Here's a fun pet project I've been working on: udreamed.com/. It is a dream analytics app. Here is the RUclips channel where we post a new video almost three times per week: ruclips.net/channel/UCiujxblFduQz8V4xHjMzyzQ Also available on iOS: apps.apple.com/us/app/udreamed/id1054428074 And Android: play.google.com/store/apps/details?id=com.unconsciouscognitioninc.unconsciouscognition&hl=en Check it out! Thanks!
Hi James.. Thanks for this tutorial. I've been reading a couple of AMOS manuals trying to work out how to do multigroup analysis without any progress, but your tutorial has clarified that and enabled me to complete the analyses.
Hi Dr Gaskin. To be able to know whether the relationship is stronger or weaker you mentioned that after constarining the specific paths we are measuring, we can compare the standardized regression weights of males or females. But in this case shouldn't we be looking at the regression weights of the unconstrained model rather than the structural weights model (which have these specifc paths already constrained). If we loo at the standardized regression weights from the estimates tab they match the numbers on the paths of the unconstrained model not those on the structural weights model. thank you.
Dear James! What an amazing series of videos! Thank you. I was wondering about running a mulitigroup (3 groups) interaction SEM . Should I just run the same?
Dear Dr. Gaskin, Thank you for the videos. From the example you explained that the number of population in each group is similar and comparable. However, in my research one group has more than 480 data while the other is only 45 data. Is it still comparable using the method you explained above? How could we justify the comparison? This happens in my research because the other group has significantly higher population than the other by more than 40x. Collecting the similar amount of data would be an uphill task if not impossible. In my research the comparison between the two groups full model comparison resulted on 'Groups are different at the model level. Check path differences' in the statwiki excel file. The path comparison shows two paths are significant. Thanks again
With such a large difference, you will likely run into trouble. The main trouble will be from the small sample size having insufficient power to minimize error. If you did find two paths different, then that is amazing with such a low sample size in one group.
Dear Prof. James, thank you very much for the helpful tutorial. For my master thesis I´m working with AMOS for the first time. To test if a model consisting of six factors with two items each (I know, it´s a bit unusual to have so few items...) differs between two groups, I followed your instructions in the video. As grouping variable I used country (1, 2). When I used the multi group analysis- bottom of AMOS, the default settings differ from the ones in the video: Instead of STRUCTURAL WEIGHTS, I can only choose MEASUREMENT WEIGHTS, structural covariances and measurement residuals. Plus, if I run the analysis I get only results in the "Model Comparison"-Section for Structural covariances and measurements residuals. How comes? Can you possibly help me with that? Thank you in advance for your support.
Hi James, thank you for your informative video. I have just one more question concerning the interpretation: when we want to check which effect is stronger and if the effects are significant (or in my case which one is positive or negativ) we have to click at the "unconstrained" model, right? thank you in advance.
Hi James, thanks a lot for great video series. I have a question about the step after the invariance testing completed. I am trying to do multigroup analysis with two groups. If I got nonsignificant p value for chi square difference between the unconstrained and measurement weights models (or in the second step, between the measurement weights and structural weights model), which model in the output should I look for the estimates? Should I choose measurement invariance and structural weights, respectively. Or do I report the standardized estimates of the unconstrained model? Thank you in advance.
If the p-value for the chi-square difference is non-significant, then you can remove all additional constraints and just report the unconstrained model.
Hello Dr. James Gaskin thanks a lot for the tutos. I've been following you for a long time. I want to know whether I should choose the Measurement Weights or Structural Weights for my multigroup analysis in CFA? Second question: Should this analysis be performed only with a path model or it's OK with a structural moedl? Third question: In case I have both significant and insignificant p-values, which ones should I assess for the MI and how to do it? Thanks in advance.
Dear James, could you please help me? I've got two groups (n=208,n=162) and when their standardized regression weights are rather different (ex: .20 versus .35 with p
Hi James, As far as I get, there are two types of invariances. One is measurement invariance (as demonstrated in your other video "Measurement Model Invariance") and the other is structural invariance. I assume what you did here is checking out structural-ish invariance (assuming that there is no latent variables in this tutorial). Am I correct? My question is what if I get a significant chi-square difference in terms of gender in structural measurement but non-significant difference in measurement model? What does it mean and do I have to report both? (I run into some studies where they just reported structural invariance). Thank you.
You are correct. And, your scenario is exactly what you want. You want measurement invariance and structural non-invariance (i.e., differences at the structural level, but not the measurement level). This would mean that your measures mean the same thing across groups, but that the relationships between constructs differ.
@@Gaskination Thank you for the reply. I'm a bit confused by the article below: www.ncbi.nlm.nih.gov/pmc/articles/PMC4900100/. They depict structural invariance as a nice thing (suitable, global, universal, cross-cultural and etc.) as opposed to your abovementioned reply. My other question (just popped into my mind right now): What if I am not developing some kind of scale? Do I still have to conduct measurement invariance calculations? Thank you so much! You are such a lifesaver!
@@berkandemir4220 They have titled their study incorrectly. I looked through the study and they are only examining measurement invariance. This should have been caught in the review process, but probably no one on the review team was an SEM expert.
