Date: Mon, 14 Jun 2004 03:59:32 -0400 Holger Bargen "SAS(r) Discussion" Holger Bargen Re: one-sided one-sample t-test To: SUBSCRIBE SAS-L statto_22

Dear Steve, first I have to say that I am not a statistican but only a biologist, who knows a little about statistics.

As far as I know I would give you the following answere to your question: If you conduct a t-test (two-sided and you get a p value of lets say p = 0.01 - I did not calculate your data - it is only an example) you would get a value of p = 0.005 for a one sided approach. The p for a one-sided test is always half the p of a two sided test. But now you have to look in which direction there is a deviation from the expected mean! In your case the p value, which is half the p-value you calculated for two-sided gives the probability that the blood preasure is smaller than the baseline. This is the other direction of that one you wanted to test! So you'll need to calculate the p-value for the other side. In theorie the p value right sided (lets call it pr) is always 1 - the p value left sided (lets call it pl) for a given hypothesis: pr = 1 - pr Or in other words: The total probabilities under a normal distribution is always 1. So, you only have to do this calculation. For the data I gave you in my example above you would get p(other side) = 1 - 0.005 = 0.995 This is clearly not significant. You can do the same with your data.

But first of all I would like to ask you the following question: You have got 12 data-pairs (if I counted it right) and in none of them you got a change in the direction you expected but to the other side. Do you really think that any statistics is necessary to verify your findings? Statistical calculations are senseful - but only if it is not clear by simply looking at the data if there is an effect or not.

Hope I could help you a little and keep in mind: I am not a statistican. Maybe a statistican at this forum could verify my words or say that I am wrong.

Best regards Holger

On Tue, 8 Jun 2004 15:28:00 +0000, statto stats <statto_22@HOTMAIL.COM> wrote:

>I'm trying to anayse %change from baseline data. > >data pressure; >input SBPbefore SBPafter @@; >per_change=SBPafter/SBPbefore - SBPbefore/SBPbefore; >datalines; >130 100 134 100 130 100 138 100 140 100 138 100 140 100 135 100 136 100 130 >100 136 100 137 100 ; run; > >As you can see most people have a considerable reduction in their SBP >(Systolic Blood Pressure). > >However, I suspected there to be an increase in SBP and so set about >intending to analyse the data using a one-sided t-test with the following >Hypotheses. > >H0: mean %change < 0.25 >H1: mean %change >=0.25 > >I still hop to use these hypothesis. Using proc ttest > >proc ttest data=pressure alpha=0.05 h0=0.25 ; >var per_change; >run; > >or proc univariate > >proc univariate data=pressure alpha=0.025 mu0=0.25 cibasic cipctlnormal >normal plots plotsize=40; >var per_change; >probplot per_change /normal(mu=est sigma=est) pctlminor; >inset mean std / format=3.0 header='Normal Parameters' position=(95,5) >refpoint=br; >run; > >However this actually test the hypotheses > >H0: mean %change = 0.25 >H1: mean %change not equal to 0.25 > >And gives a significant p-value? >Does anybody know how I can test my original hypothesis? And get a >non-significant p-value? > >Thanks >Steve > >_________________________________________________________________ >It's fast, it's easy and it's free. Get MSN Messenger today! >http://www.msn.co.uk/messenger

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