home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!paladin.american.edu!auvm!UNC.BITNET!UPHILG
- Message-ID: <STAT-L%92072109291197@VM1.MCGILL.CA>
- Newsgroups: bit.listserv.stat-l
- Date: Tue, 21 Jul 1992 09:23:00 EST
- Sender: "STATISTICAL CONSULTING" <STAT-L@MCGILL1.BITNET>
- From: "Philip Gallagher,(919)966-7275" <UPHILG@UNC.BITNET>
- Subject: Power Analysis texts compendium
- X-To: STAT-L@mcgill1.BITNET, st8330@siucvmb.BITNET
- Lines: 328
-
- Last week I requested advice on texts to study in order to
- become a Power Guru. I have had several references to
- established texts and a set of updates on what I suppose to
- be the cutting edge of power research and theory, mostly
- from the shops of Keith Muller at UNC and Ralph O'Brien at
- Florida State. (Apologies to Ralph O'Brien for having misnamed
- him in my first note.)
-
- Please note how much brand new power software is being offered
- for distribution in one form or another. If the major stat
- packages can bring themselves to ignore this, they deserve
- to go down the drain.
-
- I note that I have heard from no one from the school of thought
- that says that power analyses are totally without value because
- the distributional requirements so vital to power calculations
- are never (well, almost never) satisfied except in stat textbook
- classroom by-hand problems. (I have certainly not done justice
- to this argument, and I beg anyone who can make it better to please
- do so.) I know that persons who feel this way exist (there is
- one in our department), and, if you have never run into these
- ideas, it is only fair of me to tell you that they make a strong
- and impassioned case.
-
- I have had occasion to go back to Cohen's work since I wrote
- first, and I find it to be more useful than my first message
- conveyed. It is of only limited help with the Muller/O'Brien
- field, but much of their work has reached the literature since
- his 2nd edition was published, so it's only fair to give him
- a break.
-
- Now I will cut and paste - if I do anyone an injustice, please
- write an explanatory rejoinder directly to STAT-L.
-
- Thanks to everyone - I now have lots of EXCITING reading material
- for my vacation!
-
-
- ========================================================================
- From: phil@wubios.wustl.edu(J. Philip Miller)
- I started with Cohn's book but it is too simplified for me as well - he really
- does not ever mention the noncentral stat distributions, and that is all you
- really have to know if you are prepared to compute things themselves. I have
- worried a lot about this and and one time really wrote some good code to "do
- it right", but have become less anamored with it since most of the use of
- power that I have on a day to day basis starts with some "off the top of my
- head" numbers and to then worry about whether the math is being done "right"
- or not is not placing the emphasis where it should.
-
- ========================================================================
- From: Ken Hardy <KENHARDY@UNCVM1>
-
- A piece of software that performs many power calculations and has a nice
- bibliography on power analyses for various statistics is PASS (Power
- Analysis and Sample Size). I've got the software and will campus mail
- you the references section.
- ========================================================================
- From: "dick darlington" <dick_darlington@qmrelay.mail.cornell.edu>
-
- Power analysis is of course highly dependent on the test you're using, so I
- will answer just for my specialty: regression and linear models. Two books
- with fairly extensive discussions of power are:
- Rao, C.R. Linear statistical inference and its applications
- Graybill, F.A. Intro to linear statistical models
- I should mention that the power tables for linear models in Cohen are not
- exact, and do not distinguish between fixed and random scores, which is
- necessary for exact power analysis. My own book "Regression and Linear Models"
- gives short BASIC programs that yield exact power values for linear-model tests
- under the standard assumptions. There was a good article on linear-model power
- analysis in the Psych Bulletin a year or two ago, but I don't have the
- reference handy.
- ========================================================================
- From: WKJ <JAMES@XAVIER>
-
- Michael Korda's Book " Power " is a lot of fun.
