Assessing Measurement Bias in Substance Use Disorder Criteria Associated with Childhood Adversity and Genetic Liability

Abstract

Objective:

The diagnosis and severity of substance use disorders (SUDs) are classified by the number of criteria endorsed. However, environmental and genetic factors may influence criterion endorsement. To evaluate this, we tested for differential item functioning (DIF) of SUD criteria by adverse childhood events (ACEs) and SUD polygenic scores (PGS).

Method:

In 10,275 Yale-Penn participants (M_age = 40.59 years, 56.2% male, 47.21% of African-like genetic ancestry [AFR], 52.79% of European-like genetic ancestry [EUR]), we used item response theory models to estimate difficulty and discrimination parameters for each criterion for alcohol, cannabis, and opioid use disorders. We then tested whether these properties varied based on scores on a latent ACEs factor and ancestry-specific PGS using moderated nonlinear factor analyses.

Results:

There was variability in the difficulty and discrimination of SUD criteria. Many criteria discriminated less effectively among individuals with higher ACEs factor scores. Continued substance use despite physical/psychological problems ($\beta$ = 0.08, SE = 0.01, p < 0.0001) and withdrawal ($\beta$ = 0.07, SE = 0.02, p = 0.001) were more difficult to endorse for EUR individuals with high ACEs scores than those with lower ACEs scores. No DIF was identified by PGS.

Conclusions:

Findings highlight the impact of ACEs on SUD assessment. Considering the relative weighting of criteria or developing screening procedures that consider subgroup-specific differences in symptom functioning may help address these biases.

Introduction

Substance use disorders (SUDs) are prevalent and complex conditions characterized by a persistent pattern of harmful substance use that disrupts an individual’s life. These disorders affect millions of people annually, with the most recent national estimate being that 17.1% of the US population aged 12 or older had a past-year SUD (Substance Abuse and Mental Health Services Administration, 2024). SUD diagnoses rely on the endorsement of at least two out of eleven criteria (e.g., tolerance, failure to quit/control use) within a 12-month period, as described in the fifth edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5) (American Psychiatric Association, 2013). Each diagnostic criterion reflects a different dimension of substance use and its impact, ranging from the physiological (e.g., withdrawal) to the behavioral (e.g., using substance in larger amounts/longer than intended). However, the current diagnostic framework treats all criteria as equally indicative of SUD severity across all groups, a uniformity that may oversimplify the complexity of these disorders.

The current approach to determining SUD severity, which is based on a simple count of criteria endorsed, does not accurately achieve that goal. For example, although studies have shown that SUD criteria have a unidimensional structure (Gillespie, Neale, Prescott, Aggen, & Kendler, 2007), studies on alcohol use disorder (AUD) show that certain criteria, like giving up activities and failing to fulfill role obligations, reflect greater AUD severity than others like drinking for longer or in larger amounts than intended and experiencing craving (Saha, Chou, & Grant, 2020). Furthermore, certain AUD criteria, such as withdrawal, better predict transitions to a more severe disorder presentation after accounting for differences in the number of criteria endorsed (Miller et al., 2023). Similar findings have been observed for other SUDs, including opioid use disorder (OUD) (Kopak & Hoffmann, 2024) and cannabis use disorder (CUD) (Compton, Saha, Conway, & Grant, 2009).

Although these differences in criterion informativeness are well-documented, much less is known about whether the functioning of SUD criteria varies systematically across individuals with different environmental or genetic risk profiles. A significant environmental factor related to SUDs is exposure to adverse childhood events (ACEs), which include various forms of abuse, neglect, and household dysfunction. ACEs are associated with a range of negative mental health outcomes, including an increased risk of developing SUDs (LeTendre & Reed, 2017). This increase in risk may stem from difficulties with emotion regulation and a tendency to use substances to manage emotional distress (Kranzler et al., 2024). Individuals with greater exposure to ACEs and fewer buffering factors (e.g., religiosity and family support) may endorse certain SUD criteria more frequently, not because of a higher underlying severity of SUD, but because their substance use behaviors are more closely tied to use of maladaptive coping strategies and poorer self-regulation (Sebalo, Königová, Sebalo Vňuková, Anders, & Ptáček, 2023). Thus, it is plausible that individuals with higher ACEs scores may be more likely to endorse criteria like cravings or continued use despite social or interpersonal problems, even at similar levels of the latent SUD factor.

Similarly, genetic liability, as indicated by polygenic scores (PGS), affects an individual’s likelihood of developing an SUD and may shape how the disorder manifests. PGS provide a measure of genetic risk by aggregating the effects of many genetic variants associated with SUDs. Individuals with higher PGS for SUDs have a greater susceptibility to these disorders (SooHoo et al., 2025), which may affect how easily they endorse certain DSM-5 criteria. For example, genetic liability has been linked to differences in alcohol sensitivity (Schuckit et al., 2012), which is associated with heavier consumption. Conceptually, this could increase the likelihood of endorsing criteria like tolerance and withdrawal, even at similar levels of the underlying SUD factor. Genetic liability could also contribute to greater difficulty in reducing or controlling substance use, which may be reflected in more frequent endorsement of this criterion.

