Multinomial and ordinal logistic regression using SPSS
HTML-код
- Опубликовано: 1 апр 2018
- This video provides an overview of options available through SPSS in carrying out multinomial and ordinal logistic regression. It covers assessment of global and local model fit as well.
Side note: The very last demonstration illustrates the use of binary logistic regression with a dependent variable with ordered categories (although technically presence of children - coded 0=no, 1=yes - is not ordinal). It is worth mentioning that multinomial logistic regression could have been used, and that with 2 categories on the DV all three procedures would yield equivalent results.
You can download a copy of the data here if you want to carry out the procedures as you follow along: drive.google.com/open?id=1Vkb...
Please be check out my newest video (August 2021) on ordinal logistic regression here: • Ordinal logistic regre...
For more instructional videos and other materials on various statistics topics, be sure to my webpages at the links below:
Introductory statistics:
sites.google.com/view/statist...
Multivariate statistics:
sites.google.com/view/statist...
Wonderful job of explaining the likelihood ratio tests; will steal (with attribution ;) ) for my next multivariate course!
DEAR VIEWERS: I HAVE CREATED TWO NEW VIDEOS (JULY, 2019) ON MULTINOMIAL (ruclips.net/video/1BL5cL8_Cyc/видео.html) LOGISTIC REGRESSION AND ORDINAL (ruclips.net/video/rSCdwZD1DuM/видео.html) LOGISTIC REGRESSION. THESE VIDEOS INCLUDE LINKS UNDERNEATH THE VIDEO DESCRIPTIONS TO A POWERPOINT PRESENTATION (WITH ADDED DETAILS) AND THE RAW DATA. PLEASE CHECK THESE OUT!
nice video and can help those interested researchers in MLG model.
Hi everyone, please be check out my newest video (August 2021) on ordinal logistic regression here: ruclips.net/video/CdOHB3U5YHk/видео.html
Oh, how nice you have race in ordinal ranked measure. :) Just joking. Thx for your vid!!! Helped a lot.
This was helpful
Thank you so much Dr this is very useful . I am wondering that how can I check if my model has outliers or not.
Hello Mike, I have a question. Is there a precedence of taking a random sample of one category of the dependent variable so as to have similar proportions to the second for a three category dependent variable. My category proportions are 0.77, 0.20, and 0.027. Is there any other way to model the three category dependent variable for these proportions.
what if your p-value for pearson is 0,000 but for deviance it is 1,000? is this a problem? cannot find anything on the internet on this topic. thx
Thank you for your explanation. I have 2 nominal and 2 ordinal variables as independent variables (4), and 2 ordinal likert-scale dependent variable. In this case should i use ordinal logistic regression or nominal logistic regression? Your reply will help me a lot. Thank you in advance
Good source, is there a possibility to set control varaibles
Thank you for the video, I was curious if you had any insight on estimated respons probability using an ordinal logistic regression.
Data Analysis
Dear Mike,
May I ask a question. I have a model like this: Y (X1, X2, Gender, Age, Position)
Y: dependent variable based on 7-point Likert scale,
X1: continuous variable; X2: based on 7-point Likert scale; Age and Position: category variable; Gender: dummy variable. So which regression model is appropriate? Thanks for your help.
hi , i've run the multinominal test, I found a statistical significant relation between the reduced model and the full model.But none of the parameter values (in the parameter estimates) is significant. can you tell me what does it mean, please? thanks.
Is it possible to run a hierarchical multinomial logistic regression? If there no Blocks to add variables, do I run stepwise using SPSS?
GOOD DAY SIR. I have a little knowledge in MNL. My study is about transportation mode choice. Can i have a favor sir?
Thank you, professor. After I run the ordinal logistic regression, SPSS told me that there are 310 (73.1%) cells (i.e., dependent variable levels by observed combinations of predictor variable values) with zero frequencies. Is this variable still qualify to be a dependent variable?
I also fine the same problem. have u solved them? and how?
@@polomarco1256 I think that's because of the scale/metric predictors. Did you find out if that's a problem for the analysis?
