This was an excellent presentation of the topic, Joel. Your video was exactly what I needed to improve my understanding of how to model a categorical variable with three categories in AMOS. Thank you!
Dear Dr., how can we add multi-categorical variables (i.e. education level) in a full latent SEM model as control variable; If we have 8 categories, do we need to make 8 dichotomous - dummy variables ;
I appreciate your thoughtful answer to my past question. I have further questions. Please tell me why the results of categorical independent variables are interpreted based on unstandardized coefficients, not standardized coefficients. Moreover, can standardized coefficients be used to compare the effect sizes of the categorical independent variables?
The reason why you want to use the unstandardized values is because of the dummy coding. It is more appropriate to use unstandardized values when you dummy code variables in your model.
Hello Dr. Collier. Thank you for the video! I have a question regarding the correlation between the dummy-coded predictors. I am using a pretty similar design in my study and tested a mediation analysis with SEM in R. The graph of the SEM showed me a standardized correlation of -0.5 between the two dummy-coded predictors, which makes sense if you calculate correlations between 1 0 1 0... and 0 1 0 1... As I looked at your model, I could see the same correlation of -0.4999. Does that influence the results of the mediation analysis in any way, or is it just a relic and can be ignored?
Hello Dr. Collier, thanks for your work! Your videos are very helpful! I have one question: One goal of my study is to see how skepticism moderates the effect of different advertisements (3 categories) on purchase intention. Do i just multiply the dummy variables for 2 categories (category 1 and 2) with the mean centered variable skepticism and include these to my model? Then, the model would include the two direct effects of both categories (included as dummy variables) on purchase intention which can be compared to the reference group (category 3). This should be fine, I guess. Furthermore, there would be 2 interaction effects regarding to the moderating effect of skepticism on the effect of category 1 and 2 on purchase intention (dummy category1 X skepticism mean centered; dummy category2 x skepticism mean centered) If the interaction effects with skepticism are both negative significant, is it possible to interpret the following: "For high skepticism the effects of category 1 and 2 on purchase intention are weaker compared to category 3 (reference categroy)"? I'm not quite sure whether this approach is correct... I would appreciate a comment on this. Thank very much, especially for your great videos. Best wishes, Marcel
Marcel you are correct in how to set up this moderation test. If your results are negative from the interaction test, then you will interpret each dummy coded relationship. Skepticism would weaken the relationship of the dummy coded category to PI in comparison to the reference group. You would need to interpret this for both dummy coded variables. What you wrote was essentially correct. Hope this helps.
So great Dr.Collier. Actually, I have a close challenge here. In my model, I have an independent variable (including 19 questions!) which is comprised of 3 types of categorical solely (1_ yes/no, 2&3_ two groups of different multiple choices). Is it possible to use your approach to do the CFA test in AMOS?
No, you can't perform a CFA of an independent variable with multiple categories. A CFA is trying to assess if you are capturing the concept with your measures. Your multiple categories would present a fractured overall construct. You don't typically see a CFA done with a categorical IV because there is usually no degree of variation. It is a yes or no in the category. Hope this helps.
Thank you for creating this incredibly helpful video. I conducted a 2x2 experiment, so I created one dummy IV with 4 conditions. A reviewer said this was incorrect, and that I need to use each of the two IVs separately in the model. When I run the model that way, the results are nearly the same (just a slight difference in the estimates). What are your thoughts on the best way to approach a model with 4 conditions?
I agree with the reviewer. The best way to address your model is to use two separate IVs in the model rather than cramming them all into one IV. Hope this helps.
@@joelcollier9387 Very helpful indeed! I appreciate the quick response and I've already incorporated the new model into my revised manuscript. Thank you again.
I am really thankful to you for the video and your book about SEM. I am referring to your book (Collier, 2020) and testing categorical moderated mediation (2*3 factorial design). Is it possible to answer the hypotheses based on the output of insufficient model fit? Or the model fit must be sufficient when hypotheses are answered?
If you have poor model fit, it most likely means you have a measurement issue with one (or more) of your unobservable constructs. If the measures are in doubt then it also puts doubt into the validity of the findings via structural results. Try to clean up your measurement issues...that should help things.