ACEs and PGS might impact how SUD criteria function. Two ways of understanding how criteria function—difficulty and discrimination—can be assessed using item response theory (IRT). Difficulty reflects how severe an individual’s underlying SUD must be before they endorse a particular criterion, while discrimination is the ability of a criterion to distinguish between individuals at different levels of the underlying disorder. For example, individuals with high genetic liability for AUD may be less sensitive to the intoxicating effects of alcohol (Schuckit et al., 2012), which can lead them to consume larger quantities to achieve the desired effects. This pattern of heavier drinking may result in the earlier development of tolerance and withdrawal, reflecting lower item difficulty for these criteria among individuals with higher PGS. If these criteria are frequently endorsed regardless of one’s level on the SUD latent factor, they may also provide less information about where individuals fall on the SUD continuum, thereby reducing discrimination. Such differential endorsement patterns indicate the presence of differential item functioning (DIF).

DIF can arise due to differences in discrimination, where a criterion varies in its ability to differentiate between severity levels across subgroups, or differences in difficulty, where a criterion is more or less likely to be endorsed by certain subgroups at the same severity level. Recognizing both types of DIF is crucial because they impact the accuracy of SUD assessments. If diagnostic criteria function differently across subgroups defined by genetic risk or environmental exposures, such as ACEs, systematic biases in diagnosis and treatment may result. Identifying and addressing DIF ensures that diagnostic criteria reflect true differences in severity rather than artifacts of subgroup-specific response patterns, improving the validity and fairness of SUD assessments.

Current Study

Despite evidence that both ACEs and genetic liability shape SUD risk (Kranzler et al., 2024), no study has systematically examined whether DSM-5 SUD criteria function equivalently across these dimensions. To address these gaps, we used moderated nonlinear factor analysis (MNLFA) to test for DIF (Bauer, 2017) related to SUD PGS and exposure to ACEs. Specifically, MNLFA is a latent variable approach that can evaluate whether differences in criterion endorsement reflect true differences in the latent SUD factor or measurement bias across levels of ACEs or PGS. We hypothesized that individuals with higher ACEs scores would be more likely to endorse criteria related to coping-motivated substance use (e.g., craving and continued use despite social/interpersonal problems) at comparable levels of the latent SUD factor, indicating uniform bias (i.e., differences in item difficulty or intercept bias). We also hypothesized that individuals with higher SUD PGS would be more likely to endorse physiological criteria (e.g., tolerance and withdrawal) at lower levels of severity (i.e., uniform bias), and that these criteria would exhibit weaker discrimination at higher PGS levels (i.e., nonuniform bias, also referred to as differences in discrimination or slope bias). Additionally, by evaluating patterns of DIF across alcohol, cannabis, and opioid use disorders, we were able to examine whether these effects were substance-specific or generalized across SUDs. Thus, the current study provides a comprehensive evaluation of how SUD criteria function across PGS and ACEs, offering insights into the validity of current diagnostic practices.

Methods

Transparency and Openness

We report how we determined our sample size, all data exclusions (if any), all manipulations, and all measures in the study. All analyses were conducted using Mplus v 8.11. Analysis code and materials are available by emailing the corresponding author. The study’s design and its analyses were not pre-registered.

Participants

Participants were 10,275 individuals with genotype data from the Yale-Penn cohort, a case-control sample ascertained for genetic studies of SUDs (Gelernter, Kranzler, Sherva, Almasy, et al., 2014; Gelernter, Sherva, et al., 2014). Recruitment occurred at Yale University, UConn Health, the University of Pennsylvania, the Medical University of South Carolina, and McLean Hospital, with participants coming from addiction treatment centers and psychiatric services and through community advertisements. The University of Pennsylvania’s Institutional Review Board (IRB) approved the study (#804787 and #812856). All recruitment sites received local IRB approval as well. Given the small sizes of other genetically inferred ancestry groups, we included only individuals who were genetically similar to the 1000 Genomes African (AFR; 42.71%) and European (EUR; 52.79%) reference panels. Table 1 provides additional sample details.

Table 1.

Sample characteristics.

Phenotype	% (N) or Mean (SD)
Age	40.59 (11.72)
Sex
Male	56.20% (5775)
Female	43.80% (4500)
Genetically inferred ancestry
AFR	47.21% (4851)
EUR	52.79% (5424)
Years of Education	12.74 (2.37)
Employed	56.33% (3985)
Marital Status
Married	17.87% (1836)
Widowed	2.03% (209)
Separated	5.40% (555)
Divorced	17.80% (1829)
Never Married	56.90% (5846)
Past month alcohol use	59.06% (6067)
Past month cannabis use	22.49% (2311)
Past month opioid use	13.47% (1384)
Adverse childhood experiences factor
Number of caregivers	1.29 (1.28)
Number of moves	2.25 (2.39)
Witnessed violent crime	15.96% (1640)
Experienced sexual abuse	13.68% (1406)
Experienced physical abuse	8.08% (830)
Household substance use	48.42% (4975)
Household smoking	64.77% (6655)
Lack of religious participation	10.08% (1036)
Poor caregiver relationship	5.52% (567)
Infrequent contact with relatives	20.59% (2116)

Note: SD = standard deviation, AFR = genetically similar to 1000 Genomes African reference panels, EUR = genetically similar to 1000 Genomes European reference panels.

