Predicting the Onset of Opioid Use Disorder in the Swedish General Population

Abstract

Objective:

Given the public health importance of opioid use disorder (OUD), we sought to understand better its risk predictors in the Swedish general population.

Method:

We examined the Swedish population, born 1950–1970 (n = 2,092,359), and followed through 2018. Using Cox, logistic, and co-sibling models, we explored associations between a wide range of putative risk factors and a first onset of OUD—assessed through medical, criminal, and pharmacy registers—in the entire cohort and in the cohort wherein prior cases of drug use disorder (DUD) were censored.

Results:

OUD was predicted by the following four risk factor domains: (a) externalizing syndromes, especially prior non-opioid DUD; (b) psychopathology; (c) psychosocial factors, including social class and immigrant and marital status; and (d) serious injuries and pain diagnoses. When predicting OUD as the first form of DUD, the importance of pain diagnoses as a predictor increased. Co-sibling analyses suggested that the association of some of these risk factors with OUD onset was likely largely causal, whereas others were a mixture of causal effects and familial confounding. An aggregate risk score from these individual risk factors had reasonable receiver operating characteristic (ROC) curve performance.

Conclusions:

OUD is a multifactorial syndrome for which risk can be meaningfully predicted by prior externalizing syndromes, internalizing and psychotic psychopathology, indicators of psychosocial status, and predictors of pain diagnoses. Some important differences were seen in the prediction of any OUD onset versus OUD onset as the first form of DUD. Much of the effect of these predictors appear, in co-sibling analyses, to likely reflect causal influences.

Europe in general and Sweden in particular have seen increases in opioid consumption over recent decades (Hastie et al., 2014), as well as marked rises in opioid-related deaths (Andersson et al., 2020; Fugelstad et al., 2019). In 2016, Sweden's per capita opioid analgesic consumption was typical for Northern Europe (Degenhardt et al., 2019). Although the level of use in Europe has stabilized more recently, as in the United States (Baldwin et al., 2021), there has been a shift toward the use of short-acting opioids, especially oxycodone and fentanyl, which have a high addictive potential and risk for overdose death (Baldwin et al., 2021; Heilig & Tägil, 2018; Muller et al., 2019). Sweden also has substantial rates of nonmedical use of prescription opioids (Novak et al., 2016). These trends have increased the urgency of understanding the predictors of opioid use disorder (OUD) in general and particularly what features discriminate individuals with OUD from those with non-opioid drug use disorder (noDUD).

Therefore, we examine here a wide range of predictors of OUD in a large general population sample in Sweden, wherein OUD is ascertained from criminal and medical registries, including a recently available primary care registry (Sundquist et al., 2017). We address the following specific issues:

1. In the entire Swedish population and in this population censored for all individuals with a prior noDUD, examine the predictive associations with OUD, both in univariable and multivariable analyses, for a broad host of potential risk factors covering four broad domains: (a) externalizing disorders or behaviors; (b) other psychopathology; (c) psychosocial factors such as low social class and immigration and marital status; and (d) pain diagnoses and serious injuries.

2. Examine sex differences in these results.

3. For those predictors that would be expected to differ in siblings, examine their relationship to OUD in a discordant co-sibling design so as to gain insight into the degree to which their association with OUD results from causal effects versus familial confounding (Ohlsson & Kendler, 2020).

4. Using a split-half method, develop, from multivariable analyses of predictors of OUD, a predictive score for OUD onset and evaluate by area under the curve (AUC) analyses.

Method

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. We secured ethical approval for this study from the Regional Ethical Review Board of Lund University (No. 2008/409, 2012/795, and 2016/679).

The following sources were used to create our data set: Total Population Register, containing information about year of birth and sex; Multi-Generation Register, linking individuals born after 1932 to their parents; the Longitudinal Integration Database for Health Insurance and Labor Market Studies, containing information about education and income from 1990 to 2018; Hospital Discharge Register, containing hospitalizations for Swedish inhabitants from 1964 to 2018; Day Surgery Register, containing diagnosis from 1997 to 2000; Prescribed Drug Register, containing all prescriptions in Sweden picked up by patients from July 2005 to 2018; Outpatient Care Register, containing information from all outpatient clinics from 2001 to 2018; Crime Register, which included national complete data on all convictions in lower court from 1973 to 2018; Swedish Suspicion Register, which included 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 had medical diagnosis from Primary Health Care clinics from counties in Sweden, which counties included 87% of the Swedish population (for details, see Sundquist et al., 2017). We defied baseline as the year 2000 and included everyone residing in Sweden that year and born between 1950 and 1970.

