Re: hierarchical multiple regression question

[Art Kendall]

When some subset of the variables is perfectly predictive of some other subset of the variables, it is necessary to drop one of the set.  That is why you would enter 1 fewer dummy than you have groups to represent a nominal level variable.  The same thing applies when the variables are items and scores, ranks within cases, proportions within cases, or standardized (Z scored) within cases. I.e., you need to drop 1 variable from the set, or you have "perfect collinearity".

Stepwise leads the secondary control coping variable to be the only of the 5 coping variables that remains (whereas with entry method SPSS automatically does NOT enter only this variable despite tolerance and VIF suggesting no serious problems with collinearity).

You have no problem with the reduced subset. You would if the other variable were in.

Should I be using the total scores (vs. proportion scores here) to prevent this problem.

 

Are the alphas in your response based on your data or from outside research?

Participants rated 57 items on a 4-point scale ranging from 1 (never) to 4 (a lot) yielding five factors (3 voluntary coping strategies; 2 involuntary stress responses) and 19 subscales. 

It looks like you have 5 scales (factors) which hopefully are orthogonal. is that true?  Would you say that the 19 subscales are facets of the 5 factors?

Should I be using the total scores (vs. proportion scores here) to prevent this problem.

Is the "controlling for total response" (ipsatizing) to aid interpretation in a clinical setting?  Since the response scale is an extent scale anchored at "never", how does it come about that there are differences in the total number of items with responses?


I am posting this reply to the list. That is the netiquette because other members of the list will see it and respond to it (or not bother). Also some people watch the list to further their understanding.  People who search the archives will then have the complete thread.


Art Kendall
Social Research Consultants

Tanya Tompkins wrote:

Thanks again for your interest and help.  Please see below for answers.

 

Tanya L. Tompkins, Ph.D.

Associate Professor

Department of Psychology

LINFIELD COLLEGE

900 SE Baker Street, A570

McMinnville, OR  97128

503-883-2684 (phone)

503-883-2669 (fax)

[hidden email]

 

 

---

From: Art Kendall [[hidden email]]
Sent: Thursday, December 10, 2009 8:37 AM
To: Tanya Tompkins
Subject: Re: hierarchical multiple regression question

 

What happens if you change the word "STEPWISE" to "ENTER" for the second block?

Stepwise leads the secondary control coping variable to be the only of the 5 coping variables that remains (whereas with entry method SPSS automatically does NOT enter only this variable despite tolerance and VIF suggesting no serious problems with collinearity).

Please give us the details of how you create the set of measures for the second block?  They are the 5 major dimensions of a coping measures called the Responses to Stress Questionnaire (RSQ; Connor-Smith et al., 2000).  Here is the description from my method section:

 

Coping.  The Responses to Stress Questionnaire (Family Conflict - Version) (RSQ; Connor-Smith et al., 2000), was used to assess both involuntary stress reactivity and volitional coping efforts in response to family conflict.  Participants rated 57 items on a 4-point scale ranging from 1 (never) to 4 (a lot) yielding five factors (3 voluntary coping strategies; 2 involuntary stress responses) and 19 subscales.  The psychometric properties of the RSQ have been well established across samples of varying ages and SES levels (Connor-Smith et al., 2000). Primary Control Engagement (a = .80) includes items measuring problem solving (a = .59), emotional expression (a = .75), and emotional regulation (a = .46). The Secondary Control Engagement (a = .80) factor is comprised of items measuring strategies such as acceptance (a = .51), cognitive restructuring (a = .52), distraction (a = .44), and positive thinking (a = .69). Disengagement coping (a = .86) consists of strategies such as avoidance (a = .80), denial (a = .43), and wishful thinking (a = .72).  Involuntary Engagement (a = .92) includes rumination (a = .76), intrusive thoughts (a = .72), physiological arousal (a = .79), emotional arousal (a = .70), and impulsive action (a = .80).  Involuntary Disengagement responses (a = .85) include emotional numbing (a = .43), cognitive interference (a = .72), inaction (a = .60), and escape (a = .62).  As suggested by Connor-Smith all scores are computed as proportions of the total score to control for overall rates of responding.

it is it possible that these are some form of an ipsative set?  Does the set sum to a constant? (maybe percents or  proportions?)  Yes they are proportion scores

If you try a REGRESSION with the variable that gets kicked out as the DV and the rest of the variables in that block as IVs, what happens?