Measures

Substance Use Disorders.

SUD criteria were assessed using the Semi-Structured Assessment for Drug Dependence and Alcoholism (SSADDA), a diagnostic interview derived from the Semi-Structured Assessment for Genetics of Alcoholism (Pierucci-Lagha et al., 2005). The SSADDA has demonstrated reliability for assessing SUDs (Pierucci-Lagha et al., 2007; Pierucci-Lagha et al., 2005). For the current study, we examined lifetime criteria for alcohol (N_AFR = 4,283; N_EUR = 4,969) (Gelernter, Kranzler, Sherva, Almasy, et al., 2014), cannabis (N_AFR = 3,515; N_EUR = 4,113) (Sherva et al., 2016), and opioid (N_AFR = 1,628; N_EUR = 2,810) (Gelernter, Kranzler, Sherva, Koesterer, et al., 2014) use disorders among individuals who reported using each substance.

Adverse Childhood Experiences.

The SSADDA also assesses demographic information and environmental factors, including ACEs and related protective factors. ACEs were assessed using ten indicators: two reflecting an unstable home life (i.e., three or more main caregivers and two or more relocations), three measuring traumatic experiences (i.e., violent crime, sexual abuse, and physical abuse), two measuring substance use within the household (i.e., alcohol/drug use and smoking), and three measuring the absence of protective factors (i.e., no religious participation, poor caregiver relationship, and infrequent contact with relatives). An ACEs factor was derived from these ten items using confirmatory factor analyses (RMSEA = 0.03, CFI = 0.96, SRMR = 0.05) (Kranzler et al., 2024).

Genotyping and Imputation.

Participants provided blood or saliva samples for DNA extraction and genotyping. Genotyping was performed using the Illumina HumanCoreExome array, Illumina HumanOmni1-Quad microarray, or the Illumina Multi-Ethnic Global array. Imputation was conducted using the Michigan Imputation Server, with reference data from the 1000 Genomes Project Phase 3 (Auton et al., 2015). Individuals with genotype call rates < 98% and single-nucleotide polymorphisms (SNPs) with call rates < 98%, minor allele frequencies < 1%, imputation quality < 0.7, or Hardy-Weinberg equilibrium p < 1×10⁻⁷ were removed. The procedure for determining genetic ancestry has been described previously (Zhou et al., 2020). Briefly, using SNPs common to both Yale-Penn and the 1000 Genomes Project Phase 3 reference panels, we calculated principal components (PCs) and inferred genetic ancestry based on the distance of 10 PCs for each participant to the reference population samples.

Polygenic Scores.

In the Yale-Penn sample, we calculated PGS for AUD (Zhou et al., 2023), CUD (Levey et al., 2023), and OUD (Kember et al., 2022) using summary statistics from large-scale genome-wide association studies (GWAS). For GWAS that included Yale-Penn, we obtained summary statistics that excluded Yale-Penn to ensure independence of the target sample. Within-ancestry PGS were calculated using PRS-cs software (Ge, Chen, Ni, Feng, & Smoller, 2019), a Bayesian method that leverages continuous shrinkage priors to estimate posterior SNP effect sizes while accounting for linkage disequilibrium (LD) structure. We used ancestry-specific reference panels from the 1000 Genomes Project Phase 3 to generate within-ancestry PGS, with the “auto” shrinkage option applied in PRS-cs to optimize effect size estimation. We fixed the random seed to one to ensure reproducibility. PGS were standardized within each ancestry group to have a mean of zero and a standard deviation of one.

Data Analysis

All analyses were run within each ancestry group to account for population differences in genetic structure. First, we estimated each item’s difficulty and discrimination using 2-parameter logistic (2PL) IRT models. Difficulty refers to how severe the underlying SUD must be for an individual to endorse the criterion, while discrimination refers to how well the criterion differentiates individuals with varying levels of SUD. The difficulty and discrimination parameters can be used to directly compare the criteria within ancestry groups. To highlight the differences in diagnostic informativeness, we summarize in the results the criteria that exhibited the highest and lowest discrimination and difficulty parameters for each substance and genetically inferred ancestry group.

