Skip to main content
NCBI home page
Wiley Open Access Collection logoLink to Wiley Open Access Collection
. 2025 May 22;49(7):1489–1503. doi: 10.1111/acer.70075

Correlational and causal modeling of alcohol‐related symptoms and internalizing disorder status: Further elucidation of a harm paradox

Justin Anker 1, Bryan Andrews 1, Erich Kummerfeld 1, Paul Thuras 1, Matt G Kushner 1,✉
PMCID: PMC12285913  NIHMSID: NIHMS2082970  PMID: 40405337

Abstract

Background

Individuals with internalizing (anxiety and depressive) disorder (INTD) suffer from an alcohol‐related “harm paradox”; that is, they experience more alcohol‐related symptoms in aggregate than do others who drink at the same level. Here, we extend this earlier finding by examining the association of INTD with a wide range of individual alcohol‐related symptoms.

Methods

The study sample included respondents in the NESARC Wave 3 who reported having consumed alcohol in the past year (N = 24,485). We used logistic regression analysis to identify the association between INTD and risk for 37 individual alcohol symptoms. We used the BOSS causal discovery algorithm to identify the best‐fitting causal model for the full dataset and for 100 resampled datasets, each composed of a randomly selected 50% of the full dataset. Causal edges that appeared in at least 80% of the resampled datasets were deemed “highly stable.”

Results

After controlling for the level of daily alcohol volume and demographic variables, INTD significantly increased the relative odds of having all 37 alcohol‐related symptoms measured (ORs ranged from 1.5 to 4.6). Interactions between INTD and the level of alcohol use were largely nonsignificant. Highly stable direct (unmediated) causal edges emanated primarily from INTD to the symptoms of alcohol withdrawal and dependence.

Conclusions

Those with INTD are at greater risk for a wide range of alcohol symptoms than others who drink at the same level, even at relatively low levels of alcohol use. We consider that INTD could exert a direct causal influence specifically on withdrawal and dependence symptoms due to overlapping experiential and/or neurobiological aspects of these alcohol use symptoms and INTD. We conclude that the harm paradox likely contributes to the elevated risk of developing alcohol use disorder comorbidity among those with INTD.

Keywords: alcohol, anxiety, comorbidity, depression, harm‐paradox

We found that, after controlling for the level of daily alcohol use and demographic variables, those with a current anxiety or depression (“internalizing”) disorder (INTD) demonstrate significantly elevated relative odds of having all 37 alcohol‐related symptoms measured (a “harm paradox”). Causal discovery analysis showed stable unmediated causal edges emanated primarily from INTD to the risk for symptoms of alcohol withdrawal and dependence. The harm paradox likely contributes to the elevated risk for developing alcohol use disorder comorbidity among those with INTD.

graphic file with name ACER-49-1489-g001.jpg

INTRODUCTION

High‐quality epidemiological studies show that alcohol use disorder (AUD) is found in those with anxiety and depressive disorders (“internalizing disorder”; INTD) at a rate that is two to three times greater than that found in those without INTD (Kushner et al., 2008); a phenomenon conventionally referred to as “comorbidity.” Given the relatively high base rate of both INTD and AUD individually and the further elevated risk for either condition in the presence of the other, comorbidity affects a large number of individuals (Grant et al., 2004; Kessler et al., 1997). Not only are those with INTD at greater risk for developing AUD, they also experience worse AUD treatment outcomes than those with AUD alone (Driessen et al., 2001; Kushner et al., 2005). Despite decades of research, however, our understanding of the processes that promote comorbidity remains incomplete (Anker & Kushner, 2019; Kushner et al., 2000, 1990).

Elevated risk for AUD among those with INTD may result, in part, from higher levels of alcohol use due to negative reinforcement from alcohol's tension‐reducing properties (Menary et al., 2011). However, recent evidence suggests that those with INTD may also be more vulnerable to the development of alcohol symptoms than others who drink at the same level. For example, Kushner et al. (2011) demonstrated that the time from first regular drinking to the onset of alcohol dependence is, on average, significantly shorter (“telescoped”) among those with INTD. Kushner, Menary, et al. (2012) further demonstrated that the INTD telescoping effect is not explained by higher levels of use. Most recently, Anker et al. (2023) showed that, after controlling for the average daily amount of alcohol consumed over the past year and a number of other clinical (e.g., family history of AUD) and demographic (e.g., sex and income) variables known to be associated with AUD risk, individuals with INTD experienced more alcohol‐related symptoms than those without INTD. Referencing Anker and Kushner (2019), Anker et al. (2023) suggested that those with INTD may experience a neurosusceptibility to the development of AUD symptoms.

The purpose of this work is to extend that of Anker et al. (2023) with the goal of better characterizing the alcohol‐related harm paradox in those with INTD. The specific aims of this work are to (1) quantify the magnitude of the harm paradox effect associated with individual alcohol‐related symptoms (vs. the total count of symptoms in the earlier work); (2) document the extent to which the harm paradox effect is modified by drinking level (vs. controlling level of alcohol use in the earlier work); and (3) discover the best fitting and stable causal model of INTD status and alcohol symptoms, with a focus on direct (unmediated) causal influences (vs. exclusively correlational associations in the earlier work). As in the earlier work, the sample is drawn from the NESARC 3 restricted to those who reported using alcohol in the past year. For both Anker et al. (2023) and this work, the decision to study the full range of alcohol use (vs. only those with heavy drinking or an AUD diagnosis) aligns with our goal of uncovering processes contributing to elevated risk for the development of alcohol‐related problems among those with INTD.

METHODS

Source of the data

The study utilized data from the third wave of the National Epidemiological Survey on Alcohol‐Related Conditions (NESARC 3), conducted between 2012 and 2013 (Grant et al., 2015). The NESARC 3, sponsored by the National Institute on Alcohol Abuse and Alcoholism (NIAAA), aimed to determine the prevalence of psychiatric disorders within the US' adult civilian noninstitutionalized population (aged 18 years and above), including those in Alaska, Hawaii, and the District of Columbia in a nationally representative sample of 36,309 individuals. Institutional Review Board approval for NESARC 3 was provided by the US Census Bureau and the US Office of Management and Budget. The Institutional Review Board of the University of Minnesota reviewed this study and waived the requirement for additional informed consent by the participants.

The NESARC 3 employed multistage probability sampling to randomly select participants. The survey asked participants to self‐report their race/ethnicity, with classifications including White, Black, Native American or Alaskan, Asian, Native Hawaiian or Pacific Islander, and Hispanic or Latino. To help ensure representativeness, the study oversampled African American and Hispanic households and adults aged 18–24 and employed statistical weighting to account for the oversampling and nonresponse at the household and individual levels.

Participants

The primary analyses include NESARC 3 respondents who reported consuming at least one alcoholic drink in the last 12 months and who did not exceed the reporting cap of 14.4 ounces per day/28 drinks, for whom the average daily level of use could not be determined (N = 25,574). Additionally, 1089 cases were dropped due to missing values. This left 24,485 individuals whose data were analyzed in this study. The demographic composition of the study sample was predominantly female (54%), with ages spanning from 18 to 98 years old (Mean = 43.2, SD = 16.4).

