The prescription of a DMARD for patients with RA is considered a standard of effective care [1] and there is increasing expectation for rheumatologists to objectively measure RA disease activity and to strive to treat patients to achieve low disease activity [2, 3]. However a recent study of Medicare managed care enrollees found only 63% received a DMARD [4]. The explanation for underutilization is not fully known; however DMARD decisions are complex and require patients to consider important tradeoffs. DMARDs improve arthritis symptoms and slow the rate of joint damage, but also carry the risk of serious infections amongst other possible harms [5, 6].

When patients consider medication information they process the risks affectively as well as cognitively [7]. Differing responses like confidence or anxiety can influence willingness to take a medication and adherence [8]. A good decision is informed, consistent with patient values and acted on [9]. However, this depends on an individual’s ability to understand and evaluate options and to make judgments that are relatively free of bias [10]. There is evidence that a depressed person's diminished ability to think, concentrate or indecisiveness may impair ability to participate in decision-making [11]. Risk assessments can be biased by emotion in part through increased use of the affect heuristic in judging risks rather than more deliberate processing of data [12]. However the specific effects of depression, happiness, and cognition on risk perception have been only incompletely studied in medical decisions. In this study we evaluate how patient demographics, RA disease and treatment related experience, mood and health literacy influence risk perception and DMARD willingness.

The integrated model of behavioral prediction (IMBP) proposes that peoples’ willingness to take a medication are influenced by their beliefs about the expected outcomes of therapy, their social network’s support of and use of the medication (perceived norm), as well their perceived ability to implement it (self-efficacy) [13]. In the model, these three factors plus background variables such as demographics, individual differences, and exposure to health messages (i.e. decision aids) can be used to predict patients’ willingness to take a proposed medication. One would expect if environmental factors do not impede medication use, a patient with positive treatment expectations would have greater willingness to take a proposed medication. We hypothesized that risk perception could be a unique patient attribute that might be influenced by background factors as well as modified by varied formats of a risk presentation in a decision aid.

1436 patients with RA were surveyed. All participants were randomized and allocated to intervention. Subjects were blinded to allocation. Data were analyzed on 1009 respondents. In addition to non-respondents, we learned that 4 patients had died, 4 had changed address and could not be reached, and 5 returned the questionnaire without completion. The overall response rate was 71%. Data from all study completers were included in the analysis. In the completed sample all patients had RA. For RP and DMARD willingness, missing data were 11.2% and 10.0% respectively. To evaluate if participants with missing values were systematically different than cases without missing values [29], we evaluated means of missing values by estimation-maximization methods. Little’s MCAR test Chi-Square was P <.01 suggesting that missing values are not missing completely at random [30]. Review of estimated means of RP and DMARD willingness disclosed as expected that there were higher rates of missing values in patients with low health literacy (11.2 vs. 21.8% and 10.0 vs. 18.4% respectively), low education (11.2 vs. 17.7% and 10.0 vs. 14.5% respectively) and major depression (11.2 vs. 15.4% and 10.0 vs. 12.6% respectively), but not in female, low income or minority subjects. As missing values were not missing completely at random, multiple imputation was utilized. SPSS performed imputation for variables with > 10.0% missing values [30].

Mediation analysis

In regression analysis reduced HL was significantly associated with increased risk perception and decreased DMARD willingness. Though a cross sectional survey and correlational research methods are not optimal approaches to evaluate causation, we wanted to evaluate the degree to which HL’s effect on DMARD willingness was mediated through risk perception. We conceptualized this with a simplified causal model illustrated in Figure 1.

First we evaluated the relationship between HL and DMARD willingness. This demonstrated there was a significant effect to be mediated, β = .142 (P< 0.01). Next we evaluated the effect of the initial variable, HL, with the mediator, RP, which displayed significant relationship, β = - .082 (P<0.01). Finally we demonstrated the mediating effect of RP on DMARD willingness, the criterion variable, by simultaneously using RP and HL to predict DMARD willingness. This revealed a significant relationship between RP and DMARD willingness, β = - .471 (P<0.01). Once this mediator was taken into account, the relationship between HL and DMARD willingness dropped to β = .086 (P<.01). This indicates that RP partially mediates the relationship between HL and DMARD willingness.