Measures

The outcome measure was OUD, which we assessed using a medical registration of the ICD-10 code; F11, or prescribed drugs for opioid dependence; and N07BC. In addition, we used criminal convictions from 1996 to 2006 and in 2009 when the type of drug used was registered.

A number of possible predictors were assessed using the Swedish Register Data. Individuals’ socioeconomic level was defined using highest level of both parents’ education and categorized into high (university level), mid (upper secondary school), and low (compulsory school only). Marital status was assessed at baseline and divided into the following four categories: unmarried, married, divorced, and widowed. Immigration status was assessed using country of birth, and everyone born outside Sweden were defined as immigrants.

Neighborhood-level socioeconomic status was based on Small Areas for Market Statistics obtained from Statistics Sweden, with an average population of around 1,000 inhabitants. We assessed the socioeconomic characteristics of each neighborhood using an aggregated measure, described in detail elsewhere (Winkleby et al., 2007). The aggregated measure was based on four dimensions of deprivation in the working population ages 25–64; the proportion of people with low income, low education, unemployment, and receipt of social welfare and standardized to have mean 0 and standard deviation 1 each year. Based on the standardized score, the socioeconomic level of each neighborhood was categorized as low (deprivation score above 1), mid (deprivation score between -1 and 1), or high (deprivation score below -1).

For the specific criteria used of identifying cases of alcohol use disorder (AUD), criminal behavior (CB), drug use disorder other than OUD, depression and anxiety (DAD), bipolar disorder and other nonaffective psychosis (BPN), pain, and injuries, see Appendix Tables 1, 2, and 3.

Table 1.

Descriptive statistics for our study sample

Variable of interest	Total number	Men	Women
Total number	2,092,359	1,067,337	1,025,022
Opioid use disorder	11,721 (0.56%)	6,370 (0.60%)	5,351 (0.52%)
High education	729,780 (34.88%)	327,103 (30.65%)	402,677 (39.28%)
Mid education	1,076,262 (51.44%)	559,661 (52.44%)	516,601 (50.40%)
Low education	286,317 (13.68%)	180573 (16.92%)	105,744 (10.32%)
Unmarried, baseline	868,149 (41.49%)	498,117 (46.67%)	370,032 (36.10%)
Married, baseline	991,444 (47.38%)	469,027 (43.94%)	522,417 (50.97%)
Divorced, baseline	224,056 (10.71%)	97,949 (9.18%)	126,107 (12.30%)
Widowed, baseline	8,710 (0.42%)	2,244 (0.21%)	6,466 (0.63%)
Not Swedish born	66,756 (3.19%)	33,257 (3.12%)	33,499 (3.27%)
Low neighborhood SES	559,806 (26.75%)	277,370 (25.99%)	282,436 (27.55%)
Mid neighborhood SES	1,273,270 (60.85%)	652,583 (61.14%)	620,687 (60.55%)
High neighborhood SES	259,283 (12.39%)	137,384 (12.87%)	121,899 (11.89%)
AUD, any time	149,086 (7.13%)	107,738 (10.09%)	41,348 (4.03%)
Criminal behavior, any time	291,906 (13.95%)	226,306 (21.20%)	65,600 (6.40%)
Other DUD, any time	68,278 (3.26%)	44,334 (4.15%)	23,944 (2.34%)
DAD, any time	399,089 (19.07%)	154,693 (14.49%)	244,396 (23.84%)
BPN, any time	55,003 (2.63%)	27,040 (2.53%)	27,963 (2.73%)
Pain diagnosis, any time	716,927 (34.26%)	324,742 (30.43%)	392,185 (38.26%)
Injuries, any time	981,883 (46.93%)	525,641 (49.25%)	456,242 (44.51%)

Notes: Data are n or n (%). SES = socioeconomic status; AUD = alcohol use disorder; DUD = drug use disorder; DAD = depression or anxiety disorder; BPN = bipolar and nonaffective psychosis.

Table 2.