Hello James, Thanks so much for your wonderful videos. I have a question regarding testing hypotheses. My research paper includes 3 samples. I tested measurement invariance and its result indicates that the groups are partially invariant. Then, I tested the latent model of each sample of them separately and confirmed the model fit of each model of them. After that, I did multigroup analysis of the three datasets together and its result indicates that the groups are different at the model level. (Note: Here is my hypotheses: H1: PE positively influences BI. H2: PE positively influences PU. H3: PU positively influences BI. H4: T positively influences BI. H5: R positively influences BI. H6: SI positively influences BI.) So, my questions are: 1- I used the same steps that you have used in this video to do multi-group analysis among 3 groups, is this right? 2- which kind of results I should use it to test my hypotheses, the results of paths coefficient of the latent model of each sample separately or the results of paths coefficient of the multi-group analysis? 3- when I did multi group analysis, the chi-square difference test was significant at the whole model level, but when I did the chi-square difference test for each path separately, the result was non significant for most paths except of 2 of them, does this make sense? 4- In multigroup analysis, the result of chi-square difference test for one of my paths was significant which means that this path is not different among the three groups. whereas, this path was significant in 2 models of my latent models and non significant in the third one which means that this path is different among the 3 groups. so, I am very confused about these results, could you explain that what does it mean? Thanks in advance.
1. Yes 2. Report the results based on the hypotheses. So, you might have to run multiple models if you have different types of hypotheses. 3. Yes. It means the difference at the model-level is originating from those two paths. 4. Even when the difference is not significant, if the estimates are very different, or significant vs non-significant, this is also a form of moderation. For example, in your case, you can say that X is a good predictor for Y for two of the groups, but not for the other group. So, there is moderation, despite the lack of statistically significant difference.
Thanks Sir for your reply. I have two more questions: 1- how can I do t-statistic after using multi group analysis to make comparisons of path coefficients between the three groups? (Note: I have download your excel Stat-tools but it just calculate t-statistic between 2 samples not 3, how can I use this tool to make comparisons between 3 samples?) 2- which one of tables I should use it to write the path coefficient in t-statistic formula, regression weights or standardized regression weights?
1. Group comparisons must be done in pairs. So, you can do this as A:B, A:C, B:C or you can do it as A:BC, B:AC, C:AB. 2. The formulas always use unstandardized coefficients.
Many thanks Sir. 1-I want to know what is the better method to test multi-group analysis and to examine the significant differences of each path across my groups? is it multi-group analysis by using t-statistic or multi-group analysis by using chi-square difference test? (Note: I have 3 groups). 2- when I did chi-square differences test for each path. I found Δ ꭓ2 of one of my paths was insignificant, indicting that there is no difference of this path across the groups, while the path coefficients unstandardized estimates differs across groups (0.586, 0.371, 0.457), what I should report about that? 3- Also, there is another path that was insignificant in each model and its Δ ꭓ2 was insignificant too across groups, does it mean that this path is invariant across groups and there is no moderation effect for this path or not (because it was already insignificant in each model)? what I should report about that?
1. The chi-square difference test is the most accepted approach. 2. Differences in paths can exist, even if they are not statistically significant differences. These coefficients have sufficient levels of error to not be distinguishably different. 3. If there is no difference between paths, then the moderation is not significant for this path.
Hi James, the video was great! My exact model is this: the question is how satisfied are parents with children walking (y1 in the SEM1), School bus (y2 in the SEM2) and family car (y3 in the SEM3). SEM1, SEM2 and SEM3 all have the same structure and the same variables, but different weights of different significance. So I have three separate models. So this is not actually a multi-group analysis. But is there any solution to combine these three models in the same model? It is important as I like to compare the regression weights between the three models (to see if the weights are significantly different at a given confidence interval, when each two models are compared together). What do you suggest? Thanks in advance
If you just want to know which leads to greater satisfaction, then you should create a satisfaction score (maybe average) from the satisfaction items, and then do an ANOVA where the transportation mode is the factoring variable.
Dear Professor, thank you so much for these extremely helpful videos. It could be that I am missing something obvious. However, I was wondering whether it would be possible to keep the previous moderation indicators in the model and then proceed with the multi-group analysis. In other words, I did not understand why we would exclude Experience as a moderator from a model when testing for the male/female differences. I would appreciate any help from the community as well. Thank you in advance.
Dear Professor Gaskin, Many thanks for your superb videos! When running Multigroup Analysis, your demonstration compared Males vs. Females per path. Is it possible to compare more than two groups, e.g. Males (High) vs. Females (High); Males (Low) vs. Females (Low)? In other words, the Multigroup Analysis accommodates four groups as opposed to two. Many thanks in advance.
You would need to do the comparison in pairs, or one group versus all. So, just create a new variable in your dataset to indicate each of the four groups. Then in AMOS, compare across two groups at a time.
Dear sir, your videos are simply awesome. You are explaining simply and understandably. Expecting more videos. Sir, I have one doubt too. When performing model analysis, I got an error like the model is unidentified and adds 9 constraints. How to resolve this issue sir? Could you please help me?
Sounds like some of your latent variables are missing a constraint. Each latent variable must have at least one constraint. This constraint is placed on the path of the first indicator by default, but it can be moved around (doubleclick a path and go to the parameters tab and type 1 in the box), or it can be placed on the latent variable as a variance constraint. Either is fine and will result in very similar (usually identical) estimates. Thi might also happen if you added error variables manually. Make sure each of the error variables has a constraint on the path to the observed variable.
Hello Dr. Gaskin. Thank you so much for these life-saver videos. I also watched your measurement invariance videos. And I am just curious if we can conduct metric invariance test using this cool Multi-group analyses feature in AMOS?
Hello again. Actually, I checked both CFA videos in 2016 SEM Series but I could not find that you using multiple-group analyses button in AMOS to test metric invariance. (You are doing it with Chi-square difference test by using excel tools) Am I looking at the wrong place? Thank you.
Dang. I think you're right. Looks like I only use it for multiple groups comparisons in the causal model: ruclips.net/video/-j3LkADfgWs/видео.htmlh58m17s Sorry about that. I'm working on a plugin to do it, but it's not ready yet...
Thanks Dr Gaskin! Please could you recommend a resource for conducting SEM in AMOS with a binary outcome variable? I noted your response to another RUclipsr was to use Bayesian estimation but I also wanted to know how the interpretation of output differs for binary vs Likert scale output. Thanks again!
