Personality factors and substance use

Personality factors are often considered as strong indicators of individual differences in susceptibility to substance reinforcement in the previous theoretical framework of substance use (e.g. [3, 4]). In earlier ages, a large body of studies were interested in the relationships between personality factors of the five-factor model (FFM; e.g. [5]) and substance use problems. The FFM personality factors include openness to experience, extraversion, neuroticism, agreeableness, and conscientiousness. Many of these studies have revealed that FFM personality factors were influential for alcohol use. For example, it was consistently found that high neuroticism, low agreeableness, and low conscientiousness were significantly associated with alcohol use problems (e.g. [6–9]). In a meta-analysis examining the relationships between FFM personality factors and alcohol use using the cross-sectional data from 20 studies [10], it was found that individuals with severe alcohol use problems often showed high neuroticism, low agreeableness and low conscientiousness, consistent with findings of other meta-analysis studies (e.g. [11, 12]). Regarding illicit drug use problems, most previous studies have identified the same influential FFM personality factors. For example, [13] found that women’s drug craving was negatively related to conscientiousness and agreeableness. [14] found that compared with non-clinical participants, opioid-dependent individuals showed higher neuroticism, lower extraversion and lower conscientiousness, but similar levels of openness to experience and agreeableness. Similarly, in [15], cocaine/heroin users were found to show very high neuroticism and very low conscientiousness, and marijuana users were found to show high openness to experience, and low agreeableness and conscientiousness. In summary, regardless of the drug types, high neuroticism and low conscientiousness were found to be the common personality traits shown by users with problematic substance use across different studies [11].

Despite these broader FFM personality factors being well-examined, researchers were also interested in which non-overlapping specific personality facets were related to increased substance use. In previous studies, it was believed that individuals with different personality traits might show differences in susceptibility to the reinforcement effects of using alcohol and drugs [16]. According to a strand of studies in this theme (e.g. [17, 18]), their theoretical reasoning and empirical evidence suggested that increased substance use can be attributable to two different reinforcement processes, the positive reinforcement of substance use for a hedonic purpose and the negative reinforcement of substance use for relieving negative affective status. As such, the effect of personality factors on substance use might be mediated by these two types of reinforcements.

For example, anxiety and depression proneness were identified as individuals’ personality factors influencing their susceptibility to the negative reinforcement effects of using alcohol and drugs as a way to cope with adversities. More specifically, individuals with high anxiety proneness were more likely to seek for alcohol or anxiolytic substances to reduce the tension or other negative affect (e.g. [19, 20]), but less likely to use stimulants such as marijuana (e.g. [21]). Consistent with anxiety proneness, depression proneness was also found to be associated with alcohol use. For example, in a meta-analysis of eight longitudinal studies [22], it was found that depression predicted alcohol consumption quantity and vice versa. Moreover, it was found that depression-sensitive individuals were more likely to use analgesic drugs like opiates [23]. This is because depressive individuals would feel less pain due to affective and social problems by the inhibitory effects of using analgesic drugs or alcohol [5].

Regarding the positive reinforcement of substance use, impulsivity and sensation seeking were found to be strongly associated with substance use. For example, it was found that impulsivity-related personality factors were associated with problematic alcohol or drug use in both adolescents [17, 24, 25] and adults [26]. A recent review study also suggested that generally impulsivity was related to alcohol use, but it pointed out that different dimensions of impulsivity showed different degrees of correlation with the quantity and frequency of alcohol use, drinking problems and alcohol dependence [27]. Therefore, different impulsivity traits would affect the increased alcohol use in different ways [28]. With respect of sensation seeking, individuals with high sensation seeking were found to show elevated nicotine, alcohol, and marijuana use [29] as well as their elevated motivations for substance use [18, 30]. However, it should be noted that although whether impulsivity and sensation seeking indicate a single personality factor is controversial [31], they were found to show distinct mechanisms of influencing substance use [32], and therefore individuals with elevated impulsivity and sensation seeking should receive tailored prevention and intervention programs.

