Queries regarding ANCOVA and MANCOVA

[Dason]

And with GUIs come limitations. I think you're seeing that for yourself with this whole SPSS thing. With R if you can dream it then you can do it.

[victorxstc]

Yes you are right well said.

[With this level of enthusiasm, I think your code is as well written in R ]

[Dragan]

But the LSD is as well deactivated in SPSS.

Well, no, it's not, victor....What you want to do is go to options and bring the grouping variable over into "Display Means for:" and then click the box for "Compare Main effects" and your options are LSD, Bonferroni, and Sidak.

[victorxstc]

Well, no, it's not, victor....What you want to do is go to options and bring the grouping variable over into "Display Means for:" and then click the box for "Compare Main effects" and your options are LSD, Bonferroni, and Sidak.

That was one good hint Dragan. Thanks

So why did you mention the deactivation of the post hocs in the first place? And why SPSS developers have done this incosnsitent path to first disallow you to do something, and then provide you with the disallowed thing in a hidden place?!

I see its subcommand changes to the below code. meaning that SPSS considers it something different from a simple post hoc test.

/EMMEANS=TABLES(Var1) WITH(Var2=MEAN) COMPARE ADJ(LSD)

[Dragan]

That was one good hint Dragan. Thanks

So why did you mention the deactivation of the post hocs in the first place? And why SPSS developers have done this incosnsitent path to first disallow you to do something, and then provide you with the disallowed thing in a hidden place?!

I see its subcommand changes to the below code. meaning that SPSS considers it something different from a simple post hoc test.

I mentioned it because a lot of people don't understand that the usual post-hoc ANOVA tests e.g. Tukey, etc. are not appropriate for ANCOVA because the standard error is not computed correctly.

Briefly, in the denominator of the t-statistic for comparing adjusted means for ANCOVA, the standard error is;

Sqrt [MSw [ 1/ni + 1/nj + [ (Xbar_i - Xbar_j)^2 / SSres(X) ] ]

where MSw is the adjusted mean squares within from the overall analysis and X is the covariate.

[victorxstc]

I mentioned it because a lot of people don't understand that the usual post-hoc ANOVA tests e.g. Tukey, etc. are not appropriate for ANCOVA because the standard error is not computed correctly.

Briefly, in the denominator of the t-statistic for comparing adjusted means for ANCOVA, the standard error is;

Sqrt [MSw [ 1/ni + 1/nj + [ (Xbar_i - Xbar_j)^2 / SSres(X) ] ]

where MSw is the adjusted mean squares within from the overall analysis and X is the covariate.

Thanks a lot Dragan Then we have two quite different post hocs but with the same name (for example Bonferroni). One is usable for ANOVA, and the other one which can take the covariate into account is useful for ANCOVA. (and I couldn't understand the rest of the technical notes, but it is good to digg it some day when ready)

[statanx]

Thanks to all of you...victorxstc, Dason, Dragan..

I will try what Dragan told.

Dragan, why do you prefer LSD over Bonferroni?

[Dragan]

Thanks to all of you...victorxstc, Dason, Dragan..

I will try what Dragan told.

Dragan, why do you prefer LSD over Bonferroni?

I prefer it (LSD) because the Bonferroni adjustment is too conservative. Note that the omnibus F-test should be significant to use the LSD approach.

[statanx]

I prefer it (LSD) because the Bonferroni adjustment is too conservative. Note that the omnibus F-test should be significant to use the LSD approach.

ok.

I think I need to go through all these post hoc procedures in detail.
Can you please provide me with link (if you have) having the step wise procedure for ANCOVA/MANCOVA and the post hoc procedures?

[victorxstc]

For doing ANCOVA go through "Menu toolbar -> Analyze -> General Linera Model -> Univariate" then hit OK so that the GUI starts. In the appeared dialog box, select your dependent variable and insert it using the arrows into the box labeled "Dependent variable". Then put your independent variables into the one of the two "Factors" boxes. Then put your covariates into the "Covariate" boxes.

Now hit the button labeled "Options". Drag or insert your desired variables into a box labeled "Display means for".

Then check the option "Compare main effects" below that box.

Then from the drop-down menu below that box and that checked option, select one of the options "Bonferroni, LSD, or Sidak".

[victorxstc]

For doing MANCOVA, do the same. But when starting, instead of the path "Menu toolbar -> Analyze -> General Linera Model -> Univariate", go through the path "Menu toolbar -> Analyze -> General Linera Model -> Multivariate". The rest is the same. Note that in this dialog box, you can put more than one variable in the "dependent variable" box.

Here is the step by step instruction to do MANOVA and MANCOVA:
http://www.ucdenver.edu/academics/co...ANOVAHowTo.pdf

Note that ANCOVA is similar. But the dialog box opens from the "Univariate" option (instead of Multivariate).

[victorxstc]

Although as Dragan stated, Bonferroni might be too conservative, it is still correct (the second form that Dragan taught us [the one which can work with covariates]). I suggest you playing with both tests and see if the P values given by the Bonferroni differ considerably from those given by the LSD or not. If I had a similar situation and saw that Bonferroni and LSD are showing similar results, I would publish the Bonferroni one, because I have seen it being used more.

Besides, given the more conservative nature of Bonferroni, its significant P values are more reliable.

[statanx]

Thanks a lot victorxstc for your quick replies, clarifying a number of my doubts and providing the stepwise procedure.