Bisphosphonates or RANK-ligand-inhibitors for men with prostate cancer and bone metastases: a network meta-analysis

Abstract

Background: Different bone-modifying agents like bisphosphonates and receptor activator of nuclear factor-kappa B ligand (RANKL)-inhibitors are used as supportive treatment in men with prostate cancer and bone metastases to prevent skeletal-related events (SREs). SREs such as pathologic fractures, spinal cord compression, surgery and radiotherapy to the bone, and hypercalcemia lead to morbidity, a poor performance status, and impaired quality of life. Efficacy and acceptability of the bone-targeted therapy is therefore of high relevance. Until now recommendations in guidelines on which bone-modifying agents should be used are rare and inconsistent.

Objectives: To assess the effects of bisphosphonates and RANKL-inhibitors as supportive treatment for prostate cancer patients with bone metastases and to generate a clinically meaningful treatment ranking according to their safety and efficacy using network meta-analysis.

Search methods: We identified studies by electronically searching the bibliographic databases Cochrane Controlled Register of Trials (CENTRAL), MEDLINE, and Embase until 23 March 2020. We searched the Cochrane Library and various trial registries and screened abstracts of conference proceedings and reference lists of identified trials.

Selection criteria: We included randomized controlled trials comparing different bisphosphonates and RANKL-inihibitors with each other or against no further treatment or placebo for men with prostate cancer and bone metastases. We included men with castration-restrictive and castration-sensitive prostate cancer and conducted subgroup analyses according to this criteria.

Data collection and analysis: Two review authors independently extracted data and assessed the quality of trials. We defined proportion of participants with pain response and the adverse events renal impairment and osteonecrosis of the jaw (ONJ) as the primary outcomes. Secondary outcomes were SREs in total and each separately (see above), mortality, quality of life, and further adverse events such as grade 3 to 4 adverse events, hypocalcemia, fatigue, diarrhea, and nausea. We conducted network meta-analysis and generated treatment rankings for all outcomes, except quality of life due to insufficient reporting on this outcome. We compiled ranking plots to compare single outcomes of efficacy against outcomes of acceptability of the bone-modifying agents. We assessed the certainty of the evidence for the main outcomes using the GRADE approach.

Main results: Twenty-five trials fulfilled our inclusion criteria. Twenty-one trials could be considered in the quantitative analysis, of which six bisphosphonates (zoledronic acid, risedronate, pamidronate, alendronate, etidronate, or clodronate) were compared with each other, the RANKL-inhibitor denosumab, or no treatment/placebo. By conducting network meta-analysis we were able to compare all of these reported agents directly and/or indirectly within the network for each outcome. In the abstract only the comparisons of zoledronic acid and denosumab against the main comparator (no treatment/placebo) are described for outcomes that were predefined as most relevant and that also appear in the 'Summary of findings' table. Other results, as well as results of subgroup analyses regarding castration status of participants, are displayed in the Results section of the full text. Treatment with zoledronic acid probably neither reduces nor increases the proportion of participants with pain response when compared to no treatment/placebo (risk ratio (RR) 1.46, 95% confidence interval (CI) 0.93 to 2.32; per 1000 participants 121 more (19 less to 349 more); moderate-certainty evidence; network based on 4 trials including 1013 participants). For this outcome none of the trials reported results for the comparison with denosumab. The adverse event renal impairment probably occurs more often when treated with zoledronic acid compared to treatment/placebo (RR 1.63, 95% CI 1.08 to 2.45; per 1000 participants 78 more (10 more to 180 more); moderate-certainty evidence; network based on 6 trials including 1769 participants). Results for denosumab could not be included for this outcome, since zero events cannot be considered in the network meta-analysis, therefore it does not appear in the ranking. Treatment with denosumab results in increased occurrence of the adverse event ONJ (RR 3.45, 95% CI 1.06 to 11.24; per 1000 participants 30 more (1 more to 125 more); high-certainty evidence; 4 trials, 3006 participants) compared to no treatment/placebo. When comparing zoledronic acid to no treatment/placebo, the confidence intervals include the possibility of benefit or harm, therefore treatment with zoledronic acid probably neither reduces nor increases ONJ (RR 1.88, 95% CI 0.73 to 4.87; per 1000 participants 11 more (3 less to 47 more); moderate-certainty evidence; network based on 4 trials including 3006 participants). Compared to no treatment/placebo, treatment with zoledronic acid (RR 0.84, 95% CI 0.72 to 0.97) and denosumab (RR 0.72, 95% CI 0.54 to 0.96) may result in a reduction of the total number of SREs (per 1000 participants 75 fewer (131 fewer to 14 fewer) and 131 fewer (215 fewer to 19 fewer); both low-certainty evidence; 12 trials, 5240 participants). Treatment with zoledronic acid and denosumab likely neither reduces nor increases mortality when compared to no treatment/placebo (zoledronic acid RR 0.90, 95% CI 0.80 to 1.01; per 1000 participants 48 fewer (97 fewer to 5 more); denosumab RR 0.93, 95% CI 0.77 to 1.11; per 1000 participants 34 fewer (111 fewer to 54 more); both moderate-certainty evidence; 13 trials, 5494 participants). Due to insufficient reporting, no network meta-analysis was possible for the outcome quality of life. One study with 1904 participants comparing zoledronic acid and denosumab showed that more zoledronic acid-treated participants than denosumab-treated participants experienced a greater than or equal to five-point decrease in Functional Assessment of Cancer Therapy-General total scores over a range of 18 months (average relative difference = 6.8%, range -9.4% to 14.6%) or worsening of cancer-related quality of life.

