Abstract
We tested the hypothesis that high novelty seeking (NS—an externalizing trait), sweet-liking (SL—a phenotype that may reflect processing of hedonic stimuli) and initial insensitivity to the impairing effects of alcohol (SRE-A) act independently and synergistically to increase the likelihood of having alcohol-related problems in young adults.
A sample of 145 young adults, ages 18–26, balanced for gender and alcohol use disorders identification test (AUDIT) scores <8 or ≥8 were selected from a prior sample. NS, SL and SRE-A were assessed along with AUDIT score and family history of alcoholism (FH). The effect of phenotypes and their interaction on the likelihood of alcohol problems was assessed.
All three phenotypes contribute to the total AUDIT score. The best-fitting model explaining 35.8% of AUDIT variance includes all three phenotypes and an interaction between NS and SL/sweet-disliking (SDL) status. The addition of FH to the model explains an additional 4% of variance in both models. Classification and regression tree analysis showed that the main phenotype influencing AUDIT score is NS. The SL/SDL phenotype is a strong modifying factor for high NS. SRE-A was shown to be a weak modifier for individuals with low NS.
The evidence supports the hypothesis that the presence of multiple alcohol use disorders (AUD) risk phenotypes with different underlying neurobiological mechanisms within an individual (SL, NS and SRE-A) represents a higher likelihood for developing alcohol-related problems and may allow for a graded assessment of risk for AUD and offer the possibility for early intervention strategies.
INTRODUCTION
Alcohol use disorders (AUDs) are some of the most common medical disorders in the USA with recent epidemiological studies indicating 12-month and lifetime prevalence of AUDs of 13.9 and 29.1%, respectively (Grant et al., 2015). Identifying traits associated with risk for developing AUDs is critical for understanding, predicting and intervening in the early phases of progression of these disorders.
A number of factors have been found to be associated with risk for developing AUDs, including early onset of drinking alcohol (Dawson, 2000), a reduced behavioral response to alcohol (Schuckit et al., 2004), neurocognitive disinhibition (Kamarajan et al., 2006) and several measures of the externalizing spectrum including conduct disorder and high novelty seeking (NS; Wingo et al., 2016). However, at present, despite identification of a variety of risk factors, our understanding of the forces that lead to AUDs and our ability to predict risk in a given individual are limited.
One of the challenges for finding biological predictors for the risk of AUDs is the complexity of the disease. AUDs demonstrate variability in age of onset, symptom profiles and progression (Litten et al., 2015). The pathophysiology of AUDs is unclear with multiple genes having been identified as contributing to AUDs (Edenberg et al., 2019) with the variance contributed by any individual gene being small. Nevertheless, there has been interest in identifying phenotypes of risk as they provide integrative measures of genetic/epigenetic and environmental forces and, therefore, have particular value as measures of risk compared with individual genes.
Two of the most promising phenotypes associated with AUDs based on replication by multiple groups are neurophysiological disinhibition associated with externalizing behavior (Porjesz and Rangaswamy, 2007) and level of response to alcohol (Schuckit et al., 2014). Other promising endophenotypes include delayed reward discounting, alterations in executive function, hedonic response to sweet taste (sweet-liking) and brain structure (for review, see Salvatore et al., 2015).
Our group has significant experience with the study of hedonic response to sweet taste as a risk phenotype for AUDs (Kampov-Polevoy et al., 2001, 2003; Lange et al., 2010; Kampov-Polevoy et al., 2014). The hedonic response to sweet taste is a heritable and stable phenotype (Mennella et al., 2005; Keskitalo et al., 2007) that is thought to index dimensions of hedonic processing of rewarding stimuli mediated through opioidergic/dopaminergic systems in striatal/accumbens brain regions (Castro and Berridge, 2014). Variation in sweet taste and associated genes has been associated with differences in alcohol consumption (Lanier et al., 2005; Bachmanov et al., 2011; Choi et al., 2017). In humans, there is extensive evidence for at least two phenotypes associated with the hedonic response to sweets—a sweet-liking (SL) phenotype characterized by preference for increasing concentrations of sucrose and a sweet-disliking (SDL) phenotype characterized by an increased preference for sucrose to ~0.3 M followed by a progressive decline in preference (Thompson et al., 1976). Sweetness intensity across solutions is rated similarly by SL and SDL phenotypes (Kampov-Polevoy et al., 1999).