I don't have a good video or resource for this. Interpretation is just like for logistic regression, where the regression weights are interpreted in reference to the reference category. In the case of binary variables, it is just the lower value (usually zero) of the binary scale. So, for example, if I have a binary outcome variable that is purchase (one) or not purchase (zero), and a predictor like education. Then, if I observe a positive regression weight, I would interpret this: as education increases, the customer is more likely to purchase. If it is a negative coefficient, then it would be interpretted: as education increases, the customer is less likely to purchase.
Hello doctor James Gaskin. When I analyze many groups at the same time, AMOS give me one set of fit indices and all are good. But, when I separately analyze each groups, there are various set of fit indices, some are horrible fit. Which one should be a criteria to be written on Ph.D. dissertation?
Dear Prof Gaskin, In your 2013 multi-group video you show how to perform the critical ratios for differences test for analysing differences in beta values between variables. Which approach is better, the critical ratios for differences test or the Chi Square difference test approach as outlined in this video?
This is so helpful! I appreciate your effort in creating this video. One question I have is the reference. I'm trying to do multigroup analysis on gender and include the result in my manuscript. I wasn't able to find papers on reporting. How can I put references when including the result of p-value?
Dear James, I enjoy you lectures! What if the two groups to be compared differ in size (e.g. the sample of males is twice the size of females), this will effect standard errors of the same variable for each group. The coefficient for a particular effect may then be statistically significant for one group but statistically non-significant for the other group. Perhaps the effect is in the non-observed reality in both groups present but sample size in one group fails to reach that conclusion. I face this problem when investigating whether a particular hypothetical causal model fits to populations that are similar (same type of study population) but live in different countries. When pooling country records the model is confirmed, but when fit to country-specific situations (multi-group) than country-specific configurations emerge. Apart from substantive explanations for this, the problem is that there are major differences in country-sample size ranging from about 250-500, while the pooled sample is about 2000 cases. Anyway to control for differences in sample size in multigroup sem? Thank you for replying
The only way (that I can think of) to control for these differences is to take a random sample of the larger groups to match the lower sample size of the smaller groups. So, if you have 100 females and 200 males, just take a random sample of 100 males when doing your analysis.
Thank you James for your quick reply. It will give me a good idea about the impact of sample size on model fit, sampling errors and statistical significance of model parameters.
It's just a chi-square difference test. I like this article: Floh, A., & Treiblmaier, H. (2006). What Keeps the E-Banking Customer Loyal? A Multigroup Analysis of the Moderating Role of Consumer Characteristics on E-Loyalty in the Financial Service Industry. A Multigroup Analysis of the Moderating Role of Consumer Characteristics on E-Loyalty in the Financial Service Industry.(March 26, 2006).
Dear Dr. James Gaskin, thank you for your video. I am running a multi group analysis in AMOS. For a particular path in my model, the path coefficient is highly significant in one group and the same path coefficient is highly insignificant in the other group. However, the chi-square difference test for this path in the two groups comes out insignificant. Please advise!
This is because there is too much error in the non-significant path. So, all you can say is that the path is meaningful for one group, but not the other.
Dear James, thanks for your very informative videos! I'm a beginner at SEM and trying to do multigroup latent growth curve analysis using the "Multi-group analysis"-tool in AMOS. When attempting to do so, it appears to me as if the tool removes some of the constraints that I had imposed on my original model (equal factor loadings and intercepts over time) - is this true? If so, should I put them back manually to be able to interpret my models? Also, under Model comparison in the output I get 8 different tables, and I’m unsure which to use. The tool created 9 different models for me (named unconstrained -> measurement residuals).
Hi, Dr. Gaskin, I've recently came across a question about SEM model fit. What's the difference between Multi group (male120/female140) SEM model fit and overall (all260) SEM model fit? Why they are different? Can I use the good multigroup model fit as an argument to continue following multigroup analysis, even though my Measurement Invariance testing result is very significant?
When split into groups, it is considered a configural invariance test. When not split, then it is the goodness of fit for the model. When reporting model fit, you should report it unsplit. When doing invariance tests, you should report it split. If your invariance test is significant (i.e., the groups are different), then you need to identify a solution for those differences to achieve at least partial metric invariance (if you plan to do multigroup moderation moving forward).
Hello, Prof. Gaskin! Is it possible that the chi-square difference test results in a significant p value, meaning that there is difference between males and females at the global model level, but when testing for each path at a time, none of the single paths results in a significant p value? I mean, I found out that the model is different for males and females, but I couldn't find the path where this difference is...What could be the explanation for that? Does it mean that the answers of males and females were different for a certain latent construct, but this difference did not effect the influence of this factor in the other constructs (for example, females had stronger negative attitudes than males, but these attitudes did not effect the purchase intention differently for males and females)? Thank you very much for the videos, I am learning a lot from your channel!!
It is possible, though not common, for this to happen. Are you using a latent model or a composite model? (latent would be like CFA, but with regression lines also between latent factors, composite would be using factor scores and regression lines only between the boxes). If using a latent model, make sure you first do a measurement invariance test. This might be the reason you are finding model-level differences, but no path differences.
@@Gaskination Hi James, may I know why the model-level could be invariant, but the individual paths could be different? I have found in my results, indeed, the unconstrained and constrained model are not significantly different, but when I check individual paths, two paths are significantly different across groups. I would be grateful to know how to interpret this and if there is any reference for it. Huge thanks!!!
@@junyixie1008 When taken together, the individual paths can exhibit small enough differences that the whole model does not appear different. However, when observed in isolation (focusing only on that path), the difference can be significant. I'm really bad at remembering references... Here is a link to some potentially useful references: statwiki.gaskination.com/index.php?title=References#Moderation_and_Multigroup
Hi James, I may have missed this, but for a basic mediation model (Model 4), what is the recommended sample size for each group when doing a multigroup analysis in AMOS? Thanks!