Authors' conclusions: When considering bone-modifying agents as supportive treatment, one has to balance between efficacy and acceptability. Results suggest that Zoledronic acid likely increases both the proportion of participants with pain response, and the proportion of participants experiencing adverse events However, more trials with head-to-head comparisons including all potential agents are needed to draw the whole picture and proof the results of this analysis.

Trial registration: ClinicalTrials.gov NCT00079001 NCT00003232 NCT00019695 NCT00104650 NCT00321620 NCT00216060 NCT00181558 NCT00268476 NCT00554918 NCT00685646.

Conflict of interest statement

Tina Jakob: none known

Yonas Mehari Tesfamariam: none known

Sascha Macherey: none known

Kathrin Kuhr: none known

Anne Adams: none known

Ina Monsef: none known

Axel Heidenreich: none known

Nicole Skoetz: none known

Copyright © 2020 The Cochrane Collaboration. Published by John Wiley & Sons, Ltd.

Figures

1

Ideal network diagram of all comparisons.

2

Study flow diagram.

3

Risk of bias summary: review authors' judgements about each risk of bias item for each included study. Spaces are left blank in the case a judgement is not applicable (e.g. study reports only outcomes subjective to participants).

4

Network diagram for outcome pain response. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

5

Forest plot for outcome pain response: random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results.

6

Leaguetable of network meta‐analysis for the outcome proportion of participants with pain response. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. No. of studies: 4. No. of treatments: 5. No. of pairwise comparisons: 4. No. of designs: 4 Heterogeneity/inconsistency: Qtotal = 0, P = not available; I2 = not available; Tau2 = not available
Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

7

Forest plot for sensitivity analysis for outcome pain response (risk of bias (RoB) low): random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results.

8

Leagutetable of sensitivity network meta‐analysis including only studies with low risk of bias for outcome proportion of participants with pain response. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. No. of studies: 3. No. of treatments: 4. No. of pairwise comparisons: 3. No. of designs: 3 Heterogeneity/inconsistency: Qtotal = 0, P = not available; I2 = not available; Tau2 = not available
Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

9

Network diagram for outcome adverse event: renal impairment. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

10

Forest plot for the outcome adverse events: renal impairment. Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results.

11

Leaguetable of network meta‐analysis for outcome adverse event: renal impairment. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 6. No. of treatments: 4. No. of pairwise comparisons: 6. No. of designs: 4 Qtotal = 0.92, P = 0.82/Qwithin = 0.56; P = 0.76/Qbetween = 0.36, P = 0.55; I2 = 0.0%, Tau2 = 0
Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

12

Forest plot of splitting direct and indirect evidence for the outcome adverse event: renal impairment. In addition to the confidence interval for the network estimate, prediction intervals are shown as bars for each comparison. Local approach to check inconsistency—comparison of direct and indirect estimates for closed loops. As presented in Figure 9, there is one closed loop in the network (zoledronic acid—clodronate—no treatment/placebo). There is no significant difference between direct and indirect estimate (P value of test for disagreement: P = 0.547).