Measures

INTD

All participants were assessed using the Alcohol Use Disorders and Associated Disabilities Interview Schedule V (AUDADIS‐V); (Grant et al., 2015), conducted by trained interviewers. The AUDADIS‐V is a highly structured diagnostic interview designed to evaluate the criteria outlined in the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM‐5, research version) to generate categorical psychiatric diagnoses. It employs initial screening questions to ascertain the presence of key features of each disorder in determining whether to proceed with further diagnostic inquiries. Specifically, if these screening questions are not affirmed, participants are deemed not to exhibit the disorder in question, and no additional questions pertinent to that disorder are pursued.

Following the approach of Anker et al. (2023), individuals were categorized as having an INTD if they met diagnostic criteria for a depression (major depressive disorder, dysthymia) or anxiety (generalized anxiety disorder, panic disorder, and social anxiety disorder) disorder in the past 12 months (i.e., “current”). This resulted in the following study groups: (1) no current INTD (N = 19,908); (2) current INTD (N = 4577). In establishing the binary INTD variable, we considered epidemiological research indicating that anxiety and depressive disorders demonstrate high levels of covariance that reflect a single latent variable (“negative emotionality”) (Krueger, 1999) and that this latent variable accounts for nearly all of the elevated risk for AUD among those with any of the INTD disorders (Krueger, 1999; Kushner, Wall, et al., 2012).

Alcohol symptoms

As part of the AUDADIS‐V, respondents were presented with a list of 37 alcohol‐related symptoms/problems—primarily drawn from the DSM‐5 AUD criteria—and asked to indicate which they had experienced in the past year (Table S1).

Average daily alcohol volume

The NESARC survey collected information about the types and amounts of alcoholic beverages people consumed in the past year. Respondents were shown images of different sizes of glasses and asked to report how much they typically drink. The total amount of EtOH (pure alcohol) consumed in the past year was calculated by adding the amounts of each type of beverage. This was then divided by 365 to find the average daily intake of EtOH.

Years of drinking at the same level

After assessing drinking patterns for the past year (above), interviewers asked respondents how many years they had drank at approximately the same level and pattern as in the past 12 months.

Analyses

All data processing, statistical analyses, and model fitting were performed using the R Statistical Software (R Core Team, 2021) along with packages dplyr (Wickham et al., 2021) and haven (Wickham & Miller, 2019).

Descriptive statistics

Descriptive statistics were computed for both continuous and categorical variables to summarize the characteristics of the study sample. Continuous variables, including estimated ounces of alcohol consumption, age, the sum of alcohol‐related symptoms reported, and years of drinking at the same level as the past year, were summarized using means, standard deviations (SD), and standard errors (SE).

Bivariate inferential statistics

INTD groups were compared using independent sample Welch's t‐tests for each continuous variable—age, average daily drinking volume, the sum of alcohol‐related symptoms, and years of drinking at the same level as in the past year. For categorical variables—sex, ethnicity, education level, and income bracket—chi‐square tests of independence were used to examine the associations between each categorical variable and INTD status. Proportions (as percentages) within each level of the categorical variables were calculated to facilitate interpretation.

Logistic regression models

Logistic regression analyses were conducted to examine the effects of multiple predictors on each alcohol‐related symptom as the binary outcome (present vs. absent). Predictors in the models include INTD, age, sex, ethnicity, education level, income bracket, and years of drinking at the same level as in the past year. Interactions between each predictor and average daily alcohol volume were also included. Model summaries include coefficients, standard errors, z‐values, and p‐values for each predictor, and odds ratios (ORs) and 95% confidence intervals (CIs) are presented to interpret the magnitude and direction of associations. The risk for Type I errors associated with multiple comparisons was mitigated using the False Discovery Rate (FDR) approach, implemented via R's p.adjust function with the Benjamini–Hochberg method (Benjamini & Hochberg, 1995) at an alpha level of 0.05. This procedure reduces the expected proportion of false positives among significant findings while minimizing the loss of statistical power compared to more conservative familywise error corrections.

Causal discovery

Causal discovery refers to any of a number of algorithms that search for a causal model structure by differentiating potential causal directions, commonly based on conditional independence tests and model fit scores (Spirtes et al., 1993). We used the Best Order Score Search (BOSS) algorithm (see the method label in Table S2 for details), which has been shown to have superior performance relative to other causal discovery algorithms (Andrews et al., 2023). For the BOSS algorithm, we modeled the data with a conditional Gaussian distribution and used the Bayesian information criterion (BIC) as a goodness‐of‐fit score to evaluate the model structure (Andrews et al., 2018). Notably, modeling the data with a conditional Gaussian distribution allowed the sex, ethnicity, education, income, and INTD variables to form interaction terms with the other variables in the model. The BOSS algorithm was run subject to the constraint that age, sex, and ethnicity could not be caused by any other variables. Table 1 describes the assumptions made by the BOSS algorithm and our justifications for the present analysis.

TABLE 1.

Assumptions made by BOSS, and their justification.

Assumption Content of the assumption Justification in our use case
Causal Markov The Causal Markov assumption captures the idea that causes can be statistically “screened off” when they are completely mediated by other variables This assumption is widely accepted. The only known possible counterexample is in quantum physics (Bell's theorem)
Weak Causal Faithfulness Causal Faithfulness roughly captures the idea that if variables are causally related to each other, then they should be statistically associated with each other. BOSS is correct for a weaker version of Faithfulness that allows some unfaithful independences to occur This assumption can be violated if there are multiple causal pathways with opposing signs, which perfectly cancel each other out. The probability of a Faithfulness violation asymptotically approaches zero as the sample size grows, but finite‐sample Faithfulness violations can occur. BOSS is still correct in the presence of some types of Faithfulness violations, and simulations have shown that BOSS still has strong finite sample performance even when Faithfulness violations are allowed
Causal Sufficiency Causal Sufficiency states that there are no unmeasured common causes Both of these assumptions could plausibly be violated in our data, and would produce distributions that don't fit the conditional Gaussian distribution for any fully measured acyclic model. For that reason, such violations would plausibly result in instability in the learned model. By focusing on only highly stable edges (see below), we sought to mitigate BOSS's dependence on these assumptions
Acyclicity Acyclicity states that there are no causal feedback loops
Conditional Gaussian Distribution The data distribution follows a conditional Gaussian, where continuous variables are Gaussian, but all parameters can depend on the values of the categorical parent(s) This is more general than in logistic regression models, and is the best available. Relaxing this would require nonparametric statistics, which can not scale to large data sets such as NESARC
Open in a new tab

To determine stable edges (i.e., edges that were robust to changes in the sample composition) and, thereby, minimize the risk of violating some assumptions underlying the BOSS algorithm (Table 1), we ran BOSS on 100 resampled versions of the data (Kummerfeld & Rix, 2019). Informed by Meinshausen and Bühlmann (2010), the resampled datasets were drawn without replacement to be half the size of the original dataset. In addition to showing the full causal model identified for the entire sample, we also show a model restricted to “high stability edges” (HSEs) that directly connected INTD or daily drinking level to any of the 37 alcohol‐related symptoms in at least 80% of the 100 resampled models. Although there is no established benchmark for establishing edge stability, the 80% threshold represents a reasonable balance in mitigating the risk of false positive (accepting causal relationships that are false) and false negative (rejecting causal relationships that are true) discovery errors.

