Article type
Year
Abstract
Background: Skewed data are often characterised by a standard deviation (SD) greater than half the mean and standard deviations that increase with increasing means. A log transformation is often taken to remove the skewness. When authors are faced with summary statistics from a skewed distribution, the best method is to obtain the individual patient data (IPD) and transform the data.
Objectives: To convert summary data from a skewed distribution into summary data on the log scale, anti-logs can be taken to convert back to the natural scale, for interpretion.
Methods: Using the relationship between the parameters on the natural scale and the parameters on the log scale, formulae were derived to convert one from the other and conduct a more appropriate analysis for skewed data.
Results: The example is from a review of 'Desferrioxamine mesylate for managing transfusional iron overload in people with thalassaemia. The means and standard deviations (sd) on the natural scale of the four groups from the two studies included in this analysis were (mean (sd)), 7.9(5.52), 13.7(5.23), 3.16(2.51) and 2.34(2.19). From this it can be seen that the standard deviations are all greater than half the mean, and sds increase with the means, suggesting that the data might be skewed. The analysis using the extracted data had an overall pooled estimate of -0.35 (95% CI -1.73 to 1.02) (Standardised mean difference, random effects) with an I2 value of 87.6%. The data was converted and the ratio of the geometric means was 0.7 (95% CI 0.53 to 0.93) and the I2 value was 91.5%.
Conclusion: The new analysis highlighted the heterogeneity between the studies and the reader should therefore be cautious in drawing conclusions from these results. Analysing data within RevMan that comes from a skewed distribution may give spurious results. The analysis carried out using the log transformations was more appropriate to skewed data, and gives a way of adjusting for unequal variances, or even different measurement scales, without using SMDs.
Acknowledgements: The results used in this abstract have been taken from a systematic review undertaken by the Systematic Reviews Initiative, National Blood Service, England.
Objectives: To convert summary data from a skewed distribution into summary data on the log scale, anti-logs can be taken to convert back to the natural scale, for interpretion.
Methods: Using the relationship between the parameters on the natural scale and the parameters on the log scale, formulae were derived to convert one from the other and conduct a more appropriate analysis for skewed data.
Results: The example is from a review of 'Desferrioxamine mesylate for managing transfusional iron overload in people with thalassaemia. The means and standard deviations (sd) on the natural scale of the four groups from the two studies included in this analysis were (mean (sd)), 7.9(5.52), 13.7(5.23), 3.16(2.51) and 2.34(2.19). From this it can be seen that the standard deviations are all greater than half the mean, and sds increase with the means, suggesting that the data might be skewed. The analysis using the extracted data had an overall pooled estimate of -0.35 (95% CI -1.73 to 1.02) (Standardised mean difference, random effects) with an I2 value of 87.6%. The data was converted and the ratio of the geometric means was 0.7 (95% CI 0.53 to 0.93) and the I2 value was 91.5%.
Conclusion: The new analysis highlighted the heterogeneity between the studies and the reader should therefore be cautious in drawing conclusions from these results. Analysing data within RevMan that comes from a skewed distribution may give spurious results. The analysis carried out using the log transformations was more appropriate to skewed data, and gives a way of adjusting for unequal variances, or even different measurement scales, without using SMDs.
Acknowledgements: The results used in this abstract have been taken from a systematic review undertaken by the Systematic Reviews Initiative, National Blood Service, England.