Date: Fri, 3 Feb 2012 19:54:50 -0500
Reply-To: R B <ryan.andrew.black@gmail.com>
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: R B <ryan.andrew.black@gmail.com>
Subject: Re: 2x2 Latin square design analysis help
In-Reply-To: <1328242962015-5452678.post@n5.nabble.com>
Content-Type: multipart/alternative;
DAVID: A multivariate response along with repeated measures does make for
an intriguing model. The good news is that such a model can be employed via
the MIXED procedure. I have provided code previously on how to fit such a
model here:
http://listserv.uga.edu/cgi-bin/wa?A2=ind1104&L=spssx-l&P=R23020
One could then use the TEST subcommand to answer very interesting research
questions!
BRUCE: I see the rationale for employing an ANCOVA model as you have
parameterized it. Admittedly, because I'm often thinking about the
optimizing residual and random effects covariance matrices, I tend to
gravitate towards certain types of designs. For this particular situation,
I thought (although I never stated it) that the residual variance of
post-intervention observations could be considerably smaller than baseline,
especially for some of the conditions. The MIXED procedure would allow for
heterogeneous group*time residual variances if one were to parameterize the
model in a particular way.
Ryan
On Thu, Feb 2, 2012 at 11:22 PM, David Marso <david.marso@gmail.com> wrote:
> *AND* there are multiple outcome variables measured both pre/post and it
> seems that there might be interesting functions which could be explored.
> i.e
> ANCOVA using simple Pre/Post weight might be overly simplistic.
> One might be inclined to look at (Post-Pre)/Pre
> Proportionate weight loss. Just my intuition.
> It is easier for very heavy person to lose 15 pounds than a somewhat heavy
> person?
> OTOH, I don't know diddly squat about the physiological dynamics but it
> just
> seems that theory might toss some interesting dice into the game.
>
> Also : I would postulate directional hypotheses a-priori .
>
> Weight loss (Weight_Pre -Weight_Post)
> Control < Diet ?=? Exercise < Diet+Exercise
>
> Finally, We can consider the following models:
>
> PostWeight = B0 + B1*Diet + B2*Exercise + B3*Diet*Exercise + Residual
> vs
> PostWeight = B0 + B1*Diet + B2*Exercise + B3*Diet*Exercise +
> B4*PreWeight+Residual2
>
> If Random Assignment then one would expect that Diet,Exercise,Preweight
> would be uncorrelated
> hence Residual = B4*PreWeight+Residual2
>
> *SO* ANCOVA reduces the Error SS with a sacrifice of 1 df (assuming Pre
> and
> post are correlated).
> ----
> Enough!?
> OTOH: It becomes more interesting when we have a multivariate context.
> SEM?
> ---------------------------------------------------------------------------
>
>
> Bruce Weaver wrote
> >
> > But as David noted, the 4 conditions are formed by the 2x2 combination of
> > two binary variables. With the usual coding (i.e., 1 = Yes, 0 = No):
> >
> > D E Description
> > ---------------------
> > 0 0 Control
> > 1 0 Diet, No Exercise
> > 0 1 Exercise, No Diet
> > 1 1 Diet and Exercise
> >
> > If the outcome variable was measured at both baseline and followup, then
> > the preferred approach, arguably, would be a 2x2 ANCOVA, which is
> > equivalent to an ordinary least squares regression model as follows:
> >
> > Y1 = b0 + b1*Y0 + b2*D + b3*E + b4*D*E + error
> >
> > where
> > Y1 = outcome variable at followup
> > Y0 = outcome variable at baseline
> > D and E are as shown above, and D*E = their product.
> >
> > Of course one /could/ run this model using MIXED, but for users not
> > familiar with MIXED (and I get the impression the OP is not), it will be
> > much easier to run it via UNIANOVA (Analyze - GLM - Univariate, with D
> and
> > E entered as "fixed factors", and Y0 as a "covariate", and the D*E
> > interaction included).
> >
> > Another important question David raised (and I don't recall seeing an
> > answer) is whether there was random assignment to the 4 cells.
> >
> > There's my two cents!
> >
> >
> >
> > R B wrote
> >>
> >> From what I can tell, you have one between-subjects variable (Condition)
> >> which has four levels (Diet, Exercise, Diet+Exercise, Control) and one
> >> within-subjects variable (Time) which has two levels (baseline,
> >> post-intervention). This is commonly referred to as a mixed ANOVA (not
> to
> >> be confused with the MIXED procedure). You could analyze your data by
> >> fitting a general linear model or linear mixed model in SPSS. I don't
> >> have
> >> time to write code for you at the moment, but quite frankly, there must
> >> be
> >> numerous SPSS examples online. Bottom line is that you need to take into
> >> account correlation among residuals obtained by repeated observations of
> >> each subject. You can write specific contrasts of interest using the
> >> appropriate subcommand. I am always in favor of using the subcommand
> >> which
> >> requires that you understand the coefficient (L) matrix.
> >>
> >> Ryan
> >>
> >> On Thu, Feb 2, 2012 at 12:08 PM, Sur1605 <surabhi1605@> wrote:
> >>
> >>> Thanks David for such an amazing description. This is exactly how i
> >>> tried
> >>> doing it before i posted my question on this forum. But everytime i
> >>> analysed
> >>> my data somehow it gave me this message:
> >>> "Post hoc tests are not performed for Diet because there are fewer than
> >>> three groups.
> >>> Post hoc tests are not performed for Ex because there are fewer than
> >>> three
> >>> groups."
> >>>
> >>> Basically its not able to perform the posthoc analysis. I thought i was
> >>> doing something wrong in feeding in the data in the spreadsheet.
> >>>
> >>> And yes, i measured the same variables before and after the trial.
> Looks
> >>> like i will have to reconsider the choice of test being used here.
> >>>
> >>> Thanks a lot for your help.
> >>>
> >>> Sb
> >>>
> >>>
> >>>
> >>> --
> >>> View this message in context:
> >>>
> http://spssx-discussion.1045642.n5.nabble.com/2x2-Latin-square-design-analysis-help-tp5446710p5451327.html
> >>> Sent from the SPSSX Discussion mailing list archive at Nabble.com.
> >>>
> >>> =====================
> >>> To manage your subscription to SPSSX-L, send a message to
> >>> LISTSERV@.UGA (not to SPSSX-L), with no body text except the
> >>> command. To leave the list, send the command
> >>> SIGNOFF SPSSX-L
> >>> For a list of commands to manage subscriptions, send the command
> >>> INFO REFCARD
> >>>
> >>
> >
>
>
> --
> View this message in context:
> http://spssx-discussion.1045642.n5.nabble.com/2x2-Latin-square-design-analysis-help-tp5446710p5452678.html
> Sent from the SPSSX Discussion mailing list archive at Nabble.com.
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> LISTSERV@LISTSERV.UGA.EDU (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD
>
[text/html]
| 
