0
$\begingroup$

ologit

I hope the image is relatively clear... My dependent variable is the change in people's idea about whether it is a good idea, where it is the support they give to the same question minus the support they gave before. Frame is where 1=gain, 0=loss. PositionOWNPART is where 1=Owner/Partner, 0=Not Owner/Partner.

I interpret that the odds of being in the top category versus the combined middle and low are about 1.5 times higher for those who received the gain frame compared to the loss frame as well as for the top and middle category combined to the bottom category, at a 5% significance level.

However, I was wondering whether it is possible, from this graph, to interpret whether people who are owner/partner(PositionOWNPART) change their support for whether it is a good idea depending on which frame they receive? Thank you!

****Edited******* interaction variable and margins

Thank you for all your help, I have now added the interaction variable. For the interaction variable at (1 1), can I interpret that the odds of being in the top category versus the combined middle and low are about 2 times higher for those who received the gain frame and is an owner/partner as well as for the top and middle category combined to the bottom category, at a 5% significance level? Would this be correct?

I need help on how to analyse the margins table too, thank you!

$\endgroup$
  • 1
    $\begingroup$ For that you would need an interaction term between Frame and PositionOWNPART in your model. Try that and post again if you still have a question. $\endgroup$ – EdM Aug 21 '15 at 18:52
  • $\begingroup$ The additive marginal effects bake in this interaction. Try margins r.PositionOWNPART, dydx(Frame) or margins, dydx(Frame) at(PositionOWNPART=(0 1)). @EdM's approach can be done with i.PositionOWNPART##i.Frame as the covariates. $\endgroup$ – Dimitriy V. Masterov Aug 21 '15 at 20:37
  • $\begingroup$ @EdM Thank you! I have now edited the post where I have included the interaction variable and the margins. $\endgroup$ – Irene Aug 22 '15 at 0:07
  • $\begingroup$ @Dimitriy V. Masterov Thank you! I have now edited the post where I have included the interaction variable and the margins. I was wondering how to interpret the margins part as well? $\endgroup$ – Irene Aug 22 '15 at 0:07
1
$\begingroup$

Your understanding of how Stata reports coefficients for ordered logistic regression is correct. A coefficient for this 3-category case represents both the odds ratio for top category versus combined middle+low categories and for combined top+middle versus low. This page explains the Stata output for ordered logistic regression, and also suggests a test of whether this simple odds model is appropriate, something you probably want to examine.

Your interest in whether the influence of position depended on frame (and necessarily vice-versa) required adding an interaction term to the model. Interpretation of Stata output for interaction terms between categorical predictors is explained on this page.

You have to be careful in examining the odds-ratio coefficients. The odds ratio of about 2 for the (1,1) case in the interaction table toward the bottom right of your output is with respect to the (0,0) case, as are all the other coefficients in that table. The (1,1) coefficient is the only one significantly different from (0,0), and its 95% CI do not overlap the point estimates of the other 2 coefficients, supporting a situation where only the combination of "gain" for frame and "owner/partner" for position has an influence on your outcome measure. You should do a more formal analysis of contrasts (if that was your original hypothesis before you saw the data) or of the multiple comparisons before publishing. I'm not sufficiently familiar with Stata to give you a good explanation of the margins table. The second page I linked seems to illustrate some alternate ways to examine interactions in Stata.

$\endgroup$
  • $\begingroup$ Thank you! So when I interpret the odds ratio for the interaction variable, I will be sure to add that it is with respect to (0,0), which is where the person is not an owner/partner and received the loss frame. In regards to your mention of more formal analysis of contrasts, would this work? Hypothesis: owners/partners are more likely to respond to the frames. But, my results found that with respect to receiving the loss frame and having a position that is not owner/partner, an owner/partner do react positively to the gain frame, however its reaction to the loss frame is not significant. $\endgroup$ – Irene Aug 22 '15 at 11:34
  • $\begingroup$ if you consider the difference between the "gain" and "loss" frames as the measure of "responding to frames," then owner/partners did have a larger response. Test that contrast. It's just that the difference in response was due to a large response to the "gain" with a response to "loss" not significantly different from the other group. That seems to be a perfectly respectable finding. Next time you do this, try using labels for the groups/frames instead of the 0,1 coding; it makes the outputs much easier to read and avoids confusion about what (0,1) means, for example. $\endgroup$ – EdM Aug 22 '15 at 12:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.