Thank you for your video. I learned a lot. I ran the model with the categorical variables along with the continuous variables in the covariate box and got a different result when I ran the categorical variables in the factor box. Do you know why the difference in my models?
Hi there. When you run the categorical variable through the factor box, the program uses a default dummy coding system where the last group on your categorical variable is treated as the reference/baseline category and all groups are then compared against it. It's kind of annoying because you don't have any control over it. But there's nothing technically wrong with doing it that way. I just prefer to do my own dummy coding. Just as an FYI, I have a few newer videos on these procedures:
ruclips.net/video/6S_878RheL8/видео.html
ruclips.net/video/1BL5cL8_Cyc/видео.html
ruclips.net/video/rSCdwZD1DuM/видео.html
I hope you visit them. Cheers!
Got it. Your knowledge helped me impress my professor and get an A. Thank you very much!
Thank you Sir for this video! I have a question I would be greatful if you could answer. I have a specific IV and when I do the OLRA with different combinations of independent variables always including this specific IV, each time the significance comes different for this specific IV. What may be the reason? I dont know if I could explain my question but...
I thought the other independent variables were constant when interpreting one. Was I wrong?
For exmp if you didnt add the Locus of control variable and only added Age and Years of schooling, the significance of Age would be different. Are the significance of independent variables dependent to each other when calculating Log Reg???
what if i also have a categorical factor variable (binary) ? How do i interpret it then. Please let me know
Do you know why someone might get a warning saying that unexpected singularities are encountered in the hessian matrix?
Hi there. This should help with answering your question: www-01.ibm.com/support/docview.wss?uid=swg21480408
May I ask the multinomial model, there is a choice between factor and covariance. What is the difference between them? some people say if an IV is categorical we should put it into the factor, and the numberic IV I should put it into the covariance. Is this understanding right? cos u mention dummy variable when you expalin this in the video. I am a bit confused. Thanks for your answer
Hi there. A dummy variable is a variable with only two levels (for instance: coded 0 and 1). If you include dummy variables as predictors, then you can enter them straight as covariates (but you can't add categorical variables with more than 2 levels in the covariates box). If you use the factors option, then the program will automatically do the dummy coding for you. I tend to prefer to do my own dummy coding rather than the program doing it for me, because the program selects the category on the original variable with the highest value as the reference group by default. But it's really more of a matter of preference. Point is that when dummy coded variables are used, you can include them as covariates in the model. If your factor is not converted to dummy variables (and you are ok with the program's coding system), then put your original variable in the factors box. For more info on dummy coding, you can go here (ruclips.net/video/yGU8fWZOPjs/видео.html). This video covers dummy coding in the context of OLS regression, but the same principles apply.
@@mikecrowson2462 Thank you so much, help me a lot!
thank you for the useful information! but do you know how can we run an ordinal dependent variable with ordinal independent variable with spss?
The approach depends on how you wish to treat the ordinal variable. If there are enough values (such as with Likert-type items; e.g., 1-5 or 1-7), you might be able to treat the predictor as continuous. If you are working with very few categories on the predictor, then you probably want to treat the variable as a factor. In SPSS, you have a factor box when you run the analysis. When you put the variable in that box, SPSS will create a set of dummy variables for you (I believe it will treat the last category on your ordinal variable as the reference category by default). On the other hand, you can create the dummy variables yourself (which is typically what I prefer) and enter them as predictors in the model. I have a video on dummy coding (in the context of OLS regression, but it works the same with ordinal and multinomial LR) here: ruclips.net/video/XGlbGaOsV9U/видео.html
By the way, I have a much newer video on ordinal logistic regression at: ruclips.net/video/rSCdwZD1DuM/видео.html
@@mikecrowson2462 thank you so much for ur time answering this!
@@mikecrowson2462 Kindly do a video
Thanks a lot for the video. In your example of ordinal logistic regression explanation for increase in age means more likely to be in the given 1 and 2 thresholds, so can you please explain how to interpret for say gender (i.e. how to identify being a male or female is more likely or less likely to be in the given threshold, as the location generalise gender and not male or female separately) I hope I'm making sense...need this analysis for my research.