Dear Dr, I was wondering, if I only have binary predictors, is it compulsory to create composite variables for other constructs in the model? I didn't find answer in your book。
If your independent variable is binary there is no need to summate the other constructs in your model. It is better if you do not summate the constructs because you can assess measurement error. Saying that, the independent variable will act like a summated variable because it will only have values of 1 or 0. In essence, it will be treated as an observable variable. Hope that helps.
Hello Dr. Collier! Thank you for this video! It is a godsend. I do have a question though - In this video, you have demonstrated how to use dummy variables in AMOS for nominal variable with 3 categories. But what if I have ordinal variable with 3 categories that I want to set up as a control variable? In that case, do I still need to introduce 2 dummy variables or I can bring that ordinal variable with 3 categories as it is? Thank you!
If the variable is ordinal, then, no, you do not need to set up dummy variables. You will bring it into your model like any other regular variable. Might want to make sure you look at standardized estimates because the unstandardized might be hard to interpret with varying scales.
@@joelcollier9387 Hello Professor, sorry for the persistent troubling. Just to be sure, if I have gender as a nominal variable (male, female), age as an ordinal variable (1=20 years or below, 2=21-30 years, 2=31-40 years, 4=41 years and above) and income as another ordinal variable (1=50000 or below, 2=50001-100000, 3= 100001 or above) that I want to bring as control variables. I need to dummy code gender only and bring one dummy coded variable in the analysis, while age and income I can bring as regular variables (without dummy coding). And for analysis, I need to interpret the sign of gender, while for age and income, I need to look at standardized regression weights as suggested by you earlier. Kindly let me know if this is the right way to proceed with the analysis. Your comments will be highly valuable. Also, I would be highly grateful if you could share the reference for this as I have to justify the analysis in my thesis with support from literature. Thank you in anticipation!
I'm surprised that you seem to be using MLE for the analysis. Wouldn't you have to use a different estimator considering a multicategorical predictor is included in the model? Thanks in advance for your answer.
There is plenty of justification in the Literature for using Maximum Likelihood to assess categorical variables. You can also use GLM but it can cause model fit issues at times.
@@joelcollier9387 Hi, thank you for creating such a useful video. Would you mind sharing a reference about "using Maximum Likelihood to assess categorical variables"? I'd like to learn more before running analysis.
@@jingyiwei296 If you are looking for a simple reference on why to use Maximum Likelihood to assess categorical variables, you can use Andrew Haye's Book: Introduction to Mediation, Moderation, and Conditional Process Analysis, 2022 Guilford Press
Dear Dr, I want to ask, for dummy variable is it necessary to do normality test? How about normality test if I combine dummy variable and non dummy variable? thank you
No need for normality tests with dummy coded variables. Those dummy codes are usually "1" and a "0". No need to test normality with a bimodal response.
@@joelcollier9387 I've been looking for an answer for this question for weeks, and it's such a relief I finally found one. Thank you so much! However, I have one more question to ask. As far as I'm aware, if you check "assessment of normailty" in AMOS, you get a multivariate kurtosis in order to test multivariate normality. The questions is, can I rely on this multivariate kurtosis to test multivariate normality, even when dummy variables are included in the model? I'd really appreciate your answer!
@@user-oy4bd3mf3u You can not really assess normality with dummy coded variables or even categorical variables.. Any test of normality will be with continuous variables in AMOS. With dummy coded variables...it is just 1 and 0. There is no normality assessment in a bimodal response. Hope that helps.
@@joelcollier9387 Thank you for the reply! But I'm a little bit confused here... Given that the path analysis requires the assumption of multivariate normality to be met, if it is not possible to test multivariate normality when dummy variables are included, is it even possible to use dummy variables in a path analysis? Sorry to bother you again Dr. Collier!
@@user-oy4bd3mf3u There is an assumption of normality for all continuous variables in the model. All dummy coded variables are not included in the test of normality. Dummy coded variables will have a high amount of kurtosis because it is a "1" or "0". The assumption of normality must take place with the continuous variables. That is why there can only be continuous variables as dependent variables in AMOS. Hope that helps.
This was an excellent presentation of the topic, Joel. Your video was exactly what I needed to improve my understanding of how to model a categorical variable with three categories in AMOS. Thank you!
And I just bought your book, too. Great idea to plug it at the end. Content marketing in action!
Thanks Todd. I hope you are well.
