Online literature searching strategy

We searched the online datasets of PubMed, Web of Science, and Cochrane library using the terms of “nasopharyngeal carcinoma or cancer,” “radiotherapy,” and “induction chemotherapy or neoadjuvant chemotherapy or adjuvant chemotherapy or concurrent chemoradiotherapy” to identify all potential clinical trials. For related Chinese studies, the National Knowledge Infrastructure and WanFang datasets were searched. All studies were restricted to randomized clinical trial. Two investigators completed this process independently, and any discrepancy was solved by consensus. Our network meta-analysis was conducted in accordance with the PRISMA guidelines (PRISMA checklist).

Continue Reading

Related Articles

Clinical trial inclusion criteria

The clinical trials published between 1998 and 2018 would be included into this study for analysis if they meet the following criteria: 1) randomized phase II–III clinical trial; 2) recruiting patients with newly diagnosed, nonmetastatic, and advanced NPC; 3) RT was delivered as conventional fraction and total radiation dosage of 66 Gy or more should be scheduled; 4) either experimental or control arm in the trials should contain one of the four treatments (induction + CRT, CRT + adjuvant, CRT alone, RT alone).

Study quality evaluation

In order to select appropriate statistical method and obtain unbiased results, we employed the Jadad/Oxford quality scoring system16 to assess the study quality. The randomization procedure, sample size calculation, adoption of blinded principle, allocation concealment, and intention-to-treatment analysis of included trials were reviewed and scored according to the standard. Two authors completed this process independently and discrepancies were solved by consensus.

Study data extraction

Three investigators reviewed each included study separately to collect and extract related data including study author information, patient recruitment time period, sample size, patient tumor stage, RT and chemotherapy protocol, follow-up duration, and study endpoints. All extracted information and data were reviewed by the fourth investigator to check whether discrepancy existed between the three investigators. Otherwise, discrepancy would be solved by consensus.

Study endpoint

In our current network meta-analysis, we set OS (defined as the time interval between randomization and death from any cause) as the primary endpoint. The other two endpoints were distant failure-free survival (DFFS, defined as time interval between randomization to first distant metastasis) and locoregional failure-free survival (LFFS, defined as time interval between randomization and first local or regional or both recurrence). Given the different definitions of progression-free survival (PFS) or disease-free survival in different clinical trials, we therefore did not perform analysis on this endpoint.

Statistical analysis

Survival data were extracted from trials and expressed in our study as HRs and corresponding CIs since they are the only summary statistic allowed for censoring and time to an event. HRs and corresponding CIs were extracted from original text if they were available, otherwise they were obtained them from a previous meta-analysis15 or a pooled data analysis.17

Pairwise meta-analysis between two treatment arms was conducted first. Pooled HRs and corresponding 95% CIs of direct comparison between two treatment arms were calculated to evaluate the survival difference, and P<0.05 was considered significant. We used the chi-squared test and I2statistic to establish the heterogeneity between studies, and the χ2 P-value <0.1 or an I2 statistic >50% was considered significant. Stata statistical package 13.0 (StataCorp LP, College Station, TX, USA) was applied to complete this analysis.

For network meta-analysis, multiple treatment comparisons were conducted using netmeta package18,19 and frequentist approach18 in R software (version 3.3.5; R Foundation, Vienna, Austria). The logarithmic of HR (logHR) and its variance (selogHR) of each direct comparison were calculated for statistical network meta-analysis. Also, treatment effects of network meta-analysis were estimated by HRs and 95% CIs and presented in forest plots. Inconsistency and heterogeneity between and within different comparisons were evaluated by Q test, which was proposed by Rücker.19 The P-value of Q test >0.1 indicates no heterogeneity, and vice versa. Random-effect model would be applied and sensitivity analysis would be performed in case of significant heterogeneity. Finally, each treatment would be ranked by a P-score, which was proposed by Rücker and Schwarzer18 as a frequentist analog to surface under the cumulative ranking curve.20,21 Briefly, a P-score of 100% indicates the best treatment and 0% for the worst treatment. P<0.05 was considered significant for all analysis.