Confidence intervals permit, but don’t guarantee, better inference than statistical significance testing

Abstract

A statistically significant result, and a non-significant result may differ little, although significance status may tempt an interpretation of difference. Two studies are reported that compared interpretation of such results presented using null hypothesis significance testing (NHST), or confidence intervals (CIs). Authors of articles published in psychology, behavioral neuroscience, and medical journals were asked, via email, to interpret two fictitious studies that found similar results, one statistically significant, and the other non-significant. Responses from 330 authors varied greatly, but interpretation was generally poor, whether results were presented as CIs or using NHST. However, when interpreting CIs respondents who mentioned NHST were 60% likely to conclude, unjustifiably, the two results conflicted, whereas those who interpreted CIs without reference to NHST were 95% likely to conclude, justifiably, the two results were consistent. Findings were generally similar for all three disciplines. An email survey of academic psychologists confirmed that CIs elicit better interpretations if NHST is not invoked. Improved statistical inference can result from encouragement of meta-analytic thinking and use of CIs but, for full benefit, such highly desirable statistical reform requires also that researchers interpret CIs without recourse to NHST.