Post-traumatic stress disorder and drug use disorder: examination of aetiological models in a Swedish population-based cohort

ABSTRACT

Background: There are two primary phenotypic models of comorbidity between post-traumatic stress disorder (PTSD) and drug use disorder (DUD), i.e. self-medication (PTSD precedes and causes DUD) and susceptibility (DUD precedes and causes PTSD). We sought to clarify the longitudinal relationship between PTSD and DUD, while examining sex differences.

Method: We used approximately 23 years of longitudinal data from Swedish population registries to conduct two complementary statistical models: Cox proportional hazard models (N ≈ 1.5 million) and a cross-lagged panel model (N ≈ 3.8 million).

Results: Cox proportional hazards models, adjusting for cohort and socioeconomic status, found strong evidence for the self-medication hypothesis, as PTSD predicted increased risk for DUD among both women [hazard ratio (HR) = 5.34, 95% confidence interval (CI) 5.18, 5.51] and men (HR = 3.65, 95% CI 3.54, 3.77), and moreover, that the PTSD to DUD association was significantly higher among women (interaction term 0.68, 95% CI 0.65, 0.71). The results of the susceptibility model were significant, but not as strong as the self-medication model. DUD predicted risk for PTSD among both women (HR = 2.43, 95% CI 2.38, 2.50) and men (HR = 2.55, 95% CI 2.50, 2.60), and HR was significantly higher in men (interaction term 1.05, 95% CI 1.02, 1.08). Investigating the pathways simultaneously in the cross-lagged model yielded support for both pathways of risk. The cross-paths instantiating the susceptibility model (0.10–0.22 in females, 0.12–0.19 in males) were mostly larger than those capturing the self-medication model (0.01–0.16 in females, 0.04–0.22 in males).

Conclusions: We demonstrate that the relationship between PTSD and DUD is bidirectional, with evidence that future research should prioritize examining specific pathways of risk that may differ between men and women.

HIGHLIGHTS

• Post-traumatic stress disorder (PTSD) and drug use disorder (DUD) are highly comorbid, and few large population-based longitudinal studies have been conducted to better understand why these disorders co-occur at a rate far greater than chance.

• We used approximately 23 years of longitudinal data from the Swedish National Registries, in a sample of over 1.5 million people, to look at the prospective relationships between PTSD and DUD, and vice versa.

• We found evidence for bidirectional risk such that having one disorder increased the future risk for the other disorder, although the effect sizes were higher for PTSD’s risk on future DUD, and some patterns differed by sex.

Antecedentes: Existen dos modelos primarios fenotípicos de la comorbilidad entre el trastorno de estrés postraumático (TEPT) y el trastorno por uso de sustancias (TUS; es decir, automedicación [el TEPT precede y causa el TUS], susceptibilidad [el TUS precede y causa el TEPT]). Intentamos aclarar la relación longitudinal entre el TEPT y el TUS, mientras examinamos las diferencias por sexo.

Método: Utilizamos ∼23 años de datos longitudinales de los registros de la población sueca para realizar dos modelos estadísticos complementarios: modelos de riesgo proporcional de Cox (N∼1.5 millones) y modelo de panel con retraso cruzado (N∼3.8 millones).

Resultados: Los modelos de riesgos proporcionales de Cox, ajustados por cohorte y estatus socioeconómico, encontraron evidencia sólida para la hipótesis de la automedicación, ya que el TEPT predijo un mayor riesgo del TUS tanto entre mujeres (HR = 5.34 [5.18–5.51]) como entre hombres (HR = 3.65 [3.54, 3.77]), y además, que la asociación del TEPT con TUS fue significativamente mayor entre las mujeres (término de interacción (0.68 [0.65, 0.71])). Los resultados del modelo de susceptibilidad fueron significativos, pero no tan fuertes como el modelo de automedicación. El TUS predijo el riesgo de TEPT tanto entre mujeres (HR = 2.43 [2.38, 2.50]) como entre hombres (HR = 2.55 [2.50, 2.60]) y el HR fue significativamente mayor en los hombres (término de interacción (1.05 [1.02, 1.08]). La investigación de las vías simultáneamente en el modelo con retraso cruzado arrojó apoyo para ambas vías de riesgo. Las vías de cruces que instanciaron el modelo de susceptibilidad (0.10–0.22 en mujeres, 0.12–0.19 en hombres) fueron en su mayoría más grandes que los que capturaron el modelo de automedicación (0.01–0.16 en mujeres, 0.04–0.22 en hombres).

Conclusiones: Demostramos que la relación entre el TEPT y el TUS es bidireccional con evidencia de que futuras investigaciones deberían priorizar el examen de vías específicas de riesgo que pueden diferir entre hombres y mujeres.