Following preclinical work demonstrating an association between preference for/consumption of sweets and propensity to consume various addictive substances including alcohol (Belknap et al., 1993; Carroll and Holtz, 2014), a body of work has evolved in humans that indicates the SL phenotype is significantly and positively associated with alcohol dependence and the genetic risk of AUDs as assessed by family history of alcoholism (FH; Kampov-Polevoy et al., 2003; Krahn et al., 2006; Pepino and Mennella, 2007; Mennella et al., 2010). Furthermore, studies in young adults have shown that the SL phenotype is strongly associated with alcohol problems and that the presence of high NS with the SL phenotype greatly increases the likelihood of alcohol problems (Lange et al., 2010; Kampov-Polevoy et al., 2014). The finding of a strong interaction between high NS and SL to increase the likelihood of alcohol problems in young adults led us to hypothesize that the presence of multiple and independent risk phenotypes in the same individual may be of particular concern with regard to the development of AUDs.
To further test this hypothesis, we here report the effect of adding a third phenotype, low initial sensitivity to alcohol, to the impact of high NS and SL on drinking and alcohol-related problems in young adults. The reported sample is the same as our previous reported sample (Kampov-Polevoy et al., 2014) but we have now included the results of the subjective response to alcohol (SRE-A) questionnaire (Schuckit et al., 1997) to assess for initial insensitivity to alcohol. Additionally, we have used statistical modeling to examine the interaction of the three phenotypes and the presence of alcohol problems. Our hypothesis was that adding a third phenotype would further enhance the predictive value of our model regarding the presence of having alcohol-related problems in young adults.
MATERIALS AND METHODS
Subjects (n = 163; 18–26 years of age) were recruited via email announcement using the University of North Carolina at Chapel Hill mass email system (also see Kampov-Polevoy et al., 2014). Approximately 250 individuals were screened over the phone to yield 163 subjects balanced by gender and, within gender, balanced for alcohol use disorders identification test (AUDIT; Saunders et al., 1993) scores of <8 and ≥8. The primary reason for exclusion was that the potential subject’s cell, for example, male with high AUDIT, had been filled. All subjects were able to read and write English. They were asked to maintain abstinence from alcohol for 24 h prior to testing, fasted for at least 2 h prior to testing and abstained from smoking for at least 1 h prior to testing. Exclusion criteria included a reported history of medically diagnosed bipolar disorder, active depression, psychotic disorder or active substance use/dependence other than alcohol or nicotine according to DSM-IV criteria (First et al., 2002). None of the subjects were excluded because of these criteria. Women who were known to be pregnant or at a risk of a pregnancy, for example, missed menstrual cycle, were excluded. Subjects who did not drink alcohol for religious or cultural reasons were excluded. The study was approved by the Committee on the Protection of the Rights of Human Subjects, School of Medicine, University of North Carolina at Chapel Hill.
Procedure
Potential subjects were initially interviewed over the phone after obtaining verbal consent. The main goal of this interview was to see whether potential subjects met inclusion/exclusion criteria and to administer the AUDIT. Individuals who met criteria and were interested in participating were scheduled for a screening visit. During this visit, informed consent was acquired, and a breathalyzer (Intoximeters, Inc., St. Louis, MO) test was administered. Only those subjects with breath alcohol concentration of 0.00 g/dl were allowed to proceed with the study. No subjects were excluded because of alcohol intoxication. During this visit, all subjects completed a demographic questionnaire and the tridimensional personality questionnaire (Cloninger, 1987), a 100-item, self-administered instrument that assesses three dimensions of personality—NS, harm avoidance and reward dependence. Family history of alcoholism was obtained using family history assessment module (Rice et al., 1995) questionnaire. A positive screen in a first degree relative was considered evidence for a positive family history of alcoholism (FH+). All other subjects were considered family history negative (FH−).