-
- ========================================================================
- From: jones@reed.edu(Albyn Jones)
-
- well, you could always start with Johnson & Kotz, the set of books
- called "Continuous univariate distributions" and look at the non-central
- t, non-central chi-squared, non-central F distributions...
-
- ========================================================================
- From: "Dennis Roberts" <DMR@PSUVM>
-
- If you want, I will be glad to send you a section from a book of mine re:
- power, pretty elementary but may help. In addition, I would suggest you
- contant Jack Barnette at the Univ of Alabama who has an excellent demo
- software program on power, that is very useful for "seeing" what power is.
- His address is Behavioral Studies, Univ of Alabama, Tuscaloosa Alabama and I
- am blocking on the zip. His bitnet is JBARNETT at UAVM1 .... but he may
- not respond. If you want my material, send me your surface address and I
- will ship immediately.
- ========================================================================
- From: tgee@ccs.carleton.ca(Travis Gee)
-
- You might try Wm. Hays and Robert Winkler's _Statistics: Probability,
- Inference and Decision._ My edition's well out-of-date, but there was
- a healthy discussion of it in 1971, at least. More recently, an
- author to search in the PsychLit database is Bruno Zumbo (alas, no
- references available in my filing cabinet at the moment). He's a
- specialist on relative efficiency of nonparametric stats. His PhD
- comprised a Monte Carlo simulation which showed that under violation
- of assumptions, nonparametric stats are often _more_ powerful than
- their parametric analogues. If that fascinates you, and you can't
- access his papers, you could e-mail him at zumbo@acadvm1.uottawa.ca
- He _is_ a power guru.
- Also along those lines, Seigel and Castellan's _Nonparametric
- Statistics for the Behavioral Sciences_ discusses the power of various
- and sundry methods. If you're looking at multivariate models, I seem
- to recall that the SPSSx manuals and the SAS manuals have references
- in many sections which refer to the power of assorted statistics. But
- don't try those papers until the groundwork is done. Good luck!!
-
- ========================================================================
- From: phil@wubios.wustl.edu(J. Philip Miller)
-
- A recent reference for the simple stuff:
-
- Jurgen Bock & H Toutenburg, Sample Size Determination in Clinical Research,
- Chapter 16 in Rao & Chakraborty, eds, Handbook of statistics, vol 8, 515-38,
- 1991.
-
- ========================================================================
- From: dixon@srel.edu
-
- Phil:
- Scheffe, The Analysis of Variance, is still a good place to go
- for the theory of power for ANOVA.
- Otherwise, the literature on power is really scattered. There are
- lots of 'warm fuzzy' treatments for nonstatisticians, e.g. Cohen, as
- you have discovered. One I give to my consultees is Lipsey, Design
- Sensitivity: Statistical power for experimental research, Sage. It's
- a bit more concise than Cohen, but it still shoehorns everything into
- an ANOVA framework.
- My favorite source is the Russ Lenth's manual for his PowerPack
- program. It includes a practical guide to power calculations that
- lays everything out in a very understandable way, if you know about
- non-centrality parameters. I otherwords, I use it, but I don't expect
- my consultees to read it. PowerPack is also a very easy to use program
- that got high marks in Goldstein's review of packages to compute power
- (Am. Stat. 43:253).
- ========================================================================
- From: Jerry Dallal <jerry@NUTMEG.HNRC.TUFTS.EDU>
-
- Have a look at Odeh and Fox, Sample Size Choice, for starters. Don't buy
- it without having a look at it first. It's very expensive and mostly charts,
- but the discussion in the front of the book is reasonable.
-
- I've found Cohen's Statistical Power Analysis for the Behavioral Sciences hard
- to read. It may just be me. But if you find it hard going, don't think all
- texts are like it. (Who knows? You may like Cohen and dislike O&F!)
- ========================================================================
- From: Fred Detwiler ST8330 at SIUCVMB
-
- Lipsey, M. W. (1990). Design Sensitivity. Sage.