Excellent video
Dear Dr., how can we add multi-categorical variables (i.e. education level) in a full latent SEM model as control variable; If we have 8 categories, do we need to make 8 dichotomous - dummy variables ;
I appreciate your thoughtful answer to my past question. I have further questions. Please tell me why the results of categorical independent variables are interpreted based on unstandardized coefficients, not standardized coefficients. Moreover, can standardized coefficients be used to compare the effect sizes of the categorical independent variables?
The reason why you want to use the unstandardized values is because of the dummy coding. It is more appropriate to use unstandardized values when you dummy code variables in your model.
@@joelcollier9387 I understand. Thank you for getting back to me.
Hello Dr. Collier.
Thank you for the video!
I have a question regarding the correlation between the dummy-coded predictors. I am using a pretty similar design in my study and tested a mediation analysis with SEM in R. The graph of the SEM showed me a standardized correlation of -0.5 between the two dummy-coded predictors, which makes sense if you calculate correlations between 1 0 1 0... and 0 1 0 1... As I looked at your model, I could see the same correlation of -0.4999. Does that influence the results of the mediation analysis in any way, or is it just a relic and can be ignored?
Hello Dr. Collier, thanks for your work! Your videos are very helpful!
I have one question:
One goal of my study is to see how skepticism moderates the effect of different advertisements (3 categories) on purchase intention.
Do i just multiply the dummy variables for 2 categories (category 1 and 2) with the mean centered variable skepticism and include these to my model?
Then, the model would include the two direct effects of both categories (included as dummy variables) on purchase intention which can be compared to the reference group (category 3). This should be fine, I guess.
Furthermore, there would be 2 interaction effects regarding to the moderating effect of skepticism on the effect of category 1 and 2 on purchase intention (dummy category1 X skepticism mean centered; dummy category2 x skepticism mean centered)
If the interaction effects with skepticism are both negative significant, is it possible to interpret the following:
"For high skepticism the effects of category 1 and 2 on purchase intention are weaker compared to category 3 (reference categroy)"?
I'm not quite sure whether this approach is correct...
I would appreciate a comment on this. Thank very much, especially for your great videos.
Best wishes, Marcel
Marcel you are correct in how to set up this moderation test. If your results are negative from the interaction test, then you will interpret each dummy coded relationship. Skepticism would weaken the relationship of the dummy coded category to PI in comparison to the reference group. You would need to interpret this for both dummy coded variables. What you wrote was essentially correct. Hope this helps.
So great Dr.Collier. Actually, I have a close challenge here. In my model, I have an independent variable (including 19 questions!) which is comprised of 3 types of categorical solely (1_ yes/no, 2&3_ two groups of different multiple choices). Is it possible to use your approach to do the CFA test in AMOS?
No, you can't perform a CFA of an independent variable with multiple categories. A CFA is trying to assess if you are capturing the concept with your measures. Your multiple categories would present a fractured overall construct. You don't typically see a CFA done with a categorical IV because there is usually no degree of variation. It is a yes or no in the category. Hope this helps.
Thank you for creating this incredibly helpful video. I conducted a 2x2 experiment, so I created one dummy IV with 4 conditions. A reviewer said this was incorrect, and that I need to use each of the two IVs separately in the model. When I run the model that way, the results are nearly the same (just a slight difference in the estimates). What are your thoughts on the best way to approach a model with 4 conditions?
I agree with the reviewer. The best way to address your model is to use two separate IVs in the model rather than cramming them all into one IV. Hope this helps.
@@joelcollier9387 Very helpful indeed! I appreciate the quick response and I've already incorporated the new model into my revised manuscript. Thank you again.
I am really thankful to you for the video and your book about SEM. I am referring to your book (Collier, 2020) and testing categorical moderated mediation (2*3 factorial design). Is it possible to answer the hypotheses based on the output of insufficient model fit? Or the model fit must be sufficient when hypotheses are answered?
If you have poor model fit, it most likely means you have a measurement issue with one (or more) of your unobservable constructs. If the measures are in doubt then it also puts doubt into the validity of the findings via structural results. Try to clean up your measurement issues...that should help things.
@@joelcollier9387 OK. Then I will clean up my model. Thank you for your comment!
Dear Dr, I was wondering, if I only have binary predictors, is it compulsory to create composite variables for other constructs in the model?
I didn't find answer in your book。
If your independent variable is binary there is no need to summate the other constructs in your model. It is better if you do not summate the constructs because you can assess measurement error. Saying that, the independent variable will act like a summated variable because it will only have values of 1 or 0. In essence, it will be treated as an observable variable. Hope that helps.
