Treatment Decision Analysis Model for Prostate Cancer: A Randomized Trial

Following institutional review board approval for this study and after informed consent from patients is obtained, men referred to our institution for treatment of clinically localized prostate cancer will be interviewed at baseline following diagnosis but prior to making a treatment decision, by a research study coordinator whom we will train in utility assessment. At that initial interview, the patient's utilities will be assessed, in a private consultation room, for each important post-treatment health state that he potentially may enter. The states we would assess include living with untreated prostate cancer, impotence, urinary dysfunction, bowel dysfunction, rising PSA following definitive local therapy, and metastatic prostate cancer. We will also record characteristics of the patient's disease (pretreatment serum PSA, clinical stage, biopsy Gleason grade), demographics (age, race, level of education, employment, marital status) and comorbidity. Patient-reported urinary, bowel, and sexual function/bother will be assessed at baseline using a short form version of a 48-item validated HRQOL instrument for localized prostate cancer. All patient information will be entered into a secure, HIPAA-compliant database accessible only to the principal investigator, research study coordinator(s), data manager, and biostatistician. All patient identifying information will be removed prior to data analysis and will not be included in any publication or presentation that results from this study. For all patients, the quality-adjusted disease-free survival for each treatment modality will be calculated using the decision analysis model. For the decision analysis arm of the trial, this information will be conveyed to the patient and the physician prior to treatment decision. These patients will also receive standard treatment counseling. For the control arm, both patients and physicians will be unaware of the decision analysis recommendation and the patients will receive standard treatment counseling alone. After the patient has made a treatment decision he will be given a short questionnaire to assess decisional conflict. If after one month no treatment decision information is available, a letter and form requesting treatment information along with the decisional conflict questionnaire will be sent to patients. At 6, 12, and 24 months following the treatment decision, the patient will be sent a brief validated questionnaire which will measure his regret of treatment choice, treatment satisfaction, decisional conflict and current health state utility using validated instruments. A short form version of a 48-item validated questionnaire will also assess his HRQOL for urinary, bowel and sexual function/bother domains at these time intervals. Approximately two weeks after receiving each follow-up questionnaire and/or letter, a member of the study group will contact the patient by telephone if we have not received the completed questionnaire. S/he will inquire about any difficulties with the questionnaire and will remind the patient about sending it back.

The primary endpoint of this study (decisional regret) and secondary endpoints (treatment satisfaction, decisional conflict, and current health state utility) will be assessed by means of a patient-reported questionnaire using validated instruments assessed at 6, 12 and 24 months after randomization. A decision analysis model will be developed using a Markov modeling approach. Using Treeage Data® pro suite, the model will incorporate cancer outcomes, treatment-related morbidity probabilities derived from nomograms, and patient-specific utilities to estimate the quality-adjusted disease-free survival for each treatment option. In the model, all patients will begin with localized prostate cancer with no evidence of metastases. A 6-month transition cycle will be used for the Markov model. Bootstrap simulation with replacement will be utilized to derive mean utilities and 95% confidence intervals. Sensitivity analyses will be performed by varying the disease progression rates and utilities within their 95% confidence intervals. To investigate if the decision analytic model, compared to standard interactions alone, will influence treatment choice and lead to higher quality treatment decisions that are congruent with a patient's values, an intent-to-treat analysis will be conducted on patients who did and did not receive the decision analysis recommendation with decisional regret as the primary endpoint. In a planned secondary analysis, we will analyze both arms of our trial and test the hypothesis that a patients who chooses a treatment strategy that did not appear to maximize his quality-adjusted disease-free survival (i.e. not the one the decision analysis would have yielded) will be at increased risk of regret. This will be conducted by identifying patients in both arms who choose the treatment that is also recommended by the decision model. The probability of selecting the decision analytic treatment strategy (which is also the probability of selecting the highest quality treatment decision congruent with a patient's values) will be investigated using a logistic regression model with covariates that include treatment arm, an indicator variable to define patients whose baseline treatment choice matched the decision analysis model, an interaction between the treatment arm and this indicator variable, as well as other clinically relevant baseline covariates and two way interactions. Testing for an interaction between treatment arm and the indicator variable will allow us to test if the effect that knowledge of the decision analytic model has on the probability of selecting the highest quality treatment decision is different for patients whose baseline treatment choice matched the decision analysis model compared to patients whose baseline treatment choice did not match the decision analysis model. Each patient will also have the following outcomes measured at 6 months, 12 months, and 24 months following their treatment: decisional regret, decisional conflict, treatment satisfaction, health state utility score, and HRQOL for urinary, bowel, and sexual function domain scores. All scores will be standardized such that each patient's score is between 0 and 100. Baseline decisional regret scores, informed decision scores, and treatment satisfaction scores will be set at 100, indicating no decisional regret, a high degree of feeling as though an informed decision was made, and high treatment satisfaction. Scores for health state utility and HRQOL domains will also be obtained at baseline. Based on each patient's final treatment choice, we will also calculate the difference between the number of quality adjusted life years based on the decision analysis model and the final treatment choice.

