Logistic regression and odds ratios as means to adjust for baseline incomparability’s in randomised controlled trials: description of the disadvantages, alternatives and frequency of use
MetadataShow full item record
Background: In randomised controlled trials (RCTs) adjustment for baseline characteristics is often done. When logistic regression is used, the measure of association will be the odds ratio (OR). If the incidence of the outcome is high (>10%) the odds ratio is incomparable to the risk ratio. Furthermore, the larger the effect of the intervention (OR < 0.5 or > 2.0) the larger the difference between the odds ratio and risk ratio. Additionally, we reviewed how often ORs are reported which are incomparable to risk ratios in RCTs and how frequently researchers use logistic regression in RCTs. In this report we present alternative methods to adjust for baseline incomparability’s with which it is possible to calculate RRs directly. By analysing RCT we determine to what extent alternatives for logistic regression are used. Methods: We reviewed all 298 RCTs published between January 1st and December 31st 2008 in the British Medical Journal, Annals of Internal Medicine, The Lancet, The Journal of the American Medical Association and The New England Journal of Medicine. All articles were scored on a standardised form. We calculated the percentage of RCTs that presented an OR below 0.5 or above 2.0, an incidence above 10%, or both. Results: In 78 RCTs (26.2%) logistic regression was used for any purpose and 55 RCTs (18.5%) reported one or more ORs. ORs smaller than 0.5 or/and larger than 2.0 were reported in 37 RCTs (12.4%), and in 20 RCTs (6.7%) at least one overall incidence reported was above or equal to 10%. In 9 RCTs (3.0%) an OR below 0.5 or above 2.0 with a corresponding incidence above 10% was reported, this is 13% of the RCTs that reported ORs. Alternative methods for baseline adjustment in RCTs are used only 34 times out of the 298 publications that were analysed. Conclusion: Our results show that in 1 in 33 RCTs published in five major general clinical journals the presented odds ratio cannot be interpreted as a risk ratio. The use of logistic regression occurs in 1 in 4 RCTs while there are more appropriate models available. The available alternative methods are very scarcely used, although they can be easily applied with modern statistical computer programs. A misinterpretation of risks or benefits of e.g. surgical procedures or pharmacotherapy may have life threatening consequences and could cost huge amounts of money in clinical settings.