-
-
- ========================================================================
- From: "GEOFF SELIG - CONCORDIA U. COMPUTING SERVICES"
- <ILFC594@Vax2.Concordia.CA>
-
- You might want to have a look at a text by Neter, Wasserman and Kutner
- (_Applied Linear Statistical Models, Third Edition_ [Irwin]) for some
- fairly good discussions on power (and linear models in general). As well,
- David Howell's _Statistical Methods for Psychology, Third Edition_
- [PWS-Kent] has excellent discusions of power in a variety of t-test and
- ANOVA situations (and just about everything else).
-
- ========================================================================
- From: Ralph G. O'Brien <robrien@stat.ufl.edu>
-
- Phil: Bob Parker (Vandy) relayed a memo that you are interested in
- becoming more "power"ful. I have just finished the final, final draft of
- a long chapter on power analysis (O'Brien, R. G. and Muller, K. E.
- (early 1993), Unified Power Analysis for t Tests Through Multivariate
- Hypotheses, (to appear) in Applied Analysis of Variance in Behavioral
- Sciences, eds. L. K. Edwards, New York: Marcel Dekker.). We think our
- approach is more general and unifying, yet no harder to understand and
- apply than Cohen's. We cover some things he doesn't, and vice-versa.
- We don't use constructs such as "small," medium," and "large" effect
- sizes as this is an unjustified, oversimplified notion: Most content
- researchers will quickly opt for "medium," even though they have no idea
- what "medium" implies with respect to their proposed study. What is a
- "small" effect size in one field of study may be a "huge" one in
- another.
-
- We have SAS %INCLUDE modules to get things done, available for the
- taking via anonymous ftp from my workstation, as outlined below. Also
- included below is an example that should be almost self-explanatory if
- you like the cell-means model for ANOVA. The chapter and software
- start at a more basic level and go through the general univariate and
- multivariate linear models. We use the modules all the time here in
- developing protocols within the Clinical Research Center, and I keep
- expanding their capabilities as needs arise. Our beta version has
- options covering tests of correlations (via Fisher's r-to-z), and tests
- comparing independent logits from 2 x NumGroups contingency tables.
-
- I am doing a poster paper on power for the multivariate GLM in August at the
- Boston Joint Stat Meetings (O'Brien, R. G. and Shieh, G., "Pragmatic,
- Unifying Algorithm Gives Power Probabilities for Common F Tests of the
- Multivariate General Linear Hypothesis"). We think our algorithm offers
- an "upgrade" on the Muller-Peterson (1984) method, which will be
- discussed again in a forthcoming (Dec. 1992?) JASA paper by Muller and
- others. "O'Brien and Shieh" is now being reviewed informally by Jon
- Shuster of our group, and it should be submitted soon for journal review.
- It will NOT appear in any of the ASA Proceedings. No public software is
- being planned/developed yet for the our proposed algorithm. Actually,
- Keith Muller's SAS IML module can be altered in minor ways to do it, but
- such a change would be up to Keith, of course.
-
-
-
-
- -----------------------------------------------
- INPUT FOR 2 x 3 UNBALANCED ANOVA
- (runs in "base" SAS environment; no IML needed)
- -----------------------------------------------
- options ls=72 nosource2;
- %include "~robrien/POWER.public/OneWyPow.beta.sas"; *(UNIX format);
- * %include OneWyPow; *(CMS format);
- title1 "Example: Unbalanced 2 x 3 ANOVA";
- title2 "Order: A1B1 A1B2 A1B3 A2B1 A2B2 A2B3";
- cards;
- mu 100 110 90 105 130 80 .
- weight .2 .2 .1 .2 .1 .2 .
- sigma 25 40 .
- alpha .05 .01 .
- Ntotal 100 120 160 .
- nooverall
- contrasts
- "A Main Effect (Type III)" 1 1 1 -1 -1 -1 .
