Including control variables

Frequently asked questions about PLS path modeling.
Post Reply
cgrimpe
PLS User
Posts: 11
Joined: Tue Nov 15, 2005 12:20 pm
Real name and title:

Including control variables

Post by cgrimpe »

Hi,

there has been quite a lot of discussion on moderating effects in PLS. However, I'm still unsure how to include a bunch of control or context variables. In OLS, you typically have your hypothesized relation and then there are a number of controls that could have an impact on this relationship. These might include firm relatedness, age, size, experience and so forth that are normally simply added to the regression equation. Some of them might also be dummy variables.

My understanding up to now is that you would have to create an LV for every control measured by just one indicator. This, however, renders new problems: First, constructs shouldn't be measured by just one item and second, this could blow up your required sample size when applying Chin's rule.

What is your opinion?
User avatar
panhans
PLS Junior User
Posts: 9
Joined: Mon Jan 02, 2006 8:45 am
Real name and title:

Maybe use subgroup analysis first

Post by panhans »

... why not first consider a subgroup analysis to identify which control variables are really relevant, and then to run PLS a second time adding only those controls that showed a significant effect in the subgroup analysis?

Best,
Dirk
jjsailors
PLS Expert User
Posts: 62
Joined: Fri Oct 14, 2005 1:43 am
Real name and title:

Re: Including control variables

Post by jjsailors »

cgrimpe wrote:Hi,

My understanding up to now is that you would have to create an LV for every control measured by just one indicator. This, however, renders new problems: First, constructs shouldn't be measured by just one item and second, this could blow up your required sample size when applying Chin's rule.
First, the need to measure constructs by multiple indicators to correct for measurement error is independent of the anaylsis method, so if you would use a single indicator for these control variables in an OLS regression, it is no worse (and no better) to do so in PLS. Also, there is no modeling of measurement error in PLS anyway (regardless of whether the indicators are specified as formative or reflective).

Second, regading sample size. There was a nice article referenced the other day on the forum (~ PLS is not a Silver Bullet) that you should read. If your sample size is so small that you would have concerns by adding a few control/covariate variables, you may have issues even without adding them that you should explore.
John J. Sailors, PhD
Associate Professor of Marketing
The University of St. Thomas
Opus College of Business
Minneapolis, MN
Post Reply