Chemotherapeutic impact on pain and global health-related quality of life in hormone-refractory prostate cancer: Dynamically Modified Outcomes (DYNAMO) analysis of a randomized controlled trial

Abstract

Purpose: This paper applies the Dynamically Modified Outcomes (DYNAMO) model to a clinical trial of two chemotherapeutic regimens on global health-related quality of life (GHRQL) in hormone-refractory prostate cancer.

Methods: DYNAMO identifies the causal influences operating in a clinical trial and their mediation, moderation, and modulation by uncontrolled variables. The Southwest Oncology Group trial S9916 randomized assignment to mitoxantrone plus prednisone (M + P) versus docetaxel plus estramustine (D + E) treatments. In this application, we examine baseline-adjusted impacts of worst pain (McGill Pain Questionnaire) on GHRQL (EORTC Quality of Life Questionnaire-C30) at 10 weeks.

Results: The average treatment levels of pain did not differ, hence, the average mediated effect of treatment on GHRQL was zero. Nonetheless, M + P reduced the impact (the relational outcome) of pain on GHRQL by 54% relative to D + E. Individual variation in the relational outcome (modulation) was of the same magnitude as the average difference between the groups. Performance status moderated the direct effects of treatment, with D + E being more effective in good, but not poor, performance strata.

Conclusions: The DYNAMO approach comprehensively accounted for treatment effects. Rather than a single average effect, there were three distinct treatment effects: one direct effect for each performance status level and a direct effect on the relationship between pain and GHRQL.

Figures

Figure 1

Key results of the relational outcome model applied to the SWOG 9916 example. The Direct Causal Effect (DCE) of Treatment (TX, 0=M+P, 1=D+E) on Pain was zero, so the Average Mediated Effect was also zero. The DCE of Treatment on the relational outcome β showed that the average dependence of GHRQL on Pain was roughly twice as strong in the D+E arm (.17+.20) as in the M+P arm (.17). The βi values varied across patients, with within-group variance equal to .024 (SD=.15). The DCE of treatment depended on the level of the pre-randomization Performance Status measure. The randomized treatment intervention caused an average improvement of 5.89 GHRQL points in the good performance status stratum, but only an improvement of .25 GHRQL points in the poor performance status stratum. Since the Average Mediated Effect was zero, the Average Causal Effects equaled the Direct Causal Effects (5.89 GHRQL improvement for PS=0, .25 GHRQL improvement for PS=1).

Figure 2

Boxplots of the distributions of the relational outcome slope coefficient for individuals in the M+P and D+E treatment arms. The relational outcome is an individual’s expected change in GHRQL given a one-unit increase in Pain, E(GHRQL|Patient=i,Pain=p+1) – E(GHRQL|Patient=i,Pain=p), if other variables could be held at fixed values. This is a systematic attribute of a person, distinct from measurement error. The boxes show the 75th and 25th percentiles, with the central lines denoting the medians. The whiskers include the non-outlying values, while the isolated dots represent outliers. The interquartile range (75th percentile – 25th percentile) was approximately equal to the average treatment arm difference, so both population and individual relational effects were important.

Figure 3

Partitioned causes of observed GHRQL change in individual patients. Causal models are modular, representing the expected changes from controlled manipulations holding constant other variables. The Individual Causal Effect (ICE) summarizes relationships in Figure 1 to show how GHRQL (Y) would change for a particular patient because of treatment (if he or she were to receive the other randomized intervention, holding constant all other causes). Under these assumptions of modularity and unchanged other causes, the total cause of Y must equal the sum of a patient’s ICE and all other causes UY. Therefore one can calculate UY by simple subtraction once the ICE has been estimated.

Figure 4

Model-based estimation and counterfactual inference for individual patients. The abscissa in this scatterplot represents the model-based Individual Causal Effect (ICE) estimate, while the ordinate is the observed Y (GHRQL) adjusted for baseline (a residual “change” score). Patients A and B had both positive (beneficial) causal effects from therapy and positive (beneficial) observed change in Y. Patient A had an estimated positive ICE of 10, but an observed improvement of only 5, hence the other causes of Y amounted to -5. Holding constant the other causes, the value expected for Y if the ICE were modified to zero would be -5. Hence A’s improvement would not have happened without therapy: A improved because of therapy. Patient B’s observed improvement was 17, of which only 5 resulted from therapy. If the ICE had been zero for B, we still would have observed a positive change of 12 for B; hence B improved regardless of therapy. Patient C had a negative (harmful) ICE of -12, but a positive observed change of 7. Thus C improved despite therapy; C would have improved by even more without the harmful effect of therapy. Any patient falling in the same sector as patients A, B, or C shares the causal attribution appropriate for that sector. The causal attributions for patients who worsen (below the abscissa) follow similarly.

Figure 5

Causal attribution plots for the M+P (left panel) and D+E (right panel) treatment arms. The plots contain the specific model-based and observed data points for all patients. The sectors of the plots correspond to the definitions in Figure 4. For example, the upper left quadrants of both panels contain data for patients who improved despite therapy. The lines of identity are not represented diagonally in these plots because the ranges are greater for observed data (the ordinates) than for the Individual Causal Effects (ICEs) (the abscissas); this is a graphical display choice made to enhance point visibility and has no effect on the interpretations, which are identical to those of Figure 4.