背景：创伤后应激障碍（PTSD）和药物使用障碍（DUD；即自我药疗[PTSD先于并导致DUD]、易感性[DUD先于并导致PTSD]）之间存在两种主要的共病表型模型。我们试图阐明 PTSD 和 DUD 之间的纵向关系，同时研究性别差异。

方法：我们使用瑞典人口登记处约 23 年的纵向数据来建立两个互补的统计模型：Cox 比例风险模型（N∼150 万）和交叉滞后面板模型（N∼380 万）。

结果：根据队列和社会经济状况进行调整的 Cox 比例风险模型发现了自我药疗假说的有力证据，因为 PTSD 预测女性 (HR = 5.34 [5.18–5.51]) 和男性 (HR = 3.65 [3.54, 3.77]) 增高的 DUD 患病风险，此外，女性中 PTSD 与 DUD 的关联显著较高（交互作用项 (0.68 [0.65, 0.71]）)。易感性模型的结果很显著，但不如自我药疗模型那么强。DUD 预测女性 (HR = 2.43 [2.38, 2.50]) 和男性 (HR  = 2.55 [2.50, 2.60]) 的 PTSD 风险，并且男性的 HR 显著较高（交互项 (1.05 [1.02, 1.08])）。同时研究交叉滞后模型中的路径得到了对两种风险路径的支持。实例化易感性模型的交叉路径（女性为 0.10–0.22，男性为 0.12–0.19）大多大于自我药疗模型的交叉路径 （女性为 0.01–0.16，男性为 0.04–0.22）。

结论：我们证明 PTSD 和 DUD 之间的关系是双向的，有证据表明未来的研究应优先考查男性和女性之间可能不同的特定风险途径。

KEYWORDS:

• PTSD
• drug use disorder
• comorbidity
• epidemiology

PALABRAS CLAVE:

• TEPT
• Trastorno por uso de sustancias
• comorbilidad
• epidemiología

关键词:

• PTSD
• 药物使用障碍
• 共病
• 流行病学

1. Introduction

Exposure to traumatic events, such as natural disasters, accidents, and interpersonal violence, is a common transdiagnostic risk factor for numerous conditions (Benjet et al., 2016), including substance use disorders (SUDs), i.e. alcohol use disorder (AUD) and drug use disorder (DUD) (Jakupcak et al., 2010; Kevorkian et al., 2015; Keyes et al., 2011; Mills et al., 2006), and post-traumatic stress disorder (PTSD) (Breslau, 2009). PTSD and SUDs are frequently comorbid, and data from large epidemiological studies demonstrate that those with PTSD are at four to five times greater likelihood for meeting criteria for SUD compared to those without PTSD (for a review, see Brady et al., 2021). Much of the literature on PTSD and SUD has either examined alcohol and drug use disorders together (i.e. SUD) or examined AUD specifically. Here, we focus, where possible, on the prior studies on PTSD and DUD.

Data from epidemiological studies suggest that among individuals with PTSD, DUD co-occurrence is common; however, large ranges are present in the literature. For example, in the original National Comorbidity Study, 26.9% of women and 34.5% of men with PTSD also met criteria for a DUD (Kessler et al., 1995), whereas in a more recent National Epidemiologic Study of Alcohol and Related Conditions (NESARC) study found that among those with PTSD, 22.3% met criteria for a DUD and 46.4% met criteria for any SUD (Pietrzak et al., 2011). Among adults with SUD, estimates of the prevalence of PTSD range from 26% to 60% for lifetime and 15% to 42% for current (Back et al., 2000; Brady et al., 2004; Jacobsen et al., 2001; Mills et al., 2006; Reynolds et al., 2011) compared to 8.3% (lifetime) and 3.8% (current) in the general population (Kilpatrick et al., 2013).

Although clear sex differences exist for both PTSD (which is more common in women than in men) and SUD (more common in men than in women) independently, epidemiological studies of comorbidity do not reveal such clear patterns of sex differences (Torchalla & Nosen, 2019). For example, some studies have reported no difference in the prevalence of comorbidity (Blanco et al., 2013), while other studies find patterns of dual diagnosis consistent with the pattern of each individual disorder. Specifically, some studies suggest that men with PTSD are more prone to SUD than women with PTSD (34.5% vs 26.9%, respectively) (Kessler et al., 1995) and that women with SUD are about twice as likely to develop PTSD than are men with DUD (Goldstein et al., 2016). Further, pathways to PTSD and SUD co-occurrence may differ for men and women, underscoring the need for studies to examine sex differences in aetiological studies (Torchalla & Nosen, 2019).

Both PTSD and SUD are associated with increased personal and societal burden (e.g. Dobie et al., 2004; Henkel, 2011; McGowan, 2019). Beyond the high toll that each disorder has independently, PTSD and SUD frequently co-occur (Ouimette et al., 2005), and their comorbidity is associated with increased impairment (Simpson et al., 2019), poorer treatment outcomes (Fontana et al., 2012; Hien et al., 2000; Pimlott Kubiak, 2005; Read et al., 2004), legal problems and service utilization (Brown et al., 1999; Foa & Williams, 2010; McCauley et al., 2012), interpersonal problems (Najavits et al., 1999), unemployment (Drapkin et al., 2011), suicidality (Müller et al., 2015; Rojas et al., 2014), and reduced quality of life (Norman et al., 2018). Furthermore, most research on PTSD-SUD co-occurrence either focuses solely on AUD or includes both alcohol and illicit substances broadly under an SUD category (Brady et al., 2021). Thus, there is a need to elucidate the aetiology of PTSD-DUD comorbidity to inform prevention and intervention efforts.

Despite the high comorbidity and substantial personal and societal burden associated with these comorbid disorders, the causal relations between PTSD and DUD are unclear. Although numerous conceptual frameworks have been posited to explain the nature of this comorbidity (for reviews, see Berenz et al., 2019; María-Ríos & Morrow, 2020), two primary phenotypic aetiological models have received the most attention: the self-medication hypothesis (i.e. PTSD leads to DUD via drug use to cope with trauma-related symptoms) and the susceptibility hypothesis (i.e. DUD broadly affects the risk for PTSD via increased likelihood of trauma exposure associated with substance use behaviours or via impaired homeostatic response to stressors). Numerous methodologies have been used to examine the these models of aetiology (e.g. cross-sectional and longitudinal, self-report motives studies, ecological momentary assessment studies, and laboratory studies), and the key findings will be reviewed in brief.