Hi, when running ordinal logistic regression you aren't predicting thresholds themselves, but rather the likelihood of a case falling at a given level on your ordered variable or below a category versus falling into a higher category (see a nice discussion here: onlinecourses.science.psu.edu/stat504/node/176/). The thresholds are not really your focus during interpretation, but rather the relationship between your IV's and the likelihood of a case falling into a higher versus lower category on your DV. So try to avoid getting locked into thinking about the thresholds when interpreting your results. The regression coefficient for any predictor is interpreted as the predicted change in log odds of a case falling at a higher versus lower level on your ordered categorical variable. If it helps, think of binary logistic regression with gender as a categorical predictor. With gender as a predictor, a b=.10 would indicate that the difference in the log odds of a case falling into the target group relative to your reference category (with the positive coefficient indicating greater log odds of falling into the target group). With an ordinal DV with more two levels, this coefficient would be interpreted the predicted difference in log odds of a case falling into a lower versus higher category on the DV. If you had a DV with three levels, there would be two thresholds (demarcating transitions between (a) the lowest category and the two higher categories and (b) the middle and lower categories and the highest category. Nevertheless, the relationship between the predictor and the DV is the same, irrespective of where transitions occur on the DV. Hope this helps.
Thanks a lot for taking out your precious time in answering and explaining my query clearly and in a very detailed manner. If you don't mind, my other problem is that in my multinomial logistic regression analysis both 'Model fitting information' and 'Goodness of fit' are showing statistically significant (0.05) so as to accept the null hypothesis...
How can we run multinomial regression in spss using complex sample option?
After running several times of missing imputation, all by variables becomes non-significant, how do you think I can justify this, do you think the differences between each variable are still valuable? Thnx.
Hi. Are you saying that before you used imputation you had significant predictors, but afterwards they were non-significant? I wasn't clear on what you are asking about. By the way, if this is what happened, then the difference in results could be due to the type of imputation used (e.g., mean imputation versus regression imputation) and also possibly the pattern of missingness (i.e., if your data on key variables were not missing at random).
I'm wondering why you didn't address the warning at the top.
Anyone, can you tell the difference between ML and Ordered reg in terms of the independent variable?
Hi there, multinomial logistic regression and ordinal logistic regression refer to two types of logistic regression that differ based on the characteristics of the dependent variable (not the independent variables in the model). You can include both continuous independent variables, as well as IV's that are categorical (either as factors, or dummy coded) in either of these models. FYI, I have more recent videos on multinomial logistic regression (ruclips.net/video/1BL5cL8_Cyc/видео.html) and ordinal logistic regression (ruclips.net/video/rSCdwZD1DuM/видео.html) you might check out. cheers!
thanks for the detailed video!
I noticed when addressing the test of parallel lines for ordinal regr. you contradict yourself, saying twice "if this is non significant that would indicate we're meeting that assumption & that's a good thing" & "if this is not statistically significant then that would actually be an indicator that the relationships are not holding across ranks"...
which one is correct? I presume the 2nd should be: " *IS significant" ??
Hi there. It looks like I mispoke there. If the parallel lines test is significant, then you are violating the proportional odds assumption. If it's not significant, then the assumption is met. I have a newer video on ordinal logistic regression here (ruclips.net/video/rSCdwZD1DuM/видео.html) that you might check out. It also contains a Powerpoint you can download where I go into more detail on interpretation of results. Cheers!
@@mikecrowson2462 the link can not be opened.