Sweet taste test
To assess hedonic response to sweet taste, we used a standard sweet taste test (for methods see Kampov-Polevoy et al., 1997) that was conducted at the end of the visit and at least 90 min after the last meal. To reduce impacts on taste, subjects had not chewed gum, consumed a flavored beverage or smoked for at least 1 h prior to the sweet taste test. During the test, 2 ml each of five concentrations of sucrose solution (0.05, 0.10, 0.21, 0.42 and 0.83 M) were presented five times in a prespecified, random order. For perspective, Coca-Cola Classic™(Coca Cola Company, Atlanta, GA) is a 0.33 M sugar solution. Subjects were asked to rate: ‘How sweet was the taste?’ and ‘How much do you like the taste?’ on a 200-mm analog scale. Average rating of pleasurableness of each sucrose solution was calculated. SL was defined as giving the highest pleasantness rating to the highest sucrose concentration (0.83 M), and SDL was defined as giving the highest pleasantness rating to one of the lower sucrose concentrations (0.05, 0.10, 0.21 or 0.42 M).
To assess initial insensitivity to the impairing effects of alcohol without having to administer alcohol, Schuckit et al. (1997) developed a self-rating instrument—the self-rating of the effects of alcohol (SRE)—which demonstrated good comparability to results obtained with individuals administered alcohol in a laboratory setting. High scores on the first item (SRE-A) indicate evidence of initial insensitivity to the effect of alcohol.
Dependent variables
The AUDIT was selected as the primary dependent variable. The AUDIT is a 10-item self-administered test that was developed by the World Health Organization in 1982 as a simple way to screen and identify people who have the preliminary signs of hazardous drinking and mild dependence and are at risk of developing alcohol problems. The AUDIT has proven to be accurate across all ethnic and gender groups (Saunders et al., 1993).
Statistical analyses
Study subject characteristics were summarized for a select set of variables using sample size, mean, standard deviation, and range for continuous variables and sample size, counts and percentages for dichotomous variables. Associations between SL, NS and SRE-A were assessed using correlations on the continuous (original) scales. For modeling, NS and SL status were dichotomized primarily for consistency with our previous work (Lange et al., 2010). Another reason is that dichotomized variables provide a threshold that may have practical clinical value when assessing an individual. We kept SRE-A as a continuous variable because there is no clear criterion to dichotomize it, although when we used decision trees (as described below) the tree search algorithm finds the best dichotomization to maximize the explained variance.
Statistical modeling used linear regression and decision trees. We considered bivariate associations between the phenotypes and the outcomes. We also considered all possible combinations of variables and their two- and three-way interactions to assess which model fits the data the best. To account for overfitting, we considered Akaike information criterion (AIC) accounting for model complexity. This criterion is similar to the goodness of fit (as assessed by the likelihood function), but it also includes a penalty for the number of explanatory variables in the model. This penalty controls overfitting, because increasing the number of variables in the model almost always improves the goodness of the fit. ‘The smaller the value of AIC the better model fits the data’. AIC also has a useful application for model comparison. The quantity exp([AICmin − AICi]/2) is known as the relative likelihood of model i. It is corresponding to a probability that model i better explains the data. This approach is preferred to commonly used k-fold cross-validation because of the small sample size where splitting the sample will result in too small and unreliable subsamples. Family history was considered in the supplemental modeling because of previous work that indicated that it is strongly associated with the risk of problematic alcohol use (Kendler et al., 2015).
A classification and regression trees (CART) approach was used to estimate the best interactions and visualize the effects of the main effects and interactions in the model. Regression trees recursively partition the sample into groups, where subjects within a group are more similar to each other than they are to subjects in other groups with respect to the outcome (e.g. total AUDIT score). Unlike ordinary regression models, CART allows one to understand and visualize potential interactions between the chosen predictors regarding the risk of alcohol-related problems, as described by the AUDIT continuous score. All analysis was done using statistical packages in R.
RESULTS
Study subject characteristics
Subject characteristics are presented in Table 1. Of 163 subjects SRE-A was missing in 12, SL/SDL status was missing in 4 and AUDIT was missing in 4. Altogether 18 subjects had at least one observation missing so the analysis sample size was 145 subjects.