Studies examining self-reported use of alcohol and substances to cope with emotional distress have found a relationship between coping motives and PTSD (e.g. Berenz et al., 2016; Grayson & Nolen-Hoeksema, 2005; Kaysen et al., 2007; Stappenbeck et al., 2013; Tomaka et al., 2017; Waldrop et al., 2007). Experimental studies have also shed light on the neurofunctional mechanisms of PTSD-SUD comorbidity. In a review article, Hien et al. (2021) posit that deficits in executive function, social cognition, negative emotionality, and reward salience influence PTSD-SUD comorbidity, reviewing studies examining deficits in each disorder examined in isolation, as well as the few studies that have examined comorbid samples, which are few and far between. A number of ecological momentary assessment studies have been conducted on PTSD and alcohol use behaviours (for a review, see Lane et al., 2019), and emerging work on PTSD and drug use behaviours has also been published (e.g. Buckner et al., 2018; Sanjuan et al., 2020), largely supporting prospective relationships between PTSD fluctuations and craving or substance use. Although person-centred methods (e.g. ecological momentary assessment (EMA) and mechanistic studies) are important to understanding comorbidity, they are outside the scope of this paper, which utilizes the largest epidemiological sample to date, of over 1.5 million people, to prospectively examine the risk of comorbidity.

Cross-sectional epidemiological studies consistently find a PTSD-DUD association beyond chance expectations (odds ratios = 1.8–6.53) (Glantz et al., 2009; Kessler et al., 1995; Mills et al., 2006; Pietrzak et al., 2011), but few have tried to look at order of onset. Using retrospectively reported data on age of onset of PTSD and alcohol dependence (AD) from the National Epidemiologic Survey on Alcohol and Related Conditions (NESARC), evidence was found for both the self-medication [hazard ratio (HR) = 1.36] and susceptibility (HR = 1.27) models, with non-overlapping confidence intervals, and furthermore, bidirectional relationships were stronger for women than for men (Berenz et al., 2017).

Limited epidemiological longitudinal studies of PTSD and DUD have tested potential causal pathways, and none, to our knowledge, has done so by sex, which is a critical gap to fill given key sex differences in these disorders. Most evidence suggests a PTSD to DUD pathway; for example, PTSD predicted a 3.9-fold increased risk for substance use disorder onset at the 10 year follow-up of the National Comorbidity Study (Swendsen et al., 2010); however, the reverse direction was not tested. Similar findings in support of the self-medication hypothesis have been demonstrated in an epidemiological study of adolescents, whereby PTSD predicted SUD longitudinally, but the reverse association was not found (Wolitzky-Taylor et al., 2012). In another epidemiological sample, survival analyses suggested PTSD had a substantial association with the risk of onset of DUD (HR = 4.5), far larger than the effect of DUD predicting future onset of PTSD (HR = 1.6); however, the prevalence of baseline DUD was low in this sample, thereby limiting the power of this test (Chilcoat & Breslau, 1998). Other non-epidemiological community sample longitudinal studies have also examined the prospective relationships. Among a sample of returning veterans, specific PTSD symptom clusters were predictive of future high-risk drug use behaviours, supporting the self-medication hypothesis (Livingston et al., 2022). Further evidence for the self-medication hypothesis was found in an analysis of longitudinal data of youth, such that PTSD symptoms were associated with higher levels of alcohol and drug problems, and although adolescent substance use problems predicted future assaultive trauma, there was no relationship with future PTSD (Haller & Chassin, 2014). Another study examined latent classes of cannabis use, finding that classes associated with higher use were prospectively related to PTSD symptoms (Lee et al., 2018).

The primary objective of this study was to clarify the longitudinal relationship between PTSD and DUD to compare the evidence for the two primary phenotypic models of comorbidity between PTSD and DUD (i.e. self-medication and susceptibility), while examining sex differences. To do so, we used two complementary analysis techniques, one that examines aggregate risk (Cox models) and one that examines more nuanced effects (cross-lagged models) to model longitudinal data (over approximately 23 years) from Swedish population registries (N > 1.5 million). This powerful registry study overcomes many of the prior limitations of the extant epidemiological studies, such as limited sample size, lack of testing of sex effects, retrospective report, not testing both directions of effect between the disorders, and examination of SUD broadly rather than DUD specifically.

2. Methods

We linked nationwide Swedish registers via the unique 10-digit identification number assigned at birth or immigration to all Swedish residents. The identification number was replaced by a serial number to ensure anonymity; informed consent was not required for this study as the data are completely anonymous. The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Declaration of Helsinki of 1975, as revised in 2008. All procedures involving human subjects/patients were approved by the Regional Ethical Review Board of Lund University (no. 2008/409 and later amendments). The following sources were used to create our data set: the Total Population Register, containing information about year of birth and sex; the Multi-Generation Register, linking individuals born after 1932 to their parents; the Longitudinal Integration Database for Health Insurance and Labour Market Studies (LISA), containing information about education and income from 1990 to 2018; the Hospital Discharge Register, containing hospitalizations for Swedish inhabitants from 1964 to 2018; the Prescribed Drug Register, containing all prescriptions in Sweden picked up by patients from July 2005 to 2018; the Outpatient Care Register, containing information from all outpatient clinics from 2001 to 2018; the Crime Register, which includes complete national data on all convictions in the lower courts from 1973 to 2018; the Swedish Suspicion Register, which includes national data on individuals strongly suspected of crime from 1998 to 2018; and the Mortality Register, with dates and causes of death from 1952 to 2018. In addition, we obtained medical diagnoses from primary healthcare clinics from all counties throughout Sweden (see Supplementary information for details).