In summary, anxiety proneness, depression proneness, sensation seeking, and impulsivity were found to be four major influential personality factors of substance use in previous studies. The Substance Use Risk Profile Scale (SURPS) [33] further clarified these personality factors as follows: anxiety sensitivity, hopelessness, impulsivity, and sensation seeking. The validity of SURPS has been well investigated in both adults and adolescents and in different cultural contexts (e.g. [34, 35]).

Explanatory item response modeling

A lot of studies applied the IRT framework to DSM-5 SUD criteria by examining their severity, discrimination, and item information (e.g. [36–38]). Notably, IRT analysis has been predominant for detecting the measurement invariance of DSM-5 SUD symptoms at the item level (e.g. [36]), which is called differential item functioning (DIF) analysis.

A typical IRT analysis employs a particular item response model to examine the relationships between examinees’ responses to items and a latent construct. For unidimensional IRT models, a fundamental assumption is that a “causal” common latent variable (e.g. a disorder) influences the responses to items (e.g. criteria). As such, an individual’s responses can be used to reflect his or her position on the underlying disorder continuum. IRT models use different functional forms to link persons’ levels on disorder continuum to the probabilities of endorsing criteria. A commonly used IRT model for dichotomously scored criteria is the Rasch model [39], in which the probability of the endorsement of person j on symptom i can be expressed as follows: (1) In Eq 1, θ_j denotes the person j’s level on the disorder continuum and β_i is used to indicate the criterion severity, which is defined as the location on the disorder continuum where the probability of endorsing a criterion is 0.5 (i.e. 50%). Higher severity parameter values mean that higher disorder levels are required for the criteria endorsement. Using the Rasch model, each individual’s latent disorder level can be estimated. This is advantageous to using sum or mean score as the disorder severity measure because each criterion’s different contribution to the estimation of latent disorder level can be identified by IRT.

Despite the advantages of traditional IRT analysis, it was often used to reveal the psychometric properties or establish the measurement models of instruments. In most substantive areas, however, how to explain the individual differences on item responses or latent construct is of great interest to practitioners and researchers. The explanatory IRT models are established based on the generalized linear mixed model (GLMM), in which item predictors, person predictors, and their interactions can be included as covariates in the same model. Therefore, the explanatory IRT models can integrate the measurement analysis and the statistical analysis for an explanatory purpose within a single model [40].

Using GLMM, Rasch model can be reconstructed in the following way: η_ij = θ_j+β_i+ε_ij, in which β_i indicates the opposite meaning of symptom severity (the higher the value is, the easier the symptom endorsement will be), θ_j has a normal distribution with a mean of 0 and a variance of 1, and ε_ij has a normal distribution with a mean of 0 and a variance of π² / 3. The logit link is used to transfer the probability of endorsement into a continuous scale (η_ij) where (2) As such, item predictors can be incorporated into the model as follows: (3) where X_ik indicates the value of item i on item property k (k = 1, …, K). It should be noted that in the above expression, the individual item parameter β_i is replaced by the linear combination of item properties (e.g. item categories), so the influence of item properties on responses can be represented by their fixed effects estimated by the model. This model is also called “Linear Logistic Test Model”, which allows interaction between different properties of items as well. Similarly, the person properties can be incorporated into Rasch model as follows, (4) where Z_pj indicates the value of person j on person covariate p (p = 1, …, P), and the influence of each person covariate on item responses can be represented by the fixed effect J_p. It should be noted that θ_j in the above expression indicates the remaining person effect accounting for the effects of person covariates. This model is also called “Latent Regression Rasch Model”. Moreover, the interaction between item and person covariates can also be incorporated into the Rasch model as follows, (5) where W_ijh indicates the value of person j on the interaction between item i and a person covariate h. The influence of the interaction can be represented by the fixed effect δ_h. This model can be used to detect DIF if the fixed effect of interaction is significant, which means the item functioning differs between different subgroups (e.g., males and females) given a certain person covariate (e.g. gender). In addition, the interaction between an item covariate and a person covariate can also be incorporated into the same model, which can be used to detect the differential facet functioning.