13

Net heat plot for outcome adverse events: renal impairment (random‐effects model). There are negligible signs of inconsistency in the net heat plot. Local approach to check inconsistency—comparison of direct and indirect estimates for closed loops.

14

Forest plot for sensitivity analysis of the primary outcome adverse event: renal impairment (risk of bias (RoB) low): random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results.

15

Leaguetable of sensitivity network meta‐analysis including only studies with low risk of bias for outcome adverse event: renal impairment. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. No. of studies: 4. No. of treatments: 3. No. of pairwise comparisons: 4. No. of designs: 3 Qtotal = 0.38, P = 0.83/Qwithin = 0.00, P = 0.95/Qbetween = 0.38, P = 0.54; I2 = 0.0%, Tau2 = 0
Treatment effects + 95% confidence intervals (risk ratios, random‐effects model).

16

Forest plot of splitting direct and indirect evidence for outcome adverse event: renal impairment (risk of bias: low). In addition to the confidence interval for the network estimator, a prediction interval is shown. Local approach to check inconsistency—comparison of direct and indirect estimate for closed loops. There is one closed loop in the network (graph not shown; zoledronic acid—clodronate—no treatment/placebo). There is no significant difference between direct and indirect estimate (P value of test for disagreement: P = 0.538).

17

Network diagram for the outcome adverse event: osteonecrosis of the jaw. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

18

Forest plot for the outcome adverse event: osteonecrosis of the jaw. Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results. Since there are no closed loops in the network, no local approach to check inconsistency comparing direct and indirect estimates was done. Also, prediction intervals could not be calculated.

19

Leaguetable of network meta‐analysis for outcome adverse event: osteonecrosis of the jaw. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 4. No. of treatments: 4. No. of pairwise comparisons: 4. No. of designs: 3 Heterogeneity/inconsistency: Qtotal = 0.45, P = 0.50; I2 = 0.0%, Tau2 = 0
Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

20

Forest plot for sensitivity analysis of outcome adverse event: osteonecrosis of the jaw (risk of bias (RoB) low): random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results.

21

Leaguetable of sensitivity network meta‐analysis including only studies with low risk of bias for outcome adverse event: osteonecrosis of the jaw. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options would have been marked in yellow: treatment option in column better than treatment option in row. No. of studies: 3. No. of treatments: 4. No. of pairwise comparisons: 3. No. of designs: 3 Heterogeneity/inconsistency: Qtotal = 0, P = not available; I2 = not available, Tau2 = not available
Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

22

Network diagram for outcome total number of skeletal‐related events. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

23

Forest plot for the outcome total number of skeletal‐related events (SREs). Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields slightly different results (Figure 99).

24

Leaguetable of network meta‐analysis for outcome total number of skeletal‐related events. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 12. No. of treatments: 6. No. of pairwise comparisons: 12. No. of designs: 6 Qtotal = 11.48, P = 0.12/Qwithin = 11.38, P = 0.077/Qbetween = 0.09, P = 0.76; I2 = 39%, Tau2 = 0.0124
Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

25

Forest plot for the outcome: total number of skeletal‐related events (SREs). Fixed‐effect model.

26

Forest plot of splitting direct and indirect evidence for the outcome total number of skeletal‐related events. In addition to the confidence interval for the network estimate, prediction intervals are shown as bars for each comparison. Local approach to check inconsistency—comparison of direct and indirect estimates for closed loops. As presented in Figure 22, there is one closed loop in the network (zoledronic acid—clodronate—no treatment/placebo). There is no significant difference between direct and indirect estimate (P value of test for disagreement: P = 0.847).

27

Net heat plot for outcome total number of skeletal‐related events (random‐effects model). There are no signs of inconsistency in the net heat plot.

28

Forest plot for sensitivity analysis of outcome total number of skeletal‐related events (SREs) (risk of bias (RoB) low). Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results.

29

Leaguetable of sensitivity network meta‐analysis including only studies with low risk of bias for outcome total number of skeletal‐related events. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 7. No. of treatments: 5. No. of pairwise comparisons: 7. No. of designs: 5 Qtotal = 2.09, df = 3, P = 0.55/Qwithin = 1.95, df = 2, P = 0.38/Qbetween = 0.14, df = 1, P = 0.71; I2 = 0%, Tau2 = 0
Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

30

Forest plot of sensitivity analysis of splitting direct and indirect evidence for the outcome total number of skeletal‐related events (risk of bias low). In addition to the confidence interval for the network estimator, a prediction interval is shown. Local approach to check inconsistency—comparison of direct and indirect estimate for closed loops. There is one closed loop in the network (graph not shown; zoledronic acid—clodronate—no treatment/placebo). There is no significant difference between direct and indirect estimate (P value of test for disagreement: P = 0.710).