Because they are causally adjacent in the model, HSEs represent causal associations that are not mediated by other variables in this dataset, such as other alcohol symptoms or demographics. However, it is important to keep in mind that graphs showing only HSEs do not imply there are no mediated associations linking INTD or daily alcohol volume to alcohol symptoms, only that the associations represented by HSEs are not mediated by other variables in the model.

RESULTS

Descriptive statistics and bivariate inferential/regression analyses

Table 2 presents demographic/socioeconomic characteristics, number of alcohol‐related symptoms, alcohol intake, family history of alcohol problems, and frequency of current (past 12 months) DSM‐5 diagnoses for individuals with versus without a current INTD.

TABLE 2.

Demographic characteristics of participants with no INTD and current INTD (NESARC 3, N = 24,485).

No INTD (N = 19,908) INTD (N = 4577) Statistic, p‐value
Demographic/socioeconomic category
Age, mean (SD), years 43.7 (16.6)a 41.0 (15.1)b t(24,483) = 10.2, p < 0.001
Gender
Female, no. (%) 10,122 (50.9%)a 3026 (66.1%)b χ 2(1) = 348.4, p < 001
Male, no. (%) 9786 (49.1%)a 1551 (33.9%)b
Ethnicity, no. (%)
Black 4032 (20.3%)a 805 (17.6%)b χ 2(4) = 102.0, p < 001
American Indian/Alaska Native 243 (1.22%)a 100 (2.2%)b
Asian/Native Hawaiian/Other 920 (4.6%)a 129 (2.8%)b
Pacific Islander
Hispanic, any race 3905 (19.6%)a 785 (17.2%)b
White 10,808 (54.3%)a 2758 (60.3%)b
Annual household income, no. (%)
<$5000 to $29,999 7051 (35.4%)a 2099 (45.9%)b χ 2(2) = 209.4, p < 001
$30,000 to $89,999 9009 (45.3%)a 1893 (41.4%)b
$90,000 to 200,000 or more 3848 (19.3%)a 585 (12.8%)b
Education no. (%)
High school or less 7202 (36.2%)a 1752 (38.3%)b χ 2(3) = 58.4, p < 001
Some college 4566 (22.9%)a 1220 (26.7%)b
College grad 5327 (26.8%)a 1064 (23.3%)b
Post grad 2813 (14.1%)a 541 (11.8%)b
No. alcohol‐related symptoms (SD) 1.38 (3.38)a 3.02 (5.64)b t(24,483) = −25.7, p < 0.001
Average daily drinking volume (12 mo, oz) 0.60 (1.28)a 0.74 (1.59)b t(24,483) = −6.1, p < 0.001
Years of drinking at past year level (SD) 10.4 (11.4)a 8.8 (10.1)b t(24,483) = 8.8, p < 0.001
Current DSM‐5 diagnoses (%)
Major depression or dysthymia NA 4066 (89.0%)
Generalized anxiety disorder NA 1334 (29.1%)
Social anxiety disorder NA 672 (15.0%)
Panic disorder NA 800 (17.5%)
Alcohol use disorder 3442 (17.3%)a 1366 (29.8%)b χ 2(2) = 370.9, p < 001
Non‐alcohol related substance use disorder 712 (3.6%)a 540 (11.8%)b χ 2(2) = 516.8, p < 001
Open in a new tab

Note: Each subscript letter denotes a subset of INTD categories whose column proportions do not differ significantly from each other at the 0.05 level.

Continuous variables

Analyses of continuous variables revealed significant differences between individuals with (vs. without) INTD (Table 2). Specifically, the average daily drinking volume in the INTD group was 0.16 ounces higher than in the No INTD group (p < 0.001), despite the INTD group having been drinking at their current level for fewer years. The INTD group was, on average, approximately 3 years younger than the non‐INTD group (p < 0.001), which may explain the shorter duration of consistent drinking observed among individuals with INTD, even as they consume more alcohol on a daily basis. Reflecting findings reported by Anker et al. (2023), the INTD group also had a greater number of alcohol‐related symptoms than the group with no INTD.

Categorical variables

A significantly higher proportion of those with versus without INTD were women. Table 2 also shows that the distribution of various ethnicity categories varied significantly between the INTD groups. Regarding socioeconomic status, significant variations were noted in education levels and income brackets between the groups. These findings indicate that a lower education level and income are associated with a higher incidence of INTD. As expected, individuals with INTD had a significantly higher prevalence of current AUD than those without INTD. Current nonalcohol‐related substance use disorder was also more common in the INTD group than in the No INTD group.

Logistic regression analyses

Main effects

The findings for the logistic regression main effects of INTD on each alcohol symptom in the full model are presented in Table 3, which organizes these effects by the magnitude of the odds ratios (ORs) for INTD as a predictor of each AUD symptom with all covariates, two‐way interactions, and daily average alcohol intake in the model. As shown, individuals with a current INTD had a significantly greater likelihood of exhibiting each of the individual alcohol‐related symptoms, even while controlling for average daily alcohol volume and demographic variables that may contribute independently to an alcohol‐related harm paradox.

TABLE 3.

NESARC 3, N = 24,485: main effect results of INTD status on alcohol‐related symptoms.