We tested for DIF using MNLFA (Bauer, 2017), which evaluated for uniform and nonuniform bias based on PGS/ACEs scores. Compared to traditional IRT DIF analyses, MNLFA offers several strengths, including the ability to model multiple continuous moderators simultaneously and improved integration with latent variable modeling (Bauer, 2017). MNLFA is particularly useful in this context because it can evaluate whether criterion differences across ACEs and PGS levels reflect measurement bias or true differences in the underlying SUD factor. Uniform bias corresponds to DIF of the difficulty parameter, which occurs if an item is more or less likely to be endorsed by some individuals at the same underlying level of the SUD latent factor. Negative uniform bias indicates that the item is easier to endorse as ACEs or PGS increase, representing a shift to the left in the item characteristic curve (ICC). Nonuniform bias corresponds to DIF of the discrimination parameter, which occurs when the ability of an item to differentiate severity levels is different across values of the moderating variable (i.e., ACEs and/or PGS). Negative nonuniform DIF indicates reduced discriminative ability as PGS or ACEs increase. In terms of the ICC, this would be reflected by a flatter slope. In all models, we allowed the mean and variance of the latent SUD factor to vary as a function of ACEs and PGS by regressing the factor on these covariates and including factor-covariate interaction terms. This enabled us to separate true item-level DIF from factor differences. Prior confirmatory factor analyses in this sample demonstrated excellent fit for unidimensional SUD factors (SooHoo et al., 2025), supporting the unidimensionality and local independence assumptions of MNLFA. To account for multiple comparisons, we used a Benjamini-Hochberg false discovery rate (FDR) correction within ancestries and substances to evaluate significance. All models included age, sex, and the first ten ancestry principal components as covariates. We did not include socioeconomic status or education as covariates, because these variables are likely to be downstream correlates of ACEs and PGS.

Results

Alcohol Use Disorder

Item Response Theory.

For AUD (Table 2), the most easily endorsed criteria were: using in larger amounts/longer than intended, failure to quit/control use, and continued use despite social problems, while craving, withdrawal, and continued use despite physical problems were the most difficult to endorse. The most discriminating (i.e., particularly effective at distinguishing AUD severity) criteria across ancestry groups were role failure, continued use despite social problems, and time spent obtaining/using alcohol. In contrast, tolerance and hazardous use had the lowest discrimination values.

Table 2.

Item difficulties and discrimination for substance use disorder criteria.

Difficulty parameter (Rank Order)
Criterion	Alcohol Use Disorder			Cannabis Use Disorder		Opioid Use Disorder
	AFR		EUR	AFR	EUR	AFR	EUR
Larger/longer	−0.65 (11)	−0.50 (11)		0.52 (10)	0.60 (9)	1.24 (7)	0.49 (7)
Failure to quit/control	−0.55 (10)	−0.14 (7)		0.55 (9)	0.78 (6)	1.10 (11)	0.42 (10)
Time spent	0.32 (3)	0.33 (3)		0.42 (11)	0.46 (10)	1.24 (7)	0.51 (6)
Craving	0.53 (1)	0.54 (1)		0.89 (4)	0.95 (3)	1.37 (3)	0.57 (4)
Role failure	0.13 (5)	0.09 (6)		1.07 (1)	1.00 (1)	1.43 (2)	0.68 (1)
Social problems	−0.31 (9)	−0.27 (9)		0.79 (8)	0.73 (7)	1.27 (6)	0.54 (5)
Activities given up	0.12 (6)	0.19 (5)		0.88 (5)	0.82 (5)	1.34 (4)	0.61 (3)
Hazardous use	−0.12 (8)	−0.44 (10)		0.93 (3)	0.43 (11)	1.59 (1)	0.62 (2)
Physical/psychological problems	0.24 (4)	0.28 (4)		0.98 (2)	0.99 (2)	1.23 (9)	0.47 (8)
Tolerance	−0.10 (7)	−0.15 (8)		0.80 (7)	0.61 (8)	1.29 (5)	0.46 (9)
Withdrawal	0.47 (2)	0.37 (2)		0.85 (6)	0.83 (4)	1.15 (10)	0.40 (11)
Discrimination parameter (Rank Order)
Larger/longer	3.40 (4)	3.78 (6)		3.20 (1)	2.96 (4)	6.64 (7)	6.15 (10)
Failure to quit/control	2.00 (9)	2.85 (9)		1.71 (10)	2.09 (10)	6.16 (10)	7.16 (7)
Time spent	3.49 (3)	4.28 (4)		2.89 (4)	3.08 (3)	6.76 (4)	8.67 (1)
Craving	2.73 (8)	3.35 (8)		2.48 (8)	2.75 (7)	7.77 (1)	8.67 (1)
Role failure	4.17 (1)	5.21 (1)		2.90 (3)	3.17 (2)	6.72 (5)	7.77 (4)
Social problems	3.68 (2)	4.91 (2)		2.32 (9)	2.43 (9)	6.72 (5)	7.52 (5)
Activities given up	3.04 (5)	4.50 (3)		3.09 (2)	3.22 (1)	6.82 (3)	7.22 (6)
Hazardous use	2.00 (9)	2.80 (10)		1.63 (11)	1.90 (11)	3.09 (11)	3.84 (11)
Physical/psychological problems	2.87 (7)	3.58 (7)		2.77 (6)	2.77 (6)	6.40 (8)	6.99 (8)
Tolerance	1.93 (11)	2.25 (11)		2.58 (7)	2.62 (8)	6.31 (9)	6.46 (9)
Withdrawal	3.01 (6)	3.90 (5)		2.84 (5)	2.87 (5)	7.37 (2)	8.38 (3)

Note: AFR = genetically similar to 1000 Genomes African reference panels, EUR = genetically similar to 1000 Genomes European reference panels. Rank orders are provided in parentheses, with the most difficult/discriminating item being ranked 1 and the least being ranked 11.