Univariable and multivariable hazard ratios (HRs) with first onset of an opioid use disorder for range of putative risk factors in our total sample and in the total sample cases of prior drug use disorder (DUD) population censored

Putative risk factor	Univariable		Multivariable
	Total sample HR [95% CI]	No prior DUD HR [95% CI]	Total sample HR [95% CI]	No prior DUD HR [95% CI]
Men vs. women	1.16 [1.12, 1.20]	0.86 [0.81, 0.90]	0.86 [0.83, 0.90]	0.85 [0.81, 0.90]
Birth year	1.04 [1.04, 1.05]	1.06 [1.05, 1.07]	0.96 [0.95, 0.96]	0.97 [0.96, 0.98]
Mid vs. high parental education	1.38 [1.31, 1.46]	1.34 [1.26, 1.43]	1.05 [0.99, 1.12]	1.10 [1.02, 1.18]
Low vs. high parental education	1.30 [1.23, 1.38]	1.93 [1.78, 2.10]	1.07 [1.01, 1.13]	1.08 [1.00, 1.17]
Unmarried, baseline, vs. married	1.88 [1.80, 1.96]	1.34 [1.27, 1.42]	1.01 [0.96, 1.05]	1.01 [0.96, 1.08]
Divorced, baseline, vs. married	2.81 [2.66, 2.96]	1.89 [1.76, 2.03]	1.13 [1.07, 1.19]	1.22 [1.14, 1.32]
Widowed, baseline, vs. married	2.41 [1.90, 3.05]	1.35 [0.96, 1.92]	1.33 [1.05, 1.69]	1.14 [0.81, 1.62]
Not Swedish born	1.65 [1.51, 1.80]	1.44 [1.26, 1.64]	1.05 [0.96, 1.15]	1.03 [0.91, 1.18]
Low vs. high NSES, baseline	2.82 [2.66, 2.99]	1.19 [1.10, 1.29]	1.23 [1.15, 1.31]	1.35 [1.24, 1.47]
Mid vs. high NSES, baseline	1.46 [1.39, 1.53]	1.25 [1.16, 1.35]	1.09 [1.04, 1.15]	1.19 [1.11, 1.27]
AUD, time dependent	15.25 [14.66, 15.87]	7.19 [6.74, 7.68]	2.11 [2.00, 2.22]	3.14 [2.93, 3.38]
Criminal behavior, time dependent	7.44 [7.15, 7.74]	3.09 [2.92, 3.29]	1.72 [1.64, 1.81]	1.68 [1.58, 1.79]
Non-opioid DUD, time dependent	40.45 [38.95, 42.02]	-	11.29 [10.70, 11.90]	-
DAD, time dependent	8.38 [8.06, 8.71]	6.23 [5.90, 6.57]	2.50 [2.39, 2.62]	3.21 [3.02, 3.40]
BPN, time dependent	6.20 [5.83, 6.59]	3.81 [3.42, 4.25]	1.05 [0.98, 1.12]	1.54 [1.37, 1.73]
Pain diagnosis, time dependent	5.41 [5.20, 5.63]	6.31 [5.98, 6.66]	3.30 [3.16, 3.45]	4.59 [4.33, 4.87]
Injuries, time dependent	2.78 [2.67, 2.89]	2.40 [2.28, 2.53]	1.39 [1.33, 1.46]	1.43 [1.35, 1.51]

Notes: 95% CI = 95% confidence interval; NSES = neighborhood socioeconomic status; AUD = alcohol use disorder; DAD = depression or anxiety disorder; BPN = bipolar and nonaffective psychosis.

Table 3.

Examination of sex effects for the association between risk factors and first onset of opioid use disorder in the entire sample in univariable and multivariable analyses