Alcohol‐related symptoms OR CI p
NESARC item Abbrev 2.50% 97.5%
1. Have period when drinking interfered with taking care of home or family Drinking disrupts home/family care 4.36 3.27 5.80 ***
2. Continue to drink even though made depressed, uninterested in things, suspicious or distrustful of other people Drinking despite emotional disturbances 4.30 3.48 5.33 ***
3. Want a drink so badly couldn't think of anything else Craving alcohol intensely 4.26 3.33 5.45 ***
4. Have fits or seizures when effects of alcohol were wearing off Seizures during withdrawal 3.56 1.70 7.39 **
5. See, feel, hear things when effects of alcohol were wearing off Hallucinations during withdrawal 3.48 2.52 4.78 ***
6. Continue to drink even though causing problems at school or work Drinking affects education/work 3.34 2.26 4.91 ***
7. Give up or cut down important activities to drink Sacrificing activities for alcohol 3.28 2.44 4.40 ***
8. Feel anxious or nervous when effects of alcohol were wearing off Anxiety during withdrawal 3.02 2.53 3.59 ***
9. Give up or cut down pleasurable activities to drink Foregoing hobbies for drinking 2.99 2.17 4.08 ***
10. Continue to drink even though causing health problem Health problems from drinking persist 2.92 2.42 3.52 ***
11. Feel unusually restless when effects of alcohol were wearing off Restlessness during withdrawal 2.75 2.39 3.16 ***
12. Spend a lot of time being sick or getting over bad aftereffects of drinking Sick/recovering from drinking often 2.67 2.13 3.33 ***
13. Continue to drink despite causing trouble with family or friends Relationship issues due to drinking 2.63 2.14 3.23 ***
14. Sweat or heart beat fast when effects of alcohol were wearing off Withdrawal‐induced sweating/tachycardia 2.53 2.18 2.92 ***
15. More than once try unsuccessfully to stop or cut down on drinking Repeated failed attempts to cut back 2.53 2.06 3.10 ***
16. Ever have job or school troubles because of drinking Job/school problems due to alcohol 2.50 1.69 3.68 ***
17. Have problems with family or friends because of drinking Drinking strains family/friend relationships 2.41 2.04 2.89 ***
18. Shake when effects of alcohol were wearing off Withdrawal tremors 2.36 1.90 2.92 ***
19. Have a vehicle accident while intoxicated Alcohol‐related vehicle accident 2.34 1.65 3.28 ***,†
20. Feel a very strong urge to drink Overpowering need to drink 2.31 2.04 2.62 ***,†
21. Get into physical fight when or right after drinking Physical altercations post‐drinking 2.29 1.90 2.89 ***
22. Drink or use medicine or drug other than aspirin, tylenol or advil to get over bad aftereffects of drinking Using substances for hangover relief 2.29 1.90 2.76 ***
23. Drink or use medicine or drug other than aspirin, tylenol or advil to avoid bad aftereffects of drinking Using substances to prevent hangovers 2.27 1.81 2.83 ***
24. Have period when spent lot of time drinking Extensive time spent drinking 2.24 1.81 2.76 ***
25. Had trouble falling asleep when effects of alcohol were wearing off Sleep disturbances during withdrawal 2.24 1.94 2.57 ***
26. Continue to drink despite prior blackout Persisting after blackout experiences 2.13 1.77 2.55 ***
27. Had to drink more to get the effect wanted Increasing amounts for desired effect 2.11 1.78 2.49 ***
28. Ever increase drinking because amount formerly consumed no longer gave desired effect Escalating consumption for effect 2.01 1.63 2.46 ***
29. In situations that increased chances of getting hurt while drinking or after drinking Risky situations linked to drinking 1.99 1.63 2.41 ***
30. Ever find usual number of drinks had less effect than before Diminished response to usual amount 1.82 1.58 2.10 ***
31. Have period when ended up drinking more than intended Exceeded intended alcohol consumption 1.80 1.60 2.02 ***
32. Ever have period when kept drinking longer than intended Extended drinking sessions 1.80 1.58 2.04 ***
33. More than once want to stop or cut down on drinking Desire to stop/reduce drinking 1.70 1.51 1.89 ***
34. Have nausea when effects of alcohol were wearing off Nausea during withdrawal 1.67 1.49 1.87 ***
35. Drink equivalent of a fifth of liquor in one day Consuming large quantities at once 1.64 1.28 2.09 ***
36. More than once drive a vehicle after having too much to drink Drove drunk 1.50 1.26 1.77 ***
37. More than once drive a vehicle while drinking Drove while drinking 1.50 1.30 1.73 ***
Open in a new tab

Note: p‐Values for multiple comparisons were conducted using FDR. Results are rank‐ordered by OR magnitude. † (dagger) denotes a significant two‐way interaction with average daily drinking volume (12 mo, oz), ***p‐value<0.001.

Abbreviations: CI, 95% confidence interval; est., estimate; OR, odds ratio; p, p‐value; SE, standard error.

The highest INTD odds ratios were observed for alcohol‐related symptoms in domains indicating negative health, withdrawal‐related negative emotions, and living up to various responsibilities. For example, the INTD OR for disruptions to home or family care due to drinking was 4.36 (95% CI: 3.27–5.80, p < 0.001), and drinking despite it resulting in worsening emotional disturbances was 4.30 (95% CI: 3.48–5.33, p < 0.001). INTD was also strongly associated with craving alcohol intensely (OR = 4.26, 95% CI: 3.33–5.45, p < 0.01) and severe withdrawal symptoms like seizures (OR = 3.56, 95% CI: 1.70–7.39, p < 0.01) and hallucinations (OR = 3.48, 95% CI: 2.52–4.78, p < 0.001). Symptoms such as drinking affecting education/work (OR = 3.34, 95% CI: 2.26–4.91, p < 0.001), sacrificing activities for alcohol (OR = 3.28, 95% CI: 2.44–4.40, p < 0.001), and anxiety during withdrawal (OR = 3.02, 95% CI: 2.53–3.59, p < 0.001) also showed strong associations with INTD. Substantial ORs were also noted for symptoms associated with foregoing hobbies for drinking (OR = 2.99, 95% CI: 2.17–4.08, p < 0.001), drinking despite it causing health problems (OR = 2.92, 95% CI: 2.42–3.52, p < 0.001), and restlessness during withdrawal (OR = 2.75, 95% CI: 2.39–3.16, p < 0.001).

Interaction effects

As noted above, models tested both main effects and interactions between each main effect and average daily alcohol volume. Dagger symbols in Table 3 indicate a significant interaction between INTD and the level of alcohol use. The first thing to note is that only two symptoms demonstrated this interaction effect, and these effects, while statistically significant, were objectively small to the point of triviality. The OR was positive in one case (“Overpowering need to drink”; OR = 1.13, 95% CI: 1.05–1.22, p < 0.01) and negative in the other case (“Being involved in an alcohol‐related vehicle accident”; OR = 0.90, 95% CI: 0.82–0.98, p < 0.05); that is, INTD acted as a modest protective factor in that case.

The full model results

While Table 3 shows only effects specific to INTD and its interaction with alcohol volume, we created a heatmap of FDR‐adjusted p‐values for all coefficients in the logistic regression analysis for which there was a significant effect (Table S3). Not surprisingly, the heatmap shows that alcohol use level was positively associated with virtually all alcohol‐related symptoms assessed.

Causal discovery analyses

Figure 1 displays the complete network of edges (solid lines) in the BOSS‐generated causal model using the full dataset. Note that the opacity of the edge lines in the graphic indicates how often that edge appeared in the 100 resampled models (more opaque = higher frequency). Edges that did not appear in the model of the full dataset but did appear in at least one of the resampled models are shown in dotted lines. The two darkly shaded nodes—INTD and daily alcohol volume—are the “target” variables of primary interest.

FIGURE 1.

FIGURE 1

Open in a new tab

Full network of all edges. Depicts every directed edge in the model estimated using the full dataset as well as those identified across 100 resampled models, irrespective of stability threshold. Edge opacity reflects the frequency with which the edge appeared in the resamples; more opaque edges occurred more consistently. Solid black lines represent edges that occurred in the model estimated by BOSS on the full dataset, while dotted lines denote edges that were not retained. The darker nodes (INTD and daily alcohol volume) are the primary “target” variables.

