The prescription drug abuse survey took place in July, 2016. The IRB was approved by Michigan State University Human Research Protection Program on October 1, 2014. Survey Sample International (SSI) was used to obtain a nationally representative respondent pool as well as a rural over-draw. The SSI company maintains a large opt-in panel of survey respondents balanced on age, gender, income and region. Opt-in panels are increasingly accepted in survey work as response rates from phone and mail surveys are dwindling. Opt-in panels seem to produce results comparable to robust traditional surveys, but at lower expense [10]. We requested additional respondents from rural areas to investigate the role that rural areas may play in recognition and stigma as, seemingly, in contrast to prior outbreaks of drug abuse epidemics, there is anecdotal evidence of PDA higher incidence in rural areas. Additional rural respondents were drawn from randomly selected rural counties as defined by the 2013 USDA Rural Urban Continuum Codes (counties with a code of 7, 8 and 9) and was also balanced on age, gender and income to the extent possible. A national pretest was used to help finalize the survey design.
The survey was opened to the SSI panel until the required number of responses was obtained from each category (gender, income, region, etc.) required for a balanced response set; thus there is no response rate as classically defined in survey research. The survey instrument presented respondents with a vignette that described an individual with symptoms/behaviors commonly associated with PDA. To account for possible gender effects half the respondents (approximately 250 national draw and 63 rural draw) received a vignette about a woman (Michelle) while the other half received a vignette about a man (Michael). Generally, the vignette is based on those found in Jorm et al. [11]. The Michael version of the vignette used in the survey follows.
Vignette: Michael is 30 years old. He went to see his doctor after experiencing a work-related injury and the doctor prescribed a painkiller, hydrocodone (brand names: Vicodin, Norco, Lortab), for Michael to take. He started taking the painkiller as instructed by the doctor but felt like it was not enough to control his pain and started taking an extra pill every day. After a follow-up visit, the doctor told Michael that his injury had healed and that he should stop taking the painkiller, but he continued taking it until he ran out. At that point, he felt like he needed more of the painkiller and went to a new doctor to get a new prescription.
A total of 631 respondents completed the survey. After reading the vignette, the respondent was asked a series of questions about what might be wrong with the person described in the vignette, how that person might be helped, and a series of questions designed to measure the respondent’s stigma regarding people like the one in the vignette. We adapted stigma questions from Griffiths, Christensen, and Jorm [5] to reflect the appropriate issue and to better fit our US context, as their work originated in Australia. Given that it is likely that personal experience in dealing with prescription drug use disorders, either in oneself or a family member, will impact recognition and/or stigma responses, subsequent to the vignette and recognition, stigma, and help-seeking questions, respondents were asked about their own and family member/close friend history of PDA. In addition, standard sociodemographic questions were included to be used to further explore the determinants of recognition and stigma. All results, except the recognition and stigma response proportions, are presented in unweighted format due to our use of socio-economic controls used by SSI to recruit a balanced sample. While we constructed weights for the response proportions, using weights in a regression that also uses the variables used to construct the weights would introduce bias.
In the following analysis, the primary data from the survey is matched with regional level indicator variables as well as a selection of county-level secondary data drawn from the Robert Wood Johnson Foundation-funded County Health Rankings project at the University of Wisconsin. While the University of Wisconsin data provides counts of mental health services providers, there are many missing values so mental health coverage quartiles, which are more complete, are used instead of counts. Finally, consistent with observations by Cerdá et al. [12] we proxy state-level community health norms with each state’s position with respect to loosening of restrictions on marijuana. We draw the state’s legalization position from an inventory produced by Governing magazine [13].
Analysis of recognition uses multilevel mixed-effects logit estimation with respondent, regional and county-level sociodemographic variables as controls. Analysis of the responses to stigma questions is conducted using a generalized linear multilevel multinomial logit estimation procedure, again using controls. We selected multinomial logit for analysis of the stigma data because it allows the response categories to be unordered, and our survey allowed the “don’t know” option in the Likert-type questions, and a priori “Don’t know” is neither higher nor lower than, say, “strongly agree.” The neutral “neither agree nor disagree” category was used as the base response for the stigma analysis. The strongly (dis) agree and (dis) agree categories were combined in the multinomial logit analysis to facilitate interpretation. In the multinomial logit model the log-odds of each response is expressed as follows: \( {\mu}_{ij}=\log \frac{\pi_{ij}}{\pi_{iJ}}={\alpha}_j+{x}_i^{\prime }{\beta}_j \),
where αj is a constant and βj is a vector of regression coefficients, for j = 1, 2,..., J-1. This model is similar to the binary outcomes model except that it extends to more than two outcomes. See Aldrich and Nelson [14] for a discussion of the procedure.