- "B Main Effect (Type III)" 1 -1 0 1 -1 0 .
- > 0 1 -1 0 1 -1 .
- "AB Interaction" 1 -1 0 -1 1 0 .
- > 0 1 -1 0 -1 1 .
- "A2B1&2 vs. A2B3" 0 0 0 .5 .5 -1 .
- end
- %include "~robrien/POWER.public/FPowTab2.sas"; *(UNIX format);
- * %include FPowTab2; * (CMS format);
-
- --------------
- PART OF OUTPUT
- --------------
- ALPHA 0.05
- ----------------------------------------------------------
- | | Std Dev |
- | |-----------------------------|
- | | 25 | 40 |
- | |--------------+--------------|
- | | Total N | Total N |
- | |--------------+--------------|
- | |100 |120 |160 |100 |120 |160 |
- | |----+----+----+----+----+----|
- | |Pow-|Pow-|Pow-|Pow-|Pow-|Pow-|
- | | er | er | er | er | er | er |
- |--------------------------+----+----+----+----+----+----|
- |EFFCTITL |TESTTYPE | | | | | | |
- |------------+-------------| | | | | | |
- |A Main |2-tailed t |.156|.178|.222|.090|.099|.116|
- |Effect (Type|-------------+----+----+----+----+----+----|
- |III) |1-tailed t |.241|.270|.326|.145|.159|.185|
- |------------+-------------+----+----+----+----+----+----|
- |B Main |Regular F | | | | | | |
- |Effect (Type| | | | | | | |
- |III) | |.997|.999|.999|.811|.881|.957|
- |------------+-------------+----+----+----+----+----+----|
- |AB |Regular F | | | | | | |
- |Interaction | |.474|.554|.690|.208|.244|.316|
- |------------+-------------+----+----+----+----+----+----|
- |A2B1&2 vs. |2-tailed t |.999|.999|.999|.880|.931|.979|
- |A2B3 |-------------+----+----+----+----+----+----|
- | |1-tailed t |.999|.999|.999|.933|.965|.991|
- ----------------------------------------------------------
-
-
- -------------------------------------------------------------
- Getting SAS-based Power Shareware Modules Using anonymous ftp
- -------------------------------------------------------------
-
- The best way to obtain the current versions of these modules and
- related files is to use the Internet network to ftp-get the files.
- ftp (lower case letters may be required) stands for File Transfer
- Protocol and is now almost universally available. The details
- regarding where and how these files are stored will change over time.
- Accordingly, the instructions given here only tell how to get a short
- file that gives complete up-to-date information on this.
-
- (1) To get to the "front door" of the correct workstation in the
- University of Florida's Department of Statistics enter
- ftp banana.stat.ufl.edu
-
- (2) To login, use the name
- anonymous
-
- For the password, enter your full system name, something like
- mmouse@wdw.orlando.fl
-
- (3) To get a copy of the short information file, enter
- cd pub
- get ReadMe.power myfile.asc
- This will put the contents of a file called ReadMe.power on your
- computer using the file name of your choice ('myfile.asc').
-
- (4) Enter
- quit
- to return to your local computer, where you can study
- the instructions in ReadMe.power and proceed.
-
- (5) This method should work for a number of years, but computing
- systems have unpredictable half-lives. If you encounter problems,
- contact me.
-
- Ralph G. O'Brien, PhD
- Division of Biostatistics UF Office: 904-392-8446
- Box 100212 Home office: 904-378-7381
- Univ. of Florida Health Sci. Center UF fax: 904-392-4168
- Gainesville, FL 32610-0212 email: robrien@stat.ufl.edu
-
-
-
- =======================================================================
-
- Note from Phil Gallagher: Keith Muller gave me a photocopy of his
- paper that Ralph O'Brien mentioned, and it does indeed have "JASA,
- December 1992" pencilled on the top. What a great Christmas present
- to look forward to! (Yes, some of us really that nerdy, Virginia.)
-