Obviously, the explanatory item response modeling (EIRM) differs from the traditional item response modeling in that the item and person covariates used for explanatory purposes can be modelled together within the IRT framework. In a typical two-step approach, the measurement model defined by IRT analysis only plays a role in the first step and the outcome variable based on IRT analysis is used in following separate statistical analyses for investigating the predictive effects of other relevant variables. However, using EIRM with person covariates, the group differences can be directly examined within the underlying measurement model. In other words, the differences in the latent disorder level can be parameterized at the group level, which leads to unbiased measures of group differences, differing from the typical two-step approach. For example, using the two-step approach, the expected a priori (EAP) estimates of participants’ latent construct score are firstly obtained by Rasch modeling, which are then used for regression analyses in the following steps. However, the EAP estimates have been shrunk toward the overall population mean by Rasch modeling. Therefore, the regression coefficients by the following analyses cannot reflect the conditional means of the subgroup populations. Additionally, larger measurement error leads to a higher extent to which the EAP estimates shrink to the population mean. As such, the regression coefficients of the two-step approach are largely influenced by measurement and estimation error [41].

Current study

Several weaknesses of previous studies on DSM-5 SUD criteria can be identified. First, the question of how to investigate the influence of personality factors on SUDs was often addressed by a sequence of separate measurement and statistical analyses (e.g. [34]). However, this approach may lead to attenuated effects of personality factors because of the measurement errors of IRT analysis. Second, previous studies investigating the personality factors on substance use often used community samples rather than clinical samples (e.g. [35]). Indeed, a lot of previous studies on this topic used community samples given that recruiting community samples seems to be a standard practice in clinical research [42]. It was often believed that participants from the community samples show less severe substance use problems. Most previous studies in this area focused on how the personality factors of the five-factor model were related to substance use, and they often collected data from community samples such as adolescents and college students. For example, in some of their studies, alcohol use problems were found to be associated with high neuroticism and low agreeableness and conscientiousness [43, 44]. In contrast to studies using community samples, studies examining the effects of personality factors of the five-factor model by clinical samples were much rarer, but they demonstrated consistent findings for some non-drug substances like alcohol [45]. However, as we mentioned earlier, the roles of substance-use-related personality factors (i.e. anxiety sensitivity, hopelessness, impulsivity and sensation seeking) in influencing substance use problems were under-investigated. Although some empirical studies have revealed the effects of the four personality factors by community samples (e.g. [33, 34]), very few generalized these findings to clinical samples. A recent study found that the above four personality dimensions were sensitive factors among patients with alcohol and drug use problems [46], but it used a small inpatient sample and only half of the participants used illicit drugs. Therefore, to examine the generalizability of these findings, a clinical sample with illicit drug use problems is still needed.

This study aims to investigate the relation of personality factors to SUDs using the EIRM approach. The main purposes of this study are two-fold: the validity of DSM-5 SUD symptoms can be further examined by a clinical sample of illicit drug users; how personality factors predict SUD can be identified by incorporating them into the measurement model. In addition, because gender is an influential demographic variable on substance use disorder [47], and DIF by gender was present for some DSM-5 SUD criteria, we include gender as a person covariate in this study. Moreover, due to the fact that alcohol abuse and dependence present much higher prevalence than drug abuse and dependence in the population, and this pattern is severer for men than women (see [48]), the alcohol use and its interaction with gender should also be considered as person covariates for clarifying the unique contribution of personality factors.

The three specific research questions addressed in the present study are as follows: (1) What are the psychometric properties of DSM-5 SUD criteria, and do they show DIF by each of the predictors? (2) Do illicit drug users have latent SUD levels differing between men and women, or drinkers and non-drinkers? (3) Do anxiety sensitivity, hopelessness, impulsivity, and sensation seeking predict illicit drug users’ latent SUD levels, and if they do, how they are related to gender and alcohol use?

Sample

A total of 606 illicit drug users diagnosed with SUDs participated in our study. They were all under treatment at two drug rehabilitation centers located in southwestern China. All participants responded to the paper-form questionnaire under the guidance of one trained administrator, and they were informed that they could quit at any time during the data collection process. As such, all data used in our study was collected by self-report. Written informed consent for participation was obtained from each participant. The research protocol was approved by the ethics review committee of Beijing Normal University.