Hello Dr. Collier!
Thank you for this video! It is a godsend. I do have a question though - In this video, you have demonstrated how to use dummy variables in AMOS for nominal variable with 3 categories. But what if I have ordinal variable with 3 categories that I want to set up as a control variable? In that case, do I still need to introduce 2 dummy variables or I can bring that ordinal variable with 3 categories as it is? Thank you!
If the variable is ordinal, then, no, you do not need to set up dummy variables. You will bring it into your model like any other regular variable. Might want to make sure you look at standardized estimates because the unstandardized might be hard to interpret with varying scales.
@@joelcollier9387 Thank you for your guidance, sir. You are a saviour.
@@joelcollier9387 Hello Professor, sorry for the persistent troubling.
Just to be sure, if I have gender as a nominal variable (male, female), age as an ordinal variable (1=20 years or below, 2=21-30 years, 2=31-40 years, 4=41 years and above) and income as another ordinal variable (1=50000 or below, 2=50001-100000, 3= 100001 or above) that I want to bring as control variables.
I need to dummy code gender only and bring one dummy coded variable in the analysis, while age and income I can bring as regular variables (without dummy coding). And for analysis, I need to interpret the sign of gender, while for age and income, I need to look at standardized regression weights as suggested by you earlier.
Kindly let me know if this is the right way to proceed with the analysis. Your comments will be highly valuable.
Also, I would be highly grateful if you could share the reference for this as I have to justify the analysis in my thesis with support from literature. Thank you in anticipation!
@@tusharprabhakar9653 Hey Tushar, I want to know how did you perform this analysis, as i have same control variable: age and gender.
@@tusharprabhakar9653 Hey Tushar, will be of great help if you can please let me know if that was the right approach.
Thanks in advance for the help.
I'm surprised that you seem to be using MLE for the analysis. Wouldn't you have to use a different estimator considering a multicategorical predictor is included in the model?
Thanks in advance for your answer.
There is plenty of justification in the Literature for using Maximum Likelihood to assess categorical variables. You can also use GLM but it can cause model fit issues at times.
@@joelcollier9387 Awesome, thank you. I will use MLE then, as well.
@@joelcollier9387 Hi, thank you for creating such a useful video. Would you mind sharing a reference about "using Maximum Likelihood to assess categorical variables"? I'd like to learn more before running analysis.
@@jingyiwei296 If you are looking for a simple reference on why to use Maximum Likelihood to assess categorical variables, you can use Andrew Haye's Book: Introduction to Mediation, Moderation, and Conditional Process Analysis, 2022 Guilford Press
@@joelcollier9387 Hi Dr. Collier, thank you very much for your response! I'm checking the book!
Dear Dr, I want to ask, for dummy variable is it necessary to do normality test? How about normality test if I combine dummy variable and non dummy variable? thank you
No need for normality tests with dummy coded variables. Those dummy codes are usually "1" and a "0". No need to test normality with a bimodal response.
@@joelcollier9387 I've been looking for an answer for this question for weeks, and it's such a relief I finally found one. Thank you so much!
However, I have one more question to ask. As far as I'm aware, if you check "assessment of normailty" in AMOS, you get a multivariate kurtosis in order to test multivariate normality. The questions is, can I rely on this multivariate kurtosis to test multivariate normality, even when dummy variables are included in the model?
I'd really appreciate your answer!
@@user-oy4bd3mf3u You can not really assess normality with dummy coded variables or even categorical variables.. Any test of normality will be with continuous variables in AMOS. With dummy coded variables...it is just 1 and 0. There is no normality assessment in a bimodal response. Hope that helps.
@@joelcollier9387 Thank you for the reply! But I'm a little bit confused here...
Given that the path analysis requires the assumption of multivariate normality to be met, if it is not possible to test multivariate normality when dummy variables are included, is it even possible to use dummy variables in a path analysis?
Sorry to bother you again Dr. Collier!
@@user-oy4bd3mf3u There is an assumption of normality for all continuous variables in the model. All dummy coded variables are not included in the test of normality. Dummy coded variables will have a high amount of kurtosis because it is a "1" or "0". The assumption of normality must take place with the continuous variables. That is why there can only be continuous variables as dependent variables in AMOS. Hope that helps.