Dear All,
I have a dependent LV (satisfaction, SAT) measured reflectively with four indicators on a 7-point Likert scale in my model. SAT by itself is related to another LV (success, SUC). Now I’m interest in how the results (betas, R2) differ for groups with high / low satisfaction (above / below average).
Since SAT is used by itself in the model, is it a problem to do a multi-group analysis? If yes, what are the constraints?
Thanks,
Reinhard
Multi Group Analysis – Groups from LV in endogenous position
-
haroldwillaby
- PLS Junior User
- Posts: 3
- Joined: Sun Nov 15, 2009 11:38 pm
- Real name and title:
Re: Multi Group Analysis – Groups from LV in endogenous posi
Hi Reinhard.
First, to answer your question, there is no reason I know of that you couldn't do an MGA, so long as both betas are statistically significant (ss). If one beta is ss and the other isn't, then you don't need to run an MGA... I think you can assume they differ.
--> But you should justify splitting into groups on some theoretical or empirical grounds. Artificial grouping, even based on the mean, has a few drawbacks (including losing meaningful explanatory variance).
------
To further the discussion... As I see it, given that you have all continuous indicators, your question regarding differing betas, R2s for high and low is actually a question of non-linearity in the relationship of SAT and SUC.
Imagine that the scatterplot of SAT and SUC is asymptotic so that there is a diminishing response of SAT with SUC. (Which might not be that difficult to imagine perhaps? This would mean that people get less satisfaction out of increasing success. It certainly would be the case for many biological relations... think Satiety and Caloric intake.) In just such a case, the beta for Lo SUC would be greater than the beta for Hi SUC. You would likely also find a smaller R2 for Hi SUC as a result.
Splitting into Hi and Lo doesn't really overcome the problem of non-linearity. For a continuous variable with a spread of responses that roughly approximates a normal distribution, any artificial split into high and low still only gets you a straight line through what are two curved 'slopes' which are your betas.
Apologies for the stats theory lecture - you probably already know all of this you're here. More practically, do you have any reason to think that there is a natural grouping of Hi and Lo SUC (e.g. a bi-modal distribution)? If not, you should be clear about what you expect of the different ends of the distribution (Do you have a theory or prior that would advise you?), and consider either:
1) Going back to square one and looking at the scatterplot to see if there is any reason to think you should split the sample, and if so how into how many groups, at what point(s) that should be. I note that you can calculate factor scores for each respondents SAT composite in a separate software and plot the factor scores against SUC to get your scatterplot.
OR
2) Use alternative modelling techniques/software that allow for non-linear relationships (i.e. WarpPLS). WarpPLS is pretty clunky, but will do the job, and is pretty easy to figure out. It's torture to install though, and uses a lot of processor/memory on your computer. You'll need a decent machine to run it efficiently.
I hope that helps.
Regards,
Hal
First, to answer your question, there is no reason I know of that you couldn't do an MGA, so long as both betas are statistically significant (ss). If one beta is ss and the other isn't, then you don't need to run an MGA... I think you can assume they differ.
--> But you should justify splitting into groups on some theoretical or empirical grounds. Artificial grouping, even based on the mean, has a few drawbacks (including losing meaningful explanatory variance).
------
To further the discussion... As I see it, given that you have all continuous indicators, your question regarding differing betas, R2s for high and low is actually a question of non-linearity in the relationship of SAT and SUC.
Imagine that the scatterplot of SAT and SUC is asymptotic so that there is a diminishing response of SAT with SUC. (Which might not be that difficult to imagine perhaps? This would mean that people get less satisfaction out of increasing success. It certainly would be the case for many biological relations... think Satiety and Caloric intake.) In just such a case, the beta for Lo SUC would be greater than the beta for Hi SUC. You would likely also find a smaller R2 for Hi SUC as a result.
Splitting into Hi and Lo doesn't really overcome the problem of non-linearity. For a continuous variable with a spread of responses that roughly approximates a normal distribution, any artificial split into high and low still only gets you a straight line through what are two curved 'slopes' which are your betas.
Apologies for the stats theory lecture - you probably already know all of this you're here. More practically, do you have any reason to think that there is a natural grouping of Hi and Lo SUC (e.g. a bi-modal distribution)? If not, you should be clear about what you expect of the different ends of the distribution (Do you have a theory or prior that would advise you?), and consider either:
1) Going back to square one and looking at the scatterplot to see if there is any reason to think you should split the sample, and if so how into how many groups, at what point(s) that should be. I note that you can calculate factor scores for each respondents SAT composite in a separate software and plot the factor scores against SUC to get your scatterplot.
OR
2) Use alternative modelling techniques/software that allow for non-linear relationships (i.e. WarpPLS). WarpPLS is pretty clunky, but will do the job, and is pretty easy to figure out. It's torture to install though, and uses a lot of processor/memory on your computer. You'll need a decent machine to run it efficiently.
I hope that helps.
Regards,
Hal
reini_m wrote:Dear All,
I have a dependent LV (satisfaction, SAT) measured reflectively with four indicators on a 7-point Likert scale in my model. SAT by itself is related to another LV (success, SUC). Now I’m interest in how the results (betas, R2) differ for groups with high / low satisfaction (above / below average).
Since SAT is used by itself in the model, is it a problem to do a multi-group analysis? If yes, what are the constraints?
Thanks,
Reinhard