Is correction for attenuation in my regression analysis necessary?

#1
Hi everyone,

I'd like to know if I should consider a correction for attenuation in a multiple regression I'd like to run on my data.

Here's why:

Based on one of the predictors (identity management), people are allocated into 4 different groups. This variable has 16 items (4 items for each identity management group). The cronbach-alpha value for one of the groups is below 0.5.

My questions are:

- Should I do away with the group altogether because the cronbach-alpha is low (0.4)? i.e. I only consider the 3 other groups? Would this be correct?

- I've ready that I can run a correction for attenuation in the multiple regression analysis. Is this necessary? And is there any way I can run it on SPSS? Or would I have to run it on SPSS Macro/R/another program?

Thank you in advance!
 

Karabiner

TS Contributor
#2
Based on one of the predictors (identity management), people are allocated into 4 different groups. This variable has 16 items (4 items for each identity management group).
Not completely clear what you are doing there.
You have a variable measured with 16 items, and
after allocation into groups (how was this done?
and how large is your sample size, BTW?) you only
use 4 items in each respective group?


The cronbach-alpha value for one of the groups is below 0.5.
See above - what exactely does have a Cronbachs Alpha < 0.5?

- Should I do away with the group altogether because the cronbach-alpha is low (0.4)? i.e. I only consider the 3 other groups? Would this be correct?
You did not describe the study (topic, research question etc.)
and the technical details are not clear (see above), so this
is not easy to answer.

- I've ready that I can run a correction for attenuation in the multiple regression analysis. Is this necessary?
You did not describe what you will to regress on what, and
what for. Or, what you want to achieve by such a correction.
Is is a common practice in your field?

With Kind regards

K.
 
#3
Hi K.,

Thanks for your comments :)

You have a variable measured with 16 items, and after allocation into groups (how was this done? and how large is your sample size, BTW?) you only use 4 items in each respective group?
My sample size is about 180. After cleaning the data, I have about 160. Allocation in group was based on calculating the mean score for each subdimension. Participants were asked to indicate the frequency of their behaviour with regards to the item (never, occasionally, frequently, always). I am using this scale to determine which groups the participants can be allocated to.

A study, which uses this scale says this: "Item ratings are averaged and higher scores indicate greater levels of the corresponding behaviors."

Based on the data, all participants have been allocated to 3 groups. 1 of the groups remains empty, as no one has indicated that they have shown behaviour manifested in the items.

You did not describe the study (topic, research question etc.) and the technical details are not clear (see above), so this is not easy to answer.
The topic is on authenticity at the workplace amongst LGBT employees. The research question is about how self-rated authenticity affects the career-relevant variables (work engagement, commitment, job satisfaction, etc.) amongst LGBT employees, depending how they deal with their identity at the workplace - in other words, if they are "out" or "in the closet".

Outcome variables:
-career-relevant variables

Predictors:
-identity disclosure
-self-rated authenticity (proposed moderator)

You did not describe what you will to regress on what, and what for. Or, what you want to achieve by such a correction. Is is a common practice in your field
I plan to regress identity disclosure and authenticity on career-relevant variables. The cronbach-alpha for one of the sub dimensions in the identity disclosure scale is below 0.5.

So, given that only 2 groups are filled (group 1 is empty & the scale for group 3 has a low cronbach-alpha score), I wanted to know if I should/could conduct a correction to run an analysis, if I have 3 groups.

I won't say it's common practice.

Thanks for your feedback on this!
 

Karabiner

TS Contributor
#4
I still have some problems to understand what you are doing.
You want to perform 4 separate analyses, one for each group?
And within each analysis, you will use only one specific 4-item
subscale as predictor? Or all four 4-item-subscales? In any case,
I don't understand why the grouping is done at all.
If a group is constructed by picking participants who have high
values the items of one of the scales, then variance within
the group will be low for these items, and hence correlations
between items will decrease. Don't know how correction for
attenuation could help here.


With kind regards

K.
 

Miner

TS Contributor
#6
What is attenuation? Can you provide a link to the concept.
I believe that this is what the OP means https://en.wikipedia.org/wiki/Regression_dilution

Correction for attenuation is done on the correlation, not the regression. See https://en.wikipedia.org/wiki/Correction_for_attenuation

If you have noise in your IVs, you would use an Errors-in-variables approach. See https://en.wikipedia.org/wiki/Errors-in-variables_models and Deming regression https://en.wikipedia.org/wiki/Deming_regression
 

Karabiner

TS Contributor
#7
Correction for attenuation is done on the correlation, not the regression.
Seemingly, this statement answers the OP's question.

Admittedly, I have only seen it with correlation before, not with regression,
but I assumed that the OP had a reference for using a correction with regression.

With kind regards

K.