To provide a clearer representation of edge frequency across the 100 resampled models than can be inferred from the edge opacity in Figure 1, Table S4 shows the percentage of times specific edges appeared, along with their causal direction (“cause” vs. “effect”), but only for edges that (1) occurred in the model estimated by BOSS using the full dataset, and (2) were directly connected to one of the two target variables (INTD or daily alcohol volume). The table also highlights edges that meet the 80% stability criterion (i.e., high stability edges [HSEs], indicated by edges above the bolded black line), as further detailed in the Sensitivity Analyses section.

Adjacent (unmediated) high stability edges (HSEs)

Figure 2 depicts all HSEs identified in the 100 resampled models. Nearly all of the edges meeting the HSE stability criterion (present in ≥80% of the resampled models) were present in the whole sample model (solid lines shown in Figure 1); the small number of exceptions are indicated by dotted (vs. solid) lines in Figure 2. Figure 3 further restricts the HSEs shown to just those involving either of the two target variables (INTD and daily alcohol volume). As shown, HSEs emanating from the daily alcohol volume, but not from INTD, occurred for three alcohol‐related symptoms (desire to stop/reduce drinking; drove drunk; drove while drinking) and one demographic variable (education). HSEs emanating from both INTD and daily alcohol volume occurred for two alcohol symptoms (consuming large quantities at once; an overpowering need to drink). HSEs emanating from INTD, but not from daily alcohol volume, occurred for 13 alcohol symptoms, primarily those related to withdrawal or dependence, as outlined just below.

FIGURE 2.

FIGURE 2

Open in a new tab

Network of high stability edges (HSEs). Shows only edges meeting the HSE criterion (i.e., appearing in ≥80% of resampled models). As in Figure 1, edge opacity indicates frequency, and solid black versus dotted lines distinguish edges that do or do not appear in the model estimated by BOSS on the full dataset. The darker nodes highlight the two target variables.

FIGURE 3.

FIGURE 3

Open in a new tab

Target‐focused high stability edges. Displays only those HSEs that connect directly to one of the two target variables (INTD and daily alcohol volume). This filtered view clarifies which alcohol‐related symptoms and demographic factors are most strongly and consistently linked to these key constructs.

Of the 15 total HSEs emanating from INTD (13 linked to INTD only and two linked to both INTD and daily alcohol volume), eight were physical alcohol symptoms, six of which were related to dependence/withdrawal (withdrawal tremors, anxiety during withdrawal, restlessness during withdrawal, withdrawal‐induced sweating/tachycardia, sick/recovering from drinking often, and sleep disturbances during withdrawal) and six were behavioral alcohol symptoms, two of which were related to dependence/withdrawal (using substances for hangover relief and using substances to prevent hangovers). One HSE was found to emanate from INTD to a social/lifestyle‐related alcohol symptom (Drinking strains relationships with family/friends).

Only one variable in the model was found to serve as an HSE exerting unmediated causal influence on INTD or daily alcohol volume; that is, sex exerted a causal influence on both. Notably, INTD and daily alcohol volume were not connected by an HSE. More specifically, an edge from INTD to daily alcohol volume occurred in only four of the 100 resampled models, and an edge from daily alcohol volume to INTD occurred in only one of the 100 resampled models (not shown in Table S4).

Sensitivity analyses

Anxiety versus depression INTD

Past research informed our a priori decision to combine anxiety and depression disorders into a single variable (i.e., INTD). For example, Krueger (1999) showed that these disorders load on a single “internalizing” latent construct (“Negative Emotionality”) and Kushner, Wall, et al. (2012) showed that the shared covariance of anxiety and depression disorders captured by the single “Negative Emotionality” construct predicts AUD risk while the residual variances associated exclusively with a single anxiety or depression disorder do not. To subject this decision to more specific empirical scrutiny, however, we conducted post hoc sensitivity analyses in which we disaggregated INTD into two separate variables: one indicating the presence versus the absence of any one of the three anxiety disorders and the other indicating the presence versus the absence of either one of the two depression disorders. The results of these analyses (Table S5) show that both anxiety and depression disorders remained significant predictors of nearly all alcohol‐related symptoms when modeled together. These findings suggest that both disorder types contribute some unique variance to the harm paradox effect associated with most alcohol symptoms.

Edge stability criteria

As noted earlier, we chose 80% as a reasonable threshold to balance the risk of false‐positive discoveries (accepting false edges) and false‐negative discoveries (rejecting true edges). To aid in considering how the 80% stability threshold affected the findings, Table S4 shows how changes in the threshold level would have affected causal associations identified as HSEs. Note that edges above the bolded black line in the table show those edges exceeding the 80% stability threshold.

Table S4 also demonstrates that the 80% threshold was relatively robust to changes in terms of the study finding; only two additional HSEs would have been accepted in the final model (shown in Figure 3) if we lowered the threshold to 75% (INTD → Escalating consumption for effect and INTD → Risky situations linked to drinking), and only two accepted HSEs would have been rejected if we raised the threshold to 85% (INTD → Extensive time spent drinking, and Daily alcohol volume → Consuming large quantities at once). Importantly, this shows that adjusting the 80% stability threshold plus or minus 5% would not have meaningfully affected the core findings or how we interpret them.

DISCUSSION

Anker et al. (2023) showed that those with INTD manifest an alcohol‐related “harm paradox,” that is, they experience a greater total number of alcohol‐related symptoms than those without INTD who drink at the same level. The present study extends this finding by characterizing (1) the magnitude of the harm‐paradox effect associated with individual alcohol symptoms, (2) the magnitude of the harm‐paradox effect across levels of alcohol use, and (3) the causal relationships between INTD and alcohol symptoms as modeled by the BOSS algorithm.

Regression analyses showed that, after statistically controlling for past year's average daily alcohol volume and demographic variables understood to affect AUD risk (e.g., sex, education, and income), the harm paradox effect was evident for all 37 alcohol symptoms included in the dataset (Table 3). While the magnitude of the harm paradox effect differed between symptoms (ORs ranged from 1.5 to 4.36), the effect within symptoms across levels of daily alcohol volume was fairly stable; that is, only two of 37 alcohol symptoms were significantly predicted by the interaction of daily alcohol volume amount with INTD, and these effects were of a trivial magnitude (ORs = 0.90 and 1.13).

The absence of meaningful interactions between INTD and daily alcohol volume level serves to further characterize the harm paradox effect. To illustrate this, consider the analogy of a foot race in which (a) one runner gets a head start at the beginning of the race and wins, even while running at the same pace as the other runner, versus (b) both runners start at the same place, but one wins by running at a faster pace than the other. It appears that the harm paradox more closely resembles the first scenario, that is, there are differences in the intercept (alcohol symptom risk at the lowest level of alcohol use) but not in the slope (risk for new symptoms developing over increasing levels of alcohol use). This suggests that the processes that account for the harm paradox effect are fully potentiated in response to even low amounts of alcohol use.

Daily average alcohol intake was, not surprisingly, associated with virtually all alcohol symptoms (Table S3). Unlike INTD and alcohol symptoms, however, the vast majority of alcohol symptoms were not causally adjacent to daily alcohol volume in the HSE network (Figure 3). In fact, only three alcohol symptoms received unmediated and stable causal influence from daily alcohol volume but not INTD (desire to stop/reduce drinking, drove drunk, drove while drinking). Here, it is important to keep in mind that regression analyses, unlike the causal model, included one and only one alcohol symptom (the dependent variable). Regression models that would have been more analogous to the causal models would have examined the effect of alcohol use level on each individual alcohol symptom while controlling all other alcohol symptoms and their interactions.