Putative risk factor	Univariable analyses			Multivariable analyses
	Men HR [95% CI]	Women HR [95% CI]	Men vs. women HR [95% CI]	Men HR [95% CI]	Women HR [95% CI]	Men vs. women HR [95% CI]
Men vs. women			1.16 [1.12 1.20]			0.98 [0.73, 1.33]
Birth year	1.04 [1.04, 1.05]	1.04 [1.04, 1.05]	0.99 [0.99, 1.00]	0.96 [0.95, 0.96]	0.96 [0.95, 0.96]	1.00 [0.99, 1.01]
Mid vs. high parental education	1.50 [1.40, 1.61]	1.41 [1.32, 1.52]	1.06 [0.96, 1.17]	1.06 [0.98, 1.13]	1.13 [1.05, 1.21]	0.93 [0.82, 1.05]
Low vs. high parental education	3.00 [2.77, 3.25]	2.62 [2.40, 2.86]	1.14 [1.02, 1.29]	1.18 [1.09,1 .28]	1.27 [1.16, 1.39]	0.94 [0.85, 1.04]
Unmarried, baseline, vs. married	2.14 [2.01, 2.27]	1.60 [1.50, 1.71]	1.33 [1.22, 1.45]	0.98 [0.92, 1.04]	1.02 [0.95, 1.09]	0.97 [0.88, 1.06]
Divorced, baseline, vs. married	2.90 [2.68, 3.15]	2.70 [2.51, 2.91]	1.08 [0.96, 1.20]	1.06 [0.98, 1.15]	1.19 [1.10, 1.28]	0.90 [0.80, 1.00]
Widowed, baseline, vs. married	2.37 [1.49, 3.77]	2.31 [1.75, 3.05]	1.02 [0.60, 1.76]	1.43 [0.90, 2.27]	1.32 [1.00, 1.75]	1.08 [0.63, 1.85]
Not Swedish born	1.67 [1.51, 1.91]	1.59 [1.39, 1.82]	1.07 [0.89, 1.28]	1.09 [0.95, 1.24]	1.01 [0.90, 1.14]	0.93 [0.78, 1.12]
Low vs. high NSES, baseline	1.36 [1.26, 1.47]	1.24 [1.14, 1.35]	1.09 [0.98, 1.23]	1.03 [0.95, 1.12]	1.08 [0.99, 1.17]	0.96 [0.85, 1.08]
Mid vs. high NSES, baseline	1.34 [1.34, 1.57]	1.30 [1.20, 1.42]	1.11 [0.99, 1.25]	1.06 [0.98, 1.14]	1.08 [0.99, 1.17]	0.98 [0.87, 1.10]
AUD, time dependent	15.48 [14.72, 16.29]	14.90 [13.99, 15.87]	1.04 [0.96, 1.13]	2.50 [2.34, 2.67]	1.69 [1.56, 1.83]	1.48 [1.33, 1.64]
Criminal behavior, time dependent	7.07 [6.71, 7.45]	7.94 [7.49, 8.43]	0.89 [0.82, 0.96]	1.61 [1.51, 1.72]	1.86 [1.73, 2.00]	0.87 [0.79, 0.96]
Other DUD, time dependent	39.83 [37.86, 41.90]	41.27 [38.98, 43.69]	0.97 [0.89, 1.04]	11.57 [10.78, 12.41]	10.85 [10.01, 11.75]	1.07 [0.96, 1.19]
DAD, time dependent	8.84 [8.39, 9.31]	7.88 [7.45, 8.33]	1.12 [1.04, 1.21]	2.34 [2.20, 2.48]	2.71 [2.54, 2.90]	0.86 [0.79, 0.94]
BPN, time dependent	6.26 [5.77, 6.80]	6.13 [5.59, 6.71]	1.02 [0.90, 1.16]	1.03 [0.95, 1.13]	1.08 [0.98, 1.19]	0.96 [0.83, 1.09]
Pain diagnosis, time dependent	5.14 [4.88, 5.41]	5.76 [5.44, 6.09]	0.89 [0.83, 0.96]	3.09 [2.92, 3.28]	3.54 [3.32, 3.77]	0.87 [0.80, 0.95]
Injuries, time dependent	2.62 [2.49, 2.76]	2.98 [2.81, 3.15]	0.88 [0.82, 0.95]	1.36 [1.29, 1.44]	1.44 [1.35, 1.53]	0.95 [0.87, 1.03]

Notes: HR = hazard ratio; 95% CI = 95% confidence interval; NSES = neighborhood socioeconomic status; AUD = alcohol use disorder; DUD = drug use disorder; DAD = depression or anxiety disorder; BPN = bipolar and nonaffective psychosis.

Statistical analysis

First, we ran univariable Cox regressions for all individual predictors. Marital status and neighborhood socioeconomic status were assessed at baseline, in the year 2000. AUD, CB, other DUD, DAD, BPN, any pain diagnosis, and any injury were all included as time-dependent covariates. We then used age as the underlying time scale and followed individuals until death or end of follow-up, December 2018. We further ran the parallel models censoring at noDUD. All models included sex as a fixed effect. Next, we ran the multivariable models, including all predictors.

Thereafter, we investigated sex differences by running the corresponding univariable and multivariable models and allowed the associations of the individual predictors to vary by sex, by including an interaction term. We present the sexspecific associations and the interaction term, representing the differences between the sex-specific associations.