31

Results of pairwise meta‐analysis of denosumab versus zoledronic acid for the outcome total number of skeletal‐related events (SREs).

32

Network diagram for outcome skeletal‐related event: pathological fractures. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

33

Forest plot for the outcome skeletal‐related event (SRE): pathological fractures. Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results.

34

Leaguetable of network meta‐analysis for the outcome skeletal‐related event: pathological fractures. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 8. No. of treatments: 5. No. of pairwise comparisons: 8. No. of designs: 5 Qtotal = 1.73, df = 4, P = 0.78/Qwithin = 1.73, df = 3, P = 0.63/Qbetween = 0.00, df = 1, P = 0.96; I2 = 0.0%, Tau2 = 0 Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

35

Forest plot of splitting direct and indirect evidence for outcome skeletal‐related event: pathological fractures. In addition to the confidence interval for the network estimator, a prediction interval is shown. Local approach to check inconsistency—comparison of direct and indirect estimate for closed loops. As presented in Figure 32, there is one closed loop in the network (zoledronic acid—clodronate—no treatment/placebo). There is no significant difference between direct and indirect estimate (P value of test for disagreement: P = 0.958).

36

Net heat plot for outcome skeletal‐related event: pathological fractures (random‐effects model). There are no signs of inconsistency in the net heat plot.

37

Network diagram for outcome skeletal‐related event: spinal cord compression. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

38

Forest plot for outcome skeletal‐related event (SRE): spinal cord compression. Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results.

39

Leaguetable of network meta‐analysis for the outcome skeletal‐related event: spinal cord compression. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 9. No. of treatments: 5. No. of pairwise comparisons: 9. No. of designs: 5 Qtotal = 2.43, P = 0.79/Qwithin = 2.38, P = 0.67/Qbetween = 0.05, P = 0.82; I2 = 0.0%, Tau2 = 0 Treatment effects + 95% confidence intervals (risk ratios, random‐effects model).

40

Forest plot of splitting direct and indirect evidence for the outcome skeletal‐related event: spinal cord compression. In addition to the confidence interval for the network estimator, a prediction interval is shown. Local approach to check inconsistency—comparison of direct and indirect estimate for closed loops. As presented in Figure 37, there is one closed loop in the network (zoledronic acid—clodronate—no treatment/placebo). There is no significant difference between direct and indirect estimate (P value of test for disagreement: P = 0.822).

41

Net heat plot for the outcome skeletal‐related event: spinal cord compression (random‐effects model). There are no signs of inconsistency in the net heat plot.

42

Network diagram for the outcome skeletal‐related event: radiotherapy. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

43

Forest plot for the outcome skeletal‐related event (SRE): radiotherapy. Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results.

44

Leaguetable of network meta‐analysis for the outcome skeletal‐related event: radiotherapy. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 8. No. of treatments: 5. No. of pairwise comparisons: 8. No. of designs: 5 Qtotal = 0.29, P = 0.99/Qwithin = 0.28, P = 0.96/Qbetween = 0.00, P = 0.95; I2 = 0.0%, Tau2 = 0 Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

45

Forest plot of splitting direct and indirect evidence for the outcome skeletal‐related event: radiotherapy. In addition to the confidence interval for the network estimator, a prediction interval is shown. Local approach to check inconsistency—comparison of direct and indirect estimate for closed loops. As presented in Figure 42, there is one closed loop in the network (zoledronic acid—clodronate—no treatment/placebo). There is no significant difference between direct and indirect estimate (P value of test for disagreement: P = 0.955).

46

Net heat plot for the outcome skeletal‐related event: radiotherapy (random‐effects model). There are no signs of inconsistency in the net heat plot.

47

Network diagram for the outcome skeletal‐related event: bone surgery. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

48

Forest plot for the outcome skeletal‐related event: surgery. Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields slightly different confidence intervals (Figure 50).