The process(es) underlying the harm paradox is currently unknown. It is notable that INTD exerted a stable unmediated causal influence primarily on alcohol withdrawal and dependence symptoms, which have been associated with Koob's withdrawal‐negative affect stage of addiction (Koob & Le Moal, 2008). Anker and Kushner (2019) reviewed data showing that this stage of addiction shares more brain dysregulation with INTD. For example, anxiety and depression are associated with chronic stress and chronic stress contributes to dysregulated stress‐response system function and increased allostatic load. These stress system dysregulations are associated with neuroadaptations, particularly in the central amygdala, the hypothalamic pituitary adrenal system, and the bed nucleus of the stria terminalis, which are common to both anxiety/mood disorders and alcohol misuse (Agoglia & Herman, 2018; Gilpin et al., 2015; Vinkers et al., 2021; Zabik et al., 2024). Under Koob's neurobiological model of addiction (Koob & Le Moal, 2008), the central amygdala functions as a central hub for integrating signals related to negative affect, stress responsivity, and reward processing. Chronic alcohol use induces neuroadaptations in the central amygdala similar (if not identical) to those triggered by chronic stress. Moreover, evidence from neuroimaging, clinical, and preclinical studies supports the notion that stress‐induced neuroadaptations and increased allostatic load underpin both INTD (Juruena et al., 2020; Vinkers et al., 2021) and AUD (Casement et al., 2015; Hardee et al., 2018). Anker and Kushner (2019) posited that this overlapping neurobiology may cause individuals with INTD to be neurobiologically primed to develop alcohol‐related symptoms related to negative affect and withdrawal. This is consistent with the present findings showing that INTD exerted unmediated causal influence primarily on withdrawal and dependence‐related alcohol symptoms.

However, overlap in the neurobiological underpinnings of INTD and withdrawal/dependence symptoms is not the only possible explanation for the causal modeling results; the experiential similarity between some of these alcohol symptoms and common symptoms of INTD may also be relevant. For example, the experience of common INTD symptoms such as anxiety, restlessness, and sleep disturbance could result in a bias for reporting such symptoms during alcohol withdrawal. Similarly, such INTD symptoms could be suppressed during acute intoxication and then reassert themselves during withdrawal, which, again, could contribute to a reporting bias for some alcohol symptoms. We also considered the relevance of our findings to past studies of the role of environmental factors in the alcohol‐related harm paradox. For example, exposure to childhood trauma or adult adversity—such as illness, poverty, abuse, or other stressors—may be more prevalent among individuals with INTD and could contribute to the propensity to develop alcohol‐related symptoms (Shuai et al., 2022). Although analyses controlled for current socioeconomic factors such as income and education, we did not consider the influence of these factors prior to the last year, nor did we include measures of childhood adversity in our models.

Limitations

The cross‐sectional approach taken did not address some interesting and potentially important questions related to the harm paradox. For example, it would be potentially useful to know if the harm paradox affects individuals who do not have current INTD but are destined to develop it in the future. This would help determine whether the harm paradox emerges as part of an INTD endophenotype or, alternatively, is restricted to individuals who have experienced symptomatic INTD. Similarly, it might be informative to determine whether the harm paradox effect would differ for those whose INTD manifested prior to versus after the onset of particular alcohol symptoms or regular drinking. With that said, however, the order of onset as a means of identifying meaningful comorbidity subgroups has already been extensively studied without clear‐cut implications (Kushner et al., 2000).

Our use of cross‐sectional data also might have provided less rigorous answers to some of the study questions than would have been produced using prospective data. For example, our conclusion that the magnitude of the harm paradox effect does not change appreciably as drinking levels increase based on the absence of meaningful interactions between INTD and alcohol use level might have been more rigorously evaluated by measuring changes in the relevant variables over time in a single cohort. Similarly, causal associations between INTD and specific alcohol symptoms would be potentially more rigorously established in a prospective approach that required temporal priority as a prerequisite to identifying a causal edge. In this regard, it should be noted that imposing a temporal priority requirement was not necessary to achieve our primary goal of determining whether individuals with INTD demonstrate greater risk for specific alcohol‐related symptoms than others who drink at the same level. Nonetheless, it is reasonable to consider what temporal priority might look like under various theoretical causal scenarios. If, for example, the INTD endophenotype is sufficient to cause the alcohol‐related harm paradox, then the temporal relationship between the onsets of INTD and the harm paradox could be unsystematic. This would also be true if INTD served as a catalyst or aggravating factor in the development of alcohol symptoms in response to drinking (e.g., starting to smoke may directly exacerbate symptoms of a pre‐existing asthma condition). Alternatively, if the harm paradox results from experiencing frank symptoms of INTD (e.g., by inducing a reporting bias regarding alcohol symptoms), INTD onset would necessarily precede a manifestation of the harm paradox. These are important issues to address in future research.

Another limitation stems from the possibility that our lumping both anxiety and depression disorder status into a single INTD variable skewed our findings. For example, depression was far and away the most common INTD in the dataset and, thus, could have had undue influence on the effects of INTD status. As noted, we chose to establish the single INTD variable based on evidence of high levels of covariation among the various INTD conditions, with most of this covariance captured by a single underlying latent variable (“Negative Emotionality”; Krueger, 1999). Additionally, there is evidence that increased AUD risk is associated with the shared covariance of INTD disorders (including depression), with little to no additional AUD risk associated with the residual variances that are unique to any one of these disorders (Kushner, Wall, et al., 2012). Nonetheless, we employed sensitivity analyses of the primary regression finding to evaluate the empirical justification for this decision. These showed that both anxiety and depression (while controlling for the other) positively predicted most alcohol‐related symptoms. This suggests that both disorder subtypes contribute some independent predictive variance associated with the harm paradox effect. While this does not change our overall conclusion that INTD is associated with a significant alcohol‐related harm paradox effect, it does show that anxiety disorders and depression disorders contribute to this effect uniquely to some extent.

Another potential limitation relates to our focus on highly stable causal edges (present in at least 80% of the resampled datasets) in the causal modeling. We took this conservative approach to reduce the likelihood of reporting erroneous causal relationships. In the absence of a widely accepted criterion for stability, we chose the threshold of 80% replicability across the resampled datasets. The choice of an 80% threshold was done on an a priori rational basis, that is, before examining the data we chose this threshold as a means of balancing the competing risks for false‐positive and ‐negative discoveries. Suggesting that the 80% stability threshold was robust to modest changes, Table S4 shows that adjusting this threshold plus or minus 5% has little effect on our results and their interpretation.