Summary of subject characteristics
| Variable (dichotomous) | Sample size | n (%) |
|---|---|---|
| Female | 145 | 74 (51%) |
| Sweet-liking (SL) | 145 | 72 (50%) |
| High novelty seeking binary (19 for women and 20 for men cutoffs) | 145 | 35 (24%) |
| Variable (continuous) | Sample size | Mean (SD), range |
| Age | 145 | 21.0 (1.83), 18–26 |
| Novelty seeking score | 145 | 15.9 (5.84), 3–29 |
| SRE-A | 145 | 4.01 (1.77), 1–10.25 |
| AUDIT | 145 | 9.1 (6.1), 0–35 |
| Drinks/month (male) | 71 | 32.3 (32.5), 0–144 |
| Drinks/month (female) | 73 | 23.7 (27.9), 0–144 |
| Variable (dichotomous) | Sample size | n (%) |
|---|---|---|
| Female | 145 | 74 (51%) |
| Sweet-liking (SL) | 145 | 72 (50%) |
| High novelty seeking binary (19 for women and 20 for men cutoffs) | 145 | 35 (24%) |
| Variable (continuous) | Sample size | Mean (SD), range |
| Age | 145 | 21.0 (1.83), 18–26 |
| Novelty seeking score | 145 | 15.9 (5.84), 3–29 |
| SRE-A | 145 | 4.01 (1.77), 1–10.25 |
| AUDIT | 145 | 9.1 (6.1), 0–35 |
| Drinks/month (male) | 71 | 32.3 (32.5), 0–144 |
| Drinks/month (female) | 73 | 23.7 (27.9), 0–144 |
Summary of subject characteristics
| Variable (dichotomous) | Sample size | n (%) |
|---|---|---|
| Female | 145 | 74 (51%) |
| Sweet-liking (SL) | 145 | 72 (50%) |
| High novelty seeking binary (19 for women and 20 for men cutoffs) | 145 | 35 (24%) |
| Variable (continuous) | Sample size | Mean (SD), range |
| Age | 145 | 21.0 (1.83), 18–26 |
| Novelty seeking score | 145 | 15.9 (5.84), 3–29 |
| SRE-A | 145 | 4.01 (1.77), 1–10.25 |
| AUDIT | 145 | 9.1 (6.1), 0–35 |
| Drinks/month (male) | 71 | 32.3 (32.5), 0–144 |
| Drinks/month (female) | 73 | 23.7 (27.9), 0–144 |
| Variable (dichotomous) | Sample size | n (%) |
|---|---|---|
| Female | 145 | 74 (51%) |
| Sweet-liking (SL) | 145 | 72 (50%) |
| High novelty seeking binary (19 for women and 20 for men cutoffs) | 145 | 35 (24%) |
| Variable (continuous) | Sample size | Mean (SD), range |
| Age | 145 | 21.0 (1.83), 18–26 |
| Novelty seeking score | 145 | 15.9 (5.84), 3–29 |
| SRE-A | 145 | 4.01 (1.77), 1–10.25 |
| AUDIT | 145 | 9.1 (6.1), 0–35 |
| Drinks/month (male) | 71 | 32.3 (32.5), 0–144 |
| Drinks/month (female) | 73 | 23.7 (27.9), 0–144 |
Statistical independence between the phenotypes
Result
For this analysis we first examined correlations between the original continuous scores and found no association between NS scores and SL scores (correlation −0.09, P-value 0.44). The correlation between SRE-A and NS score was 0.18, P-value 0.02 and with SL score was 0.19, P-value 0.01. Because in the analysis we used binary scores we considered a chi-square between binary SL and NS which again did not find association between the binary indicators (P-value 0.46). We also conducted a t-test of association between the values of SRE-A in SL and NS groups and found no association between SRE-A and NS groups (P-value 0.5), but the association between SRE-A and SL indicator remained statistically significant (P-value 0.006) with the means of 4.4 and 3.6 in SL and SDL groups, respectively.
Interpretation
These analyses indicate that NS and SL are independent of one another. SRE-A score has evidence of some association with SL but not with NS.
