Potential patient participants will be identified using EMRs; International Classification of Diseases, Tenth Revision diagnosis codes; direct clinician referral; and referral from informational handouts in patient-facing platforms (eg, clinic rooms and chemotherapy teach-back packets). Study sites will develop reports to generate lists of potentially eligible patient participants in accordance with all rules of the HIPAA (Health Insurance Portability and Accountability Act), preparatory to the research exception [45]. The research coordinator will review listed patients’ EMRs to verify eligibility following institution-specific procedures for contacting potential patient participants in clinic, over the telephone, or through letters. After receiving permission using one of the methods described above, trained research staff at each site will contact potential patient participants in-clinic or over the telephone to confirm eligibility. Subsequently, research staff will obtain consent, and data will be collected using Research Electronic Data Capture (REDCap; Vanderbilt University), a free and secure web-based application to capture research data [46].

Once patients have consented and enrolled in the study, data collection will include patient-reported outcomes (PROs), chart reviews, and interviews. Table 1 provides an overview of the study procedures and time frame.

Study staff will confirm patient demographics and cancer type from the medical record. Sites will use a chart abstraction tool to gather medication-related data every 6 months. Opioid dose will also be assessed through self-report by reading back the dose and instructions on the opioid bottle and confirmed in the EMR. Patient participants will consent to receiving a phone call from the study staff for pill counts during the latter part of the month of their prescriptions every 3 months.

Patient participants will complete baseline assessments and PROs electronically using REDCap, over the phone, or during clinic visits through a tablet by the study staff. If patient participants are unable to complete PROs, care partners will be permitted to complete the PRO assessments on behalf of the patient participants. We acknowledge that based on systematic reviews in oncology, there may be differences in patient-proxy reports, with proxies having a more negative view of the patient’s well-being [47].

The outcome measure for pain will be the Pain, Enjoyment, and General Activity (PEG) scale, a 3-item commonly reliable and validated measure used in oncology populations that asks about pain severity, pain interference in enjoyment of life, and pain interference in general well-being [48-50]. The PEG meets the Initiative on Methods, Measurement, and Pain Assessment in Clinical Trials recommendations and has a 7-day recall period [51]. The PEG will be collected weekly to provide longitudinal impact on pain severity, pain interference, and whether opioids precede any change in pain.

Additional monthly PROs will include information on mood (Functional Assessment of Cancer Therapy-General Well-Being Subscale); comorbidities (Self-Administered Comorbidity Questionnaire); functioning (Carey index); and if prescribed opioids, information on opioid misuse (Patient-Reported Outcomes Measurement Information System), opioid side effects (Patient-Reported Outcomes version of the Common Terminology Criteria for Adverse Events), and substance use (Tobacco, Alcohol, Prescription Medication, and Other Substance Use Tool) [52-56]. These measures were selected because of their favorable psychometrics in the study population, brevity, and similar recall periods. This patient-participant burden is similar to other studies among patients with metastatic cancer [57].

For Aim 3, we will conduct one-on-one interviews with patients, their care partners, and clinicians with the goal of better understanding opioid-related decision-making. These interviews will occur every 3 months for 2 years or until death. The rationale for this frequency is that it strikes a balance between interviewing people often enough for them to remember aspects of opioid decision-making since the last interview but not so often as to ask about opioid-related decisions more frequently than those decisions occur or cause undue participant burden.

This analysis will draw from all patient participants, regardless of whether they are prescribed opioids, so we can compare those who are with those who are not prescribed opioids. We will estimate differences in pain severity and interference between patients treated with opioids versus patients not treated with opioids using linear mixed-effects models, which will allow us to (1) account for repeated measures over time for each patient and (2) adjust for potential confounders. In the linear mixed-effects models, the dependent variables will be based on the PEG: pain severity (first question is a continuous 0-10 value) and pain interference (mean of 3 PEG items) at each time point. Current opioid prescription, baseline pain severity or interference, and other confounders (ie, factors statistically associated with the exposure and outcome) will be included as fixed effects, while the subject will be included as a random effect.