Our decision to highlight direct (unmediated) causal edges reflected our goal of circumscribing the information within the aims of the study, but could potentially lead to some misunderstandings in the meaning of the results. For example, showing no unmediated causal paths between INTD and daily drinking level, and few edges between alcohol use level and specific alcohol symptoms, Figure 3 could be misinterpreted to mean there are no causal links between these variables. In fact, however, the causal discovery results shown only refer to causal edges remaining after all the mediated causal chains are accounted for. More complex causal paths (albeit difficult to track) are represented in the full causal discovery model shown in Figure 1.

Because we included individuals representing a full range of drinking levels, this was not a formal study of comorbidity (i.e., those with multiple diagnoses) as much as a study of a process that might account for the documented elevated risk for developing AUD among those with INTD. We did not test, for example, whether individuals with both AUD and INTD have more severe alcohol symptoms than those with AUD alone; although this may well be the case. Because of this, our findings will be more applicable to understanding INTD‐related risk for the development of AUD among those with INTD rather than distinguishing those with comorbid AUD‐INTD from those with either condition alone. With this said, we did include some individuals with very high levels of average daily alcohol consumption (over 14 drinks per day) and examined the interaction effect between INTD and drinking level in predicting alcohol symptoms. If it were the case that the harm paradox effect is quite different as alcohol use level advances to the pathological, these interactions should be significant with relatively large effect sizes, which they were not. This suggests that the harm paradox effect influences the risk for alcohol symptoms fairly equally over the full range of alcohol use levels.

It is important to note that the validity of the causal models is potentially affected by the degree to which the assumptions of our approach were met (Table 1). Several aspects of the study contribute to our confidence that these assumptions were met sufficiently to allow for reasonable confidence in the findings. For example, our conservative approach of requiring causal edges to exceed the 80% stability criterion mitigates threats to both the Causal Sufficiency and Acyclicality assumptions; that is, because violations of these assumptions would be expected to result in model instability. Additionally, the risk of cyclicality (e.g., reciprocal causal influence between INTD and alcohol symptoms), which would be more likely at relatively high levels of alcohol use (e.g., Kushner et al., 2000), is minimized by the fact that the majority of cases studied were not drinking at pathological levels. Regarding the Faithfulness assumption, we note that (1) a definitive test of this assumption is not possible; (2) BOSS does not assume full faithfulness as some causal models do; and (3) the Faithfulness assumption is not specific to causal discovery but is relevant to most statistical testing. For example, as noted in Table 1, BOSS is still correct in the presence of some types of Faithfulness violations, and simulations have shown that BOSS still has strong finite sample performance even when Faithfulness violations are allowed (Andrews et al., 2023; Lam et al., 2022). Still, it does remain possible that causal modeling results were affected by parametric model misspecification (e.g., failure to allow for enough interaction or non‐linearity) or the influence on the model of other unmodeled variables (e.g., the existence of developmental windows where the interplay between INTD and alcohol‐related symptoms may be more pronounced).

Finally, it should be noted that the sample analyzed in this study (NESARC 3) was also utilized in our previous study of the harm paradox (i.e., Anker et al., 2023). As such, this study should be understood as an extension rather than as an independent replication of that earlier work.

CONCLUSION

This study was conducted with the aim of better characterizing factors that promote alcohol‐related problems among those with INTD. Not surprisingly, we found that consuming larger quantities of alcohol promotes alcohol‐related symptoms. We also found that those with INTD drink somewhat more than their peers without INTD; however, this effect was relatively small. Given the large (double to triple) increase in risk for AUD among those with INTD, it is reasonable to consider factors beyond alcohol use level that may play a role in this marked increase in risk. Anker et al. (2023) provided evidence that the harm paradox (more symptoms resulting from the same level of drinking) is one such factor. The present study provided important new details concerning this effect including the following: (a) those with INTD are more likely to experience a wide range of specific alcohol‐related symptoms than others who drink at the same level; (b) this effect is present at low levels of drinking and persists through high levels of drinking; and (c) a smaller number of these alcohol symptoms, primarily those related to withdrawal and dependence, receive unmediated causal influence from INTD. We conclude that the harm paradox likely contributes to the elevated risk of developing alcohol use disorder comorbidity among those with INTD.

CONFLICT OF INTEREST STATEMENT

The authors have no conflicts of interest to declare.

Supporting information

Tables S1–S5

ACER-49-1489-s001.docx (5MB, docx)

Anker, J. , Andrews, B. , Kummerfeld, E. , Thuras, P. & Kushner, M.G. (2025) Correlational and causal modeling of alcohol‐related symptoms and internalizing disorder status: Further elucidation of a harm paradox. Alcohol: Clinical and Experimental Research, 49, 1489–1503. Available from: 10.1111/acer.70075

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are openly available in NESARC‐III at https://www.niaaa.nih.gov/research/data‐archive‐resources/niaaa‐controlled‐datasets/nesarc‐iii‐study.

