Date: Wed, 2 Nov 2011 06:14:39 -0400 Peter Flom "SAS(r) Discussion" Peter Flom Re: Wide CI but good power? To: Jan McClure <201111020216.pA1LWwcO000726@waikiki.cc.uga.edu> text/plain; charset="us-ascii"

Hi Jan

Power analysis is about your ability to find statistical significance given a particular sample size, effect size and alpha. Here you say you did power analysis with an OR of 3 (which is a pretty big OR in the fields I work in, but might be normal in your field). But your actual ORs in these two examples were much higher. So, you found significance. But you get wide CIs because of small cell sizes: This makes sense. You don't want to make a precise statement based on 1 person (or even 3). The second CI is narrower than the first because you have a smallest cell of 3 instead of 1.

Peter

Peter Flom Peter Flom Consulting http://www.statisticalanalysisconsulting.com/ http://www.IAmLearningDisabled.com

-----Original Message----- From: SAS(r) Discussion [mailto:SAS-L@LISTSERV.UGA.EDU] On Behalf Of Jan McClure Sent: Tuesday, November 01, 2011 10:16 PM To: SAS-L@LISTSERV.UGA.EDU Subject: Wide CI but good power?

Hi,

I am a doctoral student in epidemiology (SAS 9.1.3 on a PC)

I ran a logistic regression analysis with almost all categorical independent variables. The 95% are very wide for many of the variables. For example desperation OR 12.5 (CI 1.68, 92.9). I looked at a cross tabulation of my outcome variable and desperation (dichotomous) and found cell sizes of 142,1,307, and 30. Small cell sizes I think could be causing my wide CI. However I have similar cell sizes for the variable withdrawal (205,3,240,27) and even though its OR is 11.3 (close to desperation) the 95%CI is 2.66, 48.4) Not narrow but not as wide as desperation.

I re-did my a priori power calculation using the actual prevalence in the sample (70%), alpha 0.05, effect size 3 (OR), two sided and got 81% power for my sample size of 460.

My question is can one have wide CI and good power at the same time or am I confused? If this is possible could I get a narrower CI somehow given the same prevalence in this population?

Thanks for any feedback! Jan

Back to: Top of message | Previous page | Main SAS-L page