The results reported in table 3 and the abstract were:
- Improvement of pain while lying in bed on VAS OR 2.36 (1.13 – 4.93).
- Improvement of pain on rising OR 1.93 (0.97 – 3.86).
- Improvement of pain related disability (Roland Morris Questionnaire) OR 2.10 (1.24 – 3.56).
However, there were two sets of results. There are different results in the text page 1602 at the bottom of the first column.
The alternative results are:
- Improvement of pain while lying in bed on VAS OR 1.35 (0.77- 2.36).
- Improvement of pain on rising OR 1.48 (0.81-6.68).
- Improvement of pain related disability (Roland Morris Questionnaire) OR 2.10 (1.24-3.56)
Notice that the first set of results reach statistical significance for pain in bed and disability but the second set do not. Which do you think are the right results?
We think the second set of results are the ones that should have been reported in the abstract. The authors probably would not have got their paper published in The Lancet had they done this!
The authors may have used logistic regression inappropriately. It would be important to know if this type of analysis had been specified a priority in the protocol or whether it is an afterthought because the results were not statistically significant. We do not think that logistic regression was needed as there should not have been a need to adjust for confounding (if a trial is randomised and of sufficient size this should mean that confounders will tend to be distributed evenly across both groups and as the authors themselves note, the two groups have similar baseline characteristics. In addition, when using logistic regression the variables should be independent of each other, and they are not.
NOTE: the power calculation on page 1600 statistical analysis, when the authors said they needed to show a difference in improvement of 20%. The difference is actually 4-5% so the study is underpowered.