Effect of the combination of three phenotypes (SL, NS and SRE-A) for prediction of AUDIT score compared with one- and two-phenotype models
Result
We considered two regression models. One with two phenotypes: NS and SL, to which we refer as two-phenotype model, and a model with NS, SL and SRE-A, to which we refer as a three-phenotype model (Equations (1) and (2), respectively)
(1) Two-phenotype model: AUDIT ~ β0 + β1*Sweetliking + β2*NoveltySeeking + β3*Sweetliking* NoveltySeeking
(2) Three-phenotype model: AUDIT ~ β0 + β1*Sweetliking + β2* NoveltySeeking + β3* Sweetliking* NoveltySeeking + β4* SRE-A
Using AIC criterion we calculated the quantity exp ([AICmin − AICi]/2). The three-phenotype model has an AIC of 882, and the two-phenotypes model has an AIC of 893, which implies that the probability that a two-phenotype model is better than a three-phenotype model is exp([882–893]/2) = 0.004, which supports the assessment that the three-phenotype model is better (i.e. the probability that a two-phenotype model is superior to a three-phenotype model to explain the AUDIT score is 0.4%).
The best-fitting model included three main effects and a single interaction between NS and SL/SDL status (see a three-phenotype model). This is an improvement in AIC-adjusted fit over the model with only two phenotypes: SL/SDL status and NS (a two-phenotype model), used in Kampov-Polevoy et al. (2014) and Lange et al. (2010). As can be seen in Table 2, adding SRE-A as a third phenotype increases the percentage of explained variables.
Resulting coefficients for 2-phenotype and 3-phenotype regression models of AUDIT score. Standard errors are presented in parenthesis
| Outcome | Model | SL/SDL status (SE) | NS (SE) | SRE.A (SE) | SL*NS (SE) | Percent of variance explained |
|---|---|---|---|---|---|---|
| AUDIT | 2-phenotype | 0.23 (0.98)* | 2.65 (1.32)** | — | 8.3 (2.01)*** | 29.9% |
| AUDIT | 3-phenotype | 0.88 (0.96)* | 2.55 (1.2)** | 0.86 (0.24)*** | 8.1 (1.94)*** | 35.8% |
| Outcome | Model | SL/SDL status (SE) | NS (SE) | SRE.A (SE) | SL*NS (SE) | Percent of variance explained |
|---|---|---|---|---|---|---|
| AUDIT | 2-phenotype | 0.23 (0.98)* | 2.65 (1.32)** | — | 8.3 (2.01)*** | 29.9% |
| AUDIT | 3-phenotype | 0.88 (0.96)* | 2.55 (1.2)** | 0.86 (0.24)*** | 8.1 (1.94)*** | 35.8% |
*Denotes statistical significance at 0.06 level.
**Denotes statistical significance at 0.01 level.
***Denotes statistical significance at 0.001 level.
Resulting coefficients for 2-phenotype and 3-phenotype regression models of AUDIT score. Standard errors are presented in parenthesis
| Outcome | Model | SL/SDL status (SE) | NS (SE) | SRE.A (SE) | SL*NS (SE) | Percent of variance explained |
|---|---|---|---|---|---|---|
| AUDIT | 2-phenotype | 0.23 (0.98)* | 2.65 (1.32)** | — | 8.3 (2.01)*** | 29.9% |
| AUDIT | 3-phenotype | 0.88 (0.96)* | 2.55 (1.2)** | 0.86 (0.24)*** | 8.1 (1.94)*** | 35.8% |
| Outcome | Model | SL/SDL status (SE) | NS (SE) | SRE.A (SE) | SL*NS (SE) | Percent of variance explained |
|---|---|---|---|---|---|---|
| AUDIT | 2-phenotype | 0.23 (0.98)* | 2.65 (1.32)** | — | 8.3 (2.01)*** | 29.9% |
| AUDIT | 3-phenotype | 0.88 (0.96)* | 2.55 (1.2)** | 0.86 (0.24)*** | 8.1 (1.94)*** | 35.8% |
*Denotes statistical significance at 0.06 level.