49

Leaguetable of network meta‐analysis for the outcome skeletal‐related event: surgery. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 6. No. of treatments: 5. No. of pairwise comparisons: 6. No. of designs: 4 Heterogeneity/inconsistency: Qtotal = 2.11, P = 0.35; I2 = 5.3%, Tau2 = 0.0216 Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

50

Forest plot for outcome skeletal‐related event (SRE): surgery. Fixed‐effect model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). Since there are no closed loops in the network, no local approach to check inconsistency was conducted.

51

Network diagram for the outcome skeletal‐related event: hypercalcemia. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

52

Forest plot for the outcome skeletal‐related event (SRE): hypercalcemia. Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results. Since there are no closed loops in the network, no local approach to check inconsistency was conducted.

53

Leaguetable of network meta‐analysis for the outcome skeletal‐related event: hypercalcemia. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options would have been marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 4. No. of treatments: 4. No. of pairwise comparisons: 4. No. of designs: 3 Heterogeneity / inconsistency: Qtotal = 0.57, P = 0.45; I2 = 0.0%, Tau2 = 0 Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

54

Network diagram for the outcome mortality. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

55

Forest plot for the outcome mortality. Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results.

56

Leaguetable of network meta‐analysis for the outcome mortality. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 13. No. of treatments: 6. No. of pairwise comparisons: 13. No. of designs: 6 Qtotal = 3.35, P = 0.91/Qwithin = 3.15, P = 0.87/Qbetween = 0.20, P = 0.65; I2 = 0.0%, Tau2 = 0 Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

57

Forest plot of splitting direct and indirect evidence for the outcome mortality. In addition to the confidence interval for the network estimator, a prediction interval is shown. Local approach to check inconsistency—comparison of direct and indirect estimate for closed loops. As presented in Figure 54, there is one closed loop in the network (zoledronic acid—clodronate—no treatment/placebo). There is no significant difference between direct and indirect estimate (P value of test for disagreement: P = 0.654).

58

Net heat plot for outcome mortality (random‐effects model). There are negligible signs of inconsistency in the net heat plot.

59

Forest plot for sensitivity analysis of outcome: mortality (low risk of bias (RoB)). Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results.

60

Leaguetable of sensitivity network meta‐analysis including only studies with low risk of bias for the outcome mortality. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options would have been marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 9. No. of treatments: 5. No. of pairwise comparisons: 9. No. of designs: 5 Qtotal = 2.98, P = 0.70/Qwithin = 2.85, P = 0.58/Qbetween = 0.13, P = 0.72; I2 = 0%, Tau2 = 0 Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

61

Forest plot for sensitivity analysis of splitting direct and indirect evidence for the outcome mortality (low risk of bias). In addition to the confidence interval for the network estimator, a prediction interval is shown. Local approach to check inconsistency—comparison of direct and indirect estimate for closed loops. There is one closed loop in the network (graph not shown; zoledronic acid—clodronate—no treatment/placebo). There is no significant difference between direct and indirect estimate (P value of test for disagreement: P = 0.721).

62

Net heat plot for sensitivity analysis of the outcome mortality (low risk of bias; random‐effects model). There are negligible signs of inconsistency in the net heat plot. The fixed‐effect model yields the same results.

63

Network diagram for the outcome adverse event: grade 3 to 4. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

64

Forest plot for the outcome adverse event: grade 3 to 4. Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results. Since there are no closed loops in the network, no local approach to check inconsistency was conducted.

65

Leaguetable of network meta‐analysis for the outcome adverse event: grade 3 to 4. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 8. No. of treatments: 7. No. of pairwise comparisons: 8. No. of designs: 6 Heterogeneity/inconsistency: Qtotal = 0.44, P = 0.80; I2 = 0.0%, Tau2 = 0 Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

66

Results of pairwise meta‐analysis for the outcome adverse event: grade 3 to 4 (random‐effects model).

67

Network diagram for the outcome adverse event: hypocalcemia. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

68

Forest plot for secondary outcome: adverse event: hypocalcemia. Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). There are strong differences between the fixed‐effect and random‐effects estimates and confidence intervals. In addition, the treatments are ranked in a different order (Figure 70).

69

Leaguetable of network meta‐analysis for the outcome adverse event: hypocalcemia. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options would have been marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 7. No. of treatments: 5. No. of pairwise comparisons: 7. No. of designs: 5 Qtotal = 6.90, P = 0.075/Qwithin = 2.39, P = 0.30/Qbetween = 4.51, P = 0.034; I2 = 56.5%, Tau2 = 1.0252 Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

70

Forest plot for secondary outcome adverse event: hypocalcemia. Fixed‐effect model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending).

