Skip navigation

Swots' Corner

1 MALMAP - four legs good, two legs better



In the absence of very large randomised controlled trials, meta-analysis of the results of small trials remains one of the best types of evidence. As with any type of research, good work is best; as always it is useful to know what features of any piece of work we should look for to see if it is the best we could ask for.

In meta-analysis what is emerging is a clear distinction between two types of meta-analysis - that based on meta-analysis of individual patient data (MAP) and meta-analysis of data from the literature (MAL).

MAL is the traditional type of meta-analysis where the authors find all the trials on a particular topic in the literature, and then combine the data in the form of a meta-analysis. Most of the meta-analyses in the Cochrane Database of Systematic Reviews, and, indeed, most meta-analyses ever done are of this type. These are tremenously helpful - they identify all the papers published on a particular topic (and often unpublished material) as well as providing an overview of the results. Even if we do not agree with the result or conclusions, at least we can get the papers, read them, and try to draw our own conclusions.

In meta-analysis of individual patient data (MAP) those doing the meta-analysis ask research workers for the original patient data and work with and from it to derive their meta-analytic conclusions. The benefits of doing this are elegantly and succinctly described in a JAMA leader [1]; this is worth reading by anyone who uses reviews for policy or clinical decision making. A detailed paper on meta-analysis of individual patient data [2] is worth reading by those doing or thinking of doing meta-analysis.

2 Regression To the Mean



The efficacy of three interventions to lower plasma lipid levels was explained in a randomised controlled trial. In all three groups cholesterol concentration fell by about 10% after six months and the researchers concluded that the interventions had been the cause of the reduction. The explanation could, however, have been simply that there had been regression to the mean (RTM).

The words regression to the mean can bring an overpowering weariness to those of us unused to statistical terms. The phenomenon occurs with any physiological variable; for cholesterol, for instance, adults have an average daily peak-to-trough variation of 0.7 mmol/L (range 0.4 - 0.9 mmol/L), but the daily range in some diabetic patients can be as high as 1.9 mmol/L [3].

Blood pressure varies from day to day and hour to hour. If a sample of people is drawn from the population because they have "high blood pressure" then some of them may simply be at the upper end of fluctuation and when they are measured again will have a lower blood pressure. If people were selected for low blood pressure ( a common diagnosis in continental Europe) a proportion of them will be found to have a higher pressure at later times as they regress to their mean.

RTM is at the same time both a simple and a subtle phenomenon, but is excellently explained in an article in the Lancet [4]. This paper is a bit of a mindstretcher, and some of the maths will be beyond many of us, but none of the text is too difficult. Clinicians who deal with individual patients and/or with variable like blood pressure or lipids, (and those who do not), should read this paper and add it to their own swot's corner collection.

References:

  1. AD Oxman, MJ Clarke, LA Stewart. Meta-analyses using individual patient data are needed. Journal of the American Medical Association 1995 274: 845-6.
  2. LA Stewart, MJ Clarke. Practical methodology of meta-analyses (overviews) using updated individual patient data. Statistics in Medicine 1995 14:2057-79.
  3. RW Simpson, RD Carter, RA Moore, WAF Penfold. Diurnal changes in plasma lipoproteins in normal subjects and diabetics. Diabetologia 1980 18:35-40.
  4. P Yudkin, IM Stratton. How to deal with regression to the mean in intervention studies. Lancet 1996 347: 241-3.



previous story in this issue