**Denotes statistical significance at 0.01 level.
***Denotes statistical significance at 0.001 level.
Interpretation
The best model for the AUDIT score shows that each of the chosen phenotypes is individually associated with the AUDIT score and that the model with three phenotypes is better than the model with only two phenotypes. Model coefficients for AUDIT outcomes are presented in Table 2. Table 2 shows that although each phenotype by itself contributes to a certain extent AUDIT score, the combination of endophenotypes significantly increases such prediction in terms of percentage of explained variables. The three-phenotype model explains an additional 6% variance of AUDIT compared with the two-phenotype model.
Effect of family history of alcoholism on prediction of AUDIT score in two- and three-phenotype models
Result
When controlling for FH status, we get similar results (i.e. a three-phenotype model performs better than the two phenotypes based on AIC criteria exp ([868–880]/2 = 0.002). The odds ratios for FH were 3.07 and 3.12 for the two- and three-phenotype models respectively. FH status is correlated with the combination of SL/high NS (Chi-squared test was significant with P-value of 0.004), and when FH is added to the model, it slightly lowers the effect of the SL/high NS phenotype to 6.9 from 8.3 and to 6.4 from 8.1 for the two- and three-phenotype models, respectively. The addition of FH to the model explains an additional 4% of variance in AUDIT score.
Interpretation
The analysis indicates that the three-phenotype model adjusted for FH continues to make a better prediction of alcohol-related problems than the two-phenotype model with the probability of the opposite being very small, 0.002. The direction and the effect sizes of NS, SL, SRE-A and NS*SL are similar in both models.
Interaction of the three phenotypes and alcohol problems assessed using CART
Classification trees are used for binary and regression trees for continuous outcomes. Unlike ordinary regression models, CART allows one to understand potential interactions between the chosen predictors regarding the risk of alcohol-related problems, as described by the AUDIT continuous score. Regression trees recursively partition the sample into groups, where subjects within a group are more similar to each other than they are to subjects in other groups with respect to the outcome (e.g. AUDIT score; Breiman, 2001; Hastie et al., 2009).
Result
We present the results of a regression tree analysis in Fig. 1 below. Panel A illustrates tree analysis for two phenotypes SL and high NS. The strongest classifier is high NS, which partitions the sample into groups: low NS and high NS with mean total AUDIT scores of 7.6 (range 0–23) and 13.8 (range 4–35), respectively. The next significant split is of the high NS group into two subgroups SDL and SL with respective means of 10.4 (range 4–20) and 18.5 (range 4–35). When SRE-A is added as the third phenotype (Panel B), it splits the branch with low NS (means 6.2 vs. 8.7) but not high NS individuals. The subsample with SRE-A ≤ 3.25 has an Audit score of 6.2 vs. 8.7 among individuals with SRE-A > 3.25.
Regression tree analysis examining interaction of two and three phenotypes on total AUDIT score.
Interpretation
The main phenotype influencing AUDIT score is NS. At the same time, the SL/SDL phenotype is a strong modifying factor for high NS. SRE-A is a weak modifier for individuals with low NS. For the specific subgroup of SL and high NS, the effect of SRE-A is not clear because the size of the SL and high NS is too small to be further split to obtain reliable estimates. All three studied phenotypes—SL/SDL status, NS and SRE-A—are associated with alcohol-related problems in healthy young individuals with high NS having the strongest effect.
DISCUSSION
This study continues our efforts to investigate the contribution of various phenotypes and their combinations for predicting the risk of developing alcohol-related problems in young adults. These studies both provide a better understanding of the pathogenesis of AUDs and identify potential phenotypes for early detection of risk for an AUD in an individual, which, in turn, may allow for intervention in the early phases of progression of AUDs. Given the increased risk of poor outcomes for early alcohol initiators (Hingson et al., 2006; Hingson and Zha, 2009), characterizing phenotypic markers of adolescent substance use proneness may be particularly useful for identifying those who ultimately stand to benefit the most from youth substance use prevention efforts.
A major strength of phenotypic markers is that, unlike current diagnostic criteria, they produce less heterogeneity and a more direct link to the underlying genetic/developmental forces of disease (see Gottesman and Gould, 2003).