To assess unmeasured genetic confounding, we ran uni-variable co-sibling models in full sibling pairs discordant for risk factor exposure, or discordant for age at exposure. Each pair is treated as a stratum, and the estimated hazard ratio (HR) represents the association within full sibling pairs, sharing half of their genes. For families with multiple siblings, all possible pairs were included.

To evaluate the predicative effect of these covariates, we assessed all predictors at baseline, making a random split of the sample into two halves. We ran a multivariable logistic regression on the training half and then, based on the estimated parameters, evaluated our model using the second, testing half. First, we calculated the predicted probabilities for each individual in the testing half and divided the sample into risk deciles. For each decile, we then calculated the proportion of individuals with OUD. We further evaluated the causality of theses risk scores by conducting both a population and a co-sibling analysis, where the risk decile was included as the explanatory variable. Finally, we constructed a ROC curve and calculated the AUC to illustrate the predictive ability of our risk factors. All statistical analyses were run in SAS Version 9.4 (SAS Institute Inc., Cary, NC) and R Studio Version 4.0.3 (RStudio Team, 2020).

Results

Descriptive statistics

Our population cohort, born 1950–1970, includes 2,092,342 individuals (51.0% men) with a mean age (SD) at follow-up (death, OUD onset, or December 31, 2018) of 57.11 (6.18) years (Table 1). Individuals with an OUD diagnosis before 2000 were censored from the sample. OUD was relatively rare, diagnosed in 0.56% of the sample, and had a modest male preponderance. Of this sample, 71% had no registration for DUD before their first OUD registration.

Table 1 also contains the frequencies of the putative risk factors included in our analyses. noDUD was seen in 3.27% of the sample, whereas AUD and CB were much more common, with prevalences of 7.13% and 13.95%, respectively. All of these outcomes were considerably more common in men. We included two measures of psychopathology, internalizing psychopathology (DAD) and psychotic illness (BPN), which occurred in 19.07% and 2.63% of the population, respectively. Because of our interest in identifying potential iatrogenic forms of OUD, we also included a history of pain diagnoses and serious injuries, both quite common in the Swedish general population, as they are associated with the receipt of prescription opioids.

Univariable and multivariable analyses for OUD onset

We first examine, in our entire cohort, univariable analyses of the prediction of OUD, using time-dependent covariate modeling with temporally discrete risk factors for the time-dependent predictors (Table 2, left column). All predictors are statistically significant. Prior noDUD is far and away the strongest predictor of OUD, followed by AUD, DAD, CB, and BPN. That is, prior psychopathology and CB were far stronger individual predictors than psychosocial and marital-related predictors. The next most potent was a pain diagnosis followed by a range of other predictors, with HRs varying from 1.3 to 2.8. Men were at a moderately elevated risk for OUD compared with women.

We next examined the prediction of onset of OUD, censoring cases with prior noDUD (Table 2, second column). That is, we sought to clarify the predictors of OUD onset when it was the first form of drug use disorder. All predictors were significant, with the exception of widowhood. The HRs for AUD, CB, marital status, neighborhood deprivation, and psychotic psychiatric disorders were considerably lower than those seen in the prior analyses. The effect of DAD was modestly attenuated, and the effects of low parental education and pain diagnoses increased in strength. Furthermore, the effect of sex on risk was reversed with the risk for OUD now modestly higher in women.

The third column of Table 2 presents the results of our multivariable analyses predicting OUD in the entire population. As expected, a substantial shrinkage was seen in the HRs from the parallel univariable model. Mid versus high parental education, immigration, single marital status, and psychotic psychiatric disorders were no longer significant.

Prior noDUD remained by far the strongest predictor, followed by a pain diagnosis, DAD, and AUD.

In the fourth column of Table 2, we present the multivariable analyses given censoring of individuals with prior noDUD. Being unmarried, widowed, or an immigrant no longer significantly predicted OUD. The remaining HRs, as expected, were considerably reduced from the parallel uni-variable analyses. The most significant predictors of OUD were, in order of magnitude, pain diagnoses, DAD, AUD, and BPN.