There are lots of ways to calculate it. I prefer a basic formula: 50+5x, where x is the number of observed variables. This applies to all groups individually.
Hi dear James Gaskin, Firstly i should tell you thanks for videos. I 'd like to ask you a question. How to analyze formative constructs in AMOS if our scale's items were treated as a formative measure of the construct (in contrast to the Likert statements which are reflective measures (true or false etc.))? Thank you in advance
+ibrahim arısal AMOS won't allow for formative measures. Instead you will need to use a PLS software like SmartPLS. An alternative is to turn the formative measures into a score (sum or average), and then use the resultant new variables in a path analysis in AMOS.
Thanks, a lot Dr. James. You did a great work. Could you please tell me how to write research questions and hypothesizes if I have multigroup (i.e. High achievement and low)
Usually the RQ would be about how some grouping variable will affect the relationship of interest. The hypotheses will explicitly state the direction and strength of the effect for one group versus the other. Here is an example: statwiki.kolobkreations.com/index.php?title=Structural_Equation_Modeling#Multi-group_effects
Dear Prof. James Gaskin, thank you very much for your helpful videos. I'm having some doubts as to how to test indirect effects, using your MyindirectEffects estimand, in a multigroup analysis. I tried to use it but the result was the same for the two groups...is there a way to test it separately for each group? Thank you in advance.
You would have to uncheck the box for "all groups" in the object properties window, parameters tab. Uncheck it before naming the paths. Then name the paths for one group. Then run it and then switch to the other group (also make sure to remove the names for the first group). Or you can use my moderated mediation estimand. This allows you to name the paths for both groups and run them together.
Dear Dr. Gaskin, I want to know whether the technique you use in the video is applicable for structural model with latent construct? What i mean is, my model have multiple factors with multiple items. If this technique is applicable for latent construct, what is the procedure to analyze specific path? Which path in the structural weight model that I need to deleted? Since for latent construct, there also path for each items (factor loadings). Do i need to delete that too? Thank you in advance.
Hi James, just a question, do we have to do both interactions and multigroup analysis or we can do one of them? I am confused with what are they. If not wrong interaction is a control or moderator with no categories, while, multigroup has various categories, e.g. age, gender etc.?
Professor Gaskin, Thank you for the tutorial. I am using AMOS 24 and when I do multigroup analysis the program "creates" 6 models (unconstrained, measurement weights, structural weights, structural covariances, structural residuals and measurement residuals). Why do we only need the unconstrained and the structural weights models in the multigroup analysis? Thank you
Hi James, Surprisingly, I have found different significant results using this technique as demonstrated in this video and by using the plugin from statwiki page. Although the effect size is quite similar yet there are some different results in terms of significant results. Could you please advise on this problem? Many thanks - James. I highly appreciate it.
The plugin uses a chi-square difference test as well. However, we have noticed that in version 26 and 27 of AMOS that it has trouble saving the chi-square and df of the second model. This results in a zero difference, which is wrong. If this is the issue you are experiencing, then I recommend using the approach in this video instead of the plugin.
@@Gaskination Heaps of thanks, James, for this clarification. I highly appreciate it. Looking forward to meeting you on May 11 as I will be one of the participants from RMIT University, Australia, attending the SEM online course. Have a great weekend.
Hi Dr Gaskin, I was wondering what is an orthodox way to approach the EFA and CFA prior to multigroup analysis. Should I run two distinct EFAs and CFAs for each group?
It is customary to run them together. You can conduct invariance tests as well if you suspect there will be measurement differences. You can search my channel for "invariance": ruclips.net/user/Gaskinationsearch?query=invariance
Hi James, thank you for your great videos. Unfortunately, I could not quite find the solution to my current problem. My model is based on a 2x2 mixed design. One factor (scenario) splits my sample into two unbalanced groups (288/312) and the second factor (stimuli) is measured with two conditions within-subjects. I can only compute a variable assigning the subjects in one of the two conditions of the first factor (scenario). So for testing the difference between the two conditions of the second factor (stimuli), I would need an according dummy-variable which I don't have since each subject was measured on both conditions. Does AMOS provide any similar solution comparable to a mixed ANOVA that is able to split the within-subject factor? Would be great to get some help as I've been search the web for a solution for quite a while. I am fairly new to AMOS and I would very much appreciate your help.
AMOS does not, but you can just create a variable that splits into four (to indicate each group). Then you can do comparisons between pairs at a time: A:B, A:C, A:D, B:C, etc.
Hi James, very nice and informative video about the model comparison. I loaded 4 Groups into my Amos Dashboard, but Amos still gives me only 1 "Nested Model Comparsion" Output like you did with 2 Models. Shouldn't i get more Information because i used 4 models? I would like to see if there are any significant differents between the 4 models in one of the variables. Maybe you can give me some information in how to handle that. Cheers and thanks in advance.
We can only calculate the differences between two at a time. The human brain (and consequently our tools) can only conceive of pairwise comparisons. This can be A:B, A:C, A:D, B:C, B:D, C:D, or it can be one vs all: A:BCD, B:ACD, C:ABD, D:ABC.
Dear Prof James Gaskin Thank you so much for your informative video. I have a question about the confusing results of the structural invariance in my project. When I tested the full model by following your video, the p-value was insignificant. Thus, all paths are invariance across the groups. However, when I checked the estimated loadings in each group, one path was significant in group A and NOT significant in group B. I wonder if something wrong with my data. Thank you so much in advance.
Thank you so much for your reply. In my study, I have 250 in sample A and 242 in sample B. It is not a significant difference in sample size. Could you please give me some advice on how I could solve the problem?
