An easy method to robust a stepwise regression model at SPSS?

[dermot]

Hi everyone, (excuse me if I ask a question that doesn't make much sense)

I'm working on a regression model for exploratory purposes, so a series of independent variables (which found significant based on bivariate analyses) were put into a stepwise regression model.
Some of the variables may have a problem with heteroscedasticity, so I would like to do a "robust" by bootstrapping in SPSS, but it doesn't work with automatic selection method (e.g. stepwise), so my question is:

Does it make sense to put the predictors selected based on the final stepwise model into a new regression model with the "enter" method with bootstrapping? (then report the values from this "bootstrapping enter model")

P.s. as this is just for a working paper, I don't want to spend time on checking heteroscedasticity based on every independent variable and correct them, etc.

Thanks!

[Maguin, Eugene]

I don't think this plan is going to work well for you and I don't think it's needed. As I read your message you have a single dependent variable (DV) and multiple IVs that have been found to be associated with the DV. Turns out that those IVs have different variances. Ok. What's new? How does it matter if the IVs have different variances?

But, let's suppose you want to persist in this plan. Regression has a stepwise subcommand. How is this different from what you refer to as an "automatic regression routine"? Each bootstrap cycle will yield a different covariance matrix and, thus, the retained predictor set may/will differ. How do you resolve that variability?
Gene Maguin



-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of dermot
Sent: Thursday, February 26, 2015 7:47 AM
To: [hidden email]
Subject: An easy method to robust a stepwise regression model at SPSS?

Hi everyone, (excuse me if I ask a question that doesn't make much sense)

I'm working on a regression model for exploratory purposes, so a series of independent variables (which found significant based on bivariate analyses) were put into a stepwise regression model.
Some of the variables may have a problem with heteroscedasticity, so I would like to do a "robust" by bootstrapping, but SPSS doesn't work with automatic selection method (e.g. stepwise), so my question is:

Does it make sense to put the predictors selected based on the final stepwise model into a new regression model with the "enter" method with bootstrapping? (then report the values from this "bootstrapping enter
model")

P.s. as this is just for a working paper, I don't want to spend time on checking heteroscedasticity based on every independent variable and correct them, etc.

Thanks!



--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/An-easy-method-to-robust-a-stepwise-regression-model-at-SPSS-tp5728821.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]

If you have so many trivial variables that you will not bother to figure
the proper transformations for them, then you have too many variables
to tackle with stepwise in any guise.

Consider:  If there are 100 variables, with 20 "real" predictors, then the
real predictors will tend to be correlated with each other.  For p=100,
there will be about 5 variables with p < .05  that are "random", i.e., only
associate with the outcome by chance.   These 5 artifacts will tend to be
NOT correlated with the real predictors, so they will have a good chance of
dominating the results of any stepwise procedure where the "partial
coefficients" are selected on the basis of having independent contributions
to the prediction.   I once showed this with simulations that gave me two or
three times fake predictors for every real one.  Your variable pre-selection
("use all the ones that are significant") is designed to create this effect,
without actually offering all 100 variables to the regression procedure. 
 - If that is what your data look like, you would be able to obtain more "robust"
prediction (as confirmed by other samples) if you insist on creating your final
score from *all* the bivariate predictors, perhaps with equal weights.  - You
want to effectively out-vote the random predictors with the real ones. 

Otherwise, depending on the problem, it may make for better logic if you
*insist* that the predictors be inter-correlated before you use them, instead
of letting stepwise regression favor the ones that are not intercorrelated.


The small-sample way of coping with this is to use factor analysis, by eye
or by computer, to create a much smaller number of reasonable predictors.

The only way to actually test your selections from hundreds of predictors is
by cross-validation -- usually, a lot of it.  And not the bootstrap variety.
Cross validation requires relatively huge samples, if you want to be able to
see *large* effects in each of several sub-samples.

--
Rich Ulrich

> Date: Thu, 26 Feb 2015 05:46:36 -0700

> From: [hidden email]
> Subject: An easy method to robust a stepwise regression model at SPSS?
> To: [hidden email]
>
> Hi everyone, (excuse me if I ask a question that doesn't make much sense)
>
> I'm working on a regression model for exploratory purposes, so a series of
> independent variables (which found significant based on bivariate analyses)
> were put into a stepwise regression model.
> Some of the variables may have a problem with heteroscedasticity, so I would
> like to do a "robust" by bootstrapping, but SPSS doesn't work with automatic
> selection method (e.g. stepwise), so my question is:
>
> Does it make sense to put the predictors selected based on the final
> stepwise model into a new regression model with the "enter" method with
> bootstrapping? (then report the values from this "bootstrapping enter
> model")
>
> P.s. as this is just for a working paper, I don't want to spend time on
> checking heteroscedasticity based on every independent variable and correct
> them, etc.
>
> Thanks!

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

[dermot]

Thanks a million!

[Muir Houston-3]

I take it you are aware of the numerous robust critiques of using stepwise entry methods?



Muir Houston, HNC, BA (Hons), M.Phil., PhD, FHEA
College of Social Sciences Ethics Officer
Social Justice, Place and Lifelong Education Research
School of Education
University of Glasgow
0044+141-330-4699

Silver bullet or red herring? New evidence on the place of aspirations in education
R3L+ Project - Adult education in the light of the European Quality Strategy
http://www.learning-regions.net/

GINCO Project - Grundtvig International Network of Course Organisers
http://www.ginconet.eu/

THEMP - Tertiary Higher Education for People in Mid Life
http://themp.eu/ 
 

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of dermot
Sent: 26 February 2015 12:47
To: [hidden email]
Subject: An easy method to robust a stepwise regression model at SPSS?

Hi everyone, (excuse me if I ask a question that doesn't make much sense)

I'm working on a regression model for exploratory purposes, so a series of independent variables (which found significant based on bivariate analyses) were put into a stepwise regression model.
Some of the variables may have a problem with heteroscedasticity, so I would like to do a "robust" by bootstrapping, but SPSS doesn't work with automatic selection method (e.g. stepwise), so my question is:

Does it make sense to put the predictors selected based on the final stepwise model into a new regression model with the "enter" method with bootstrapping? (then report the values from this "bootstrapping enter
model")

P.s. as this is just for a working paper, I don't want to spend time on checking heteroscedasticity based on every independent variable and correct them, etc.

Thanks!



--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/An-easy-method-to-robust-a-stepwise-regression-model-at-SPSS-tp5728821.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