Differential Item Functioning.

In AFR individuals, only one criterion—activities given up—showed uniform bias (β₀ = 0.08, SE = 0.02, p_FDR = 0.002), indicating that it was harder to endorse for individuals with higher ACEs scores (Table 3). In EUR individuals, two criteria—continued use despite physical problems (β₀ = 0.08, SE = 0.01, p_FDR = 0.0001) and withdrawal (β₀ = 0.07, SE = 0.02, p_FDR = 0.0006)—showed uniform bias by ACEs. In AFR individuals, tolerance (β₁ = −0.09, SE = 0.02, p_FDR = 0.002), failure to quit/control use (β₁ = −0.10, SE = 0.02, p_FDR = 0.0004), and activities given up (β₁ = −0.12, SE = 0.02, p_FDR = 0.0004) exhibited nonuniform bias, meaning that the criteria discriminated less effectively among individuals with higher ACEs scores. In EUR individuals, all criteria except three—larger/longer, time spent, and withdrawal—showed nonuniform bias by ACEs in the same direction as observed in AFR individuals. No DIF by PGS was observed for AUD (Supplementary Tables 1 and 2).

Table 3.

Differential item functioning by adverse childhood experiences for alcohol use disorder.

		AFR Individuals (n = 4,283)				EUR Individuals (n = 4,969)
Criterion	DIF Type		β	SE	p-value *	β	SE	p-value *
Larger/longer	Uniform		−0.01	0.02	0.72	−0.03	0.02	0.20
	Nonuniform		−0.09	0.03	0.07	−0.03	0.02	0.40
Failure to quit/control	Uniform		0.00	0.02	0.96	−0.03	0.02	0.20
	Nonuniform		−0.10	0.02	0.0004	−0.09	0.02	0.001
Time spent	Uniform		−0.02	0.03	0.72	−0.01	0.02	0.74
	Nonuniform		−0.03	0.03	0.72	−0.05	0.03	0.40
Craving	Uniform		0.00	0.03	0.96	0.03	0.02	0.40
	Nonuniform		−0.04	0.03	0.51	−0.11	0.02	0.001
Role failure	Uniform		0.02	0.03	0.74	0.01	0.01	0.74
	Nonuniform		−0.04	0.04	0.72	−0.14	0.03	<0.0001
Social problems	Uniform		0.02	0.02	0.72	0.01	0.01	0.74
	Nonuniform		0.04	0.04	0.72	−0.09	0.04	0.048
Activities given up	Uniform		0.08	0.02	0.002	−0.01	0.02	0.74
	Nonuniform		−0.12	0.02	0.0004	−0.10	0.03	0.01
Hazardous use	Uniform		0.04	0.02	0.24	0.04	0.02	0.07
	Nonuniform		−0.05	0.03	0.30	−0.09	0.03	0.001
Physical/psychological problems	Uniform		0.05	0.02	0.26	0.08	0.01	<0.0001
	Nonuniform		0.02	0.04	0.85	−0.08	0.03	0.02
Tolerance	Uniform		0.00	0.02	0.96	−0.01	0.02	0.74
	Nonuniform		−0.09	0.02	0.002	−0.09	0.02	0.001
Withdrawal	Uniform		0.05	0.03	0.30	0.07	0.02	0.001
	Nonuniform		−0.03	0.04	0.72	−0.03	0.04	0.69

Note: Uniform bias refers to differential item functioning in the difficulty parameter, and nonuniform bias to differential item functioning in the discrimination parameter.

^*indicates false discovery rate (FDR) adjusted p-value. Bold indicates significance at p_FDR < 0.05. AFR = genetically similar to 1000 Genomes African reference panels, EUR = genetically similar to 1000 Genomes European reference panels

Cannabis Use Disorder

Item Response Theory.

Across ancestry groups (Table 2), larger/longer was among the most easily endorsed criteria, while craving and withdrawal were among the most difficult. The largest discrepancy across groups was seen for hazardous use, which was the easiest criterion to endorse among EUR individuals, but one of the most difficult to endorse in AFR individuals. Across groups, hazardous use and failure to quit were the least discriminating criteria.

Differential Item Functioning.

In AFR individuals, continued use despite physical problems exhibited uniform bias (β₀ = 0.08, SE = 0.02, p_FDR = 0.004), with the criterion being harder to endorse for those with higher ACEs scores. For EUR individuals, time spent exhibited uniform bias in the opposite direction (β₀ = −0.07, SE = 0.02, p_FDR = 0.002), meaning that at the same level of CUD severity, individuals with higher ACEs scores were more likely to endorse the criterion (Table 4). Similarly, role failure (β₀ = −0.05, SE = 0.02, p_FDR = 0.02) was more likely to be endorsed by those with higher ACEs scores at the same level of CUD severity. In AFR individuals, only one criterion—time spent—exhibited nonuniform bias (β₁ = −0.08, SE = 0.03, p_FDR = 0.04), discriminating less effectively among individuals with higher ACEs scores. For EUR individuals, four criteria—time spent, tolerance, hazardous use, and craving—all exhibited nonuniform bias and had less discriminative ability among individuals with higher ACEs scores. There was no uniform or nonuniform bias by PGS (Supplementary Tables 1 and 2).