2.1. Sample

The study population consisted of males and females, born in Sweden (Table 1). For the first set of analyses we included everyone born between 1975 and 1990 (n = 1,541,883 for the self-medication model and n = 1,540,979 for the susceptibility model; 49% female) and for the second set of analyses individuals born between 1960 and 1990 (n = 3,829,660; 49% female). Analyses covered 23 years of longitudinal data.

Table 1. Frequency of key variables stratified by sex and drug use disorder (DUD) and post-traumatic stress disorder (PTSD) status.

DUD stratified
	Males		Females
	Total number (column %)	With DUD (row %)	Total number (column %)	With DUD (row %)
Individuals in the cohort	791,751	72,069 (9.10%)	750,132	27,997 (3.73%)
No PTSD during follow-up	727,118 (91.84%)	67,659 (9.31%)	590,749 (78.75%)	22,240 (3.76%)
PTSD during follow-up	64,633 (8.16%)	4,410 (6.82%)	159,383 (21.25%)	5,757 (3.61%)
Low parental education	51,280 (6.48%)	5,676 (11.07%)	48,871 (6.51%)	2,216 (4.53%)
Mid parental education	355,185 (44.86%)	34,552 (9.73%)	336,983 (44.92%)	13,576 (4.03%)
High parental education	385,286 (48.66%)	31,841 (8.26%)	364,278 (48.56%)	12,205 (3.35%)
Follow-up time, mean (SD)	17.44 (5.38)	8.11 (4.72)	17.79 (5.05)	8.75 (5.44)
PTSD stratified
	Males		Females
	Total number (column %)	With PTSD (row %)	Total number (column %)	With PTSD (row %)
Individuals in the cohort	791,596	74,115 (9.36%)	749,383	165,100 (22.03%)
No DUD during follow-up	723,777 (91.43%)	63,728 (8.80%)	726,831 (96.99%)	157,395 (21.65%)
DUD during follow-up	67,819 (8.57%)	10,387 (15.32%)	22,552 (3.01%)	7,705 (34.17%)
Low parental education	51,273 (6.48%)	5,958 (11.62%)	48,802 (6.51%)	12,552 (25.72%)
Mid parental education	355,123 (44.86%)	35,252 (9.93%)	336,587 (44.92%)	78,004 (23.17%)
High parental education	385,200 (48.66%)	32905 (8.54%)	363,994 (48.57%)	74,544 (20.48%)
Follow-up time, mean (SD)	17.80 (4.72)	14.10 (4.99)	16.97 (5.12)	14.12 (5.03)

2.2. Measures

DUD was assessed from medical registers, using codes from the International Classification of Diseases, Ninth Revision (ICD-9): 292, 304, and 305C–305I; ICD-10: F11–F16 and F18–F19; in the Suspicion Register by codes 3070, 5010, 5011, and 5012; and in the Crime Register by references to law 1968:64, paragraph 1, point 6, and/or law 1951:649, paragraph 4, subsection 2, and paragraph 4A, subsection 2. Individuals must have been over 15 years old to have a record in the registers, and we used date of crime, not date of conviction, as has been done in prior reports using DUD in the registries. Crimes clearly involving substance misuse, but excluding drug dealing, were included (Kendler et al., 2012, 2015).

PTSD was assessed from medical registers, using codes from the ICD that capture the traumatic stress disorders (i.e. ICD-9: 308 and 309; and ICD-10: F431, F430, F438, F439, and F620), as has been done by our group in prior analyses (Amstadter et al., 2023). While we recognize that this captures disorders beyond PTSD, for simplicity of communication, any participant meeting criteria for any of the traumatic stress disorder codes is referred to as having PTSD.

Individual socioeconomic status was assessed using parents’ highest level of education, and categorized into low (compulsory school only), mid (upper secondary school), and high (university level).

2.3. Statistical methods

In the first set of analyses, to estimate the association between PTSD and DUD, we used Cox proportional hazard models to understand the degree of aggregate risk that one disorder has prospectively on the other disorder. We followed the population from 1 January 1995 to 31 December 2018 and required that they should be at least 15 years old at the start of follow-up and without a prior DUD registration. To investigate the self-medication hypothesis, we included PTSD as a time-dependent variable and followed individuals until first registration of DUD, censoring at death or end of follow-up. Both males and females were included in the model, although we allowed for different associations between the predictors and the outcome by including the corresponding interaction terms. The interaction terms represent the difference between males and females. First, we estimated the crude, unadjusted association between PTSD and future DUD. Next, to account for confounding caused by cohort or socioeconomic status, we adjusted for birth year and parental education. To investigate the proportionality assumption, we allowed the association between PTSD and DUD to depend linearly on follow-up time. To illustrate how the associations are modified during the follow-up time, we present the associations at three time-points: at the time of PTSD diagnosis, 1 year after, and 5 years after the PTSD diagnosis. Next, to investigate the susceptibility model, we flipped the outcome and predictor and performed the same set of analyses as described above. We performed these statistical analyses using SAS/STAT^® software, version 9.4.

