Message/Author 

scrumpy posted on Thursday, October 27, 2005  2:26 pm



Hi I am running a latent variable mixture model that yields 2 classes from 8 indicators. The output provides the odds ratio comparing the classes on each indicator. How would I calculate the confidence intervals to know which odds ratios are significantly different from one another? Is this in the output somewhere? Or is there a formula to calculate CI from SE in odds ratios? Thanks very much! 


If you ask for CINTERVALS in the OUTPUT command, you will obtain confidence intervals for the odds ratios. 

scrumpy posted on Friday, October 28, 2005  12:01 pm



I had completely missed that option in the manual. Thank you! 


Hi what is the formula to calculate the confidence interval for an odds ratio? which distribution does this follow?? And is an 'adjusted odds rato' different from odds ratio? Thank you very much. 

BMuthen posted on Saturday, November 12, 2005  6:15 pm



These are given in the Hosmer and Lemeshow logistic regression book cited on the website or other logistic regression books. Essentially, these build on the lower and upper confidence interval limits for the logistic regression coefficients(log odds) which are then exponentiated to give you the corresponding odds. An adjusted odds ratio is an odds ratio for a binary x variable where you have other x variables in the logistic regression. 


How would I calculate the odd ratios confidence intervals in MULTINATIONAL LOGISTIC REGRESSION whereby imputation is used. without imputation CINTERVALS in the OUTPUT command yields the CI. why MPLus dont give the CI with imputation. 


To get the confidence interval of the odds ratio, create the confidence interval for the relevant logit and then exponentiate to confidence interval. 

CB posted on Tuesday, May 19, 2015  6:50 am



Hello Drs. Muthen, I'm running an LCA with 2 classes and I'm interested in obtaining 95% confidence intervals for the itemresponse probabilities and latent class probabilities. I know that I can add the CINTERVAL option to the output. However, can you describe how to obtain 95% confidence intervals for the latent class probabilities, since these are means rather than thresholds? I know only one class has an estimated parameter and the other class is the complement, so I'm unsure how to obtain the confidence intervals for both classes. Any references/resources that you could share would be helpful too. Thanks! 


Your output gives the [c#1] estimate which is a logit. You also have the 95% CI for that logit estimate. The probability is P=1/(1+exp(logit)). Inserting the lower and upper limits of the logit CI in the expression for P gives you the CI for P. The P for the other class is 1P so inserting logit in that formula is how to get the CI for that. 


Hello, I am having difficulty calculating the CI for the ORs of DCAT output. I am using the formula as noted in the document in the FAQ section 'logOR ± 1.96*SE(logOR). Then exponentiate those two limits to get the OR limits' The problem is either that the CI does not contain the OR and\or that it is indicating a significant result when the p value is highly insignificant or vice versa. Here is an example. Comorbidities  Est= 0.029 SE=0.068 p=0.673 OR= log(OR) of 0.029 = 1.03 CI = 0.029+\ 1.96*0.068 = (0.16, 0.10) exp CI (1.11, 1.18) So as you can see the OR is not within the CI and the CI is in conflict with the p value. Am I doing something incorrectly or have I missed something? thank you 


Try again  the lower CI limit for the logOR is 0.0219  1.96*0.068 etc. 

aurora posted on Tuesday, April 05, 2016  9:46 pm



Hello Drs. Muthen, I am running a discrete time survival analysis with timevarying/independent covariates. In your board, adjusted odds ratio is explained as an odds ratio for a binary x variable where there are other x variables in the logistic regression. Then how could I obtain an adjusted odd ratio? Thank you! 


You exponentiate the slope for x. 


Hi there I am running a multinomial logistic regression (4 levels on DV, 4 continuous IVs, 2 binary. TYPE=COMPLEX. 1) How do I determine the overall 'fit' of the model (e.g. in SPSS I get a log likelihood and Chi square of the final model compared with the intercept)? 2) How do I get confidence intervals? tried CINTERVALS in my output, but can't bootstrap with clusters 3) If the model runs without warnings can I assume it is identified/reliable (even with only 6 in one group of DV)? I screened for too small expected cell frequencies, linearity of the logit, multicollinearity, etc 5) How do I get a pseudo R2? 6) How do I interpret the intercept given for each level of my DV. Is it the probability of being in that level of the DV vs the reference group given all the predictors in the model (e.g. U1 vs U4)? Or not accounting for any predictors? Or something else? 


1) Run 2 models where one has all the slopes fixed at zero. 2) You should get CIs. If not, send output to Support along with license number. 3) Identified, yes, but the question is how good the estimates and SEs are  a Monte Carlo simulation would be needed. 5) You get McFadden's Rsquare by the runs in 1). 6) See books on multinomial regression, including our new book. 

Amber Fahey posted on Friday, April 20, 2018  10:48 pm



In growth modeling, does the pvalue for log odds provided in the output mean anything? I have several instances where the pvalue was > .05 for my logistic regression coefficient but the calculated odds fell within the lower and upper bounds of CI 95%. In the same model, I also calculated very large log values e.g., 111.05 and 39.65. I'm wondering if this is an indication that there is a problem with my model. Any feedback is greatly appreciated! Example: Log Odds = 0.19, p=.06 SE = (0.10) Odds =1.21 CI 95% = (0.99, 2.70) 


You have CI 95% = (0.99, 2.70) which says that the norelationship value of 1 is included in the CI, so you don't have significance. 

Amber Fahey posted on Sunday, April 22, 2018  10:30 am



Thank you for the prompt feedback! In your estimation, Should I be concerned about the high odds I received once I took exp(b)? Ive read that was an indication of poor model fit. 


No, it is not an indication of poor fit  it is ok. 

Back to top 