Table 4.

Differential item functioning by adverse childhood experiences for cannabis use disorder.

		AFR Individuals (n = 3,515)				EUR Individuals (n = 4,113)
Criterion	DIF Type		β	SE	p-value *	β	SE	p-value *
Larger/longer	Uniform		−0.03	0.02	0.21	−0.01	0.02	0.83
	Nonuniform		−0.07	0.03	0.15	−0.02	0.03	0.90
Failure to quit/control	Uniform		−0.04	0.02	0.24	−0.01	0.02	0.90
	Nonuniform		−0.04	0.02	0.32	−0.04	0.03	0.58
Time spent	Uniform		−0.01	0.02	1.00	−0.07	0.02	0.002
	Nonuniform		−0.08	0.03	0.04	−0.10	0.03	0.003
Craving	Uniform		0.03	0.02	0.37	−0.03	0.02	0.22
	Nonuniform		−0.08	0.03	0.06	−0.09	0.03	0.02
Role failure	Uniform		−0.03	0.02	0.58	−0.05	0.02	0.02
	Nonuniform		0.00	0.04	1.00	−0.06	0.03	0.23
Social problems	Uniform		−0.01	0.02	1.00	0.00	0.02	0.99
	Nonuniform		−0.04	0.03	0.52	−0.06	0.03	0.11
Activities given up	Uniform		−0.03	0.02	0.36	−0.03	0.02	0.39
	Nonuniform		−0.02	0.04	1.00	−0.02	0.04	0.83
Hazardous use	Uniform		0.03	0.02	0.36	−0.02	0.02	0.63
	Nonuniform		−0.06	0.02	0.12	−0.10	0.02	0.002
Physical/psychological problems	Uniform		0.08	0.02	0.004	0.00	0.02	0.99
	Nonuniform		−0.03	0.03	0.73	−0.08	0.03	0.06
Tolerance	Uniform		−0.01	0.02	0.85	−0.03	0.02	0.45
	Nonuniform		−0.08	0.03	0.06	−0.09	0.03	0.003
Withdrawal	Uniform		0.01	0.02	0.92	0.01	0.02	0.83
	Nonuniform		−0.05	0.03	0.32	−0.02	0.03	0.79

Note: Uniform bias refers to differential item functioning in the difficulty parameter, and nonuniform bias to differential item functioning in the discrimination parameter.

^*indicates false discovery rate (FDR) adjusted p-value. Bold indicates significance at p_FDR < 0.05. AFR = genetically similar to 1000 Genomes African reference panels, EUR = genetically similar to 1000 Genomes European reference panels

Opioid Use Disorder

Item Response Theory.

In AFR individuals, failure to quit/control use was the easiest criterion to endorse, while withdrawal was the easiest among EUR individuals. The most difficult criterion to endorse was hazardous use in AFR individuals and role failure in EUR individuals. The criterion with the greatest discriminative ability in both ancestry groups was craving, while hazardous use showed the least discriminative ability.

Differential Item Functioning.

In AFR individuals, there was no DIF for OUD criteria (Table 5). In EUR individuals, larger/longer exhibited uniform bias by ACEs (β₀ = 0.04, SE = 0.01, p_FDR = 0.01), meaning that it was harder to endorse among individuals with higher ACEs scores. Four criteria—tolerance, activities given up, hazardous use, and craving—had nonuniform bias by ACEs, with the items discriminating less well among individuals with higher ACEs scores. There was no DIF by PGS (Supplementary Tables 1 and 2).

Table 5.

Differential item functioning by adverse childhood experiences for opioid use disorder.

		AFR Individuals (n = 1,628)				EUR Individuals (n = 2,810)
Criterion	DIF Type		β	SE	p-value *	β	SE	p-value *
Larger/longer	Uniform		0.03	0.04	0.86	0.04	0.01	0.01
	Nonuniform		−0.03	0.04	0.94	−0.05	0.03	0.25
Failure to quit/control	Uniform		0.05	0.03	0.56	0.01	0.01	0.56
	Nonuniform		−0.10	0.03	0.09	−0.06	0.04	0.21
Time spent	Uniform		−0.01	0.04	0.96	0.00	0.01	0.86
	Nonuniform		0.00	0.05	0.96	−0.08	0.03	0.06
Craving	Uniform		−0.01	0.05	0.96	0.00	0.01	0.83
	Nonuniform		−0.02	0.04	0.96	−0.09	0.03	0.01
Role failure	Uniform		0.06	0.04	0.61	0.04	0.02	0.05
	Nonuniform		−0.04	0.04	0.86	−0.08	0.03	0.05
Social problems	Uniform		−0.06	0.06	0.86	0.01	0.01	0.49
	Nonuniform		0.10	0.07	0.61	−0.09	0.03	0.05
Activities given up	Uniform		−0.05	0.05	0.86	0.03	0.01	0.10
	Nonuniform		0.11	0.06	0.43	−0.10	0.03	0.02
Hazardous use	Uniform		0.05	0.04	0.82	0.04	0.01	0.05
	Nonuniform		−0.03	0.04	0.86	−0.13	0.02	<0.0001
Physical/psychological problems	Uniform		0.00	0.05	0.96	0.03	0.01	0.06
	Nonuniform		0.02	0.05	0.96	−0.04	0.04	0.56
Tolerance	Uniform		0.07	0.03	0.43	0.00	0.01	0.87
	Nonuniform		−0.08	0.04	0.40	−0.13	0.03	0.0002
Withdrawal	Uniform		−0.01	0.04	0.96	0.01	0.01	0.56
	Nonuniform		−0.01	0.05	0.96	−0.04	0.04	0.56

