Abstract / Description of output
Background
Effects on continuous non-normally-distributed outcomes can be arduous to pool in meta-analyses as the mean is often not reported. We assessed three methods that estimate the missing mean using other summary statistics (Hozo 2005: median, minimum, maximum, sample size; Bland 2014: median, minimum, maximum, sample size, upper and lower quartile; Wan 2014: median, upper and lower quartile, sample size).
Methods
We assessed their performance on a skewed outcome (hospital length of stay), comparing their impact on the overall meta-analysis results. Using individual patient data from two sources (the GALA carotid surgery trial and a published meta-analysis of early supported discharge (ESD) after stroke) we simulated a wide range of scenarios, varying size and number of the included studies and size and number of studies not reporting a mean.
Results
The Wan approach most closely estimated the actual treatment effect, mean difference (SD) 0.07 (0.11) days in GALA and 0.38 (0.45) days in ESD, with little effect on the confidence interval (CI) width, mean CI ratio 0.97 (0.08) in GALA and 0.98 (0.03) in ESD., The Hozo and Bland approaches led to larger differences in effect (respectively 0.91 (2.41) and 0.40 (1.08) days in GALA and 1.58 (1.76) and 0.73 (0.83) days in ESD) and much wider CIs (ratios 2.05 (1.87) and 1.49 (0.89) in GALA and 1.14 (0.22) and 1.06 (0.09) in ESD).
Conclusions
The Wan method performed best and is also more practical, since it uses the median and quartiles which are more often available than other summary statistics, especially for non-normally distributed outcomes. As the Hozo and Bland formulae include also the minimum and maximum, their estimates are more susceptible to outliers.
Effects on continuous non-normally-distributed outcomes can be arduous to pool in meta-analyses as the mean is often not reported. We assessed three methods that estimate the missing mean using other summary statistics (Hozo 2005: median, minimum, maximum, sample size; Bland 2014: median, minimum, maximum, sample size, upper and lower quartile; Wan 2014: median, upper and lower quartile, sample size).
Methods
We assessed their performance on a skewed outcome (hospital length of stay), comparing their impact on the overall meta-analysis results. Using individual patient data from two sources (the GALA carotid surgery trial and a published meta-analysis of early supported discharge (ESD) after stroke) we simulated a wide range of scenarios, varying size and number of the included studies and size and number of studies not reporting a mean.
Results
The Wan approach most closely estimated the actual treatment effect, mean difference (SD) 0.07 (0.11) days in GALA and 0.38 (0.45) days in ESD, with little effect on the confidence interval (CI) width, mean CI ratio 0.97 (0.08) in GALA and 0.98 (0.03) in ESD., The Hozo and Bland approaches led to larger differences in effect (respectively 0.91 (2.41) and 0.40 (1.08) days in GALA and 1.58 (1.76) and 0.73 (0.83) days in ESD) and much wider CIs (ratios 2.05 (1.87) and 1.49 (0.89) in GALA and 1.14 (0.22) and 1.06 (0.09) in ESD).
Conclusions
The Wan method performed best and is also more practical, since it uses the median and quartiles which are more often available than other summary statistics, especially for non-normally distributed outcomes. As the Hozo and Bland formulae include also the minimum and maximum, their estimates are more susceptible to outliers.
| Original language | English |
|---|---|
| Publication status | Published - 12 Jul 2017 |
| Event | International Society for Clinical Biostatistics - Vigo, Spain Duration: 9 Jul 2017 → 13 Jul 2017 Conference number: 38 |
Conference
| Conference | International Society for Clinical Biostatistics |
|---|---|
| Abbreviated title | ISCB |
| Country/Territory | Spain |
| City | Vigo |
| Period | 9/07/17 → 13/07/17 |