Article type
Year
Abstract
Background: In randomised trials with count outcomes (such as number of falls/exacerbations of asthma) it is possible to analyse the results in many different ways. It is not known whether it is possible to combine the results of these different methods of analysis in a meta-analysis.
Objectives: To see if different methods of analysis provide results that are alike enough to combine in a meta-analysis.
Methods: Simulation study with count data with a range of means and amount of overdispersion. Different analyses were rate ratios, poisson regression, negative binomial regression, dichotomising the data (with and without the event), three forms of survival analyses, ratio of means and ratio of medians. Confirmation by the analysis of individual patient data from at least 14 randomised trials.
Results: With a low mean, (about 20% have the event) most methods of analysis produced similar answers, except ratio of medians, which was not possible to calculate. As the mean increased, dichotomising the data increasingly underestimated the treatment effect, as did time to first event and this effect was noticeable even when 50% had the event. Survival models allowing for multiple events helped, but not completely. Negative binomial gave results very similar to poisson regression, even for considerable amounts of overdispersion, where these models were better fits to the data. These results were confirmed by the analyses of the individual patient data.
Conclusion: It should be possible to combine more studies in a meta-analysis than was previously expected. Relative risks from dichotomised data and hazard ratios from time to first event analyses should only be included if the event rate is very low. Means can be converted to ratios and an approximate standard deviation calculated, but for the ratio of medians it is not possible to work out the standard deviation from published data.
Objectives: To see if different methods of analysis provide results that are alike enough to combine in a meta-analysis.
Methods: Simulation study with count data with a range of means and amount of overdispersion. Different analyses were rate ratios, poisson regression, negative binomial regression, dichotomising the data (with and without the event), three forms of survival analyses, ratio of means and ratio of medians. Confirmation by the analysis of individual patient data from at least 14 randomised trials.
Results: With a low mean, (about 20% have the event) most methods of analysis produced similar answers, except ratio of medians, which was not possible to calculate. As the mean increased, dichotomising the data increasingly underestimated the treatment effect, as did time to first event and this effect was noticeable even when 50% had the event. Survival models allowing for multiple events helped, but not completely. Negative binomial gave results very similar to poisson regression, even for considerable amounts of overdispersion, where these models were better fits to the data. These results were confirmed by the analyses of the individual patient data.
Conclusion: It should be possible to combine more studies in a meta-analysis than was previously expected. Relative risks from dichotomised data and hazard ratios from time to first event analyses should only be included if the event rate is very low. Means can be converted to ratios and an approximate standard deviation calculated, but for the ratio of medians it is not possible to work out the standard deviation from published data.