For this study, 33 cases with missing data on the covariate measures were removed from the original dataset, and thus, a total of 573 cases were used for analysis. The ages of participants ranged from 16 to 62 (M = 28.94, SD = 11.08). Nearly 70% of the participants were male. 71% and 96% of the participants were drinking and smoking regularly before the treatment, respectively. Regarding their ethnic groups, 73% of the participants are Han Chinese and others are from other Chinese ethnic minorities.

It should be noted that participants in our study were dependent illicit drug users rather than recreational drug users. However, it was believed that the recreational use of substances like alcohol, cannabis and ecstasy might influence their dependence on other severe substances like heroin [49]. In our study, heroin and methamphetamine were two primary drugs used by participants with endorsement percentages of 65.79% and 61.08% respectively, and 35.78% of participants used at least two different illicit drugs (see Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Distribution of drug categories.

https://doi.org/10.1371/journal.pone.0217630.t001

Measures

Substance use risk profile scale.

The personality factors relevant to substance use was measured by the 23-item Substance Use Risk Profile Scale (SURPS) [33], which is composed of four subscales: (1) Hopelessness, a depression-associated personality factor which negatively reinforces the drug or alcohol use (seven items, e.g. “I am content.”; (2) Anxiety Sensitivity, a personality factor motivating substance use by reducing the anxiety symptoms (five items, e.g. “It scares me when I'm unable to focus on a task.”); (3) Sensation Seeking, a personality factor associated with thrill-seeking and the desire for stimulating, mood-enhancing effects of substance use (six items, e.g. “I would like to skydive.”); and (4) Impulsivity, a personality factor related to heavier, unconstrained drug use due to a minimal tolerance for negative emotions and rapid response to rewards of substance use (five items, e.g. “Generally, I am an impulsive person.”). The SURPS items were rated on a four-point Likert scale, six of which were inversely scored. The background variables including age, sex, smoking and alcohol use, and ethnicity were also measured.

The SURPS scale used in our study was a Chinese-translated version of the original scale. To make the instrument conceptually equivalent and perform in the same way in the Chinese context, we adopted the approach of forward-translations and back-translations to minimize the measurement bias. Specifically, two translators who were clinical psychology professionals and were also familiar with the terminology of SUDs, were invited to conduct the translation. Thereafter, an expert panel of the original translators and another two clinical psychology professors reviewed the initial translations and address the inadequacy of expressions and any discrepancies in the forward translation, which resulted in a complete Chinese-translated version of SURPS. Finally, the instrument was translated back to English by two researchers living in the English-speaking countries and familiar with the English-language culture. They were asked to focus on the conceptual and cultural equivalence of the instrument. Any discrepancies in the back-translation were further addressed by the expert panel, which resulted in the final version of Chinese SURPS.

To evaluate the psychometric properties of SURPS, we conducted the unidimensionality examination for each personality factor of SURPS given that the four personality factors were analyzed separately in our study (see Table 2). The comparative fit index (CFI), the Tucker Lewis index (TLI), the root mean square error of approximation (RMSEA), and the standardized root mean square residual (SRMR) were used as the model-data fit indices. According to previous studies [51, 52], the following cut-off values for these indices were used to indicate an acceptable fit: above 0.90 for CFI and TLL, below 0.08 for RMSEA and SRMR. We did not present the chi-square statistics because it was found to be too sensitive to large samples [53]. According to the model-data fit indices, the personality factor of hopelessness could obtain a good fit only if the the third and the seventh indicators were set to be correlated with each other. These two items both describe participants’ positive attitudes toward the future, which might be conceptually overlapping to some extent. The coefficient omega was used to indicate the internal consistency for each personality factor (see Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Psychometric properties of personality factors.

https://doi.org/10.1371/journal.pone.0217630.t002

Data analysis strategy

First, the item parameter and participants’ latent scale scores of DSM-5 SUD symptoms were estimated by Rasch modelling without any person covariates. The exploratory factor analysis (EFA) was employed to check the unidimensionality of SUD criteria by the psych package in R [54]. Second, the person covariates predicting illicit drug users’ SUD level were investigated by the model comparison approach using EIRM. A null model without any person predictors would be firstly established, followed by increasingly complex models with more and more person predictors. The model-data fit of the new model was compared to the previous one. Because the previous models are nested within the new models, a likelihood ratio (LR) test can be used for model comparison. A significant LR test statistic indicates the improvement in goodness of model-data fit. At last, to detect the DIF symptoms, the interactions between item and each person predictor were included in EIRM. The EIRM analyses were conducted using the lme4 package in R [55].

