Multinomial logistic regression using SPSS: Covid-19 threat perceptions from Pew data
HTML-код
- Опубликовано: 21 авг 2024
- This video demonstrates the use of SPSS to perform multinomial logistic regression using data from the Pew Research Center. The model being tested is aimed at predicting perceived threat to personal safety as a result of the Covid-19 virus.
You can download a copy of the SPSS data file here:
drive.google.c...
You can obtain a copy of the Powerpoint referenced in the video here:
drive.google.c...
Newer video (2023) on multinomial logistic regression: • Modeling multicategori...
Please see my NEW 2023 VIDEO on multinomial regression with multicategorical outcome variable: ruclips.net/video/UIZHNTSpED8/видео.html .
FOR ANOTHER DEMONSTRATION OF MULTINOMIAL LOGISTIC REGRESSION, SEE MY PREVIOUS VIDEO: ruclips.net/video/1BL5cL8_Cyc/видео.html
Hello Mike, How would you treat "age"? Age is continues but most of the time they are integer in data ...would you consider age discrete dependent variable in multinomial LR if let say 5 categories (e.g., 31-35).
P.s. I LOVE your videos!
The dataset file above does not include the dummy variables, so thank you for the opportunity to create dummy variables as well:)
Hi Murat, totally missed that. Was working with a lot of files. I've uploaded a new file with the dummy variables. But you can still have fun creating them if you like :)
Mike Crowson of course! I already had fun creating my own dummies😀
THANKS TEACHER
Thank you so much. I have seen only a few interpretations of MLR speak about percentages: Subtract 1 from the Exp(B) value, and this is the % likelihood increase/decrease of, for instance, a female thinking Covid is a major threat as opposed to a male thinking covid is a major threat. Do you think interpreting odds-ratios as percentages is 'safe' ? Is it conventional?
Hi there. With the method you are speaking of, you aren't interpreting the odds ratios as percentages. Rather you are interpreting the percentage CHANGE in the ratio of the odds with increments on the predictor variable. I recently put together some new materials on logistic regression that go into these issues in more detail. Here's a video (and there is a powerpoint to download to) on binary logistic regression: ruclips.net/video/vab9NezxpBc/видео.html . The Powerpoint speaks to your question. Also, I put together something last week on communicating about effect size with logistic regression using an Excel spreadsheet I created. You might find it helpful. Go to: ruclips.net/video/W8ktaSKVCL0/видео.html
Best wishes!
Hi. Your explanations are perfect. I have a concern. Why do you convert numerical covariates (i.e., age) in categorical? I have read that is recommended to keep them in continuos scale.
Hello MaWe. Ordinarily I would not encourage recoding numerical covariates into categorical variables. To tell the truth, it's been awhile since I put this video together and can't recall if either (a) the categories were already formed in the dataset itself (it was a public dataset - Pew Research - I believe, so the age variable might have already been coded using the age categories) or (b) I was creating the example and trying to demonstrate how you could include a factor variable in the model. If it was (a), then I had no control over the data that was made available. If it was (b), it would've been for demonstration only and not something I would ordinarily encourage. I hope that makes sense. Best wishes!
@@mikecrowson2462 Perfect. And thanks for your videos. They help me a lot.
Hi Mike,
Why do you not need to create an 18-29 category? Does SPSS automatically assigns it to be the reference? Thank you
Hello sir and thank you for the great presentation.
I understand that when coding dummy variables, we have to create k-1 variables in order to avoid dummy-trap.
One thing I can not understand is when we create dummy variables, how do we understand if our baseline dummy (in this case I'm talking about the variable "age18to29") contributes to the model.
It does not appear in the "parameters estimates" table (or any other table). How do we interpret the value of "age18to29" variable?
May I ask, why is the ratio scale not nominal for femaleid? and why do you include all variables as covariates instead of factors? I'm still confused, thank you.
Thanks Mike, this is a really useful demonstration. I have a question with reporting; With the reference category set to first I have significant overall effects yet when looking at where these lie between the groups none are significant. However when I switch the reference category to last there are then significant differences between the groups? Should I report both?
Hi Jacqueline, the apparent 'differences' between the two outputs stem from differences in which groups are being compared. If the first group was set as a reference group in your first model, then all regression slopes reflect differences between each subsequent group and that first one. If you run your analysis with the last group serving as the reference group, then the remaining test involve comparisons of the remaining groups with that group.
The choice of which group to use as the reference group depends on a couple of things. First and foremost, did you have specific a priori hypotheses you were wishing to test pertaining to group differences? If you did, then you should report on the contrasts pertaining to those hypotheses. In effect, those comparisons are functioning in a manner akin to planned orthogonal contrasts as a follow up to a significant ANOVA test result. And you must use the parameterization that tests your hypotheses. If you used dummy coding to simply to account for group differences as a set, but you didn't have any specific a priori hypotheses in mind about which groups might differ (having a mindset akin to post-hoc tests such as Tukey's following ANOVA), then I suppose you could treat the comparisons from a more exploratory standpoint and report on the results from the parameterization that yielded the more 'interesting' findings. The regression coefficients and tests for the other predictors in the model will remain the same; however, the intercept and slopes of your dummy variables will differ depending on how you coded them. I suppose you could report on both sets of findings, noting that you ran the model via these two different parameterizations; however, to my way of thinking it begs the question, 'Why not report on all possible differences using each group as a reference category?' (not that I would make that an issue in a manuscript review, but it is a logical question to ask).
