Odds ratio p-values
Message/Author
 Soo Owen posted on Sunday, September 30, 2018 - 5:27 pm
Hi,

I ran a multinomial logistic regression analysis in mplus. I am wondering why the p-values for coefficients don't align with those for odds ratios and which p-values I have to use to interpret and report the the results?

MODEL RESULTS
Two-Tailed
Estimate S.E. Est./S.E. P-Value

FNLCLASS#1 ON
ASIAN -3.129 0.934 -3.351 0.001
URMS -2.095 0.988 -2.120 0.034
INTLS -1.824 0.546 -3.340 0.001
GEN -0.700 0.533 -1.313 0.189

FNLCLASS#2 ON
ASIAN -2.750 0.920 -2.989 0.003
URMS -1.872 0.996 -1.880 0.060
INTLS -0.436 0.542 -0.806 0.421
GEN -0.567 0.542 -1.046 0.296
 Soo Owen posted on Sunday, September 30, 2018 - 5:30 pm
Hi,

I ran a multinomial logistic regression analysis in mplus. I am wondering why the p-values for coefficients don't align with those for odds ratios and which p-values I have to use to interpret and report the the results?

MODEL RESULTS
Two-Tailed
Estimate S.E. Est./S.E. P-Value

FNLCLASS#2 ON
ASIAN -2.750 0.920 -2.989 0.003
URMS -1.872 0.996 -1.880 0.060
INTLS -0.436 0.542 -0.806 0.421
GEN -0.567 0.542 -1.046 0.296

LOGISTIC REGRESSION ODDS RATIO RESULTS

(Est. - 1) Two-Tailed
Estimate S.E. / S.E. P-Value

FNLCLASS#2 ON
ASIAN 0.064 0.059 -15.915 0.000
URMS 0.154 0.153 -5.526 0.000
INTLS 0.646 0.350 -1.010 0.313
GEN 0.567 0.308 -1.407 0.159
 Bengt O. Muthen posted on Sunday, September 30, 2018 - 6:03 pm
Both parts of the outputs assume approximate normality for the statistic (logit/OR). That holds fairly well for the logits but less well for the ORs. For ORs, I would recommend requesting confidence intervals - they will then allow for the non-normality by using non-symmetric intervals.
 Soo Owen posted on Sunday, September 30, 2018 - 6:46 pm
Thanks for your prompt reply. So, can I report the significant results based on the coefficient p-values (rather than the OR p-values) and also report the ORs (also CIs for ORs) of the significant coefficient?

For example, from my result, the coefficient of URMS is not significant (0.06) but its OR is significant (0.00). So I got confused how to interpret this result. Thanks.
 Bengt O. Muthen posted on Tuesday, October 02, 2018 - 5:48 am
If you are going to report ORs, I would report only them and not the logits and I would report the CIs for the ORs and judge whether the CI covers 1 (the neutral point) or not.
 shonnslc posted on Wednesday, February 19, 2020 - 9:03 am
Hi,

I would like to follow up previous conversation on requesting "confidence intervals - they will then allow for the non-normality by using non-symmetric interval." In this case, should I simply add cinterval in the Output section in my code or I should request bootstrapping CI (Analysis: bootstrap = 1000; Output: cinterval(bootstrap)). Thanks!
 Bengt O. Muthen posted on Thursday, February 20, 2020 - 4:33 pm
You should request bootstrapping CI by

Analysis: bootstrap = 1000;
Output: cinterval(bootstrap)).