71

Forest plot of splitting direct and indirect evidence for the secondary outcome adverse event: hypocalcemia. In addition to the confidence interval for the network estimator, a prediction interval is shown. Local approach to check inconsistency—comparison of direct and indirect estimate for closed loops. As presented in Figure 67, there is one closed loop in the network (zoledronic acid—clodronate—no treatment/placebo). There is no significant difference between direct and indirect estimate (P value of test for disagreement: P = 0.20).

72

Net heat plot for secondary outcome adverse event: hypocalcemia (random‐effects model). There are signs of inconsistency in the net heat plot in the comparison clodronate versus no treatment/placebo. The fixed‐effect model shows even stronger inconsistency (not shown).

73

Results of pairwise meta‐analysis for secondary outcome adverse event: hypocalcemia (random‐effects model).

74

Network diagram for secondary outcome adverse event: fatigue. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

75

Forest plot for the secondary outcome adverse event: fatigue. Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results.

76

Leaguetable of network meta‐analysis for the outcome adverse event: fatigue. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options would have been marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 7. No. of treatments: 5. No. of pairwise comparisons: 7. No. of designs: 4 Heterogeneity/inconsistency: Qtotal = 2.36, P = 0.50; I2 = 0.0%, Tau2 = 0 Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

77

Network diagram for the secondary outcome adverse event: diarrhea. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

78

Forest plot for secondary outcome adverse event: diarrhea. Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). There are slight differences between the fixed‐effect and random‐effects estimates and confidence intervals (Figure 80).

79

Leaguetable of network meta‐analysis for the outcome adverse event: diarrhea. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 6. No. of treatments: 6. No. of pairwise comparisons: 6. No. of designs: 5 Heterogeneity/inconsistency: Qtotal = 1.23, P = 0.27; I2 = 18.7%, Tau2 = 0.0197 Treatment effects + 95% confidence intervals (risk ratios, random‐effects model)

80

Forest plot for secondary outcome adverse event: diarrhea. Fixed‐effect model. Reference treatment: No treatment/placebo. Treatments are ordered by P‐score (descending).

81

Network diagram for the secondary outcome adverse event: nausea. Any two treatments are connected by a line when there is at least one study comparing the two treatments. Line width: number of participants.

82

Forest plot for secondary outcome adverse event: nausea. Random‐effects model. Reference treatment: no treatment/placebo. Treatments are ordered by P‐score (descending). The fixed‐effect model yields similar results.

83

Leaguetable of network meta‐analysis for the outcome adverse event: nausea. Treatment options are ranked as indicated by the arrow from top: greater chance of being the best treatment (higher P‐scores) to bottom: lower chance of being the best treatment (lower P‐score). Network estimates with 95% confidence intervals indicating an effect between two of the treatment options are marked in yellow: treatment option in column better than treatment option in row. Global approach to check inconsistency/heterogeneity: Q‐statistics, I2. No. of studies: 9. No. of treatments: 7. No. of pairwise comparisons: 9. No. of designs: 6 Heterogeneity/inconsistency: Qtotal = 2.33, P = 0.51; I2 = 0.0%, Tau2 = 0 Treatment effects + 95% confidence intervals (risk ratios, random‐effects model).

84

Results of pairwise meta‐analysis for outcome adverse event: nausea (fixed‐effect and random‐effects).

85

Ranking plot representing simultaneously the efficacy (x axis, total number of skeletal‐related events (SREs)) and the acceptability (y axis, adverse event: renal impairment) of all bone‐modifying agents for patients with prostate cancer and bone metastases. Optimal treatment should be characterized by both high efficacy and acceptability and should be in the right upper corner of this graph. Only studies reporting both efficacy (total numbers of SREs) and acceptability (adverse event: renal impairment) are considered in the ranking plot. Studies reporting only one of the two are not included in the statistical analysis for this plot.

86

Leaguetable with network estimates of all pairwise comparisons for efficacy (total number of skeletal‐related events) and acceptability (adverse event: renal impairment). Treatments are presented in alphabetical order. Data are risk ratios (RRs) with corresponding 95% confidence intervals. For both efficacy and acceptability, RRs lower than 1 favor the first treatment in alphabetical order. To obtain RRs for comparisons in the opposite direction, reciprocals should be taken. Results not including the line of no effect in their confidence intervals and therefore suggesting evidence for a difference, are marked bold.

