Most of results are shown in a normal/normal plot. However there are some studies that are in log scale, if I transform back the results, it looks like an outlier and compromises the results.
Should I keep the log scale for these studies only, log transform everything or discart these observations?
I'm quite lost...
It is always better to present data in log scale. So you will have to convert the results to log scale and then do the meta-analysis.
If the effect estimate is rate ratio and if you have number of events and number of patients in both groups, you can calculate the log of rate ratio.
Analysing data in log-scale has two main advantages:
- dealing with typical skewed distribution of outcomes (such as RR and OR)
- raw RR and OR are not symmetric about 1
you can read these two links for some examples
http://handbook.cochrane.org/chapter_9/9_2_7_expressing_intervention_effects_on_log_scales.htm
http://stats.stackexchange.com/questions/158985/meta-analysis-of-trials-given-hazard-ratios
Hey Julian,
I would humbly disagree that log is always better, as most people have difficulties imagining differences in log terms (making the data much less intuitively understandable). However, Giorgio's arguments still hold - log transforming, from an analysis standpoint, makes your data more "normal" and thus better fit for many tests.
I assume you have made sure you used the same transform. e.g. same base (2,10,e) and offset (log(c+X))?
For your analysis, it is advised to use the same scale for everything (be it log or linear).
Greetings, David
- Badges
- Science topic