DEAR VIEWERS: I have recently uploaded a new video (2019) on binary logistic regression in STATA that INCLUDES MUCH MORE INFORMATION ON RUNNING AND INTERPRETING RESULTS. There are also ADDITIONAL SUPPLEMENTAL LEARNING MATERIALS THAT ARE LINKED UNDER THE VIDEO DESCRIPTION. PLEASE CHECK OUT: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-W53aMFOCgzM.html
Hi folks, I have just added a new 2021 video on binary logistic regression using Stata version 17 (ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-PvEjbhnIFic.html), with a focus on using the drop-down menus in the program. This video (and the accompanying Powerpoint) demonstrates how to run the analysis and how to interpret results. I have an additional video (ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-2vYqYmK6bkE.html) demonstrating how to obtain fully and partially standardized regression coefficients for your logistic regression model. I hope you try these out!
Thank you! The first video I have found that not only goes over how to run the code, but how to interpret results. My Econometrics Capstone project is entirely better because of this video.
Glad to be of help Will! Also, although you may be finished with your project, you might visit this newer video I put together in 2019 on binary logistic regression in Stata. It also has a Powerpoint that goes into more details (see under video description). Please share with your fellow students! Best wishes!
Great video but there was some mixed information regarding the negative coefficient. First, you said that males would be more likely to fall in the intent category than females and then later on you said that a negative coefficient means it is less likely to fall in the target category. So a negative coefficient means more likely or less likely to fall in the target category?
I'm so confused as to why when I my run my logit model (with results as coefficient), I get different coefficients and standard errors depending on how I order the independent variables. Each of these variables is binary.
There are only two binary variables and actually gender is nominal and the rest treated as continuous, i think that should pretty give the same results
Hi Mike! First of all: thank you very much for sharing your knowledge with us. However, I have a small doubt over the definition of some exogenous variables you have considered. For istances, when you say DOG you mean the extent to which a person perceives herself as being dogmatic, right? And regarding EPE, what really does it mean? I cannot understand at 1:26... Thank you in advance and sorry for bothering
Hi Teresa. Yes, higher scores on DOG indicate a person who is higher in dogmatism (lower scores indicate lower levels of dogmatism). EPE refers to External Political Efficacy. Higher scores indicate greater efficacy and lower scores indicate lower efficacy. Thanks for visiting :)
Hi Mike! Thanks for the video!! Do you have any recommended reading on when to treat ordinal items as metric? I am writing my final paper and ne some source on that... best
Thanks Mike for sharing document and analysis on logit output and its interpretation. What dose indicate the result of ldfbeta? additionally, if you have on multivariate probit analysis please share as.
Thank you for going over it and explaining. Question, let say in the polint variable, if it was to increase by a factor of 1, 2, 3, etc. How would the odds change? How can one model or calculate that? Thank you.
In my case I have a multinomial logit and my problem is that STATA has much limited options regarding goodness-of-fit comparing with binary logit (basically only AIC and BIC)
Excellent and fun explanation. I would like to test if a series of continuous IV predict two categorical variables (income level and gender). As you see, the DV are two, one of them is binary (gender) and the other has several levels (low, medium, high, extreme). What type of regression shall I use? I considered bivariate logistic regression, but here one of the DV is not binary.
Thanks for the video! If I have a unbalanced dataset, I have different observations for two different logistic regressions. How can I do to check whether there is an increase of model fitness? I am looking forward to your response. Thanks in advance!
Thanks so much for this video which helps my research. But, I am still confusing the Stata result table from logistics regression of "Odds ration" and "Coefficient". You mean, in Stata result, the interpretation of "coefficient" differs with OLS, but "Odds Ration" shows the same interpretation with OLS? Is it? Sorry for silly questions as I confused a lot on that point. Because I got my Stata result with "Odd rations" but I am now facing difficulties how to interpret my logistics regression model. Could you please give me more explanation and guidance how to interpret this "Odds ratio" please. Thanks in advance.
The coefficient in OLS is not the same as the coefficient in Probit/logit. In order to get the same meaning, you need to calculate the marginal effects. In Stata, you can do this by using the ''mfx'' command. Odds ratio is basically the ratio of the probabilities occurring vs not occurring and this can only be done for the logit model
Very helpful video. Many thanks! I am a bit confused about the nature of data to use in BINARY log regression. My concern is with the basic of data type. while binary suggests dichotomous values, you used continous values - POLINT, DOG & EPE. will it make any difference in the result if these values a dichotomized prior to regression?