Analyses of predictors by sex

Table 3 depicts the comparison of the performance of the predictors of OUD in men and women in our entire cohort for both the univariable and multivariable analyses. For the univariable analyses, three results are of interest. First, as suggested by the prevalence results, OUD was significantly but only modestly more common in men than in women. Second, only 6 of the 16 risk factors differed significantly across sexes. They were evenly divided with three risk factors (unmarried, community deprivation, and DAD) significantly stronger in men and three (CB, pain diagnosis, and injuries) significantly stronger in women. Third, even for these predictors, most of the differences were modest in magnitude with only one—unmarried—producing an HR outside the range of 0.88–1.14.

In the multivariable analyses, risk for OUD no longer differed by sex, and only 4 of the 16 risk factors were now significantly different in men and women. AUD was a more potent predictor of OUD in men, and CB, DAD, and pain diagnoses were significantly stronger OUD predictors in women.

Co-sibling analyses

In an effort to determine the degree to which our risk factor associations with OUD outcomes were potentially causal versus a result of familial confounding, we conducted a co-sibling analysis of those risk factors that would differ in siblings. These analyses compared the risk for OUD in siblings discordant for exposure to the specific risk factor. Of the 13 co-sibling univariable analyses conducted on our full sample, all but one (sex) declined in magnitude from the standard univariable results and all remained significant (Table 4). Examining the reduction in HR in the co-sibling design versus standard analyses—which reflects the level of familial confounding—our predictors can be usefully placed into three categories. Five predictors (sex, mid vs. high community deprivation, injuries, unmarried, and pain diagnoses) all had ratios greater than 0.75, suggesting minimal levels of familial confounding. Five putative risk factors (divorce, widowed, community deprivation, DAD, and BPN) had ratios from 0.60 to 0.70, consistent with moderate levels of familial confounding. Finally, three predictors—the externalizing disorders of CB, AUD, and noDUD—had ratios ranging from 0.50 to 0.60, demonstrating considerable, but far from complete, confounding.

Table 4.

Co-sibling analyses of risk factors for opioid use disorder

	Standard analyses	Co-sibling analyses	Ratio co-sibling/
Putative risk factor	HR [95% CI]	HR [95% CI]	standard
Men vs. women	1.16 [1.12, 1.20]	1.18 [1.11, 1.25]	1.02
Unmarried, baseline, vs. married	1.88 [1.80, 1.96]	1.63 [1.52, 1.76]	0.87
Divorced, baseline, vs. married	2.81 [2.66, 2.96]	1.90 [1.74, 2.08]	0.68
Widowed, baseline, vs. married	2.41 [1.90, 3.05]	1.58 [1.03, 2.42]	0.66
Low vs. high NSES, baseline	2.82 [2.66, 2.99]	1.80 [1.62, 2.00]	0.64
Mid vs. high NSES, baseline	1.46 [1.39, 1.53]	1.36 [1.25, 1.49]	0.93
AUD, time dependent	15.25 [14.66, 15.87]	8.12 [7.29, 9.04]	0.53
Criminal behavior, time dependent	7.44 [7.15, 7.74]	4.17 [3.83, 4.54]	0.56
Other DUD, time dependent	40.45 [38.95, 42.02]	20.36 [17.49, 23.70]	0.50
DAD, time dependent	8.38 [8.06, 8.71]	5.62 [5.15, 6.13]	0.68
BPN, any time	6.20 [5.83, 6.59]	3.90 [3.44, 4.42]	0.63
Pain diagnosis, time dependent	5.41 [5.20, 5.63]	4.21 [3.89, 4.55]	0.78
Injuries, time dependent	2.78 [2.67, 2.89]	2.44 [2.28, 2.61]	0.88

Notes: HR = hazard ratio; 95% CI = 95% confidence interval; NSES = neighborhood socioeconomic status; AUD = alcohol use disorder; DUD = drug use disorder; DAD = depression or anxiety disorder; BPN = bipolar and nonaffective psychosis.

Construction of an aggregate risk score

Finally, to evaluate the predictive effect of the explanatory variables, we constructed an aggregate risk score for the prediction of OUD both in the entire sample and censoring out individuals with prior noOUD. Detailed results of this are presented in Appendix Table 1. Individuals’ risk decile was determined based on the result from the logistic model on the training half, and Figure 1a shows the performance of this score. The top panel shows the observed probabilities of OUD for each estimated risk deciles in the testing half, and the bottom panel shows the results from a ROC analysis. We saw a large increase in risk in the upper decile, which contains 39.0% of affected individuals in the population. The AUC from the ROC analysis equaled 0.708. Per unit of this risk score, the OR for OUD [with 95% CIs] was 1.21 [1.20, 1.21]. In our co-sibling analyses, this was estimated at 1.25 [1.22, 1.28], suggesting only modest familial confounding in the observed association.