There is no way to solve it because it is not a problem. It just means that it is significant for one group and not the other. This is perfectly possible. Just means more error in the non-significant group. You can try bootstrapping to reduce that error, but I doubt it will make a lot of difference.
Thanks for the video James, I am wondering about testing multigroup analysis and moderation using critical ratio in ur stat tools, is it still valid? and what difference between this new way and the old one (critical ration "z-score")? Many thanks
+Riad Cheikh The critical ratios approach suffers from family-wise error. In order to correct for this bias, you have to include a Bonferroni's correction, which is to divide the desired p-value (e.g., 0.05) by the number of hypotheses being simultaneously tested. The test p-values must then be less than this new corrected p-value in order to be significant at the original level of confidence (e.g., 95%). The chi-square difference approach is far more accepted. I actually just came up with a new way that is very direct, but before make a video about it I'm going to get it approved through some experts.
Dear Dr. James Gaskin. Because there are many types in testing invariance across groups (e.g. metric variance, configural variance, scalar invariance...), Is the "Structural Weights" model in video above called "congfigural invariance" (same structure across groups)? Or what is it called, when we want to imposed equal structural paths for groups? Thanks and best regards
Structural invariance refers to multigroup comparisons. In this case, it is not a measurement invariance. And, whereas in measurement invariance we hope for invariance (no difference between groups), in structural invariance we hope for differences between groups.
Thanks again for your enthusiastic response. However, I still have several problems with my analysis. the CFA and SEM model are fine with model fit indices (e.g. TLI, GFI, CFI >0.9). Nevertheless, when it comes to multi-group analysis, these indices drop down poorly ( just around 0.87). Furthermore, according to your recommendation, global test must take place first, then P value. In my case, the model fit indices are underrated although the p-value somehow indicates significant difference between groups. I test with totally 3 groups: age, income and gender; and all of them revealed underrated indices. So, what should I report these outcomes in my paper? Do I necessarily get rid of this multi-group analysis due to poor fit? Thanks and best wishes
Keep the analysis if you passed measurement invariance. Report the drop in model fit. You might also look at the regressions table in the modification indices. This table will change based on which group is showing. Maybe address any regression it suggests if these will have a major impact on your model and can be theoretically justified.
Dear Sir Thanks very much for sharing such great tutorial. I was wondering if you please guide me about co-varying the error terms. In my model, I have two exogenous variables (one 2nd order factor and one simple reflective construct), one endogenous/mediator ( 2nd order factor) and one endogenous/dependent (2nd order factor). My question is whether I can co-vary the error terms of the exogenous variables (2nd order factor) with other mediator and endogenous variables (2nd order factor). Thanking you very much for continuing such wonderful job.
It is generally considered best practice to not covary error terms. However, error term scan be covaried in special circumstances, such as for variables at the same level (e.g., two mediators, or two DVs). Dave Kenny has a website that explains this in more detail: davidakenny.net/cm/respec.htm
Good afternoon Professor Gaskin. I did a multi group analysis via spss Amos. One of my paths showed to be significantly different for men and women. However, both betas are insignificant. Did moderation still occur? Kind regards
Good question. This happens occasionally. You can say there is moderation, but that you cannot determine with confidence what the individual effects are, since the individual effects cannot be distinguished from zero (i.e., they are non-significant). However, you can at least say that the effects (although undetermined) are different for men and women.
Dear James, thank you. What sample size do we need to conduct path analysis in Amos? and what about the sample size of the groups we compare? Should they be equal? Any assumptions?
The minimum that I like to follow is 50+3x, where x is the number of included variables. Group sample sizes need to meet this minimum as well, but they do not need to be equal. If they are vastly different, like 65 and 500, then it may be best to randomly sample from the larger group to more closely match the smaller group.
@@Gaskination Thank you, it helps a lot. I read something everywhere saying min sample size needs to be 200 and you need 10 participants per parameter. It s confusing if you have just begun to learn something! Btw, I agree with everyone, you and your videos are just amazing. I need to take more time for your channel.
The moderator in this case is gender. As this is a single observed variable, rather than a latent factor, it does not belong in the measurement model. Only latent factors belong in the measurement models (EFA/CFA).
Dear Dr. James, thank you very much for this helpful video. My question is whether multigroup analysis with the binary outcome variable can be performed in Amos? If the Bayesian and MCMC approaches are used, no chi-square is reported. How can we compare the model difference? Thanks.
Amos should be able to run an MGA with a binary outcome, as long as the grouping variable is not also the outcome variable, and as long as all variables included in the model have variance within each of the groups. If no Chi-square is produced, then maybe it is because the model is saturated (i.e., every possible relationship is drawn). You need degrees of freedom to produce the chi-square.
@@GaskinationDear Dr. James, thank you very much for the prompt reply. I think my model has qualified for your mentioned conditions. The problem is by using the Bayesian approach, only "posterior predictive p", "DIC", and "Effective number of parameters" are reported in the fit measures. It looks like this approach doesn't produce Chi-square. Therefore, there is no way to do model comparison. The Maximum Likelihood Estimate can be used to run an MGA with a binary outcome and get a result, from technologically speaking. However, the AMOS tutorial says that "In Amos, models involving categorical responses require the use of the Bayesian and MCMC option" (No. 10, www.structuralequations.org/). Is that true? If it is true, how to do MGA analysis with a binary outcome? Thanks again?
@@jinxinliu7822 In this case, you would have to compare the outcomes without the chi-square. This can be done by comparing the regression weights and also by using an estimand that subtracts the regression weight of one group from another group. This will produce a confidence interval around the difference and estimate a p-value. I have an estimand that could be used for this. It can be found here: statwiki.kolobkreations.com/index.php?title=Plugins#LIST_OF_ESTIMANDS: It is called "MyGroupDifferences". There are instructions there on how to use it. I've never tried running an estimand with Bayesian estimation though, so I don't know for sure if it will work... Good luck!
