Consider the Stata command: reg var1 var2 var3, b.
Such command offers the possibility to have the predictors' coefficients standardised. They are thus expressed in terms of standard deviations, which seems to be fairly hard to comment or interpret. I would like to have different opinions and arguments in this respects since I heard many diverging opinions, most of them concluding to their uselessness.
This site points out why the apparent advantages of the standardized betas are illusory
http://www.jerrydallal.com/LHSP/importnt.htm
I would also add that I often use dummy 0/1 predictors and it is does not seem natural to think about a SD change in such variables.
I think that you need to evaluate effect size in relation to the specific problem you are working on and keeping the response in the same units that you measured it in, helps with this. It really matters if the change in the responses is measured/estimated in cms, rather than metres.
So how do you compare effect sizes?
I make predictions for a sensible range of values for the predictors (holding other predictors at their mean value) and graphically compare these predictions using the natural scale of measurement of the response. Comparing graphs with the same scale for the predictions on the Y axis helps you think and interpret how big these differences are.
In MLwin, we have built a tool (based on simulation) to do this and it works for normal response models as well as counts, binary and ordinal outcomes - you can put confidence intervals around the predictions and you can handle and visualize the effects of interactions in the predictions. ( For advanced readers it does this for both population average and subject-specific situations.)
You can see this tool in use in all of the following which are on ResearchGate
A multilevel analysis of the role of the family and the state in self-rated health of elderly Chinese. - on the graphs show a complex comparison of a lot of predictors (see Figure 1)
Zhixin Feng, Wenfei Winnie Wang, Kelvyn Jones
Developing multilevel models for analysing contextuality, heterogeneity and change using MLwiN, Volume 1 - search for customised
Kelvyn Jones, V. S. Subramanian
Developing multilevel models for analysing contextuality, heterogeneity and change using MLwiN, Volume 2- search for customised
and this manual explains the logic behind the simulations/predictions
Manual supplement for MLwiN Version 2.14
Jon Rasbash, Chris Charlton, Kelvyn Jones, Rebecca Pillinger
While these tools help - in the end it comes down to interpretation and you knowing the context in which you are modelling.
Dear all,
I wish to thank you very much for your helpful answers. I will still have to learn more about it.
- Badges
- Science topic
- Similar topics
- Counseling Psychology
- Confusion