In the second analysis, we explored the nuanced associations between DUD and PTSD over time, using a cross-lagged model to simultaneously investigate the self-medication and susceptibility hypotheses. We assessed DUD and PTSD before the age of 15 years and during 5 year intervals after 15 years and until 40 years of age for the respective age brackets corresponding to the period between 1990 and 2018. In other words, we based the model on the following observed variables: DUD before the age of 15 years, DUD age 15–19, DUD age 20–24, DUD age 25–29, DUD age 30–34, and DUD age 35–40 years. PTSD was assessed during the same age periods. We included paths from the variable in the earlier age period on to the two variables, DUD and PTSD, in the succeeding age period. Furthermore, in each period, we allowed for a correlation between DUD and PTSD. Males and females were both included in the model, but the paths and correlations were estimated separately for the two groups. We used Mplus software version 7.31 (Muthen & Muthen, 2012) and the WLSMV estimator to estimate the model parameters. The model was evaluated using the root mean square of approximation (RMSEA), comparative fit index (CFI), and Tucker–Lewis index (TLI). R-squared values are reported in the supplementary information.

3. Results

Descriptive statistics for the sample used in the Cox proportional hazard models are shown in Table 1. Note that the follow-up period is shorter for those with DUD than for those with PTSD. The prevalence rates of DUD were increasing in the registries during this time period (Giordano et al., 2014), which probably accounts for the descriptive finding of a higher prevalence of PTSD among those without DUD. When constraining the sample by age (see Supplementary Table 1), PTSD prevalence is higher among those with DUD. Among comorbid cases, we examined the time to diagnosis. If the PTSD diagnosis was established after the DUD diagnosis, the median time to PTSD diagnosis was 6.2 years [interquartile range (IQR) 3.0, 9.9] for men and 5.9 years (IQR 2.6, 9.8) for women. If, instead, the PTSD diagnosis was made first, the median time to DUD diagnosis was 2.0 years (IQR 0.5, 4.6) for men and 2.4 years (IQR 0.7, 5.4) for women. Thus, there may be a higher risk of undetected DUD before PTSD compared to the reverse direction.

3.1. Cox proportional hazards models: self-medication model (n = 1,541,883)

Unadjusted analyses indicated that men were at increased risk for DUD compared to women [HR = 2.86, 95% confidence interval (CI) 2.82, 2.90], that PTSD was associated with increased risk for DUD among both males (HR = 4.26, 95% CI 4.13, 4.40) and females (HR = 6.26, 95% CI 6.07, 6.45), and that the HRs in the two sexes differed significantly (interaction term 0.68, 95% CI 0.65, 0.71).

Table 2 provides the results of the model testing the self-medication hypothesis, adjusted for covariates of birth year and parental educational attainment. The estimated association between PTSD and risk of subsequent DUD remained robust despite modest attenuation in magnitude in the model adjusted for covariates. Furthermore, the HRs for PTSD remained significantly different between the sexes. The effect of parental education on risk for DUD did not differ by sex. Birth year had a substantial effect on DUD for both sexes, with little evidence for the HRs being different between the sexes.

We tested the proportionality assumption by estimating the HR for PTSD at various follow-up times, adjusted for covariates, and present in Supplementary Table 2 the estimated associations immediately after diagnosis, 1 year after diagnosis, and 5 years after diagnosis. These analyses showed evidence of a time-varying association, such that the impact of PTSD on the risk of DUD registration was highest immediately after a diagnosis of PTSD and decreased over time.

Table 2. Adjusted hazard ratios for Cox models testing the self-medication and susceptibility models adjusted for covariates.

Self-medication model (PTSD predicting DUD)
	Males	Females	Male vs female ratio
Males vs females			2.81 (2.75, 2.87)
PTSD	3.65 (3.54, 3.77)	5.34 (5.18, 5.51)	0.68 (0.65, 0.71)
Low vs high parental education	1.48 (1.44, 1.53)	1.41 (1.34, 1.47)	1.06 (1.00, 1.11)
Mid vs high parental education	1.18 (1.17, 1.20)	1.18 (1.15, 1.21)	1.01 (0.98, 1.04)
Birth year	1.08 (1.08, 1.08)	1.07 (1.06, 1.07)	1.01 (1.01, 1.02)
Susceptibility model (DUD predicting PTSD)
	Males	Females	Male vs female ratio
Males vs females			0.36 (0.36, 0.36)
DUD	2.55 (2.50, 2.60)	2.43 (2.38, 2.50)	1.05 (1.02, 1.08)
Low vs high parental education	1.23 (1.20, 1.27)	1.17 (1.15, 1.20)	1.05 (1.01, 1.09)
Mid vs high parental education	1.12 (1.10, 1.14)	1.10 (1.09, 1.11)	1.02 (0.999, 1.04)
Birth year	1.13 (1.13, 1.13)	1.13 (1.13, 1.14)	0.998 (0.996, 01.000)

Note: PTSD, post-traumatic stress disorder; DUD, drug use disorder.

Table 3. Descriptive table for the cross-lagged panel model.