Based on examination of variation of practice patterns in the Dartmouth Atlas, John Wennberg, M.D. has suggested three general approaches to reduce inappropriate variation and increase the equity of medical care: increase the amount of effective care, reduce supply sensitive care and increase preference sensitive care [32]. The underuse of DMARDs by patients with RA, is an example of underutilization of effective care. This may be explained by healthcare system failures, individual physician behavior, or in some cases subsets of patients who are informed but elect not to take medications recommended in national standards of care. A recent computer based, simulated decision, discrete choice experiment of 144 RA patients’ preferences of specific risks and benefits of treatment found that when compared to white adults with RA, African American RA patients were more likely to be classified as risk averse (OR 8.4 [3.1-23.1]) [33]. In general, it is unknown if risk aversion is an inherent culturally based trait, influenced by affect, the result of modifiable deficits in knowledge or arises from potentially correctable cognitive bias [34]. The purpose of our study was to identify, in a large community cohort of RA patients, if a limited number of patient characteristics could predict patient risk perception and how this influenced the likelihood to take a proposed DMARD.

In our analysis, guided by the Integrative Model of Behavioral Prediction [13], we developed regression models to identify the determinants of risk perception and predictors of patent's willingness to take the proposed medication (DMARD willingness). In these we considered background variables (demographics, socio-economic, disease and DMARD experience, disease activity and related disability, support to acquire and administer DMARDS) and patient beliefs about outcomes of their current DMARD use (adverse events and satisfaction with their disease control). The results create a valid model, which accounts for significant proportion of the variance of risk perception and likelihood to take a proposed medication. RP and DMARD willingness are highly correlated, but distinct attributes, which are predicted by health literacy as well as different aspects of patients’ DMARD and disease experience. We found that when risk level and context were controlled for, RP was influenced by patient functional impairment / RA severity, past experience of a DMARD related serious adverse event, current bother from DMARD related side effects, as well as HL (see Table 3). In contrast DMARD willingness, when controlled for risk level and context, was predicted by satisfaction with RA control, post-decision regret related to the most recent DMARD choice, and HL (see Table 5). HL was the only thread that contributed to both predictive models. Importantly, depression and happiness did not demonstrate a significant effect on either criterion variable. It is notable age, gender, minority status, low education, low-income status and TNF knowledge did not contribute to the prediction of DMARD willingness. Although our data creates a significant model, it is important to note that it explains only 12.7% of the variance of willingness to take a proposed medication. This suggests that other unmeasured variables i.e. doctor patient relationship, more nuanced measurement of age related cognitive impairment, perceived costs, individual values of benefit compared to risk could significantly add to the predictive power of the model.

The results do demonstrate that HL, independent of low educational achievement or other demographics, is a common predictor of both RP and DMARD willingness. Our meditational analysis suggests that much of the effect of HL on DMARD willingness was mediated through RP. The clinical implication is that HL, as an indicator of cognition, may provide an identifiable predictor of risk aversion that could be accommodated through decision support to minimize cognitive bias [35–37]. In our study population, the 8.8% prevalence of low or marginal HL, as measured by a simple 3 question survey, mirrors Shin and colleagues who found utilizing the ACR Neuropsychological Battery the prevalence of cognitive impairment in RA patients ranged from 8% (semantic fluency test) to 29% (visuo-spatial learning/memory test) [38]. This highlights the frequency with which rheumatology providers encounter such high-risk patients, who may not be identified by demographic characteristics.