The current study allowed investigation of the interaction of three phenotypes suggested by Salvatore and colleagues (for overall review, Salvatore et al., 2015) as most promising for detecting an AUD risk. Each of the three phenotypes independently and jointly contributes to the increase in total AUDIT total score and is consistent with the literature. However, the purpose of the present study was to investigate how the phenotypes interact with one another to affect level of alcohol problems in young adults. Here several interesting findings emerged.
First, with regard to the overall impact of having more than one phenotype on likelihood of alcohol-related problems, the evidence indicates that the more phenotypes present in an individual the higher likelihood of alcohol-related problems. This indicates that the assessment of multiple phenotypes may allow for assessing level of risk above and beyond a simple high/low risk category. Of course, how many phenotypes to assess and how they interact over time to lead to alcohol problems awaits additional work including longitudinal studies, but the evidence presented here is very clear—more risk phenotypes means a greater likelihood of alcohol problems. It is noteworthy that the assessment of each of the phenotypes in the current study is simple, safe, fast and inexpensive so operationalizing this to a clinical setting, including youth, is quite reasonable. Indeed, NS and the SL/SDL phenotype can be assessed in pre-adolescence raising the possibility of very early assessment of risk (for discussion, see Mennella et al., 2010; Tomko et al., 2016).
To explore how the ‘interaction’ of the phenotypes contributes to pathways to alcohol problems we conducted classification tree analyses (Fig. 1). Examining the effect of a two-phenotype model using SL and high NS (Panel A) reveals that the main factor influencing AUDIT score is high NS score. For example, in this sample an individual with low NS regardless of SL status has an average AUDIT-total score of around 7.5 compared with scores of 18.4 in subjects with high NS/SL phenotype and 10.3 in high NS/ SDL phenotype. These findings indicate that individuals with high NS, an externalizing behavior shown to be associated with early onset heavy drinking (Laucht et al., 2007; Wichers et al., 2013), are more likely to develop problematic drinking if they have the SL phenotype and that the SDL phenotype may be weakly protective. In our previous study (Kampov-Polevoy et al., 2014), we suggested that high NS may primarily be a driver of drinking behavior and earlier initiation of drinking that, for reasons still to be determined, in the presence of the SL phenotype are more likely to evolve into an AUD.
When the SRE-A phenotype is brought into the classification tree analysis (Fig. 1, Panel B) a slightly different pathway emerges. High NS modified by the SL phenotype continues to represent the most significant path to problem drinking with SRE-A not having a significant effect on this pathway. However, in individuals with low NS, the presence of high SRE-A (indicating higher innate tolerance) is associated with higher AUDIT scores. Furthermore, it is of interest that SRE-A seems to be connected to a less severe level of alcohol problems in this population < 25 years of age. This hypothesis is consistent with Schuckit’s findings that low level of response to alcohol is associated with positive family of alcoholism that contributes to a transition from lighter to heavier drinking later in life (e.g. low response to alcohol reported at the age of 20 explains 35% of AUD related variance at the age of 35 and 48% of variance at the age of 40; Schuckit et al., 2004). Accordingly, one could hypothesize that the high NS/SL phenotype points to an early onset AUD whereas low SRE-A points to a later onset AUD—although, again, the presence of all three phenotypes in a young adult is associated with the highest level of alcohol-related problems.
Limitations to the current study include the modest sample size, absence of racial diversity and constricted social/educational level. Additionally, the impact of varying definitions of ‘sweet-liking’ was not explored (e.g. see Bouhlal et al., 2018).
In summary, the evidence supports the hypothesis that the presence of multiple risk phenotypes within an individual represents a higher likelihood for developing alcohol-related problems then the presence of single phenotypes and raises the possibility that assessment of risk phenotypes at ages <15 years or so could play a role in guiding counseling with regard to alcohol problems in youth and their families.
DATA AVAILABILITY STATEMENT
The data underlying this article will be shared on reasonable request to the corresponding author.
FUNDING
This work was supported in part by the National Institute on Alcohol Abuse and Alcoholism at the National Institutes of Health (grant number 1RC4AA020096-01).