@@Gaskination Dear Dr. Gaskin, thanks for this helpful suggestion. I'm wondering whether the maximum likelihood estimation can be used for a binary outcome in Amos. In another word, is the Bayesian approach necessary for a binary outcome? Thanks so much!
@@jinxinliu7822 I think Bayesian is preferred but not required. ML can be used. Just interpret your regression weights as 'more associated with the higher binary value' if positive and 'more associated with the lower binary value' if negative.
Hi Dr Gaskin, Thanks for your highly informative videos! Please could you clarify if it is okay to use the "groups" feature in AMOS when I have multiple categorical moderators? For example, I have gender (male,female) and residency (resident, non-resident). Can I create those two groups as moderators in the same model or do I have to run separate models, one with gender as moderator, and a second with residency as moderator? If I can use one model, then can I interpret results for only one (of four) group at a time or is it possible to interpret for 2 groups each, e.g. male AND resident? Thanks a lot!
Good questions. AMOS can handle as many groups as you create. However, for simplicity of interpretation, I'd recommend only having two groups at a time.
I preassume that hostile sexism positively strengthen the relationship between low self-esteem and organizations, following your lesson if HS dampens the negative relationship, means do not support my hypothesis
You can, but you are likely to find no differences (that are statistically significant), due to the amount of error in a small sample size. I would recommend assessing differences with less emphasis on p-values.
Dear Prof. Gaskin, I'm trying to develop two models (one for local students and one for foreign students) based on a structural model. Examples of my hypotheses are H1: Peer influence has a direct effect on local student achievement H2: Peer influence has a direct effect on foreign student achievement. I using multigroup analysis to test the hypothesis and to determine the sig. path. Am I doing the right thing?
The baseline hypothesis would be something like: H1. Peer influence has a positive effect on student achievement. And then the moderating hypothesis would be something like: H1a. The positive effect of peer influence on student achievement is stronger for foreign students. Or maybe it is for local students; I don't know... And then you would test one model, with peer influence having an effect on student achievement, but you would test it across this grouping variable (local/foreign).
Hello again Prof. Gaskin, thank you so much for the prompt reply. I have two more questions: 1. If I have H1 and H1a. as mentioned. Do I need to find additional theory/studies to support H1a? What if my study objective is making a comparative path model between these two groups but not testing the moderation effect? 2. If I'm testing H1 and H1a. kind of hypotheses using multigroup analysis, is it right to present a path model for overall and for each group in the end? Thank you
@@noraasikinabubakar9514 1. You would need to supply some abductive reasoning and some supporting literature to justify the moderation effect. Comparing a path or model based on group membership is the definition of multi-group moderation. 2. Yes, this is a good approach.
If the relationship between two variables is moderated by three moderators and there are interaction effects among moderators, should we break the models so as to study two moderators and their interaction and then drop on moderator, introduce the third one and then study the interaction between the first and third moderator?
It depends on your theory. If you are theorizing simultaneous moderated effects, then do them all at the same time. If these are not confounding, then you can do them separately. As for a reference, Hayes, Andrew. F., 2013. Introduction to Mediation, Moderation, and Conditional Process Analysis: A Regression-Based Approach, New York: The Guilford Press
Dear James, just another question, what if my hypothesis is: "Female consumers have more positive effect of satisfaction on loyalty, when the role model is female, while male consumers have more positive effect of satisfaction on loyalty, when the role model is male"... Could you kindly suggest, what shall I do in this case, please? Do I have to create 4 multigroups, male-like-male, male-like-female, female-like-male and female-like-female or is there any other way to resolve this issue. Kindly suggest please.
Thanks for the reply, very kind of you. If I create four groups, it means I will be doing changes in my data set according to male-likes-male, male-likes-female, female-likes-male and female-likes-female and will be creating one construct (with scores from 1 to 4), right? And thankyou so much for all the movies and information, you have provided. You are a real Angel and God Bless You for that :-)
Hi, thank you very much for this video. I do have two data set one from France and the other from Pakistan. i want to compare the results but the measurement models are different e.g. only 4 out of 8 observable loads on affective commitment from that data from Pakistan while from french data 7 variables of affective commitment out of 8 loads. so how to compare the results?
One way to do this would be to just use the variables that load on both. Another way is to run the model with all data (from both countries) and see what loads with all data.
Dear James, many thanks for this (and all your other) very helpful videos! One question remains, could you please help me with the following? I compared two groups with the X2 test but am wondering how to report on it. As far as I know, a X2 test would normally be reported the following way: Χ2(df, n) = ..., p < ... What would be the n in this case? As I have two, that of group 1 and that of group 2. Thank you very much!
Report it as Group 1 n and Group 2 n, and then the X2 and the df are for the difference between the constrained and unconstrained models (rather than the actual X2 and df). Then the p-value is for the test.
Greetings. I am conducting multigroup moderation analysis with three categorical groups. Now my question is since i need to do comparison in Pairs A:B, A:C, & B:C. Do i need to add all the three groups at once in the Model, run it and then do pairwise comparison. Or I need to run the model three times with only two groups in my Structural Model each time.
You can add all three groups and then just make paired comparisons. As for model fit, it should not change between these models since the models are the same. If there are large measurement differences between groups (i.e., measurement non-invariance), then perhaps you'll observe very small model fit differences. So, you only need to report model fit for the model once.
@@Gaskination by making paired comparison do you mean that we will end up having same chi square difference for all the groups(in amos Model output Comparison option) and then look for path level difference to make inference ?
