoia and cognitive disorganization. Partial AE models dropping genetic correlations had significantly worse fits compared to ACE models.
For tobacco use and hallucinations, ACE models dropping either r A or r C , but not both, were not significantly worse in terms of fit compared to the full ACE model. The AIC values for the ACE models that dropped either r A or r C were equally low and could not be distinguished in terms of fit. These results indicated overlapping familial (genetic and/or shared environmental) influences between tobacco use and hallucinations. Thus, the full ACE model is reported.
Parameters from the most parsimonious models (Table 4; see Figure S1, available online) suggested significant genetic overlap between tobacco use with paranoia (r A ¼.37) and with cognitive disorganization (r A ¼ 0.45), no shared environmental overlap, and modest overlapping unique environmental influences (r E ¼ 0.09À0.12). As a proportion of the phenotypic correlations (Table 3), genetic influences accounted for 84% of the covariance between tobacco use and paranoia and 81% between tobacco use and cognitive disorganization. The ACE model for hallucinations and tobacco indicated familial influences (A or C) accounted for 80% of the phenotypic correlation. See Table S6, available online, for bivariate statistics for full ACE models. Sensitivity analyses employing a dichotomous definition of tobacco use corroborated these findings (see Tables S7 and S8; Figure S2, available online).

DISCUSSION
This study investigated two separate questions about the relationship between PE and tobacco use. First, we established that associations between tobacco use and specific types of PE were present during adolescence and remained for most PE after controlling for several covariates. Second, this was the first twin study to investigate the degree to which genetic and environmental influences explain the association between tobacco use and PE. Tobacco use was associated with paranoia and cognitive disorganization largely due to overlapping genetic influences, and with hallucinations due to familial influences that may include genes and shared environment.
Our results suggest that some of the same genetic factors that influence tobacco use in adolescents also influence PE. These results may also indicate geneÀenvironment correlations whereby the genetic factors that influence PE create environments that make tobacco use more likely, and vice versa. Although the genetic correlation could indicate a causal association between PE and tobacco use in either direction, this was beyond the scope of this paper.
Our finding that tobacco use is associated with PE in adolescents is in line with previous findings. Gage et al. 5 found an association between tobacco at age 16 years and PE at age 18 after controlling for cannabis use, familial depression, maternal education, IQ, bullying, and childhood psychosocial issues. McGrath et al. 6 reported that those who commenced smoking before the age of 15 years were more likely to experience hallucinations at age 21. We found similar evidence while accounting for confounders not previously considered, such as sleep disturbance and SLE.
Tobacco use predicted PE to a lesser extent after including covariates, and adjusted models explained more variation in PE than models for tobacco use only. Therefore, some of the association between tobacco use and PE was attributable to the covariates tested, most notably to sleep disturbance. Considering these results and the known etiological association between PE and sleep disturbance, 24 future studies could explore the relationship between PE, tobacco use, and sleep disturbance. A previous TEDS study concluded that PE and cannabis use co-occurred due to shared environmental factors. 21 In our analyses, we controlled for confounding by several other (environmental) factors, which may explain why cannabis use did not predict PE here.
We found regular smoking to be associated with paranoia, hallucinations, cognitive disorganization, and parentrated negative symptoms, the latter two being less commonly studied compared to positive PE. The relationship between tobacco use and anhedonia diverged from this trend. Occasional smoking appears to be associated with lower anhedonia, but regular smoking, albeit not significantly, with higher anhedonia. Perhaps adolescents who experimented with tobacco, but not those who habitually smoked, were more likely to engage in pleasure-seeking behavior. Negative symptoms were parent-rated, whereas anhedonia was self-rated, which may partly account for different effect sizes between these PE types.
Another finding was that 50% of variation in adolescent tobacco use was due to common environmental influences and a third due to additive genetics. Twin studies have previously investigated the heritability of adolescent smoking. 30 Estimates of heritability and environmental influences can change over time and are context and population specific. 39 Anti-smoking regulations, legislation, and social attitudes toward smoking have changed rapidly over recent years. Between the early 2000s and the current study, the United Kingdom has banned smoking in public places, increased the legal age for purchasing tobacco products, introduced bold health warnings on tobacco products, restricted advertising, increased prices, and made nicotine replacement therapies more accessible. 40 A consideration is our operationalization of the tobacco use variable. We chose a three-level rather than a binary definition to improve power and to distinguish between the effects of regular and occasional smoking, in line with previous studies. 2,5 Our twin models assume linearity among not smoking, occasional smoking, and regular smoking. The prior regression analyses showed a linear association between tobacco use with paranoia, hallucinations, and cognitive disorganization, which supports this assumption. We note that anhedonia showed a nonlinear pattern, which could be explored in future research.
Our study was not designed to assess nicotine dependency, as never-smokers, for whom susceptibility to dependency is unknown, were included in our tobacco use measure. It is known that there may be some etiological differences between smoking initiation and nicotine dependency. 41 We also cannot rule out attrition bias, because participating families were more likely to report higher socio-economic status than nonparticipating families.
The association between PE and tobacco use is significant and modest, and findings should be viewed in this context. We could not distinguish between genetic and common environmental influences shared between tobacco use and hallucinations, likely because the phenotypic correlation was lower than for our other bivariate models. Low phenotypic correlations also meant that we could not perform bivariate models between tobacco with grandiosity, anhedonia, and parent-rated negative symptoms.
Adolescent tobacco use is modifiable risk factor. Understanding the nature of the association between PE, a possible early manifestation of psychiatric disorder in some individuals, and tobacco use is of great interest. We have contributed to this field by showing that significant associations exist between specific PE domains and tobacco use in mid-adolescence that are not fully accounted for by confounding factors. We have provided novel insights into the etiology of the covariation between some PE and tobacco use that may inform further molecular genetic studies and developmental models.    Note: Both twins (using exclusion criteria as described in the Methods for bivariate twin analyses) included in Generalized Estimating Equation (GEE) models. GEE models were conducted in the drgee package for R. Reference group for sex: "female," for ethnicity: "white," for cannabis use: "no," and for maternal smoking during pregnancy: "no." a Parent-rated.    Note: Saturated models constrained means, thresholds and phenotypic correlations across twin order. Nonsmokers defined as those who had never smoked or who smoked less than 50 lifetime cigarettes. Smokers were those who smoked 50 or more lifetime cigarettes and who smoked in the past year. -2LL ¼ minus 2 log-likelihood; Ddf ¼ difference in degrees of freedom compared to the base model; D-2LL¼log-likelihood ratio chi-square test comparing the -2LL of each model to the -2LL fit of the base model; A ¼ additive genetic influences; Base ¼ comparison model; AIC ¼ Akaike's Information Criterion (lower values reflect a more parsimonious and therefore preferred fit); C ¼ common environmental influences; df ¼ degrees of freedom; E ¼ unique environmental influences; EP ¼ estimated parameters; r A ¼ genetic correlation; r C ¼ common environmental correlation; r E ¼ unique environmental correlation a Most parsimonious model.

FIGURE S1 Path Diagrams and Parameter Estimates for Most Parsimonious Models: Psychotic Experiences and Tobacco Use
Note: Ninety-five percent CI displayed in parentheses. Dashed lines indicates nonsignificant correlations. A ¼ additive genetic influences; C ¼ common environmental influences; E ¼ unique environmental influences; r A ¼ genetic correlation; r C ¼ common environmental correlation; r E ¼ unique environmental correlation.