REFERENCES

  1. Agoglia, A.E. & Herman, M.A. (2018) The center of the emotional universe: alcohol, stress, and CRF1 amygdala circuitry. Alcohol, 72, 61–73. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Andrews, B. , Ramsey, J. & Cooper, G.F. (2018) Scoring Bayesian networks of mixed variables. International Journal of Data Science and Analytics, 6, 3–18. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Andrews, B. , Ramsey, J. , Sanchez‐Romero, R. , Camchong, J. & Kummerfeld, E. (2023) Fast scalable and accurate discovery of DAGs using the best order score search and grow‐shrink trees. Advances in Neural Information Processing Systems, 36, 63945–63956. [PMC free article] [PubMed] [Google Scholar]
  4. Anker, J.J. & Kushner, M.G. (2019) Co‐occurring alcohol use disorder and anxiety: bridging psychiatric, psychological, and neurobiological perspectives. Alcohol Research: Current Reviews, 40(1), arcrv40.1.03. Available from: 10.35946/arcr.v40.1.03 [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Anker, J.J. , Thuras, P. , Shuai, R. , Hogarth, L. & Kushner, M.G. (2023) Evidence for an alcohol‐related “harm paradox” in individuals with internalizing disorders: test and replication in two independent community samples. Alcoholism, Clinical and Experimental Research, 47, 713–723. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Benjamini, Y. & Hochberg, Y. (1995) Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society. Series B, Statistical Methodology, 57, 289–300. [Google Scholar]
  7. Casement, M.D. , Shaw, D.S. , Sitnick, S.L. , Musselman, S.C. & Forbes, E.E. (2015) Life stress in adolescence predicts early adult reward‐related brain function and alcohol dependence. Social Cognitive and Affective Neuroscience, 10, 416–423. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Driessen, M. , Meier, S. , Hill, A. , Wetterling, T. , Lange, W. & Junghanns, K. (2001) The course of anxiety, depression and drinking behaviours after completed detoxification in alcoholics with and without comorbid anxiety and depressive disorders. Alcohol and Alcoholism, 36, 249–255. [DOI] [PubMed] [Google Scholar]
  9. Gilpin, N.W. , Herman, M.A. & Roberto, M. (2015) The central amygdala as an integrative hub for anxiety and alcohol use disorders. Biological Psychiatry, 77, 859–869. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Grant, B.F. , Goldstein, R.B. , Saha, T.D. , Chou, S.P. , Jung, J. , Zhang, H. et al. (2015) Epidemiology of DSM‐5 alcohol use disorder: results from the National Epidemiologic Survey on alcohol and related conditions III. JAMA Psychiatry, 72, 757–766. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Grant, B.F. , Stinson, F.S. , Dawson, D.A. , Chou, S.P. , Dufour, M.C. , Compton, W. et al. (2004) Prevalence and co‐occurrence of substance use disorders and independentmood and anxiety disorders: results from the national epidemiologic survey on alcohol and relatedconditions. Archives of General Psychiatry, 61(8), 807–816. [DOI] [PubMed] [Google Scholar]
  12. Hardee, J.E. , Cope, L.M. , Martz, M.E. & Heitzeg, M.M. (2018) Review of neurobiological influences on externalizing and internalizing pathways to alcohol use disorder. Current Behavioral Neuroscience Reports, 5, 249–262. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Juruena, M.F. , Eror, F. , Cleare, A.J. & Young, A.H. (2020) The role of early life stress in HPA axis and anxiety. Advances in Experimental Medicine and Biology, 1191, 141–153. [DOI] [PubMed] [Google Scholar]
  14. Kessler, R.C. , Crum, R.M. , Warner, L.A. , Nelson, C.B. , Schulenberg, J. & Anthony, J.C. (1997) Lifetime co‐occurrence of DSM‐III‐R alcohol abuse and dependence with other psychiatric disorders in the National Comorbidity Survey. Archives of General Psychiatry, 54, 313–321. [DOI] [PubMed] [Google Scholar]
  15. Koob, G.F. & Le Moal, M. (2008) Addiction and the brain antireward system. Annual Review of Psychology, 59, 29–53. [DOI] [PubMed] [Google Scholar]
  16. Krueger, R.F. (1999) The structure of common mental disorders. Archives of General Psychiatry, 56, 921–926. [DOI] [PubMed] [Google Scholar]
  17. Kummerfeld, E. & Rix, A. (2019) Simulations evaluating resampling methods for causal discovery: ensemble performance and calibration. In: 2019 IEEE international conference on bioinformatics and biomedicine (BIBM), pp. 2586–2593. IEEE.
  18. Kushner, M.G. , Abrams, K. & Borchardt, C. (2000) The relationship between anxiety disorders and alcohol use disorders: a review of major perspectives and findings. Clinical Psychology Review, 20, 149–171. [DOI] [PubMed] [Google Scholar]
  19. Kushner, M.G. , Abrams, K. & Thuras, P. (2005) Follow‐up study of anxiety disorder and alcohol dependence in comorbid alcoholism treatment patients. Alcoholism, Clinical and Experimental Research, 29, 1432–1443. [DOI] [PubMed] [Google Scholar]
  20. Kushner, M.G. , Krueger, R. , Frye, B. & Peterson, J. (2008) Epidemiological perspectives on co‐occurring anxiety disorder and substance use disorder. In: Stewart, S.H. & Conrod, P.J. (Eds.) Anxiety and substance use disorders: the vicious cycle of comorbidity. Boston, MA: Springer US, pp. 3–17. [Google Scholar]
  21. Kushner, M.G. , Maurer, E. , Menary, K. & Thuras, P. (2011) Vulnerability to the rapid (“telescoped”) development of alcohol dependence in individuals with anxiety disorder. Journal of Studies on Alcohol and Drugs, 72, 1019–1027. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Kushner, M.G. , Menary, K.R. , Maurer, E.W. & Thuras, P. (2012) Greater elevation in risk for nicotine dependence per pack of cigarettes smoked among those with an anxiety disorder. Journal of Studies on Alcohol and Drugs, 73, 920–924. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Kushner, M.G. , Sher, K.J. & Beitman, B.D. (1990) The relation between alcohol problems and the anxiety disorders. The American Journal of Psychiatry, 147, 685–695. [DOI] [PubMed] [Google Scholar]
  24. Kushner, M.G. , Wall, M.M. , Krueger, R.F. , Sher, K.J. , Maurer, E. , Thuras, P. et al. (2012) Alcohol dependence is related to overall internalizing psychopathology load rather than to particular internalizing disorders: evidence from a national sample. Alcoholism, Clinical and Experimental Research, 36, 325–331. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Lam, W.‐Y. , Andrews, B. & Ramsey, J. (2022) Greedy relaxations of the sparsest permutation algorithm. arXiv [csAI]. Available from: 10.48550/arXiv.2206.05421 [DOI]
  26. Meinshausen, N. & Bühlmann, P. (2010) Stability selection. Journal of the Royal Statistical Society. Series B, Statistical Methodology, 72, 417–473. [Google Scholar]
  27. Menary, K.R. , Kushner, M.G. , Maurer, E. & Thuras, P. (2011) The prevalence and clinical implications of self‐medication among individuals with anxiety disorders. Journal of Anxiety Disorders, 25, 335–339. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. R Core Team . (2021) R: a language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. Available from: https://www.r‐project.org/ [Accessed 28th August 2024]. [Google Scholar]
  29. Shuai, R. , Anker, J.J. , Bravo, A.J. , Kushner, M.G. & Hogarth, L. (2022) Risk pathways contributing to the alcohol harm paradox: socioeconomic deprivation confers susceptibility to alcohol dependence via greater exposure to aversive experience, internalizing symptoms and drinking to Cope. Frontiers in Behavioral Neuroscience, 16, 821693. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Spirtes, P. , Glymour, C. & Scheines, R. (1993) Causation, prediction and search, lecture notes in statistics. Berlin, Germany: Springer. [Google Scholar]
  31. Vinkers, C.H. , Kuzminskaite, E. , Lamers, F. , Giltay, E.J. & Penninx, B.W.J.H. (2021) An integrated approach to understand biological stress system dysregulation across depressive and anxiety disorders. Journal of Affective Disorders, 283, 139–146. [DOI] [PubMed] [Google Scholar]
  32. Wickham, H. & Miller, E. (2019) haven: import and export “SPSS”, “Stata”, and “SAS” files. R package version 2.1.1. Available from: https://CRAN.R‐project/package=haven [Accessed 28th August 2024].
  33. Wickham, H. , François, R. , Henry, L. & Müller, K. (2021) dplyr: a grammar of data manipulation. Version 1.12. Available from: https://dplyr.tidyverse.org [Accessed 28th August 2024].
  34. Zabik, N.L. , Flook, E.A. , Feola, B. , Benningfield, M.M. , Silveri, M.M. , Winder, D.G. et al. (2024) Bed nucleus of the stria terminalis network responses to unpredictable threat in early alcohol abstinence. Alcoholism, Clinical and Experimental Research, 48, 1716–1727. [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

Tables S1–S5

ACER-49-1489-s001.docx (5MB, docx)

Data Availability Statement

The data that support the findings of this study are openly available in NESARC‐III at https://www.niaaa.nih.gov/research/data‐archive‐resources/niaaa‐controlled‐datasets/nesarc‐iii‐study.

Articles from Alcohol, Clinical & Experimental Research are provided here courtesy of Wiley

RESOURCES