Note: Uniform bias refers to differential item functioning in the difficulty parameter, and nonuniform bias to differential item functioning in the discrimination parameter.

^*indicates false discovery rate (FDR) adjusted p-value. Bold indicates significance at p_FDR < 0.05. AFR = genetically similar to 1000 Genomes African reference panels, EUR = genetically similar to 1000 Genomes European reference panels

Discussion

Consistent with prior work (Compton et al., 2009; Kopak & Hoffmann, 2024; Saha et al., 2020), we found that diagnostic criteria varied in their difficulty and discrimination. Higher difficulty indicates that a criterion reflects a more severe underlying SUD, because it is endorsed more commonly at higher levels of the latent trait. In contrast, discrimination reflects how well a criterion differentiates among individuals at similar levels of underlying SUD severity. For example, using more than intended (i.e., larger/longer) was among the easiest criteria to endorse for AUD and CUD, consistent with prior research on AUD (Kervran et al., 2020). In contrast, failure to fulfill obligations (i.e., role failure) was among the most difficult criteria to endorse for CUD and OUD, indicating that functional impairment may suggest the presence of a more severe disorder. This is consistent with a study of individuals from a French outpatient clinic, in which role failure was the most difficult OUD criterion to endorse (Kervran et al., 2020). Together, these findings suggest the need to reconsider the assumption that DSM-5 SUD criteria should be weighted equally, as some may signal a more severe disorder.

Prior work has evaluated alternatives to simple symptom counts by grading individual criteria based on item-level severity. For example, Boness and colleagues (2019) used IRT-derived severity rankings of AUD indicators to construct diagnostic severity scores, but these did not consistently outperform the DSM-5 approach across external validators. Thus, weighting criteria by severity is a promising direction, but current implementations still require refinement before they can be adopted in clinical practice or research.

Some criteria were also consistently better at discriminating among levels of the latent SUD factor, including role failure and craving. Other studies support craving as being among the most discriminating of all SUD criteria (Gilder, Gizer, Lau, & Ehlers, 2014; Kervran et al., 2020). Often defined as a strong desire to use, craving plays a key role in the development and maintenance of substance use (Koob & Moal, 1997; Robinson & Berridge, 1993). Unlike craving, hazardous use was consistently the least discriminating criterion, meaning that it was less effective at identifying where an individual falls on the SUD severity continuum. Thus, although in the DSM-5 each SUD criterion is considered to provide the same amount of information about the severity of disorder, some items are clearly more informative than others.

In addition to variation within SUDs, the relative ranking of criteria by difficulty and discrimination often varied across the three SUDs. For example, withdrawal was among the most difficult criteria to endorse for AUD but among the least difficult for OUD. This is likely a consequence of the predictable development of physical dependence with regular opioid use (Pergolizzi Jr, Raffa, & Rosenblatt, 2020). For this reason, among individuals who are taking opioids as prescribed, withdrawal and tolerance are not considered toward the diagnosis of OUD in the DSM-5 (American Psychiatric Association, 2013). Another notable difference was that craving, which was the most discriminating criterion among OUD levels, was among the least discriminating criteria for AUD and CUD. These rank-order differences suggest that the same DSM-5 criteria do not perform equivalently across substances. Thus, the informativeness of individual criteria appears to be substance-specific in some cases.

Furthermore, there was systematic measurement bias in how criteria functioned based on exposure to ACEs. The presence of significant DIF suggests that early life adversity influences how SUD symptoms manifest and are reported. Across all three SUDs, ACEs influenced symptom reporting by altering how easily criteria were endorsed (uniform bias) and by reducing the ability of some criteria to distinguish between mild and severe cases (nonuniform bias). Among individuals with higher ACEs scores, withdrawal, continued use despite physical health problems (in AFR and EUR individuals), and larger/longer, were less likely to be endorsed at the same level of SUD severity. Although we did not hypothesize that these criteria would show DIF by ACEs, the lower endorsement of physiological consequences, such as withdrawal and continued use despite physical health problems, is consistent with potential alterations in interoceptive accuracy among individuals who experienced childhood adversity. Individuals who experience ACEs have heightened baseline unpleasantness, meaning that they exist in a chronic state of discomfort or dysregulation (Schaan et al., 2019), which may make it more difficult for them to recognize withdrawal or other physical distress and report these as SUD-related, a possibility that requires further investigation. This also aligns with research showing that individuals who experience ACEs have blunted awareness of physiological signals, including an impaired ability to accurately detect heart rate changes under stress (Schaan et al., 2019).