Power calculations were performed by simulation to allow a range of parameters to vary in all scenarios. Unlike a randomized trial, in this study we will not have control over the ratio of patients allocated to the 2 comparison groups (eg, there will not be precisely a 1:1 ratio of patients prescribed opioids vs patients not prescribed opioids for comparison). Scenarios were considered where the proportion of patients prescribed opioids was allowed to vary from 30% up to 70%, all assuming a mean baseline PEG of 6 in the patients not prescribed opioids, a reduction in the PEG of 1 point for patients that are prescribed opioids, and a common standard deviation in PEG scores of 2 points [51]. All scenarios considered had at least 90% power to detect a difference of 1 point between groups with a total sample size of 630 patients (allowing for 30%-70% of the population to be prescribe opioids, meaning that with a range of 189 up to 441 patients prescribed opioids during the study period, there is ample power to detect differences between groups). All scenarios had at least 99% power to detect a 2-point difference on the PEG. We acknowledge that up to 40% of patients may drop out during the study due to the morbidity and mortality of patients in palliative care [58]. Data up until the point of attrition will be included in the analysis, and this has also been accounted for in our calculations.

In general, linear mixed-effects models provide unbiased estimates when the data are assumed to be missing at random, meaning that the likelihood of a missing value is not related to the values of the outcome data (in this case, the assumption that missing values for pain are not related to the severity of the patient’s pain). The setting of this study will have two basic missingness problems: (1) data that are missing while the patient still lives, either because the survey could not be completed or the patient did not wish to complete the pain instrument; and (2) pain data that are missing because the patient is deceased. The missing at random assumption can never be fully confirmed, so we will perform three sensitivity analyses to see how our results vary under different approaches to handling the missing pain values: (1) multiple imputations; (2) imputing the worst pain value for patients that fail to complete the pain instrument for any reason (including death before the scheduled assessment); and (3) imputing the worst pain value for patients that fail to complete the pain instrument, but omitting patients that have died. It should be noted that in this patient population, death is an expected outcome and not necessarily a negative outcome with respect to end-of-life suffering and pain management.

To test our hypothesis that coprescribed medications will be associated with an increased risk of opioid side effects (eg, benzodiazepines and somnolence), we will include only study patient participants who are prescribed opioids. Linear mixed-effects models will be used to assess the relationships between selected risk factors and the corresponding suspected opioid side effects. Each side effect is scored on the respective weekly or monthly assessments. Analyses will be conducted to examine the relationship between coprescribed medications and each of the respective opioid side effects (constipation, nausea, dry mouth, and somnolence) as well as targeted relationships specifically of interest for individual exposures or side effects combinations, including (1) coprescription of other constipating medications and constipation, (2) coprescription of other constipating medications and nausea, (3) coprescription of anticholinergic medications and dry mouth, and (4) coprescription of benzodiazepines with somnolence. We will also examine the association between which opioid is prescribed (ie, morphine vs oxycodone) and each of the same respective opioid side effects (constipation, nausea, dry mouth, and somnolence). The use of linear mixed-effects models will provide estimates of the effect of each risk factor of interest on each of the respective side effects while also allowing the inclusion of longitudinal measurements from patient participants.

With an overall sample size of 630, we anticipate approximately 50% (315/630) of the study sample to be prescribed opioids, leaving 315 patient participants available for the analyses pertinent to this hypothesis. We examined a range of scenarios allowing anywhere from 10% to 30% of the eligible participants to have each exposure, performing power analyses to detect a 0.5-point difference in means for each side effect on the 1-5 scale (assuming a SD of 1 point). Power is about 65% if the exposure is rare (10% of patients prescribed opioids) but increases to 87% if the exposure is present in 20% of patients prescribed opioids and 93% if exposure is present in 30% of patients prescribed opioids. Again, we account for up to 40% attrition during the study. Note that patient participants will contribute data until they drop out, and this data will be included in the analysis.

To test our hypothesis that younger age, history of substance use disorder, and history of mood disorders will be associated with a greater risk of opioid misuse and use disorder, we will perform multivariable Cox regression analyses to assess the respective risk factors of interest. We will use a single prespecified model that includes age, history of substance use disorder, and history of mood disorders using baseline PRO data and chart review.

With a total sample size of 630 patients recruited in Aim 1, the power calculation for this subaim assumed a scenario where approximately 50% (315/630) of the patients are prescribed opioids, leaving 315 patients available for this aim. We anticipate opioid misuse will occur in 20% to 30% of patients. For a rarer event such as opioid use disorder, which is estimated to occur in approximately 10% of the patients, a sample size of 315 patients will be sufficient to support a model with 5 candidate predictor variables.

Unlike Aim 1, missing data should not affect analyses in this aim. The development of opioid-related complications will also be considered an event if it is known to occur and otherwise assumed not to occur during the period patients are under observation.