@@polomarco1256 Hi there. It should be fixed now. But here it is again (not sure what happened before): ruclips.net/video/rSCdwZD1DuM/видео.html
@@mikecrowson2462 thank you very much. ur so kind that u share these all knowledge. :)
I run the multinomial logistic regression. This message appeared "Floating point overflow occurred while computing this statistic. Its value is therefore set to system missing." How should I deal with this?
the data on last presentation is binary right not ordinal?
btw, do u have ppt on this just like your other videos? thanks
Hi there, yes, the very last presentation is using binary logistic regression and not ordinal logistic regression. Also, unfortunately I had not begun to make powerpoints when I made this video. However, I do have newer videos that do have accompanying powerpoints related to logistic regression in SPSS. See below. Cheers!
ruclips.net/video/vab9NezxpBc/видео.html
ruclips.net/video/CdOHB3U5YHk/видео.html
ruclips.net/video/6S_878RheL8/видео.html
ruclips.net/video/rSCdwZD1DuM/видео.html
ruclips.net/video/1BL5cL8_Cyc/видео.html
Please help!- does the stats work for my question? stats.stackexchange.com/questions/339581/does-this-multinomial-regression-stat-results-answer-my-question
I took a look. I don't know the research design, so my feedback is based on me guessing how you set up the study. I'm assuming you assigned people to either a "far" versus a "near" task, and determined hand preference? The way you have it laid out, your dependent variable is hand preference (with none as the reference category), and your predictor is whether a person received a far versus near task (I'm assuming this was a between subjects factor). You treated "distance" as a factor, so SPSS is treating the group coded 2 (the far group) your reference category. If your question is whether a task that involves varying distances (far versus near) impacts the likelihood of a person choosing a particular hand when responding to the task, then the model looks like it's laid out ok. Of course, the downside is that the likelihood ratio test is nonsignificant and the dummy variable "distance" does not predict likelihood of using the left hand (as opposed to the no handedness group; b=-.288, p=.730) nor does it predict the likelihood of choosing the right hand (as opposed to the no handedness group; b=-.405, p=654). Your regression coefficients were negative and your odds ratios are < 1, which is consistent with a tendency towards no preference over left or right handedness. However, again the differences were not significant. Wondering what your sample size is and its contribution to low power. At any rate, those are my thoughts.
thank you! my sample size is very small (30) and like you said I am trying to determine whether varying distances (far versus near) impacts the likelihood of a person choosing a particular hand, however, all the subjects undertook both tasks. In fact, the question I need to answer is if there is a significant difference in hand preference when doing a distance task vs a near task. I was hoping multinomial regression would be the ideal test however I am unsure
So it sounds like you have a repeated measures design then since folks are exposed to both treatments (i.e., near versus far tasks). Thus, the standard multinomial logistic regression is not the way to go. Also, I wasn't clear if you were comparing use of dominant hand - which often, but not always, is right hand - versus non-dominant hand; or if you are literally comparing use of right, versus, left, versus nonpreference. If you are doing the former your life will be a heck of a lot easier because you can basically do a mixed effects binary logistic regression in SPSS by going through Mixed Effects Generalized Linear Models. If you are doing the latter, it also looks like there may be a mechanism to do a mixed effects multinomial logistic regression. But I am not exactly clear what a repeated measures version of this would look like in SPSS. You might check out this response: www.researchgate.net/post/Can_I_perform_a_Multinomial_logistic_regression_for_repeated_measures_design_in_SPSS. Hope this helps. Good luck with your work.
Maybe an Option B: Use Stata and run mutinomial regression through that program and request Clustered Robust standard errors (with person as a clustering variable). I could see this as a possibility
i need clarification concerning on the value in spss data.
What does the value mean 4.478E+146(shown in upper 95%CI)? And
How and which value need to report on the table for in this case?
Really difficult to listen to with all the racket....
Hello, thanks for your feedback. This is really an older video. Everything done before May 2019 was done with different equipment. However, I have newer and - in my opinion - better quality videos that also include additional links to data and powerpoints. Give them a try. I'm sure you'll be happier with them.
Ordinal logistic regression (2019 video): ruclips.net/video/rSCdwZD1DuM/видео.html
Multinomial logistic regression (2019 video): ruclips.net/video/1BL5cL8_Cyc/видео.html ) These are the same links provided in the comment above DEAR VIEWERS.
cheers.
Very very sorry for mee.....I not undestang english
Why doesn't any of these people explain how to do interaction (like in anova) in ordinal regression? Has everyone gone mad? A massive dislike.