Even though in the protocol we stated that in cases where we would include more than one comparison from a trial with multiple arms in the same meta‐analysis, we would halve the numbers of control group participants to prevent them from being included twice, this does not work for the inverse variance input method. Neuhaus 2014a reported only the results from Ancova and could not provide us with the raw data. For this trial we modelled the means and standard deviations from the intervention and the control group in RevMan as closely to the real data as possible to achieve the same mean difference and standard error. Then we halved the number of participants in the control group and entered the resulting standard errors into RevMan.

We judged studies to be at low risk for selective outcome reporting if the final publications of the trial reported what had been planned and registered in international databases (trial registries), such as ClinicalTrials.gov, ANZCTR.org.au (Australia and New Zealand), NTR (Netherland’s Trial Registry). We judged those studies that were not registered in trial registries as being at low risk for selective outcome reporting if they reported all the outcomes mentioned in the methods section.

Initially we planned to pool interventions that were categorised under broad headings like physical changes in workplace environment, policy changes and information and counselling, but later we found that the interventions were quite different from one another and decided not to combine them under these broad headings. We also added a new category consisting of approaches that used multiple categories of interventions at the same time.