Rasch model

Regarding the unidimensionality assumption, the EFA results show that the first five largest eigenvalues of factors are 4.88, 1.01, 0.24, 0.12, and 0.01, and one-factor solution accounts for 44% of the data variance. In addition, according to the scree plot, the eigenvalues drop largely from the first factor to the second factor. Moreover, the factor loadings of SUD criteria range from 0.44 to 0.82. Accordingly, it can be concluded that the illicit drug users’ responses to SUD criteria are mainly influenced by a single factor. According to the results of Rasch modeling, all symptoms present negative severity parameter values (see Table 3) and the histogram shows that the distribution of latent scale scores of SUD is negatively skewed (Fig 1), indicating that most participants presented very high latent SUD level.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Histogram of participants’ latent scale score of SUD by the Rasch model.

https://doi.org/10.1371/journal.pone.0217630.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Descriptive summary of the DSM-5 SUD criteria.

https://doi.org/10.1371/journal.pone.0217630.t003

EIRM

Modelling main effects of gender and alcohol use.

Model 1 was established by adding gender and alcohol use (see Table 4). The difference in deviance values between the null model and model 1 is significant (p < .01), therefore we could statistically infer that model 1 fit data better compared to the null model. Moreover, the fixed effects represented by latent regression coefficients show that both gender and alcohol use significantly predicted the latent SUD level. Specifically, compared to men, women have a lower likelihood of symptom endorsement and thus have lower SUD level (p < .05), and compared to drinkers, non-drinkers have lower latent SUD level (p < .01).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Fixed effects and model comparison results from the explanatory IRT models.

https://doi.org/10.1371/journal.pone.0217630.t004

DIF analyses

The group comparison by person properties assumes that the DSM-5 SUD measure is not biased, and each symptom functions in the same way across different subgroups. The DIF analysis, which investigates whether an item demonstrates differential psychometric properties for different sub groups, is conducted to examine this assumption. In EIRM, it can be addressed by including the person by item interaction. For instance, a significant effect of gender by item means how the item contributes to the estimation of latent construct level differs between men and women, which is not expected due to its confounding influence on group comparison.

Table 5 shows the identified DIF items for each person predictor. Most DSM-5 SUD symptoms show DIF for gender. Specifically, compared to men, the symptoms “craving”, “hazard”, “tolerance” and “withdraw” are severer and thus harder to endorse for women, and the symptoms “give up” and “consistent use” are less severe and thus easier to endorse for women. The item characteristic curve (ICC) of symptom “withdrawal” is shown in Fig 2 as an example of visual presentation for DIF. According to Fig 2, given the same latent SUD level, men consistently have higher endorsement probabilities of symptom “withdrawal”, which indicates that this symptom functions differently between men and women. For the alcohol use, given the same SUD level, women have a lower likelihood of endorsing the symptoms of “social” and “give up” than men do. Another example ICC plot of symptom “social” is presented in Fig 2. Regarding the personality factors, generally, most DSM-5 SUD symptoms show no DIF. However, the symptom “social” shows DIF for impulsivity, which indicates that the symptom “social” is harder to endorse for illicit drug users with higher impulsivity levels. In other words, the “social” symptom is less common for highly impulsive illicit drug users. In addition, the symptom “give up” is harder to endorse and the symptom “tolerance” is easier to endorse for the illicit drug users with higher sensation seeking levels.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. ICC plots of two example DIF symptoms.

The left panel presents the ICC of “withdrawal”, and the right panel presents the ICC of “social”.

https://doi.org/10.1371/journal.pone.0217630.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Identified DIF items by each person predictor.

https://doi.org/10.1371/journal.pone.0217630.t005