Date: Thu, 27 Dec 2001 13:51:02 -0500
Reply-To: pjw@ACSU.BUFFALO.EDU
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: pjw@ACSU.BUFFALO.EDU
Subject: Re: How to calculate the sample size?
In-Reply-To: <670440D45C5ED31180A6005004116476816201@namrid1.namrid.sld.pe>
Content-Type: text/plain
Quoting Christian Bautista <cbautista@namrid.sld.pe>:
> I want to evaluate two treatments and I going to do a
> pharmaconekinetic analysis. So, I would like to know how to calculate
> the sample size?. Is there any software?
>
> What kind of method I have to use my sample size? group sequencial
> (proportions) or equivalence (proportions) or maybe other kind of
> procedures.
If a "pharmaconekinetic" analysis involves repeated measures of
variables potentially affected by the pharmacologic agent(s) under
evaluation, then I would suggest
1) using some sort of hierarchical linear analysis software, so that
you can treat time as a continuous variable, and avoid "missing data"
problems;
2) designing the study to block on important nominal level antecedents
(such as diagnosis, if that varies) and to match and/or counterbalance
as much as possible on important continuous variables (such as baseline
scores on the response variables that you are tracking, if applicable);
3) controlling in the data analysis for blocked and/or matched
variables, plus any other antecedents that are known to have important
effects on the response and/or outcome variables.
Controlling for known sources of error variation in the data analysis
(especially in combination with conterbalancing in the design) can
greatly decrease the sample size necessary to demonstrate that
treatment effects exist, and it usually results in much more accurate
estimates of their magnitudes. In order to take the implications of
such control procedures appropriately into account in power estimates,
think of the magnitude of treatment effects in terms of partial
correlations -- the so-called (Sir Ronald)Fischer approach. You might
also wish to think more in terms of putting confidence intervals around
partial regression coefficients, and less in terms of statistical
significance per se. (See Darlington's recent text for an excellent
discussion of how to use this approach to kake decisions concerning
sample size.)
Good luck with your study,
Pow
