Work

Zanamivir for the treatment of influenza in adults: a systematic review and economic evaluation

Year: 2002
Type: review
Abstract: Many continuous biomedical outcome measures have naturally skewed distributions. Current methodologies for pooling continuous data are limited to methods for combining differences in means or standardised differences in means. These approaches assume normality of the data, and require arithmetic means and SDs to be available for each arm of the study. However, skewed outcomes are most appropriately summarised using medians or geometric means. Meta-analyses of skewed outcomes, therefore, typically necessarily exclude studies that report the most appropriate summary (medians) whilst including studies that use potentially misleading summaries (arithmetic means). It is possible to pool any summary statistic using inverse variance weights if a standard error of the summary statistic is available. Whilst there is no formula for calculating the standard error of a difference in medians, estimates of a pseudo standard error may be obtained from measures of uncertainty and statistical significance that may be obtained from measures of uncertainty, which may be presented in conjunction with the medians. Through simulation, we are investigating the relationships between sample size, the degree of skew and: • the bias and performance of arithmetic means for summarising skewed continuous data • the bias and performance of methods of pooling medians using alternative estimates of standard error. Introduction • Weighted mean difference (WMD) and standardised mean difference are two methods used to combine continuous data. • Standardised mean difference is used in preference to WMD when different scales have been used to measure the same outcome and is the difference in means divided by the pooled SD. • Both methods require means, SDs and sample sizes to be reported and use an inverse weighted method. • The underlying assumption behind the use of these methods is that the data are normally distributed. Reporting means and SDs as summary statistics also relies on an assumption of normality. • If the underlying data of the individual trials is not symmetrically distributed but is positively skewed, what implications does this have for: – the results of a single trial – the results of a meta-analysis combining these single trials? • If the data has a symmetrical shape then estimates of the mean and median will be approximately equivalent. If the data are positively skewed estimates of means and medians separate such that estimates of the mean are greater than the median. The mean is strongly affected by skew, and is not a good measure of central tendency. • Figure 18 illustrates that as the level of skewness increases the mean can over-estimate the treat- ment effect by up to 120% relative to the median. How can you tell if the aggregate data from the trials for meta-analysis might be positively skewed? • If your data are naturally bounded at zero and the mean/SD positively skewed. • If the SDs of the trials increase with estimates of the means then this may indicate that the data have a log-normal distribution. This is because the formula for the SD of the log-normal distribution is given by m(w 2 – w) 1 /2 where m is the median, w = exp(s ) and s 2 is the variance of the log-transformed data. Hence, the SD increases with increasing medians. • Although the spread of the data is controlled by the magnitude of the median, the shape or the skew of the data is determined by (w 2 – w) 1 . The larger this quantity, the more positively skewed the data. Appendix 2 Method for pooling medians
Language: English
Subfield: Epidemiology
Field: Medicine
Open Access status: diamond
OpenAlex ID: W2324712985
FWCI: 1.06
Cites: 40
Cited by: 123
Related to: 9