How to Read a Paper: The Basics of Evidence-Based Medicine by Trisha Greenhalgh

looking at is a non-randomised controlled clinical trial, you must use your common sense to decide if the baseline differences between the intervention and control groups are likely to have been so great as to invalidate any differences ascribed to the effects of the intervention. This is, in fact, almost always the case [7]. Sometimes, the authors of such a paper will list the important features of each group (such as mean age, sex ratio and markers of disease severity) in a table to allow you to compare these differences yourself.
    Cohort studies
    The selection of a comparable control group is one of the most difficult decisions facing the authors of an observational (cohort or case–control) study. Few, if any, cohort studies, for example, succeed in identifying two groups of subjects who are equal in age, gender mix, socioeconomic status, presence of coexisting illness and so on, with the single difference being their exposure to the agent being studied. In practice, much of the ‘controlling’ in cohort studies occurs at the analysis stage, where complex statistical adjustment is made for baseline differences in key variables. Unless this is performed adequately, statistical tests of probability and confidence intervals (see section ‘Probability and confidence’) will be dangerously misleading [6] [7].
    This problem is illustrated by the various cohort studies on the risks and benefits of alcohol, which have consistently demonstrated a J-shaped relationship between alcohol intake and mortality. The best outcome (in terms of premature death) lies with the cohort group who are moderate drinkers [8]. Self-confessed teetotalers, it seems, are significantly more likely to die young than the average person who drinks three or four drinks a day.
    But can we assume that teetotallers are, on average , identical to moderate drinkers except for the amount they drink? We certainly can't. As we all know, the teetotal population includes those who have been ordered to give up alcohol on health grounds (‘sick quitters’), those who, for health or other reasons, have cut out a host of additional items from their diet and lifestyle, those from certain religious or ethnic groups which would be under-represented in the other cohorts (notably Muslims and Seventh Day Adventists), and those who drink like fish but choose to lie about it.
    The details of how these different features of teetotalism were controlled for by the epidemiologists are discussed elsewhere [8] [9]. Interestingly, when I was writing the third edition of this book in 2005, the conclusion at that time was that even when due allowance was made in the analysis for potential confounding variables in people who described themselves as non-drinkers, these individuals' increased risk of premature mortality remained (i.e. the J curve was a genuine phenomenon) [8].
    But by the time I wrote the fourth edition in 2010, a more sophisticated analysis of the various cohort studies (i.e. which controlled more carefully for ‘sick quitters’) had been published [9]. It showed that, all other things being equal, teetotallers are no more likely to contract heart disease than moderate drinkers (hence, the famous ‘J curve’ may have been an artefact all along). Subsequently, a new meta-analysis purported to show that the J curve was a genuine phenomenon and alcohol was indeed protective in small quantities [10]—but a year later a new analysis of the same primary studies came to the opposite conclusion, having placed more weight on so-called methodological flaws [11]. Depending on your perspective, this could be one to discuss with your EBM colleagues over a beer.
    Case–control studies
    In case–control studies (in which, as I explained in section ‘Case reports’, the experiences of individuals with and without a particular disease are analysed retrospectively to identify exposure to possible causes of that disease), the process most open to bias is not the assessment of