87

Ranking plot representing simultaneously the efficacy (x axis, total number of skeletal‐related events (SREs)) and the acceptability (y axis, adverse event: osteonecrosis of the jaw) of all bone‐modifying agents for patients with prostate cancer and bone metastases. Optimal treatment should be characterized by both high efficacy and acceptability and should be in the right upper corner of this graph. Only studies reporting both efficacy (total numbers of SREs) and acceptability (adverse event: osteonecrosis of the jaw) are considered in the ranking plot. Studies reporting only one of the two are not included in the statistical analysis for this plot.

88

Leaguetable with network estimates of all pairwise comparisons for efficacy (total number of skeletal‐related events) and acceptability (adverse event: osteonecrosis of the jaw). Treatments are presented in alphabetical order. Data are risk ratios (RRs) with corresponding 95% confidence intervals. For both efficacy and acceptability, RRs lower than 1 favor the first treatment in alphabetical order. To obtain RRs for comparisons in the opposite direction, reciprocals should be taken. Results not including the line of no effect in their confidence intervals and therefore suggesting evidence for a difference, are marked bold.

89

Ranking plot representing simultaneously the efficacy (x axis, total number of skeletal‐related events (SREs)) and the acceptability (y axis, grade 3 to 4 adverse events) of all bone‐modifying agents for patients with prostate cancer and bone metastases. Optimal treatment should be characterized by both high efficacy and acceptability and should be in the right upper corner of this graph. Only studies reporting both efficacy (total numbers of SREs) and acceptability (grade 3 to 4 adverse events) are considered in the ranking plot. Studies only reporting one of the two are not included in the statistical analysis for this plot. Results not including the line of no effect in their confidence intervals and therefore suggesting evidence for a difference, are marked bold.

90

Leaguetable with network estimates of all pairwise comparisons for efficacy (total number of skeletal‐related events) and acceptability (grade 3 to 4 adverse events). Treatments are presented in alphabetical order. Data are risk ratios (RRs) with corresponding 95% confidence intervals. For both efficacy and acceptability, RRs lower than 1 favor the first treatment in alphabetical order. To obtain RRs for comparisons in the opposite direction, reciprocals should be taken. Results not including the line of no effect in their confidence intervals and therefore suggesting evidence for a difference, are marked bold.

91

Ranking plot representing simultaneously the efficacy (x axis, total number of skeletal‐related events (SREs)) and the acceptability (y axis, adverse event: hypocalcemia) of all bone‐modifying agents for patients with prostate cancer and bone metastases. Optimal treatment should be characterized by both high efficacy and acceptability and should be in the right upper corner of this graph. Only studies reporting both efficacy (total numbers of SREs) and acceptability (adverse event: hypocalcemia) are considered in the ranking plot. Studies reporting only one of the two are not included in the statistical analysis for this plot.

92

Leaguetable with network estimates of all pairwise comparisons for efficacy (total number of skeletal‐related events) and acceptability (adverse event: hypocalcemia). Treatments are presented in alphabetical order. Data are risk ratios (RRs) with corresponding 95% confidence intervals. For both efficacy and acceptability, RRs lower than 1 favor the first treatment in alphabetical order. To obtain RRs for comparisons in the opposite direction, reciprocals should be taken.

93

Overview of included studies and comparisons for outcome proportion of participants with pain response (A) and adverse event: renal impairment (B).

94

Overview of included studies and comparisons for the outcome adverse event: osteonecrosis of the jaw (A) and total number of skeletal‐related events (SREs) (B).

95

Overview of included studies and comparisons for the outcome skeletal‐related event (SRE): pathological fractures (A) and SRE: spinal cord compression (B).

96

Overview of included studies and comparisons for the outcome skeletal‐related event (SRE): bone surgery (A) and SRE: hypercalcemia (B).

97

Overview of included studies and comparisons for the outcome mortality (A) and grade 3 to 4 adverse events (B).

98

Overview of included studies and comparisons for the outcome adverse event: hypocalcemia (A) and adverse event: fatigue (B).

99

Overview of included studies and comparisons for the outcome adverse event: diarrhea (A) and adverse event: nausea (B).