Well, that's all I have for you. I hope this helps!
@@mikecrowson2462 Thanks for your speedy response. I did have a planned hypotheses which shows that one of the variables is significant in the Likelihood Ratio Test output but in the the parameterization none of the groups differ from the reference group which was what confused me. Just to confirm; I'll be reporting that a variable is statistically significant for the model but there are no significant differences from the reference group?
@@jacquelinefrancis8127 Hi Jacqueline, if you had targeted hypotheses for the group comparisons involving your reference group then you would be saying none of those comparisons resulted in a significant difference between the groups. And you would be reporting any other significant findings pertaining to your other predictors as usual too. Cheers!
@@mikecrowson2462 Many thanks!
love it!
murat bey burada covid threat ve madeup çevirileri nedir acaba bilgi verebilirseniz çok mutlu olurum.
selçuk seyhan Covid threat, virusun ne kadar tehdit olarak algilandigini (tehdit degil, onemsiz tehdit, onemli tehdit), madeup ise, kisilerin izledikleri okuduklari haberlerin uydurma olduguna inanma seviyeleri
@@turkmura81 çok teşekkür ederim hocam gerçekten selamlar.
selçuk seyhan rica ederim. Iyi calismalar.
This is soooo coool!!!
Thanks, Murat! Glad you liked it. This was one of my favorites to put together :) cheers!
Hello, so what is to be interpreted if my significance of a predictor in the likelihood ratio test is non significant but it gives me significant in the parameter estimates. Can I say that it is not significant in the final model but does have a significant impact when comparing the groups?
Thanks. Very informative. Is it necessary to test Mann-Whitney U and Kruskal-Wallis test for independency before perform multinomial logistic regression
hi there. No, there's no requirement for MW or KW before doing multinomial LR. Best wishes.
@@mikecrowson2462 Thanks..
What statistical analysis we may do in this data except Logistic regression.
If i want to find the risk in terms of profession than we use it as a covariate (dummy variables) or factor.
Can you please advise how to write the equation for a multinomial logistic regression?
What is the justification of using multinomial logistic regression instead of multiple liner regression..
The dependent variable is categorical, instead of continuous and will likely violate assumptions of normality and homoscedasticity of residuals, as well as potential non- linearity between IVs and DV. Had the assumption of parallel lines been met, I would have used ordinal logistic regression instead. Cheers.
Sir, please post stepwise ARIMA analysis for covid prediction in SPSS
Hi sir
Thanks a lot
I have a question;
I want to control two variables in this regression; how can do it?
nice you are. I apprciate
Hi may I know if it is necessary to make dummy variables for education level? Isn't it an ordinal variable instead of continuous variable? Thanks.
Hi there. This basically came down to a judgment on my call in the interest of simplifying the analysis. Technically, education level (as in this video) can be thought of as ordered categorical and treated as such. If I'd wanted to treat the variable in that fashion, I would've included it in the model using dummy variable coding. That said, it is recognized that when ordered categorical variables have more than a few categories (generally, > 5 categories), then the variable may 'behave' (for lack of a better way of putting it) as a continuous variable. Adopting that perspective - and given that there were 6 categories for education level in my example - I opted to simplify the analysis a bit by treating education level as continuous. Other folks might reasonably disagree, however, and opt for treating it as ordered categorical and use dummy coding instead. I hope this helps!
@@mikecrowson2462 thank you Sir. Very well explained. Your videos are a lifesaver to even beginners in SPSS statistics. Thanks again.
Hi sir, I couldn't understand why did you use dummy variables?
Hi there. Dummy variable coding (and other types of effects coding) is used for the purposes of including a categorical (or factor) variable in a regression model as a predictor. I have a video on this in the context of OLS regression (ruclips.net/video/XGlbGaOsV9U/видео.html). The concepts behind this type of coding are demonstrated in this video.
@@mikecrowson2462 Thank you for your quick reply. I will try to get the explanation from that video.
Sir
Can you please share the data link
Hi there. Underneath the video description are links to the SPSS data file and a Powerpoint. (but the link to the data is drive.google.com/file/d/14WPIEuXbRW8qI_0du8-hlWQRED0TUUpu/view
Sir
This link is not working
Ok, try the link now.
Yes it works now
Thank you so much
A really wonderful video
Hello, so what is to be interpreted if my significance of a predictor in the likelihood ratio test is non significant but it gives me significant in the parameter estimates. Can I say that it is not significant in the final model but does have a significant impact when comparing the groups?