Hi Chris, thanks for visiting! I'm not sure I totally understand your question. Are you asking about dichotomizing the IV's or the DV? The DV is binary. The IV's can be either continuous OR binary. For the record, it is permissible to include a binary predictor directly into the model as a predictor (moreover dummy coding the predictor to 0 and 1 can aid in interpretation). You could also, if you wish, treat the binary predictor as a factor and use the 'i.' prefix. [Because I tend to like to have more control over the dummy coding (rather than relying on program defaults for dealing with factor variables, which can be annoying in programs such as SPSS where the default may select the reference category for you ), I often just do the recoding myself and then enter the dummy variable directly into the model. If you have a factor with more than two levels, then could still manually choose to do the dummy coding yourself (and enter the dummy variables directly, as described above), or you could (again) use the 'i.' prefix and allow the program to dummy code your variables (but with the limitation that the default will select your reference category).] By the way, iif you want a more "up-to-date" presentation on binary logistic regression (and includes some additional supplemental learning materials under the video description), check out this RU-vid video I've recently put together ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-W53aMFOCgzM.html . I hope I have successfully answered your question!
@@mikecrowson2462 Many thanks for the feedback, Yes indeed, i was referring to dichotomizing the IVs. now, i understand that the IV can be continous given a binary DV, and one can play safe by dummy-coding to 0 and 1 (for ease of interpretation, yea?). Although i'm a bit fuzzy on the usefulness of having a "factor" predictor in the model. does it help in the model "fit" or ensure the analyst (not the program) has control on the reference category choosen? I will most definitely view your recent videos as suggested. Thank you.
Hi Chris, yes using dummy coding (0 and 1) can facilitate interpretation - particularly of the intercept in regression models. What I mean with 'factor' variables is a variable that is nominal or ordinal. With appropriate re-coding of a variable of this type, you can include it in a regression model (including binary logistic regression). Here's another video I have made on that topic (although the demonstration uses SPSS): ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-XGlbGaOsV9U.html Best wishes.
@@mikecrowson2462 Hi Mike. thanks once more for your awesome videos, they were helpful. Please, I have details of a BLR analysis; DV (outcome) = transition (0=unsuccessful, REF; 1=successful) Predictor = religion (0=christian, REF; 1=islam)...inserted in model as "i." factor variable coefficient -0.0524; p>(z) 0.909 odds ratio 0.948; p>(z) 0.909 I concluded that "religion" is non-significant predictor of "successful" transition, loosely. in further explaining the output; odds of successful transition is .909 for practicing Islams than for Christians. Do you think my explanations of the results are good? Thank you.
So Mike, what do we do if our specificity or sensitivity is really low. I ran into a low specificity (18%), while my sensitivity was 92%. What would one do in this situation?
Hi David, the difference in sensitivity and specificity generally indicates that the model is doing a better job of predicting one outcome than the other. Your model does not necessarily correctly predict group membership on the DV with the same level of accuracy. If you are wanting to increase accuracy of true negatives (specificity) then you might consider other theoretically relevant predictors that might do a better job of predicting group membership. I don't know what the nature of your research is, but the difference in classification accuracy could be less important than the type of classification errors. Just an FYI, I have a much new video on logistic regression with Stata - and there are accompany learning materials you can download at links underneath the video description. I hope you check it out: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-W53aMFOCgzM.html Best wishes!
Thank you Prof for a clear explanation of the Logistics regression model. I have a request, If you give time stamps for important parts of the video, it would be helpful.
@@sandeepsivakumar4959 Thanks for your posting. I'll have to look into the time stamps. I've never done that. However, what I have been doing over the last year is putting together Powerpoints and supplemental info with my videos. There is a much newer video here (ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-W53aMFOCgzM.html), and the Powerpoint that accompanies it can be downloaded here (drive.google.com/open?id=1HLjPcZZqXGwwa7Uee1A38L2WWoRbz5HK). Cheers!
Hi there, and thanks for your question. The fact is that the choice of statistical approach is heavily dependent heavily on the type of data you are working with. If you have a dependent variable that is binary and you are seeking to predict the probability of a case falling into one category or the other on that DV, where the probability of a given outcome is modeled as a function of one or more predictors, then binary logistic regression is a viable approach. By the way, I do have a more recent video that you should check out on this topic. I hope you check it out! ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-W53aMFOCgzM.html Cheers!
My understanding is that you don't compute inverse mills ratio when using binary logistic regression. It appears to be used in the context of probit regression. But here's a link that might give you the information you are looking for: statalist.org/forums/forum/general-stata-discussion/general/1348264-probit-inverse-mills-ratio