Actually, it is a widespread myth that overlapping 95% CI's of two independent means indicate that the means are not significantly different at alpha=0.05. When the CI's do not overlap, the means are indeed significantly different at alpha=0.05, but when the CI do overlap this does NOT indicate that the means or not significantly different at that alpha level. Therefore you can question the informativeness of this method (it might be more informative to just give the effect size and its confidence interval).
For further reading, see attached pfd.
Best,
Jurriaan Witteman
PhD-student
-----Oorspronkelijk bericht-----
Van: AUDITORY - Research in Auditory Perception namens Rowan D.
Verzonden: ma 12/19/2011 5:59
Aan: AUDITORY@xxxxxxxxxxxxxxx
Onderwerp: Re: Error Bars ( 95 % CI or SE)
A belated addition: it is sometimes the case that neither the SEs nor the CIs of the means of two conditions are particularly important or informative (over and above a description of the distributions) but rather that the SE or CI of the mean *difference* between the two conditions is important and informative thing to present and discuss. Cheers, Daniel
From: AUDITORY - Research in Auditory Perception [mailto:AUDITORY@xxxxxxxxxxxxxxx] On Behalf Of Kim White
Sent: 11 December 2011 09:53
To: AUDITORY@xxxxxxxxxxxxxxx
Subject: Re: Error Bars ( 95 % CI or SE)
Coming back to the original question:
CI's are calculated with a formula with the CE in it... They are different ways of plotting the same results.
Best, Kim White
2011/12/11 Stuart Rosen <s.rosen@xxxxxxxxx<mailto:s.rosen@xxxxxxxxx>>
Let me put forward a dissenting opinion. What you should display in your graph is neither of these but a boxplot which will give a true picture of the distribution of values that you found rather than a statistical inference which depends as much on the number of participants you used as any difference in performance between the two groups. You are going to do a statistical test anyway (and you could quote an effect size) so why waste the opportunity to give more information?
Yours - Stuart Rosen
On 11/12/2011 01:18, Vijay M R Marimuthu wrote:
When we have a d' group result (between 2 experiments),
A. Which ERROR BAR is most appropriate to use (Binomial Distribution) ?
95 % confidence Interval or Standard Errors of the Mean.
Attachment:
Payton_et_al_JIS_3.34_2003.pdf
Description: Payton_et_al_JIS_3.34_2003.pdf