Thanks to everyone who responded to my query. The overall opinion is
that the exact method should definitely be used in my case, and due to
the ease of calculation, should probably be used all of the time - the
method I had been using is an approximation which holds best when both
proportions (agree, non-agree) are large. Thanks also for the
spreadsheets that a few of you sent for performing these calculations -
I will incorporate the formulae into SAS.
-----Original Message-----
From: Inglis, Susan
Sent: 27 August 2004 11:34
To: [log in to unmask]
Subject: QUERY: Exact Confidence Intervals
Dear All,
I have a query which I hope someone can shed some light on:
When comparing 2 groups (paired) with a binary outcome of 0=positive
1=negative, I have calculated the percentage agreement between the 2
groups (i.e. if group1=0 and group2=0 then agree=1, if group1=1 and
group2=1 then agree=1, all other cases agree=0), and calculated the 95%
confidence intervals for this percentage using the basic formula for 95%
confidence intervals for proportions, amended for percentages
(pctagree+/- 1.96*sqrt(pctagree*(100-percent)/N)). However, I have also
seen cases where the exact binomial confidence interval is calculated in
this instance - thus if I have 95 observations that agree out of 95 in
total, I have 100% agreement, and using my formula the CIs are 100,100%.
But the exact binomial CIs (calculated using a software package) are
96.2,100%. Are both approaches correct? I thought exact tests were
mostly used when the sample size was very small?
Thanks,
Susan
|