Additionally, receiving a larger number of days’ supply of prescription opioids was a predictor of an opioid use disorder diagnosis , as was having a higher average daily dose . In addition to demographic and other markers, behaviorally-based criteria have been successfully used to identify problematic cases of prescription drug misuse . In a recent study, clinical expert raters identified key indicators of misuse, including interpersonal problems, arrest history, multiple opioid use, use for no identifiable reason, and comorbid other substance misuse, and used these indicators along with known indicators of misuse to improve accuracy in identifying misuse. This study indicates that multiple sources of data, particularly those regarding different domains of functioning, may best identify those at risk for opioid abuse and dependence.Previous studies have also linked problematic use of prescription drugs and mental health diagnoses. Nonmedical use of opioids has been associated with panic, depressive, social phobic or agoraphobic symptoms, and the overall number of psychiatric symptoms endorsed . Development of opioid abuse and dependence has also been associated with non-opioid substance use and mental health disorders . Recent prospective research has indicated that non-medical use of prescription medications,grow lights shelving including opioids, places individuals at risk for unipolar depressive, bipolar, and anxiety disorders . The converse relationship may also be true: othermental health conditions may predispose individuals to misuse opioids.
In a recent review of the known factors predicting opioid misuse, the authors caution that although many mental health diagnoses may be risk factors for opioid misuse, these conditions are likely to be concealed due to stigma, and some individuals may choose to take prescription opioids to treat undiagnosed cooccurring disorders rather than the appropriate psychiatric medication . This study seeks to identify demographic and healthcare related variables that predict the development of opioid abuse or dependence, utilizing data obtained from the Thomson Reuters MarketScan Commercial Claims and Encounters database, which contains information about commercially insured and Medicare eligible patients. The use of a large sample, physician-diagnosed disorders, and comprehensive demographic and health care utilization data enable detailed analysis of individuals at risk for the development for opiate abuse or dependence. First, individuals diagnosed with opioid use disorders will be compared with those who are not given opioid use diagnoses on a variety of domains. Second, the use of mathematical modeling techniques will aid in identifying people who are at risk for the development of opioid abuse or dependence. Patients within the CCAE database who had at least one opioid prescription claim between January 1, 2000 and December 31, 2008 were identified. Patients were included if they maintained continuous insurance eligibility for 6 months prior to, and 2 years beyond, this initial prescription claim . Individuals who subsequently received an ICD-9 CM diagnosis of opioid abuse or dependence were classified as those with opioid use disorders, hereafter referred to as OUDs, , and individuals who did not receive a subsequent opioid abuse or dependence diagnosis were classified as those without opioid use disorders, hereafter referred to as non-OUDs . Of the OUDs, 266 received a diagnosis of opioid abuse, and the remaining 2,647 received a diagnosis of opioid dependence. Abuse and dependence cases were therefore grouped together for the purpose of analyses from this point for the following reasons: 1) over 90% of the cases fell into the more serious category of dependence, 2) an abuse diagnosis is often a precursor to dependence, and 3) the clinical distinction between abuse and dependence is less important than the presence or absence of an addictive condition.
Furthermore, the distinction between abuse and dependence has been eliminated in the Diagnostic and Statistical Manual of Mental Disorders , and replaced with opioid use disorders. The first set of planned comparisons involved conducting either t-tests or chi-square analyses to test for statistically significant differences between cases and controls on a variety of variables present in the database. These analyses also served the purpose of identifying variables of interest for the mathematical modeling to be conducted in the next step. With regard to mental health diagnoses and co-occurring substance use disorders, the predictor variables were not time dependent . Once the variables that statistically discriminated cases from controls were identified, significant interactions between these variables were identified using CHAID analyses. The goal of a CHAID analysis is to find homogeneous clusters of a response variable where clusters are defined by the levels in a set of predictor variables. Particular emphasis is placed on the interaction of the predictor variables. The algorithm splits the population according to levels in the predictor variable, which make the responses within the resultant groups as similar as possible and the average between groups as different as possible . Significant interactions detected through CHAID were reviewed by the research team and included in the subsequent logistic regression model if the following conditions were met: 1) the levels in the predictor variable split the groups such that there was at least a 10% difference between the resultant groupings, and 2) a minimum of 10 participants per cell would need to result from the interaction split in order to be meaningful for future modeling. Once the significant variables identified through the bivariate analyses were selected and the CHAID analyses performed, we divided the sample into a build set comprised of 70% of the participants and a validation set comprised of the remaining 30%. The research team devised and tested a series of 18 logistic regression models to fit the data, with variables selected on the basis of several criteria: 1) varying degrees of parsimony, from the simplest demographic variables only to the all-inclusive model using every significant variable and interaction; 2) clinical setting, with models comprised of all mental health variables, or all pharmacy data, for example; and 3) all models were tested both with and without the interaction variables found through the CHAID analyses. Each model was tested using the validation set.
