2. 1 Review methods.

Study selection	The titles and abstracts of search results are scanned and compared to predetermined selection criteria for studies. Articles written in foreign languages are translated to English. Two authors working independently select trials. The authors are blinded to the trial uthors and institutions by asking the trial search co‐ordinator to block this out when sending abstracts. Where there is uncertainty about inclusion, the full text of the article is obtained and read. If additional information is needed about the trial in question, it is categorized as awaiting assessment until this is obtained. If differences arise regarding trial inclusion, they are discussed and if still unresolved, Dr Taryn Young at the South African Cochrane Centre is asked to take a decision.
Data extraction	Data are collected and recorded in similar data extraction forms. The following information is extracted; citation details, study eligibility, study quality and study characteristics. The study characteristics include methods, participants, interventions and outcomes.
Assessment of methodological quality of included studies	The methodological quality of the included studies is assessed to determine validity. Selection bias is checked by assessing whether the generation of a random allocation sequence and allocation concealment were performed. Randomization is considered adequate if the allocation sequence is generated from a table of random numbers or by computer. Allocation concealment is deemed adequate if undertaken by means of sequentially prenumbered sealed opaque envelopes, a centralised system or prenumbered coded identical containers. In addition, assessment of blinding, losses to follow‐up and whether the analysis was by intention‐to‐treat (ITT) is undertaken. The definition of ITT is the requirement that participants be analyzed in the groups to which they were randomized, regardless of which intervention they actually received.
Data analysis	Measures of treatment effect:The effect measures of choice here are relative risk (RR) for dichotomous data and hazard ratio (HR) for time‐to‐event data with a 95% confidence interval.
	Dichotomous data include: ‐ number of people who experience local control in each comparison group ‐ number of people who experience recurrence during the specified time periods in each comparison group ‐ number of deaths in each comparison group ‐ number of people who experience adverse events in each comparison group ‐ number of people who experience local control in each comparison group ‐ number of people who experience recurrence during the specified time periods in each comparison group ‐ number of deaths in each comparison group ‐ number of people who experience adverse events in each comparison group
	Time‐to‐event data include: ‐ time to recurrence
	Dealing with missing data: Participants lost to follow up are censored in the survival analysis.
	Assessment of heterogeneity: To identify statistical heterogeneity we look at the forest plot for overlapping confidence intervals and test it using the Cochrane Q test with a p‐value of 0.10. The impact of heterogeneity on the meta‐analysis is measured using the I‐squared test. If I‐squared is greater than 70% we will investigate by checking the trials for data entry errors and looking for existence of subgroups. Subgroup analysis by age, sex, geographical location and diagnostic method will be done. These factors have been shown to influence the occurrence of OSSN and we anticipate that they may also influence treatment effects. If no explanation for heterogeneity is found or its correction is not possible, meta‐analysis will not be done.
	Meta‐analysis: A fixed‐effect model is used for meta‐analysis but when there is heterogeneity that cannot be readily explained, a random‐effects model is incorporated.
	Sensitivity analysis: We examine how the magnitude of effect differs according to study quality or trial size and also the results of per‐protocol analysis with those of intention‐to‐treat analysis.
	Assessment of reporting bias: Publication bias is assessed by using a funnel plot to look for asymmetry.