Variable selection t multivariate analysis

Before  recruiting variables to multivariate analysis,some says do first chi-square then select those are significant (p<0.05) here i.e.  in chi-square.....then do multivariate...
others say do bi-variate analysis for each variable then select to multivariate if they are significant at p-value of 0.2.....then analyse them to find independtly associated variables.[let remind u that when i see p-value of chi-square and bi-variate,is not same....why comes?not clear]

need your idea

[Art Kendall]

Please supply more detail.  There are many kind of multivariate analysis.
What is you research question?
What data do you have?
What constitutes a cases?
How did you select the cases?
How did you gather the data?
What are your variables?  What are their roles? Manipulated random assignment to treatment?  Measured independent? Possible sources of variation?
Dependent? Repeated?
What kind of analysis are you considering?

Is this a class assignment? Self study?  Paid work?

Art Kendall
Social Research Consultants

On 12/4/2011 5:43 AM, [hidden email] wrote:

Before  recruiting variables to multivariate analysis,some says do first
chi-square then select those are significant (p<0.05) here i.e.  in
chi-square.....then do multivariate...
others say do bi-variate analysis for each variable then select to
multivariate if they are significant at p-value of 0.2.....then analyse them
to find independtly associated variables.[let remind u that when i see
p-value of chi-square and bi-variate,is not same....why comes?not clear]

need your idea

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Variable-selection-t-multivariate-analysis-tp5046017p5046017.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (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

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (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

[Rich Ulrich]

This strategy inherits the worst flaws of "stepwise regression".
Google for <Frank Harrell stepwise regression> - which finds
  http://www.stata.com/support/faqs/stat/stepwise.html

I don't know who your "some" consists of, but statisticians are
consistently and uniformly skeptical about pre-testing of variables
for an analysis, if you want to preserve any value in interpreting
p-values later on.  A better way to choose variables for a substantial
analysis is to rely on the opinion of experts by referring to literature.

Pre-testing puts you in the position of doing an analysis that is
frankly exploratory.  The main cure for this is to do massive
amounts of cross-validation.  I expect that different sets of
predictors, in different areas, will be more or less congenial to
the two strategies you outline.  For instance, if you have a really
huge N,  you might do pre-selection with a p-value even smaller
than 0.05.

An alternative to this strategy - again, depending on the sort of
data you have - would be to use factor analysis to reduce the
set of independent variables to a few composite scores.  Or do
some sort of reliability analysis to discard variables that are poorly
measured or otherwise tend to be meaningless.

--
Rich Ulrich

> Date: Sun, 4 Dec 2011 02:43:27 -0800

> From: [hidden email]

> Subject: Variable selection t multivariate analysis
> To: [hidden email]
>
> Before recruiting variables to multivariate analysis,some says do first
> chi-square then select those are significant (p<0.05) here i.e. in
> chi-square.....then do multivariate...
> others say do bi-variate analysis for each variable then select to
> multivariate if they are significant at p-value of 0.2.....then analyse them
> to find independtly associated variables.[let remind u that when i see
> p-value of chi-square and bi-variate,is not same....why comes?not clear]
>
> need your idea
>

[...]

 Dear Art Kendall, Thank u.
This is  a self study for second degree fulfillment.To give a summarized idea,this is a study dealing on determinants of adherence status(dependent variable is categorical-expected to analyzed by binary logistic regression)/done by hospital based cross-sectional study/participants selected consecutively/multi-colliniarity is considered/

so before i just inserted the variables to multivariate logistic regression,what precondition should be done in due with what i said.Why comes the difference of p-value while doing by a chi square and univriate binary log.regeression i.e.a dependent variable and one independent variable.(what is their difference).

[Art Kendall]

As Rich pointed out using preconditioning to filter variables is much the same as stepwise methods.  Stepwise methods are highly undesirable (unadvisable).

In logistics regression you want to have something like 10 or 20 cases is the smaller side of the dichotomy.  You did not detail what your current potential IVs are, so all I can point out is 1) maybe some kind of factor analysis for data reduction and/or 2) pick a few variables on theoretical grounds.

A dichotomy seems a very coarse way to measure adherence.  Is there no way you could get something like a 5-point extent scale? Categorical regression could be used to see how much difference level of measurement makes.

Art Kendall
Social Research Consultants

On 12/5/2011 3:08 AM, [hidden email] wrote:

 Dear Art Kendall, Thank u.
This is  a self study for second degree fulfillment.To give a summarized
idea,this is a study dealing on determinants of adherence status(dependent
variable is categorical-expected to analyzed by binary logistic
regression)/done by hospital based cross-sectional study/participants
selected consecutively/multi-colliniarity is considered/

so before i just inserted the variables to multivariate logistic
regression,what precondition should be done in due with what i said.Why
comes the difference of p-value while doing by a chi square and univriate
binary log.regeression i.e.a dependent variable and one independent
variable.(what is their difference).

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Variable-selection-t-multivariate-analysis-tp5046017p5048118.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (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

===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (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