Variable	Males (N = 1,961,711)		Females (N = 1,867,949)
	Eligible	With outcome	Eligible	With outcome
DUD before 15 years	1,246,538	186 (0.01%)	1,177,964	321 (0.03%)
DUD 15–20 years	1,482,961	25,153 (1.70%)	1,404,163	9,119 (0.65%)
DUD 20–25 years	1,877,452	53,854 (2.87%)	1,781,916	14,848 (0.83%)
DUD 25–30 years	1,957,594	52,527 (2.68%)	1,857,293	14,337 (0.77%)
DUD 30–35 years	1,554,841	34,903 (2.24%)	1,474,325	10,816 (0.73%)
DUD 35–40 years	1,170,824	23,338 (1.99%)	1,114,791	9,121 (0.82%)
PTSD before 15 years	1,246,538	532 (0.04%)	1,177,964	977 (0.08%)
PTSD 15–20 years	1,482,961	3,738 (0.25%)	1,404,163	8,818 (0.63%)
PTSD 20–25 years	1,877,452	16,882 (0.90%)	1,781,916	35,726 (2.00%)
PTSD 25–30 years	1,957,594	35,732 (1.83%)	1,857,293	78,240 (4.21%)
PTSD 30–35 years	1,554,841	39,313 (2.53%)	1,474,325	87,785 (5.95%)
PTSD 35–40 years	1,170,824	30,990 (2.65%)	1,114,791	61,378 (5.51%)

Note: DUD, drug use disorder; PTSD, post-traumatic stress disorder.

3.2. Cox proportional hazards models: susceptibility model (n = 1,540,979)

Descriptive statistics for the Cox models are provided in Table 1. Unadjusted analyses demonstrated that males were at decreased risk for PTSD compared to females (HR = 0.37, 95% CI 0.36, 0.37), that DUD conveyed increased risk for PTSD among both males (HR = 2.98, 95% CI 2.92, 3.04) and females (HR = 2.84, 95% CI 2.78, 2.91), and that the HRs in the two sexes differed significantly (interaction term 1.05, 95% CI 1.02, 1.08).

We next adjusted for covariates of birth year and parental educational attainment (Table 2). The addition of parental education and birth year to the model resulted in modest declines in the HRs of DUD of subsequent PTSD; however, for both sexes the relationship remained robust. Similarly to the unadjusted results, sex was a significant moderator of the relationship between PTSD and DUD, but the magnitude of the sex difference was small. Low parental education predicted increased risk for PTSD, and more so for males compared to females. Birth year was also associated with substantially increased risk for PTSD, and was not differentially associated between the sexes.

Tests of the proportionality assumption, shown in Supplementary Table 3, suggested no evidence of a time-varying association between DUD and PTSD, suggesting that DUD has a stable effect on the risk of subsequent PTSD over time.

3.3. Cross-lagged panel model: simultaneous test of self-medication and susceptibility hypotheses (n = 3,829,660)

As shown in Table 3, the prevalence of DUD (top half of the table) among both males and females increased with age until 20–25 years, then began to decrease. The pattern was more divergent by sex for PTSD, such that among women, the prevalence increased more sharply than for men. In contrast to DUD, the prevalence of PTSD did not appear to decline across the study period.

Our model yielded a good fit to the data (RMSEA = .009, 95% CI .009, .010; CFI = .992; TLI = .988); R-squared statistics are shown in Supplementary Table 5. As shown in Figure 1(a), this model estimated three types of paths: (1) two-headed correlational paths between PTSD and DUD across all time-points, included to control for the impact of potentially confounding variables (‘c’ paths); (2) directional within disorder paths, to control for the temporal stability of PTSD (‘a’ paths) and DUD (‘b’ paths); and (3) cross-lagged standardized regression paths, which go both from PTSD at time x to DUD at time x + 1 (‘d’ paths), and from DUD at time x to PTSD at time x + 1 (‘e’ paths).

Figure 1. (a) Cross-lagged panel model of registrations of post-traumatic stress disorder (PTSD) and drug use disorder (DUD), starting at the age of 15 years, in 5 year epochs between the ages of 35 and 50 years. Five types of paths are shown: the ‘a’ paths refer to the stability paths for PTSD, the ‘b’ paths refer to the stability paths for DUD, the ‘c’ paths denote the within-time correlations between PTSD and DUD, the ‘d’ paths represent the self-medication pathway, and the ‘e’ paths denote the susceptibility pathway. (b) Standardized path coefficient results from the cross-lagged panel model, by sex.

As shown in Supplemental Table 4 and Figure 1(b), the cross-paths testing the self-medication hypothesis (PTSD → DUD) had a small effect, and ranged from .04 to .22 in males and from .01 to .16 in females. The cross-paths testing the susceptibility model (DUD → PTSD) were of modest strength and ranged from .12 to .19 in males and from .10 to .22 in females. The stability paths for both PTSD (males: .35–.46; females: .36–.48) and DUD (males: .54–87; females: .48–.86) were high. The within-time correlations between PTSD and DUD were highest at the earliest time-point and decreased across time, similarly so between males (.11–.41) and females (.11–.44).

Follow-up constraints to the cross-lagged model were fitted. First, ‘d’ and ‘e’ paths were constrained over time; this model provided a significantly worse model fit compared to the unconstrained model [χ²(26) = 2708.795, p < .0001] suggesting that the cross-lagged paths were not stable over time. Secondly, a model testing whether ‘d’ and ‘e’ paths could be equated for each age span was fitted and compared to the unconstrained model [χ²(10) = 105.596, p < .0001]; there was clear evidence of a difference between the ‘d’ and ‘e’ paths.