The purpose of Aim 3 is to identify what patients, care partners, and clinicians already know and how they make opioid-related decisions to develop the lay model of the BDR. To determine decisional influences, the BDR framework uses the “mental models approach” to elicit these views [59]. To create the BDR framework lay model, we will structure the patient and care partner interviews by focusing on (1) what they already know about opioids, (2) how they make decisions, (3) examining whether this is consistent with or contrary to the normative model of evaluating risks and harms, and (4) how they think decisions around opioid use should ideally be made. The interviews begin with general questions and then become more specific, asking about each topic from the expert model. We will structure the clinician interviews by focusing on (1) their assessment of the patient’s pain; (2) approaches to best treating it, including opioids; and (3) their decision-making around opioid prescribing. We will also evaluate whether different patient, care partner, and clinician perspectives are consistent with or contrary to one another.

For the semistructured interviews (40 participants with metastatic cancer from our larger cohort, 20 who are prescribed opioids at the time of enrollment, and 20 who are not), we aim to understand perspectives on deciding to initiate or continue opioid therapy or not. This sample size is based on evidence that suggests thematic saturation is achieved at 8-16 interviews; since we expect up to 40% attrition, this approach will ensure we have at least 20 participants [60]. We will purposively sample [61] to ensure diversity on race, site, and baseline PEG value; we will also recruit some individuals who have a history of a substance use disorder and are therefore at particularly high risk for negative consequences when prescribed opioids.

Similarly, we will recruit patient participants’ care partners and clinicians. We will aim for at least 10 people who have care partners who consent to the group prescribed opioids at enrollment and 10 care partners of patient participants who have been prescribed opioids at enrollment. This approach will ensure that we have enough care partners to provide adequate perspective but not severely bias our sample by only allowing the enrollment of individuals who have care partners willing and able to enroll in this study. As described above, we will also purposefully sample 25% of patient participants who do not have care partners so that these perspectives are represented. Interviewing patients and their care partners s and clinicians will allow for triangulation of findings. Additionally, following patients, care partners, and clinicians longitudinally will allow for a detailed understanding of how decision-making changes over the course of metastatic cancer, allow participants to reflect on what they would have liked to know earlier based on what they know now, and provide real-time data on actual decisions (eg, initiation or discontinuation of opioids) on which participants can be asked to comment. Given the challenges inherent in engaging clinicians, we will aim for 15-minute focused interviews.

Trained qualitative researchers will collect and analyze the qualitative data. Interviews will be conducted remotely (through phone or Zoom [Zoom Video Communications], according to the interviewee’s preference), and interview recordings will be transcribed verbatim (with details that might identify the interviewee redacted). Coding the data will require a hybrid deductive-inductive approach. A prespecified set of codes will be identified from the normative model. The presence and absence of these codes (eg, the accuracy of knowledge about the relationship between opioids and addiction) will be used to document the consistency between the normative model and stakeholder perspectives. In addition, we will take an inductive analytical approach to identify emergent values and beliefs that are not present in the normative models. Codes will be developed through open coding of the transcripts to determine topics and themes that emerged in the interview transcripts with input on topics or themes that the study team anticipates being relevant, resulting in simultaneously inductive and deductive development of the codebook. A draft codebook including detailed code definitions will be discussed with the study team to ensure that codes reflect both the data as they are emerging as well as relevant topical themes that are important to the team. Codes will be thoroughly checked to determine that definitions are distinct enough to reduce ambiguity in the coding process.

About 25% of the transcripts will be coded by 2 independent, trained qualitative coders to ensure quality and consistency in coding. Coding will then be compared for the purposes of calculating Cohen kappa inter-coder reliability scores [62]. Any coding discrepancies identified during this comparison will be adjudicated by the coders until full agreement on coding is achieved. The finalized coding on these transcripts will be recorded in a single ATLAS.ti file, following which the primary coder will complete the coding of the remaining transcripts according to the codebook and through the adjudication process and discussion by the study team with content and methodologic expertise. This process allows for consistency and quality across transcripts. Once coding is complete and uploaded to ATLAS.ti with quotes associated with codes, we will thematically analyze and determine the relative frequencies of various codes and the most salient themes. The resulting thematic analysis will be discussed with the entire investigator team to use the findings to build lay models of patient, care partner, or clinician decision-making around opioid use for patients with metastatic cancer.