@@Gaskination if we will include all the three groups in amos in a single model and run it for moderation then in output we have only single model comparison chi square for a particular path for all the three groups. Having said that will that one model comparison chi square difference suffice for making comparison for all the three groups ( by comparing their regression coefficient and significance level in output option- "estimates" in Amos. ( Here i am running model only once with three different categorical groups. Afterwards i check model comparison in amos to see the chi square which is common for all the three groups- since amos shows only one cmin/df for a single path for all the three groups. Thereafter i see the regression estimate of the three groups to make pairwise comparison) Is this right way to do it. Or there is any issue with this approach?
Sir, Greetings. I have enjoyed your video very much. However, I have collected data on a study entitled "Relationship between employees Satisfaction and Loyalty: Comparative study of Public vs Private bank". 4 factors in satisfaction (Data were collected in Likert scale). Data were collected from two group of respondents (Private bank; Public Bank)? I had also run Structural Equation Model in AMOS. When i inputted data in SPSS; Type of bank was considered as nominal Variable (1 for private; 2 for public bank). All data from the respondent ( employees of public bank and private bank) were recorded in a one data set (that means, i have created one data set). I would like to test the hypothesis: there is no significant difference in employee satisfaction and loyalty in case of private bank and public bank. Can I run multiple Group analysis in AMOS? Does require two set of Data or Single Set of data (where both employees from public and private)? How can i prepare two set of data from single set of data? Could you please help me? How i can do this? please suggest me sir. Thanks in Advance
Just use one dataset and do a t-test or ANOVA where the factoring variable is BankType and the dependent variables are satisfaction and loyalty. This will show you whether there is a difference between private and public bank employees in terms of satisfaction and loyalty.
Respected Sir Thank you so much for your valuable guidelines. Just one more question can I run Multi group analysis in amos for testing the hypothesis whether there is a significant difference between private bank and public bank in terms of satisfaction and loyalty? Thanks
Multigroup in amos will let you determine if the relationship between satisfaction and loyalty is different when comparing public to private banks. Again, the multigroup is for testing differences in the *relationship* between variables, but will not tell you if public or private banks have more or less loyalty or satisfaction.
Dear James, thanks for the helpful instructions! I would highly appreciate if you could help me with the following problem. I did a normal path analysis which run smoothly and then I conducted a group analysis with the same model with 4 different groups. The notifications "sample moment matrix is not positive definite" pops up and suggests 4 reasons for this problem. How is this possible that only by adding the groups to the path analysis this problem comes up? I was desperately looking for a solution and would be really happy if you could help me - Thanks a lot!!!
My path model includes 1 independent, 3 mediator and 5 dependent variables. The groups have a size of 48, 72, 75 and 90 participants - is this already too small? Thanks for your support!
48 is pretty small. See what happens when you exclude that group. The problem with low sample size is that there is insufficient data to reduce error and thus minimize (execute) the model.
You might also try looking for clues in any results it gives (if it gives any results) - look for high standard error. Try removing a single variable at a time until you find the one that is giving you the trouble.
hello dear James. i would like to know that if we have one model and we aimed to comparison study between two states , how we can compare our results ? could we run SEM separately for states ?
Dear Professor Can you please make a whole video lecture about multilevel analysis with smart PLS, as in my thesis there are 3 models two based on individual level and one is on team level please guide please
Funny story... I've never done multilevel analysis. I've just never been involved in such a study. So, I'm not sure how to do it. It's on my list of things to learn eventually.
@@qaisermohiuddin413 Having never done it, I'm not sure who to point you toward. I checked another channel I follow, and they don't have it either. I guess you'll just need to google it or search on RUclips.
Hi James, i'm wondering if i should use this one or the another one in this link ruclips.net/video/mirI5ETQRTA/видео.html that you post earlier. - This one is using the fully constrained model, - The one in the link as i see used the partly constrained model Some of my friend thesis are using partly constrained model, but i based mine on fully constrained. is there many different between two way? or i can use both?
oops! I should have left the whole model constrained EXCEPT the path I was testing... Sorry about that. Good catch. Sorry also about the late reply. Somehow this comment didn't make it to my inbox. I only just stumbled upon it while responding to another comment.
@@Gaskination last thing, at the end of the clip, when we check if the relation is stronger from male to female, you said the p-value < 0,05 means the regression p-value, right? since the chi-squared test's value already < 0,05 mean they are different.
No. These are different. MGA is to compare multiple groups. Multilevel modeling is usually referring to modeling that includes group levels, such as country, state, city.
Here's a fun pet project I've been working on: udreamed.com/. It is a dream analytics app. Here is the RUclips channel where we post a new video almost three times per week: ruclips.net/channel/UCiujxblFduQz8V4xHjMzyzQ
Also available on iOS: apps.apple.com/us/app/udreamed/id1054428074
And Android: play.google.com/store/apps/details?id=com.unconsciouscognitioninc.unconsciouscognition&hl=en
Check it out! Thanks!
Creative app. Good one
@Cholpon Cholpon Sounds complex, but definitely testable in SEM. Just as a warning, culture is almost never a mediator and almost always a moderator.
I was so pessimistik about how to make excel page, then i realized you share it. I can't believe you are such an amazing guy!! Thanks a lot.
I have decided to first like your videos before watching
You are a genius, You never disappoint
Hi James..
Thanks for this tutorial. I've been reading a couple of AMOS manuals trying to work out how to do multigroup analysis without any progress, but your tutorial has clarified that and enabled me to complete the analyses.
Hi Dr Gaskin. To be able to know whether the relationship is stronger or weaker you mentioned that after constarining the specific paths we are measuring, we can compare the standardized regression weights of males or females. But in this case shouldn't we be looking at the regression weights of the unconstrained model rather than the structural weights model (which have these specifc paths already constrained). If we loo at the standardized regression weights from the estimates tab they match the numbers on the paths of the unconstrained model not those on the structural weights model. thank you.