4. Discussion

Through the use of population-based registry data, we sought to clarify the longitudinal relationship between PTSD and DUD to compare the evidence for the two primary phenotypic models of comorbidity, i.e. self-medication (PTSD precedes and causes DUD) and susceptibility (DUD precedes and causes PTSD). We also aimed to examine sex differences, given that the pathways of risk for these conditions may differ between males and females. To do so, we used data from Swedish population registries (N > 1.5 million), which were analysed in two complementary statistical methods: Cox proportional hazards models for an estimate of aggregate risk and a cross-lagged panel model for a nuanced picture of risk across development.

4.1. Cox proportional hazard models: self-medication model

PTSD was associated with a high risk for future DUD, providing strong support for the self-medication model. The magnitude of the effect was significantly stronger for females compared to males. Specifically, among females with a PTSD diagnosis, the HR for future DUD was 5.34, whereas among males, PTSD was associated with an HR of 3.65 for future DUD. However, as DUD is more prevalent among males than females, the baseline risk for females is smaller, and thus, the relative risk increase after a PTSD diagnosis may be greater than that for males, despite the absolute risk increase being similar. The findings of the self-medication model were robust in that they showed only modest attenuation in strength after the inclusion of covariates. Low parental education, a proxy for low socioeconomic status, was associated with a higher risk of DUD for both females and males, without evidence of functioning differently by sex. This finding is consistent with a wealth of literature showing that lower socioeconomic status is associated with risk for substance use outcomes (Dohrenwend et al., 1992; Nagelhout et al., 2017). Birth year was also associated with risk for DUD, such that younger individuals had a higher HR for DUD, with little evidence of a sex difference in this effect. Our prior analyses in the registry on PTSD and AUD also found strong support for the self-medication hypothesis, and contrary to the results for DUD, the strength of the relationship was stronger for males compared to females (Amstadter et al., 2023), underscoring the need for sex-specific research on comorbidity that examines AUD and DUD separately rather than as a unitary construct (Simpson et al., 2019). Violation of the proportionality assumption suggests that the nature of the association of PTSD conferring risk for DUD is not stable but rather decreases across time, suggesting that prevention of DUD may be most effective in close proximity to a PTSD diagnosis.

4.2. Cox proportional hazard models: susceptibility model

The results of the susceptibility model differ from those of the self-medication model in a number of key ways. First, the results of the susceptibility model demonstrating that DUD prospectively increased the risk for PTSD, although highly significant, were not as strong as the effects in the self-medication model. Secondly, contrary to results of the self-medication model, which found that the pathway was more pronounced among females, DUD conveyed an increased risk for future PTSD among males to a significantly greater degree than among females (HR = 2.55 and 2.43, respectively), consistent with our prior registry analyses with AUD (Amstadter et al., 2023). Thirdly, in contrast to the magnitude of the sex difference found for the self-medication model, the difference in effect between sexes was more modest for the susceptibility model. Lastly, whereas the relationship between PTSD and risk for DUD was not static over time (i.e. the proportionality assumption was violated), the relationship of DUD on conferred risk for PTSD did not decay over time, suggesting a static relationship, with the implication that timing of secondary prevention may be effective at any point following a DUD registration. Similarly to the results of the self-medication model, the results were robust, such that the magnitude of effect declined only slightly when adjusting for covariates. Similarly to the results of the self-medication model, low parental education was associated with increased risk of PTSD among both sexes, but more so for men than for women. This covariate effect is consistent with numerous studies that have reported an association between indicators of social disadvantage and liability for PTSD (Brewin et al., 2000; Ozer et al., 2003).

Our findings of a stronger association between PTSD and subsequent DUD are supportive as evidence for the self-medication model, which has broad clinical intuitive appeal (Waldrop et al., 2007). The self-medication model has also been supported in prior analyses testing these models in this registry (Amstadter et al., 2023) and with PTSD and AD in a population-based sample of US adults (Berenz et al., 2017). Our results are also consistent with prior work finding that PTSD had a substantial association with risk of onset of DUD (HR = 4.5), far larger than the effect of DUD predicting the future onset of PTSD (HR = 1.6); however, the prevalence of baseline DUD was low in this sample, thereby limiting the power of this test (Chilcoat & Breslau, 1998); our analyses overcome this limitation. Few studies have tested both directions of effect (Swendsen et al., 2010), and even fewer studies have examined sex differences therein. In contrast to prior research finding that for females the pathway was stronger from AD to PTSD than from PTSD to AD (Berenz et al., 2017), our findings suggest that the self-medication pathway was stronger for females than for males, and that the susceptibility pathway was stronger for males than for females. In sum, our results demonstrate that these models are not mutually exclusive, such that both disorders confer prospective risk for the other disorder, albeit differentially in strength and differentially by sex.

4.3. Cross-lagged panel model

Using a complementary method to that of the Cox regressions, which only provide aggregate risk of one disorder on the other, we employed a cross-lagged structural equation model across adolescence and young adulthood (ages 15–40 years) to simultaneously estimate the two hypothesized causal pathways (i.e. self-medication and susceptibility) in a more nuanced developmentally sensitive fashion. We chose this wide age range to clarify the developmental trajectory of these relationships through emerging adulthood. The cross-lagged model suggests support for both pathways, although the strength of the paths testing the self-medication hypothesis was modest and those testing the susceptibility hypothesis were generally of moderate size. Estimates for both models were similar for males and females. The stability paths for DUD were of high strength, whereas the stability paths for PTSD were of moderate strength. Both stability paths stabilized with age. Similarly, the results of the constrained model suggested that the cross-lagged paths are not constant over time, suggesting a differential impact of each disorder on the other across development. These results highlight the bidirectional nature of the relationship between these phenotypes.

