wilk's lamda

[vardhan]

My wilk's lamda score for discriminant analysis is very high i.e. .851. now there are 2 predictors out of six significant. How should i go for analysis.Can i use them in indicating differences in groups. If my null hypothesis that group do no differ would be rejected.What are scores of lamda at which hypothesis is rejected.

vardhan

[Rich Ulrich]

I don't see what you confused by.  Where do you have any actual problem?
The procedure provides you with an F-test on the discrimination.  That is the
only test on lambda. 

When there are two groups, discriminant function is mathematically
the same problem as a regression on a 0/1 criterion -- And there is
just one test.

When there are several groups and several predictors, then it is possible
to consider more than one way to evaluate the several lambda's for the
several roots, but you don't have that problem.

--
Rich Ulrich

> Date: Tue, 23 Aug 2011 23:56:44 -0700

> From: [hidden email]
> Subject: wilk's lamda
> To: [hidden email]
>
> My wilk's lamda score for discriminant analysis is very high i.e. .851. now
> there are 2 predictors out of six significant. How should i go for
> analysis.Can i use them in indicating differences in groups. If my null
> hypothesis that group do no differ would be rejected.What are scores of
> lamda at which hypothesis is rejected.
> [snip]

[SR Millis-3]

Why didn't you use logistic regression instead of DFA?  Logistic regression has fewer restrictive assumptions than DFA---and the results are much easier to interpret.

~~~~~~~~~~~

Scott R Millis, PhD, ABPP, CStat, PStat®
Professor
Wayne State University School of Medicine

Email: [hidden email]
Email: [hidden email]
Tel: [phone]

---

> Date: Tue, 23 Aug 2011 23:56:44 -0700

> From: [hidden email]
> Subject: wilk's lamda
> To: [hidden email]
>
> My wilk's lamda score for discriminant analysis is very high i.e. .851. now
> there are 2 predictors out of six significant. How should i go for
> analysis.Can i use them in indicating differences in groups. If my null
> hypothesis that group do no differ would be rejected.What are scores of
> lamda at which hypothesis is rejected.
> [snip]

[Rich Ulrich]

Scott,

I was never so gung-ho about logistic as some other people

"... [R]esults are easier to interpret"  is not my experience.

I find mean-differences between groups and simple regression-
type coefficients easy to grasp, in comparison to dealing with
odds-ratios and Log-odds.  I like the ancillary statistics that are
typical with DFA (and not with LR).

LR has a theoretical advantage that seldom makes a practical difference.

There is a trade-off for absorbing certain outliers -- logistic regression
is less robust at small Ns.  Also, LR induces users to ignore the fact
that the immediate result is supposed to be a well-behaved
linear equation -- which suggests to me, immediately, that I want
predictor variables that are pretty well-scaled and well-behaved.

Users of LR are not issued a free pass on these issues, despite
any proclamation of "less restrictive."  - I suppose what I *dislike*
about LR is mainly that its users seem to be encouraged to be
careless, whereas users of multiple regression and DFA are
encouraged to worry about assumptions.

--
Rich Ulrich

---

Date: Thu, 25 Aug 2011 12:16:50 -0700
From: [hidden email]
Subject: Re: wilk's lamda
To: [hidden email]

Why didn't you use logistic regression instead of DFA?  Logistic regression has fewer restrictive assumptions than DFA---and the results are much easier to interpret.

~~~~~~~~~~[snip]