Global null hypothesis tests were used to determine presence of one or more significant predictor variables. This index was selected a priori over other fit indices because the models were not nested,plant drying rack built from the same database, and based on a large sample size. In considering fit indices, we did not want to eliminate variables that could be of potential use in future models, which the BIC is prone to do, particularly in large sample sizes . We wanted to produce a model that favored sensitivity over parsimony, as the AIC does . The choice of the best fitting model was based on the AIC, the relative overall parsimony of the model, and the predictive ability of the model to identify OUDs in the validation set.For the first series of analyses, bivariate comparisons of OUDs and non-OUDs were completed on variables related to demographics, medical service utilization, cooccurring conditions, and concomitant medication usage. The exact variables of interest were chosen by a team of researchers with expertise in pharmacoeconomics, public health, substance misuse, and mental health. Table 1 presents demographic metrics for OUDs and non-OUDs. As expected, OUDs were more likely than non-OUDs to be younger and male . OUDs were also more likely to be a spouse or dependent, rather than the primary insured individual in the plan; 60.1% of non-OUDs were the primary insured person, whereas 43.2% of OUDs were the primary insured. Data regarding participants’ opioid utilization is presented in Table 2. The number of opioid classes differed significantly between OUDs and non-OUDs . OUDs also had a higher mean count of both short acting and long acting opioids . Similarly, the number of days of opioid supply that individuals were prescribed during the study period was also different between groups, with OUDs receiving an average of 272.5 days’ supply of opioids, and nonOUDs an average of 33.2 days’ supply of opioids. Identified OUDs also had a higher number of opioid units dispensed . Notably, the number of pharmacies visited to fill opioid prescriptions differed significantly between groups, with OUDs visiting an average of 3.3 pharmacies per year, compared to 1.3 for non-OUDs. Annual medical service utilization rates also differed significantly between groups, as shown in Table 2. OUDs had significantly more physician visits , outpatient mental health visits , inpatient admissions , inpatient mental health admissions , hospitalization days , mental health hospitalization days , and emergency department encounters . Mental health diagnoses also significantly differentiated the two groups, as shown in Table 3. OUDs were more likely to have diagnoses of anxiety, mood, pain, personality, somatoform, and psychotic disorders than non-OUDs. Whereas 57.7% of OUDs had another substance use disorder diagnosis, only 3.4% of non-OUDs did. In descending order of frequency, the most commonly given substance misuse diagnoses for OUDs were : 20.7% alcohol dependence ; 16.3% other, mixed, or unspecified drug abuse ; 13.6% unspecified drug dependence ; 12.5% combinations of drug dependence excluding opioid ; 12.1% tobacco dependence ; 9.4% alcohol abuse ; 7.1% cocaine dependence ; 4.1% cannabis dependence ; 3.9% cocaine abuse ; and 3.0% cannabis abuse .
All between-group differences were significant at the p < 0.001 level. Other drug abuse or dependence categories occurred in less than 3% of either group. Medication utilization also significantly differentiated non-OUDs from OUDs. Commensurate with the findings regarding elevated rates of mood and anxiety disorders among OUDs, these individuals were more likely to use SSRI medications and benzodiazepines than non-OUDs. Tricyclic antidepressant use was also much greater for OUDs than non-OUDs ; much of this difference was accounted for by the rates of trazodone use , which is often prescribed for insomnia. Rates of anticonvulsant use were also significantly greater among OUDs, with gabapentin accounting for much of this difference . Medications related to pain also differentiated the two groups; OUDs were more likely to be prescribed skeletal muscle relaxants , including cyclobenzaprine hydrochloride and carisoprodol, than non-OUDs. The receipt of nonsteroidal anti-inflammatory medications, or NSAIDs, was also more common among OUDs than non-OUDs . Medications are listed by category in Table 4.The research team devised a series of models using the build set, which were then used to test for predicting OUD status within the validation set. These models were developed to include variables that could be reasonably expected to be present in other data sets. For example, one model was a “diagnostic data only” model including solely ICD-9 diagnoses, which are coded the same as DSM-IV-TR diagnoses for mental health conditions . Another model was comprised of “medical utilization data only” measures and was designed to use variables that might be available in other insurance or health data sets. A “pharmacy data only” model included utilization variables related to medications and opioid use data, similar to what might be available to a pharmacy benefit manager researcher. The CHAID procedure identified a number of significant interactions that met the criteria outlined in the Methods section. These interactions were added to the core variables for each model, but only if the variables involved in the interaction were also included in the model individually. The model that was selected as the best fit for the data was comprised of mental health variables; specifically, this model included the diagnostic status for ICD-9 mental disorders and the health service utilization variables that focused on mental health . This model provided the best fit for the data, as defined by AIC values and by overall parsimony. The log likelihood ratio test for the selected model was 12,695 . The remaining models had LR values ranging from 5,785 to 13,095 ; all of these ratios were also statistically significant . Likewise, the results for the Wald and Score tests were significant across all models. In comparison to the selected model, the only other model with a lower AIC value contained all of the variables presented in the bivariate comparisons above, as well as all significant interactions involving these variables, while only providing a modest decrease in AIC. The resulting model is presented in Table 5. This predictive model was 79.5% concordant with actual OUDs in the validation data set, meaning that almost four-fifths of the OUDs were correctly identified when the model was applied to a different sample of participants.