4.4. Future directions

Given evidence from twin and molecular studies that PTSD and DUD are both moderately heritable, with correlated genetic risk (Xian et al., 2000), future studies should incorporate shared genetic risk as well as other potential shared environmental mechanisms. A number of genetically influenced traits (e.g. personality) are important in the aetiological relationship and should be investigated further (Kramer et al., 2014). There is also growing evidence of potential neurofunctional mechanisms that may partially underlie PTSD-SUD co-occurrence, such as deficits in executive function and reward salience, that could be important targets for future research and may be genetically influenced (Hien et al., 2021). In addition, future work that can disaggregate within- from between-persons effects, such as work from an EMA design, will shed light on the functional relationships between these conditions across time. As discussed in Section 4.5, we are unable to isolate the effect of trauma from that of PTSD, and similarly, we are unable to isolate the effect of drug use from that of DUD. While the registry study design has numerous advantages, including unparalleled sample size, and lack of recall and reporting bias, it does preclude answering some questions that necessitate self-report data. Future studies would benefit from examining the effects of trauma and drug use separately from the disorders, as well as other potential moderators, such as treatment seeking.

4.5. Limitations

Several limitations with regard to our reliance on registry data are worth noting. DUD in the Swedish registries is not based on self-report or clinical interview data, but rather on seeking medical attention or on being involved with the criminal justice system. Extensive analyses using this definition have been conducted within the registries, supporting the validity of the definitions, with evidence of high concordance across methods (Kendler et al., 2015), and biometric findings of familial associations being similar to those of generated from clinical interview studies (Prescott & Kendler, 1999; Tsuang et al., 1996). While this has the advantage of limited biases of self-report and recall, it presents other limitations. For example, our findings may be more generalizable for severe DUD, as the threshold for being classified with DUD via registration is high (i.e. sought medical attention or interfaced with the criminal justice system). Following this, future studies should examine the relationship between milder forms of substance use and SUD and PTSD, although samples including milder forms have less power and are more difficult to follow over time. Cases of DUD in the milder and more moderate range may not have been detected with our registration system. Furthermore, there are limitations in our PTSD variable, in that cases were primarily detected from the Primary Care Registry, which is likely to represent an underestimate of the disorder as many do not seek care and/or do not report psychiatric symptoms to their primary care physician. However, analyses from NESARC suggest that a minority of those with PTSD and DUD have never had treatment (12.6%) (Simpson et al., 2019). Related to this, there are limitations as the primary care physicians often do not have the time to carry out a comprehensive psychiatric evaluation. However, analyses of the prevalence of common psychiatric diagnoses in the Primary Care Registry yielded a prevalence estimate of 12.5% for major depression, giving support to the estimates from this registry (Sundquist et al., 2017). Our PTSD variable is limited in that it included PTSD and other traumatic stress diagnoses (e.g. acute stress disorder). In addition, the registry data report the first onset (e.g. first diagnosis) and there may be a lag between the onset of the disorder and registration. We examined the time to diagnosis among comorbid cases, and, based on the pattern of findings suggesting a shorter lag between PTSD and future DUD compared to the other way around, it is likely that there is a higher risk of undetected DUD before PTSD compared to the reverse. These potential misclassifications should be examined in future research using self-report or interview methodology.

Our sample also has limitations of generalizability beyond native-born Swedes born between 1974 and 1990 (Cox models) and between 1960 and 1990 (cross-lagged model), a relatively homogeneous culture with regard to race/ethnicity. However, approximately 20% of the population are first or second generation immigrants. In addition, we did not test the shared liability model. Given that both DUD and PTSD share genetic factors as well as common environmental vulnerabilities, testing this model is an important next step in this line of work to fully examine possible models that might explain the association between these two disorders. We also note that the cross-lagged panel models used herein, while providing an intuitive approach that was well matched to our study questions, have been the subject of criticism (Hamaker et al., 2015), and that future research should incorporate other statistical designs to cross-reference the findings herein.

4.6. Conclusions

This study represents the largest population-based study yet to examine the two main aetiological models proposed for DUD and PTSD, using two complementary methods. Support for both pathways of risk was found from both approaches; however, some differences emerged. Although the strength of the association was higher for the self-medication model than for the susceptibility model, the pathways were not found to have differential impact in the cross-lagged models. In addition, key sex differences were found in the Cox models, such that the self-medication pathway may be more pronounced for females whereas the susceptibility pathway may be more pronounced for men. Sex differences were less pronounced in the cross-lagged models. Although caution should be used when interpreting causal effects from non-experimental studies, we have increased confidence in the findings reported herein as they were largely congruent across both models. However, we cannot rule out the impact of residual confounding background factors, such as genetic and shared familial factors.

Author contributions

All authors meet the four ICMJE criteria for authorship. Drs Amstadter and Kendler conceptualized the project and research questions. Dr Larsson conducted the analyses. Drs J. and K. Sundquist acquired the data and assisted in the interpretation of study findings. Drs Amstadter and Larsson wrote the manuscript, and Drs J. Sundquist, K. Sundquist, and Kendler provided substantive edits.

Disclosure statement

No potential conflict of interest was reported by the authors.

Data availability

The data are not publicly available owing to restrictions with regard to the nationwide Swedish registers.

Additional information

Funding

This project was supported by the National Institutes of Health [grant number R01DA030005] and the Swedish Research Council [grant number 2020-01175], as well as Avtal om Läkarutbildning och Forskning (ALF) funding from Region Skåne.