<- file stat 97power.html -> Subjects included below are: Tolerance intervals REF DeCicco Complex samples (software) Shah N for equivalency Ulrich 3x2x2 X^2 Ulrich Repeated measures Ulrich

Tolerance intervals

=======================Barry DeCicco, 14 Mar 1997==========ssc From: bdecicco@sunm4048az.sph.umich.edu (Barry DeCicco) Subject: Re: Sample size for tolerance intervals Message-ID: <5gc8ur$7s1@lastactionhero.rs.itd.umich.edu> In article <19970311165901.LAA01899@ladder01.news.aol.com>, cwbern@aol.com (CWBern) writes: |> |> I'm looking for a reference (or good explanation) of finding the |> appropriate sample size required for tolerance intervals. |> Im refering to the 95%, 95% criteria that seems to be popular these days. |> ie. 95% confidence that the interval contains 95% of the population. |> |> Thanks Try 'Statistics for Engineering Problem Solving', by Stephen B. Vardeman. It covers tolerance intervals for normal data, nonparametric tolerance intervals, and tolerances intervals for linear regression data. *--------

Software for complex sample estimation

=======================Babu Shah, 07 May 1997==========ssc From: shah@rti.org (Shah, Babu) Subject: Re: Software for Sample Estimation Message-ID: <113@usenet.rti.org> The possible software packages are: Of which only WESVAR is free. www.westat.com PCCARP is from Iowa State University. STATA is by STATA corporation. SUDAAN is developed by Research Triangle Institute. www.rti.org In article <336939df.228202954@news.otago.ac.nz>, agray@commerce.otago.ac.nz says... > > Does anyone know of any good cheap (or free) software for sample >estimation? I'd like something to calculate estimates for various >designs (stratified, network, etc.) and with ratio and regression >estimation. Sample size calculations would be nice. Any pointers >would be much appreciated! > *--------

N for Equivalence (survivorship)

=======================Rich Ulrich, 05 Mar 1997==========ssc Subject: Re: Survival data: sample sizing for an equivalence hypothesis? Message-ID: <5fk72l$r2r@usenet.srv.cis.pitt.edu> Nelson Kinnersley (nelson@dircon.co.uk) wrote: : My sample sizing software tells me there is no theory to help determine : sample sizes with survival data when the hypothesis is one of equivalence. : Anybody know any different? -- I think you forgetting an essential parameter in any "power analysis", which is, "the effect-size [of interest]". That is not just survival data. What sample size do you need to detect a difference of means that is zero? For a good test, the power is only equal to the alpha, when you assume no difference. The power TO DETECT a greater difference is going to increase with the difference, which is where you get a "power curve." ---------Mar 8, FAQ addition If you want to show Equivalence, the best a single study can show is that the Confidence Interval for the difference is small -- In other words, the power to detect a *small* difference must be large, if you want narrow limits. There have been a couple of references posted in the past; I found the articles unhelpful, and I don't have the references here, right now. *--------

Sample size for 3x2x2 contingency

=======================Rich Ulrich, 13 May 1997==========ssc Subject: Re: Sample Size Calculation for Chi-square Message-ID: <5l9u9p$ctm@usenet.srv.cis.pitt.edu> Sally Vegso (sally.vegso@yale.edu) wrote: : I need to do a sample size calculation for a chi-square. The outcome : variable is dichotomous, and we will be using three equal sized sample : groups. The groups will be matched on age and sex. Are there any special : considerations I have to make because the groups are age and sex matched, : and not just completely random samples? Or can I just do the sample size : calculation in the usual way for independent samples? : Thanks for any advice -- You can just do the sample size calculation in the usual way. -- If age and/or sex do make a huge difference, then there would be a gain in power if you could take them into account. -- Three groups is a lot tougher to model than two groups. *--------

Power for repeated measures design.

=======================Rich Ulrich, 01 Aug 1997==========ssc Subject: Power for Repeated Measures ANOVA Message-ID: <5rsvrr$sal@usenet.srv.cis.pitt.edu> Mark Sherfy (you@somehost.somedomain) wrote: : Does anyone know how to compute power for the main effects in a : repeated measures ANOVA? All I've ever done power analysis for is : one-way layouts; if I plug my sample size, variance etc from a repeated : measures design into power equations for a one-way layout, will I get : the right answer? If I need to compute it differently, does anyone : have SAS code that will do it? Respond to newsgroup or msherfy@vt.edu. : Thanks! You will NOT get the right answer by plugging stuff from your repeated measures design into equations for a one-way layout. In fact, it may be doubtful that you can plug your numbers into any package and get the right numbers - I would like to hear it if there has been an improvement, but someone reported a couple of years ago about getting three different answers from the three packages he had on hand, for his simple test-problem. And then: How simple is YOUR problem? For any power analysis, I would want a package with plenty of documentation, so I know that they THINK they are giving me. And how accurate I might expect it to be. And most important, what quantities it is that they DO require for input, and what they give for output. A big thing about repeated measures is the correlation between periods, which you may not have well-quantified, especially if it is not quite homogeneous. I advise: construct something simpler. Look at a change score, if you can, for First vs Last, or First-half vs Second-half.... The powerful test for a gradual change across time is the one-DF test on the slope, and certainly NOT the multiple-DF test across (unordered) repetitions. That should be approximated, conservatively, by the First vs Last. * * * * * * * * * * * * * * * * * * * * * * * * * * * * *

Document by Rich Ulrich. E-mail to wpilib+@pitt.edu

FAQ top.

Ulrich home page.

Ulrich FAQ. http://www.pitt.edu/~wpilib/stats99.html