For each response, separate linear mixed models that include both fixed and random effects will be applied. To meet the distributional assumptions of this model, responses may be transformed prior to model fitting. Fixed effects that will be included in the model are clinically relevant baseline and time dependent covariates and two way interactions as well as linear or quadratic time trends. In particular, a covariate for treatment arm, an indicator variable to define patients who chose the decision analytic model based treatment choice, and an interaction between these two variables will be included. In this case, the interaction will allow us to test if the effect that knowledge of the decision analytic model has on various responses is different for patients who chose the decision analytic model treatment choice compared to patients who did not chose the decision analytic model treatment choice. A covariate for the difference between the number of quality adjusted life years based on the decision analysis model and the final treatment choice and an interaction between this variable and the treatment arm indicator variable will also be included in the model to examine if the effect that knowledge of the decision analytic model has on various responses is effected by the theoretical degree of mistake. To fully account for the variation in subject specific trajectories over time, a random intercept will be included in this model although linear and quadratic random effects will also be considered if appropriate. In addition, the residual covariance structure for each model will also be investigated to avoid misspecification and invalid inferences. In particular, correct specification of the covariance structure will be investigated using different ways to model the measurement error (due to variation in the measurement of the response) and serial correlation (due to decreasing correlation between measurements over time). For instance, structures for serial correlation will be investigated using a semi-variogram. Although the problem of a misspecified covariance structure can and will be avoided by using a robust variance estimator when making inferences about the fixed effects of interest based on the marginal model, misspecification also leads to an inability to properly account for the effect of missing data, a common occurrence in longitudinal studies. Note that once an appropriate model is fit to each response of interest, inference regarding the fixed treatment and covariate effects on the response will be based on marginal models using the robust variance estimator to protect against model misspecification of the covariance structure. Statistical significance of fixed effects will be based on a 0.05 significance level. As this study is exploratory, no attempt will be made to account for multiple comparisons.

As discussed previously, missing data is a common problem in longitudinal studies. Typically, patients who drop out of the study are not comparable to patients who remain in the study at all times. Consequently, the likelihood of dropout is correlated with the underlying unobserved data. To account for this problem, pattern mixture models will also be fit in addition to the previous models that ignore the effect of missing data on resulting estimates. Pattern mixture models are based on fitting separate linear mixed model to each of the seven dropout patterns that can occur in the data such that a separate time trend is estimated for each pattern. The marginal treatment effect is then calculated by a weighted average of the pattern-specific treatment effects, with weights given by the probability of occurrence of the various patterns. Because the estimated marginal treatment effects rely on the extrapolation of fitted average profiles to time points where data is not observed, sensitivity analyses will also be performed to examine the robustness of the results to alternative assumptions. For instance, rather than estimating a separate time trend for each pattern, the model could be simplified so that drop-out pattern is used as a covariate in the model. In this case, the time trend within a pattern is unstructured but parallel across patterns. If estimated marginal treatment effects agree among the different models that are investigated, this will suggest that the results are not artifacts of the particular model used to account for dropout.