Figure 1A.

Upper panel: In entire cohort: Risk of opioid use disorder (OUD) on the y-axis as a function of the decile of the risk score (on the x-axis) calculated from an analysis of the entire cohort. Lower panel: Excluding individuals with non-opioid drug use disorder: A receiver operator curve for the risk score for the prediction of opioid use disorder (OUD) calculated from an analysis of the entire cohort (solid line) compared with the null (dotted line). Sensitivity of the score is on the y-axis and 1-specificity is on the x-axis. The area under the curve (AUC) value is 0.708.

Figure 1b presents results from the prediction of OUD as the first form of drug abuse. In the upper panel, the risk curve shows a more gradual increase across the deciles, with 34.7% of the affected individuals in the top decile. In the lower panel, the AUC is slightly greater than that seen in the prediction of all OUD: 0.745. Per unit of this risk score, the OR for OUD [with 95% CIs] was 1.88 [1.82, 1.93].

Figure 1B.

Upper panel: Risk of opioid use disorder (OUD) on the y-axis as a function of the decile of the risk score (on the x-axis) calculated from an analysis of the cohort censoring all cases with prior non-opioid drug use disorder. Lower panel: A receiver operator characteristic curve for the risk score for the prediction of OUD calculated from an analysis of the cohort censoring all cases with prior non-opioid drug use disorder (solid line) compared with the null (dotted line). Sensitivity of the score is on the y-axis and 1-specificity is on the x-axis. The area under the curve (AUC) value is 0.745.

Discussion

We sought to address four main issues in this article. We first examined, in univariable and multivariable analyses, the prediction of onset of OUD in our entire Swedish cohort. Consistent with prior findings, we showed that OUD was most strongly predicted by prior non-opioid substance abuse and dependence (Blanco et al., 2019; Green, 2017; Huang et al., 2006; Kerridge et al., 2015; McCabe et al., 2008; Webster, 2017). The next strongest predictors were the other externalizing syndromes of AUD and CB (Blanco et al., 2019; Kerridge et al., 2015; Saha et al., 2016; Webster, 2017). Consistent with prior epidemiological studies, we also found a substantial association between OUD risk and prior internalizing and psychotic psychiatric disorders (Grella et al., 2009; Kerridge et al., 2015; Regier et al., 1990; Saha et al., 2016). Next in importance was a diagnosis of a pain syndrome, suggesting the possible role of iatrogenic OUD developing after first opioid exposure as a part of their health care (Beauchamp et al., 2014; Blanco et al., 2016; Higgins et al., 2018; Moe et al., 2019).

Consistent with other studies and prior Swedish data on the impact of marital status on risk for AUD, being single or being divorced increased the risk for OUD (Kendler et al., 2016, 2017; Kerridge et al., 2015; Saha et al., 2016). We also found modest effects for immigration status and, congruent with a range of prior studies, low socioeconomic status (Green, 2017; Kerridge et al., 2015; Saha et al., 2016; Webster, 2017). In the multivariable analyses, all HRs declined substantially, but it is noteworthy that the two strongest were noDUD and a pain diagnosis. Of note, most of the above-cited literature comes from interview-based surveys where the associations were assessed based on current and lifetime retrospective reports. However, our study also replicated risk factors for OUD demonstrated in a prospective study of OUD incidence in the National Epidemiologic Survey on Alcohol and Related Conditions, including marital status, sex, ethnicity, and prior psychopathology, especially prior other substance use disorders (Katz et al., 2013).

Our subsequent analyses addressed a related but somewhat distinct question—what predicted OUD onset as the first form of drug abuse? Of note, the strongest univariable HRs differed from that seen in our first analyses, with AUD, pain diagnosis, and DAD producing the highest risks. In moving to the multivariable analyses, the history of a pain diagnosis became the strongest single risk factor for the development of OUD.

Our second major issue was the degree to which the associations found between our predictors and OUD differed across sexes. Consistent with prior reports (Blanco et al., 2019), we found considerably lower male excess in OUD in our sample (Table 1, male:female ratio of 1.15:1) compared with noDUD (1.77:1). Despite our large samples and associated high statistical power, significant differences in the impact of our risk factors on OUD risk in men versus women were uncommon, and when present of modest magnitude. Consistent with results from the previously cited incidence study (Katz et al., 2013), we found that the large majority of predictors of OUD seem to act similarly in the two sexes.