When making treatment decisions, patients may use one or more cognitive strategies. Prospect theory suggests that decision makers carefully weigh multiple attributes of available treatment options before arriving a rational choice that maximizes benefit [39]. This is theorized to rely on “System 2” which is the mode of thinking that directs conscious, deliberate, effortful activities like choice [40]. Alternatively patients may utilize a heuristic approach, making simpler more cognitively efficient decisions using rules of thumb. This is thought to be directed by “System 1” which is the mode of thinking that is unconscious, fast, intuitive and the source of impressions and feelings at the source of many of our explicit beliefs [40]. In the setting of patients evaluating a new DMARD, this might be conceptualized as a tension between risk as feelings versus risk as analysis [41]. When a patient with reduced HL engages in a conversation with a physician evaluating medication options, the constrained time increases cognitive effort. Cognitive overload could influence their evaluation of risk and willingness to take the proposed DMARD leading them to substitute rational deliberation of the facts with the use of a easier, faster heuristics based on their past RA and DMARD experiences [40]. Examples might include: “if I had a side effect before it will happen again” or accepting the default “I’ll do what the doctor suggests”. This proposal is consistent with the findings that our patients’ risk perception was heavily influenced by experience of greater functional disability as well as past and present DMARD side effects. In a previous study we also found that patients’ trust in physician had nearly seven times the effect on their confidence in a DMARD decision than any other predictor including numeric literacy and DMARD knowledge [18]. This offers further evidence that many patients rely on non-deliberative decision strategies when choosing a DMARD.

The findings should be interpreted given the limitations of the study. Most notable is that our survey provided a simulated decision rather than observing an actual DMARD choice prospectively in the clinic. However, vignettes have been used in many previous studies to validly simulate patient and physician decisions [14, 42]. In addition if a prospective study design had been utilized, time and cost would have made it impossible to accrue 1009 patient responses, which provided the statistical power to evaluate multiple predictor variables each with small effects. The selection of background and predictor variables was guided by a widely accepted psychological/health behavior theory [13] and included an objective physician collected RA activity measure, the CDAI, plus patient reported outcomes of disability, appraisal of disease control as well as DMARD side effects, TNF knowledge, and validated screening measures for depression and HL. To differentiate patients with varied perception of medication risk, we utilized a simulated decision scenario with high and low risk as well as two different safety monitoring framing contexts. The response rate was high at 71%, however as the survey was anonymous, we were not able to compare the characteristics of responders to non-responders, and it remains possible that there was a differential non-response in subgroups of patients. Still the findings are important as with 6.5% minority and 18.6 % Medicaid respondents, our completed sample was demographically similar to a previous large random sample of RA patients receiving care in five geographically dispersed community rheumatology practices in Michigan [18], thus the results are likely generalizable to similar populations. Finally a cross-sectional survey allows the description of the association of criterion and predictor variables but is not the strongest design to assert causation between variables.

In our large community-based sample, medication risk perception seemed to be driven by negative RA disease and treatment experience, while DMARD willingness was predicted by perceived disease control. In the comparison of intenders and non-intenders, non-intenders were significantly older, had experience using fewer DMARDs, had lower health literacy, less knowledge about TNF inhibiting DMARDs, had higher levels of post-decision regret related to the DMARD they had most recently initiated, as well as higher perception of medication risk. Minority, low-income, and depressed patients were no more likely to be non-intenders. While clinicians should be alert to mood disturbance, depression may have less effect on risk perception and decision-making than previously proposed [10–12]. While racial disparities with the under-utilization of DMARDs exist in the US, our data support a premise that risk aversion could be, in many cases, the result of a potentially recognizable and correctable cognitive defect or bias rather than an inherent, cultural trait.

When proposing to initiate a new DMARD, clinicians must be alert for the possibility of cognitive impairment whether derived from educational, language, age related decline in information processing or sensory defect that could effect the patient’s decision making. Time pressures increase cognitive load, which might justify extending the time of deliberation beyond the constraints of the office visit by scheduling a follow up visit, using decision supports like decision aids, involving family members, or providing post-visit coaching by other health professionals [43]. Heightened clinician awareness, formal screening for low health literacy or cognitive impairment in high-risk populations, may identify patients could benefit from additional decision support. Further investigation is needed to evaluate if the use of patient decision aids to extend and structure deliberation, can increase the use effective care by the adoption of recommended medications.