Types of studies

We included RCTs in all languages. A RCT was defined as a study in which participants were allocated to treatment groups on the basis of a method to generate a random sequence (for example, using random‐number tables).

We did not include studies with cross‐over designs because these are not appropriate designs for the clinical condition of interest in this review and for this research question.

Types of participants

We included adults aged 18 and over with traumatic corneal abrasion(s) (including corneal abrasions arising from foreign body removal).

Types of interventions

The target intervention was topical NSAIDs (dose as defined by study authors, either overall daily dose or number of drops per day) in adults with traumatic corneal abrasions (including corneal abrasions arising from foreign body removal), compared to the following interventions:

1. Administration of cycloplegics (e.g. cyclopentolate drops, homatropine drops).

2. Administration of oral analgesics (e.g. NSAIDs, opioids, paracetamol/acetaminophen).

3. Administration of ocular lubricants (e.g. artificial tears (hydrogels)).

4. Administration of topical antibiotics (e.g. chloramphenicol, fusidic acid, trimethoprim/polymyxin).

5. Eye patching.

Types of outcome measures

Electronic searches

The Cochrane Eyes and Vision Information Specialist conducted systematic searches in the following databases for randomised controlled trials and controlled clinical trials. There were no language or publication year restrictions. The date of the search was 30 March 2017.

• Cochrane Central Register of Controlled Trials (CENTRAL; 2017, Issue 2) (which contains the Cochrane Eyes and Vision Trials Register) in the Cochrane Library (searched 30 March 2017) (Appendix 1);

• MEDLINE Ovid (1946 to 30 March 2017) (Appendix 2);

• Embase Ovid (1980 to 30 March 2017) (Appendix 3);

• LILACS (Latin American and Caribbean Health Science Information database (1982 to 30 March 2017) (Appendix 4);

• OpenGrey (System for Information on Grey Literature in Europe) (www.opengrey.eu/; searched 30 March 2017) (Appendix 5);

• ZETOC (1993 to 30 March 2017) (Appendix 6);

• ISRCTN registry (www.isrctn.com/editAdvancedSearch; searched 30 March 2017) (Appendix 7);

• US National Institutes of Health Ongoing Trials Register ClinicalTrials.gov (www.clinicaltrials.gov; searched 30 March 2017) (Appendix 8);

• World Health Organization International Clinical Trials Registry Platform (www.who.int/ictrp; searched 30 March 2017) (Appendix 9).

Searching other resources

We made additional efforts to identify potential RCTs relevant to the topic from the references (and references of references) cited in primary sources. We did not impose any language restriction.

Selection of studies

Two review authors (RM and OG) independently assessed the titles and abstracts of studies identified by relevance and design. We obtained full‐text versions of the articles if they appeared to meet the inclusion criteria in the initial assessment of studies. A third review author (AW) evaluated any discrepant judgements.

Data extraction and management

Two review authors (MB and MQ) independently extracted data using a standardised data collection form that included information on the name of the first author, year of publication, study design, study population and study setting. In addition to information pertaining to participant characteristics, study inclusion and exclusion criteria, details of the interventions compared and study outcomes, we extracted information on study methodology. This included the method of randomisation, allocation concealment, frequency and handling of withdrawals, and adherence to the intention‐to‐treat principle. We resolved disagreements through discussion and in consultation with a third review author (AW) as required.

Assessment of risk of bias in included studies

Two review authors (MB and MQ) independently assessed and rated the methodological quality of each trial using the Cochrane tool for assessing risk of bias as in Chapter 8 of the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011). We judged the quality of the studies by evaluating them for the following domains:

1. Random sequence generation.

2. Allocation concealment.

3. Masking of participants and personnel, and outcome assessment.

4. Incomplete outcome data.

5. Selective outcome reporting.

6. Funding source.

7. Other potential sources of bias.

We evaluated each study and assessed it separately for these domains. We judged each explicitly as follows:

• Low risk of bias.

• High risk of bias.

• Unclear risk (lack of information or uncertainty over the potential for bias).

We entered the data on what was reported to have happened in the study in the 'Risk of bias' table in Review Manager 5 (Review Manager 5 2014). We present summary figures of the 'risk of bias in included studies' in the review. These provides a context for discussing the reliability of the results of this review. We resolved any disagreement by referring to a third review author (AW) to reach a consensus.

Measures of treatment effect

We calculated summary estimates of treatment effect with 95% confidence intervals (CIs) for each comparison. Our measure of treatment effect was the risk ratio (RR) for dichotomous outcomes and the mean difference (MD) for continuous outcomes. Currently the review only includes analysis of dichotomous outcomes.

Unit of analysis issues

The unit of randomisation was the eye of individual trial participants. We did not anticipate that studies would have more than one eye affected in each individual; however, if this occurred we planned to note it in the review. If studies using a paired design were eligible for inclusion (i.e. studies assigning one eye to treatment and the fellow eye to control), we planned to use the generic inverse variance method to combine the results of such studies with those of studies randomising only one eye for each participant.

Dealing with missing data

No simple solution exists for the problem of missing data. We planned to handle this problem by contacting the investigators, whenever possible, to ensure that no data were missing for their study. We also planned to make explicit the assumptions of whatever method we used to cope with missing data.

Assessment of heterogeneity

We evaluated clinical heterogeneity (differences between studies in key characteristics of the participants, interventions or outcome measures). In the absence of clinical heterogeneity, we used the I² statistic to describe the percentage of total variation across studies that is due to heterogeneity rather than chance (Higgins 2003). An I² greater than 50% may represent substantial or considerable statistical heterogeneity (Higgins 2011). The importance we placed on the observed value of I² depended on (i) magnitude and direction of effects, and (ii) strength of evidence for heterogeneity (P value from the Chi² test and confidence interval for I²).

We also used visual inspection of the graphic representation of studies with their 95% CIs to assess heterogeneity. We generated tables and graphs using the analysis module included in RevMan (Review Manager 5 2014). We represent pooled risk ratios pictorially as a 'forest plots' to permit visual examination of the degree of heterogeneity between studies.

Assessment of reporting biases

We assessed reporting bias through careful attention to quality assessment, particularly methodology. We planned to use funnel plot analysis to assess publication bias if there were more than 10 studies included in the meta‐analysis. We also planned to use the Egger test (Egger 1997) to assess funnel plot asymmetry. A thorough search for unpublished studies through grey literature searches and contact with known experts in the field also helped to reduce the risk of publication bias.

Data synthesis

The results concentrate on the objectives and comparisons specified in the protocol for our review. We pooled data using a random‐effects model, because it was likely that the effects of topical NSAIDs may vary between studies. The random‐effects model takes into account between‐study variability as well as within‐study variability. When there were three or fewer trials, we used a fixed‐effect model. We performed meta‐analyses using RevMan 5 software (Review Manager 5 2014).

Subgroup analysis and investigation of heterogeneity

We planned to investigate heterogeneity by performing two subgroup analyses based on intuitive reasons. Firstly, we planned to perform subgroup analysis of different types of topical NSAIDs (for example, subgroup analysis of topical diclofenac and topical ketorolac). Secondly, we planned to perform subgroup analysis of traumatic corneal abrasions with different aetiologies, based on whether the abrasions are iatrogenic (arising from foreign body removal) or non‐iatrogenic in origin.

Sensitivity analysis

Finally, we planned to perform sensitivity analyses to test how sensitive the results were to reasonable changes in the assumptions that we made and in the methods for combining the data (Lau 1998). We planned to perform sensitivity analysis for randomised versus quasi‐randomised studies and eventually good‐quality studies versus poor‐quality studies.