Home › Forums › Onyx Helpdesk › pvalues for regression weight estimates
Tagged: pvalues, significance
This topic contains 4 replies, has 3 voices, and was last updated by sebwin 1 year, 4 months ago.

AuthorPosts

September 27, 2017 at 3:14 pm #773
Does Onyx calculate these? If not, is there a substitute/workaround I could use?
Thanks.January 31, 2018 at 2:11 pm #783In teaching sessions I use the z values that appear in the window at the side of the model panel; this video may help:
https://youtu.be/uvHuF0fpWoApril 15, 2018 at 8:13 pm #803Really, I need to find out which coefficients are significant.
I would suppose that mostly everyone is interested in knowing which ones of the path values in a model’s estimate are significant, but I have yet to find this information displayed in onyx.April 15, 2018 at 8:28 pm #805Hi sebwin,
the path coefficients per se cannot be significant, but I guess you mean significantly different from zero, which makes a lot of sense of course. Robin’s way is certainly quick. You just check the zvalue whether it is above 1.96 or below 1.96, respectively, and if it is, the path coefficient is significantly different from zero. Advantage of this method of displaying the result (contrary, say, to giving the pvalue for this test, which some other programs do) is that you can also check if the zvalue is, say, above 2.96; if it is, the path coefficient is significantly different from 1 instead of 0 (which sometimes can be more interesting than the test against zero, although 0 is of course more frequently useful).
If you want the “best possible pvalue”, which in this case would mean one that also includes the crossinformation from other parameters, I suggest to set up a likelihood ratio test; clone your model, and in the clone, fix the path you are interested in to zero (or one or any other value). You can then connect the two models (by dragging a path from one to the other) and check in the little ball on the edge between the models (by hovering over it with the mouse) what the pvalue for this comparison is. This is a Likelihood Ratio test, which is provably the best test asymptotically for normal distributed data.
BTW (I see we crossposted, and you asked about modification indices), you can also use the LR between these two models as a modification index if the path you restricted is a factor loading.
Cheers,
Timo
April 27, 2018 at 10:38 am #806Thank you, Timo, for your always elaborate answers.
It always takes me a while to process when you suggest an alternative approach to what I am trying to accomplish because I am not really “literate” in these matters and, thus, lack the necessary flexibility to do (and understand) A’ when I was set on doing A.
Your advice is greatly appreciated all the same — just with a lag of several days.
My thanks also to Robin, who suggested the same earlier. 
AuthorPosts
You must be logged in to reply to this topic.
Recent Comments