I get the following warning: “For the final model with dependent variable cRSQsecondarycontrolpropcorr, the variance- covariance matrix is singular. Influence statistics cannot be computed.”

 

 

 

Thanks again for your help!

Art Kendall
Social Research Consultants

Tanya Tompkins wrote:

On Dec 10, 5:04 am, Art Kendall [hidden email] wrote:

please post the syntax you are using.  If you are using the GUI (menus

etc), exit via <paste>.

What is the level of measurement of each of your variables?

How many cases do you have?  Is there enough residual variance after

block 3 that interactions might be of interest?

Art Kendall

Social Research Consultants

Tanya Tompkins wrote:

I am conducting a HMR with the following variables:

DV = youth self-report of internalizing problems

IVs = 1st block: self-report of externalizing problems

         2nd block: individual coping (primary control, secondary

control, disengagement, involuntary engagement, involuntary

disengagement)

         3rd block:  co-rumination (conceptualized as social coping

process characterized as problem-focused talk that shares with

rumination a negative and repetitive focus)

The idea here is wanting to find out if co-rumination predicts any

unique variance in internalizing symptoms above and beyond co-morbid

problems and individual coping efforts.

My challenge is how to handle this second block as there are strong

positive and negative bivariate relationships among coping strategies/

responses (.30 to .70).  When I use a simple forced entry method in

this second block it kicks out secondary control coping but the

collinearity diagnostics do not seem to suggest problems with this

specific variable.  It has one of the strongest bivariate

relationships with my IV (r= - .48). Any ideas about what is going on

and what to do about it (I've been warned about a stepwise approach so

am hesitant to use this as a solution)?  I imagine part of the problem

is that it is most strongly associated with involuntary engagement (r

= -.73) which is the coping variable that is also most strongly

related with my IV, but in the opposite direction (r = .47).

Thanks for any ideas/guidance you can suggest.

Tanya

Here is the syntax I am using (SPSS):

REGRESSION

  /DESCRIPTIVES MEAN STDDEV CORR SIG N

  /MISSING LISTWISE

  /STATISTICS COEFF OUTS CI R ANOVA COLLIN TOL CHANGE ZPP

  /CRITERIA=PIN(.05) POUT(.10)

  /NOORIGIN

  /DEPENDENT ysranxiousdepressedraw_sqrt

  /METHOD=ENTER ysraggressivebehaviorraw_sqrt

  /METHOD=STEPWISE cRSQprimarycontrolpropcorr

cRSQsecondarycontrolpropcorr

    cRSQdisengagementpropcorr cRSQinvoluntaryengagementpropcorr

cRSQinvoluntarydisengagementpropcorr

  /METHOD=ENTER corumination

  /SCATTERPLOT=(ysranxiousdepressedraw_sqrt ,*SDRESID)

  /RESIDUALS DURBIN HIST(ZRESID)

  /SAVE RESID ZRESID SRESID.

Here are scales of measurement:

DV - interval (square root transformation to correct positive skew)

Coping IVs - proportion scores representing the total endorsed for

those group of strategies divided by total coping responses

Co-rumination IV - interval (likert-type scale - mean of 27 items)

I have 108 cases with complete data so adding any more predictors is

simply not advisable.  In fact, for theoretical reasons I am actually

considering simply including in block 2 strategies that represent

engagement or approach coping of any kind (primary control, secondary

control and involuntary engagement).  I am (as noted above) however

just confused about what is going on and wanting to see if anyone else

might have ideas

===================== 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