Several criteria, including some related to physiological effects, also showed nonuniform bias among individuals with higher ACEs scores. For example, in EUR individuals, across all three SUDs, tolerance was a less effective indicator of SUD severity among individuals exposed to ACEs. The same was true for AUD in AFR individuals. One potential explanation for this finding is that physiological adaptation to substances may be perceived differently when an individual’s baseline stress and discomfort levels are elevated due to a history of ACEs. This pattern of reduced discrimination means that tolerance provides less information about where a person falls on the underlying SUD severity continuum as ACEs scores increase. Thus, the endorsement (or non-endorsement) of tolerance provides less meaningful information for accurately characterizing SUD severity. As a result, individuals with high ACEs scores may appear more similar across severity levels, increasing the risk of misclassifying severity.

The assessment of AUD was particularly affected by DIF in relation to ACEs. In EUR individuals, as ACEs increased, eight of the 11 criteria showed reduced discrimination, indicating widespread measurement bias. In AFR individuals, AUD was also the disorder with the greatest DIF, with three criteria showing DIF by ACEs compared to just one for CUD and none for OUD. Thus, AUD criteria may be uniquely sensitive to ACEs-related differences in symptom reporting. When such a large proportion of criteria show reduced discrimination at higher ACEs levels, AUD criterion count becomes a less precise indicator of underlying AUD severity. These findings underscore the importance of considering trauma history and protective factors when interpreting AUD symptom profiles.

In contrast to the findings for ACEs, we did not identify DIF as a function of PGS. This suggests that, at least using data from current GWAS, SUD PGS primarily index an individual’s overall liability to developing SUD rather than systematically altering the difficulty or discrimination of criteria. Consistent with this interpretation, a study that was conducted in the University of California-San Francisco Family Alcoholism Study incorporated AUD PGS in Multiple Indicators Multiple Causes (MIMIC) models and found that, although the PGS robustly predicted a latent AUD factor, it showed no residual associations with individual DSM-5 AUD criteria after accounting for the AUD factor (Sarles, Lane, Miller, Wilhelmsen, & Gizer, 2025). Together with our findings, this suggests that genetic risk captured by current SUD PGS mainly shifts individuals along a severity continuum, while environmental exposures (e.g., childhood experiences) may play a larger role in shaping which specific criteria are endorsed at a given level of the latent SUD factor.

Limitations

The current study has several methodological limitations. First, endorsement of lifetime SUD criteria may be affected by recall bias, and differential recall across groups could contribute to some of the observed variations in endorsement. For example, individuals with a history of trauma exhibit impaired memory for non-trauma-related events (Pitts, Eisenberg, Bailey, & Zacks, 2022), which could affect their ability to recall SUD criteria. Research using prospective, longitudinal datasets could mitigate this limitation but poses practical difficulties. Second, PGS are a valuable but incomplete measure of genetic liability, as they do not capture the effects of rare variants (in our implementation) or gene-by-environment interactions. Additionally, the predictive power of PGS varies across genetically inferred ancestry groups and is dependent upon the power of the associated discovery GWAS. For example, in prior analyses using this sample (SooHoo et al., 2025), only the AUD and CUD PGS significantly predicted their corresponding DSM-5 diagnoses in EUR individuals, and only the AUD PGS was predictive in AFR individuals, reflective of the current modest predictive utility of PGS. The use of improved multi-ancestry PGS methods warrants consideration in future studies to enhance power.

There are also interpretative limitations to consider. Our data are cross-sectional, so although we can identify differences in item functioning, we cannot ascertain when or how these differences develop. Additionally, because DSM-5 SUD severity is based on criterion counts rather than latent factor scores, and no validated approach exists for applying DIF-adjusted factor scores clinically, the current study focused on identifying measurement bias rather than modifying existing procedures for evaluating SUD severity.

Conclusions

DSM-5 SUD criteria vary both in how readily they are endorsed and how well they distinguish among individuals with different levels of the latent SUD factor. Some criteria (e.g., craving and role failure) were generally more informative about an individual’s level of the latent SUD factor, whereas others (e.g., hazardous use) exhibited lower discrimination. Notably, there was systematic measurement bias related to ACEs but not genetic liability, such that individuals in with higher ACEs scores were less likely to endorse certain physiological criteria at the same latent SUD level. Several criteria also showed lower discriminative ability among individuals exposed to ACEs, with these effects being the most pronounced for AUD. Our findings highlight the need to refine SUD assessment to account for differences in endorsement and symptom expression, particularly among individuals who experience early-life adversity. In practice, such refinement could involve considering relative weighting of criteria or developing screening procedures that consider subgroup-specific differences in symptom functioning. Future work using longitudinal designs or trauma-informed assessment frameworks will be important for evaluating how these measurement differences emerge over time and whether they meaningfully influence outcomes.

Public health significance statement:

At the same level of substance use disorder severity, individuals who experienced childhood adversity are less likely to endorse certain physiological symptoms. In addition, many symptoms also offer less precision about disorder severity for these individuals. These patterns suggest a need for more accurate screening and assessment approaches.