Published research contains "high level of statistical errors"

Evidence based practice is currently in vogue, and basing medical practice on published evidence is clearly a good idea, but what if the published findings are inaccurate? An article published this week in BMC Medical Research Methodology shows that a large proportion of articles in top science and medical journals contain statistical errors, 4% of which may have caused non-significant findings to be misrepresented as being significant.

Thirty-eight percent of the Nature papers and a quarter of the BMJ articles studied contained at least one statistical error, according to Emili Garcia-Berthou, a lecturer on biostatistics at University of Girona, Spain, and Carles Alcaraz. In total, more than 11% of the statistical results published in the two journals during 2001 were incongruent.

“Our findings confirm that the quality of research and scientific papers needs improvement and should be more carefully checked and evaluated in these days of high publication pressure,” write the authors.

The errors seen could have been caused by transcription or typesetting errors, for example if a repeated zero was omitted. Alternatively, researchers may have rounded up figures incorrectly.

The researchers showed that some numbers, four and nine, were seen less often than would be expected at the end of a given test statistic or P-value, suggesting that researchers were rounding up numbers incorrectly, possibly so that they looked ’neater’. For example study authors might round up 2.38 to 2.5 rather than 2.4.

“Although these kinds of errors may leave the conclusions of a study unchanged, they are indicative of poor practice,” say the researchers. “Our concern is that these kinds of errors are probably present in all numerical results and all steps of scientific research, with potentially important practical consequences.”

The researchers suggest that one way to minimise the effect of these errors would be for published authors to make their raw data freely available on the Internet. This would allow other researchers to check for themselves whether the results of the study are correct and the conclusions justified. “Also, fraud and sloppiness may be more easily detected,” they say.

This press release is based on the following article:

Incongruence between test statistics and P values in medical papers
Emili Garcia-Berthoug and Carles Alcaraz
BMC Medical Research Methodology 2004, 4:13