Although the demonstration of prospective associations between putative risk factors and key outcomes is an important interim goal of epidemiology, to inform prevention programs, information about the potential causal nature of these associations is crucial. For virtually all of the risk factors examined here, a randomized controlled trial is either impractical and/or unethical. Of the many possible methods for causal inference in non-experimental samples, we here use the discordant co-sibling method (see Ohlsson & Kendler, 2020, for more details). Instead of contrasting the risk for an exposed individual to unexposed members of the general population, we instead, in our third set of analyses, compare an exposed individual to his or her unexposed full sibling. This controls not only for both the family and community-rearing environment but also for genetic effects.

Our co-sibling findings grouped our individual risk factors into three groups. Those that had the last evidence of confounding included both variables in our analyses—pain diagnoses and major injuries—which potentially indexed iatrogenic OUD, suggesting that these risk factors were likely largely causal in their impact. The middle group included single and divorced marital status, which prior co-relative analyses in Sweden suggested were also largely, but not entirely, causal in their effect on risk for AUD (Kendler et al., 2016, 2017). Consistent with prior evidence for substantial genetic correlations between risk for OUD and other externalizing syndromes (Kendler et al., 2003; Tsuang et al., 1998), especially other forms of DUD, the highest, but still moderate, level of confounding was seen for noDUD, CB, and AUD.

Fourth, to assess the magnitude of predictive power contained in our set of risk factors for OUD, we constructed, from them, a risk score in a random half of our sample and tested its performance in the second half. The total predictive power was substantial and showed that, as expected, a large proportion of the affected cases in the population came from those with high aggregate risk.

Limitations

These results should be interpreted in the context of three potentially important methodological limitations. First, we detected subjects with both non-opioid and opioid DUD from medical, legal, and pharmacy records. Although this method does not require respondent cooperation or accurate recall, it could produce false-negative and false-positive diagnoses. Although large interview studies of DUD prevalence do not exist in Sweden, lifetime prevalence in nearby Norway was similar to the rates we detected in Sweden (Kringlen et al., 2001). This suggests that we did not markedly under-ascertain cases of DUD in Sweden, although we would have missed all cases who never came to medical attention or had DUD-related legal problems. The validity of our definition of DUD in Sweden is also supported by the high rates of concordance for registration observed across our different ascertainment methods (Kendler et al., 2015) and findings from twin studies, which suggest heritability estimates similar to those obtained from personal interview–based twin studies (Kendler et al., 2000, 2013; Tsuang et al., 1996). However, access to treatment for OUD is lower in Sweden than in other Nordic countries (Gedeon et al., 2019), so that could result in some under-ascertainment of OUD in our sample.

Second, although it would have seemed logical to examine, in these analyses, the receipt of an opioid prescription as a potential risk factor, such data in Sweden are available only from 2005. Since it is crucial in these analyses to examine the first such prescription and the subsequent risk for first onset of OUD, this would need to be studied in a much more restricted and younger birth cohort than used here to give us maximal power to examine other risk factors. Also, we do not have country-wide information about smoking, so we could not include that as a potential risk factor for OUD.

Third, our measure of age at onset was only approximate as it reflected age at first registration from OUD-associated medical or legal problems. This could have results in some misclassifications of our time-dependent covariate results. Where we assumed that a risk factor preceded OUD onset, it might have occurred in the early stages of OUD between “true” onset and first registration.

Finally, our results might not fully extrapolate to other countries. Compared with the United States, Sweden has a fully socialized health care system.

Conclusions

Consistent with the prior literature (Green, 2017; Webster, 2017), in the Swedish general population, we found that OUD was predicted by four domains of risk factors: (a) prior externalizing disorders or behaviors, especially noDUD; (b) prior psychopathology; (c) a range of psychosocial factors reflecting low social class and marital status; and (d) predictors of potential receipt of opioid prescriptions. We found some similarities but important differences when we predicted OUD onset in the entire population versus in those without prior DUD, with predictors of receipt of opioid prescriptions becoming more important in the latter analyses. Co-sibling analyses suggested that the association of some of these risk factors with OUD onset were likely largely causal (including indices of possible opioid prescriptions), whereas others, including past externalizing syndromes, were a mixture of causal effects and familial confounding. Finally, we constructed a risk score from these individual risk factors with reasonable ROC performance.