Yes, you are correct.
Dear James! What an amazing series of videos! Thank you. I was wondering about running a mulitigroup (3 groups) interaction SEM . Should I just run the same?
Yes, just run it in pairs. A:B, A:C, B:C or A: BC, B:AC, C:AB.
Dear Dr. Gaskin,
Thank you for the videos.
From the example you explained that the number of population in each group is similar and comparable. However, in my research one group has more than 480 data while the other is only 45 data.
Is it still comparable using the method you explained above? How could we justify the comparison?
This happens in my research because the other group has significantly higher population than the other by more than 40x. Collecting the similar amount of data would be an uphill task if not impossible.
In my research the comparison between the two groups full model comparison resulted on 'Groups are different at the model level. Check path differences' in the statwiki excel file. The path comparison shows two paths are significant.
Thanks again
With such a large difference, you will likely run into trouble. The main trouble will be from the small sample size having insufficient power to minimize error. If you did find two paths different, then that is amazing with such a low sample size in one group.
Dear Prof. James, thank you very much for the helpful tutorial. For my master thesis I´m working with AMOS for the first time. To test if a model consisting of six factors with two items each (I know, it´s a bit unusual to have so few items...) differs between two groups, I followed your instructions in the video. As grouping variable I used country (1, 2). When I used the multi group analysis- bottom of AMOS, the default settings differ from the ones in the video: Instead of STRUCTURAL WEIGHTS, I can only choose MEASUREMENT WEIGHTS, structural covariances and measurement residuals. Plus, if I run the analysis I get only results in the "Model Comparison"-Section for Structural covariances and measurements residuals. How comes? Can you possibly help me with that? Thank you in advance for your support.
Hi James,
thank you for your informative video. I have just one more question concerning the interpretation: when we want to check which effect is stronger and if the effects are significant (or in my case which one is positive or negativ) we have to click at the "unconstrained" model, right? thank you in advance.
Hi James, thanks a lot for great video series. I have a question about the step after the invariance testing completed. I am trying to do multigroup analysis with two groups. If I got nonsignificant p value for chi square difference between the unconstrained and measurement weights models (or in the second step, between the measurement weights and structural weights model), which model in the output should I look for the estimates? Should I choose measurement invariance and structural weights, respectively. Or do I report the standardized estimates of the unconstrained model? Thank you in advance.
If the p-value for the chi-square difference is non-significant, then you can remove all additional constraints and just report the unconstrained model.
Hello Dr. James Gaskin thanks a lot for the tutos. I've been following you for a long time. I want to know whether I should choose the Measurement Weights or Structural Weights for my multigroup analysis in CFA? Second question: Should this analysis be performed only with a path model or it's OK with a structural moedl? Third question: In case I have both significant and insignificant p-values, which ones should I assess for the MI and how to do it? Thanks in advance.
Structural if you want to know causal differences across groups. Structural model will be more accurate. Not sure I understand the third question.
Dear James, could you please help me?
I've got two groups (n=208,n=162) and when their standardized regression weights are rather different (ex: .20 versus .35 with p
Hi James,
As far as I get, there are two types of invariances. One is measurement invariance (as demonstrated in your other video "Measurement Model Invariance") and the other is structural invariance. I assume what you did here is checking out structural-ish invariance (assuming that there is no latent variables in this tutorial). Am I correct?
My question is what if I get a significant chi-square difference in terms of gender in structural measurement but non-significant difference in measurement model? What does it mean and do I have to report both? (I run into some studies where they just reported structural invariance).
Thank you.
You are correct. And, your scenario is exactly what you want. You want measurement invariance and structural non-invariance (i.e., differences at the structural level, but not the measurement level). This would mean that your measures mean the same thing across groups, but that the relationships between constructs differ.
@@Gaskination Thank you for the reply.
I'm a bit confused by the article below: www.ncbi.nlm.nih.gov/pmc/articles/PMC4900100/.
They depict structural invariance as a nice thing (suitable, global, universal, cross-cultural and etc.) as opposed to your abovementioned reply.
My other question (just popped into my mind right now): What if I am not developing some kind of scale? Do I still have to conduct measurement invariance calculations?
Thank you so much! You are such a lifesaver!
@@berkandemir4220 They have titled their study incorrectly. I looked through the study and they are only examining measurement invariance. This should have been caught in the review process, but probably no one on the review team was an SEM expert.
Hello James,
Thanks so much for your wonderful videos.
I have a question regarding testing hypotheses. My research paper includes 3 samples. I tested measurement invariance and its result indicates that the groups are partially invariant. Then, I tested the latent model of each sample of them separately and confirmed the model fit of each model of them. After that, I did multigroup analysis of the three datasets together and its result indicates that the groups are different at the model level. (Note: Here is my hypotheses: H1: PE positively influences BI. H2: PE positively influences PU. H3: PU positively influences BI. H4: T positively influences BI. H5: R positively influences BI. H6: SI positively influences BI.)
So, my questions are:
1- I used the same steps that you have used in this video to do multi-group analysis among 3 groups, is this right?
2- which kind of results I should use it to test my hypotheses, the results of paths coefficient of the latent model of each sample separately or the results of paths coefficient of the multi-group analysis?
3- when I did multi group analysis, the chi-square difference test was significant at the whole model level, but when I did the chi-square difference test for each path separately, the result was non significant for most paths except of 2 of them, does this make sense?
4- In multigroup analysis, the result of chi-square difference test for one of my paths was significant which means that this path is not different among the three groups. whereas, this path was significant in 2 models of my latent models and non significant in the third one which means that this path is different among the 3 groups. so, I am very confused about these results, could you explain that what does it mean?
Thanks in advance.
1. Yes
2. Report the results based on the hypotheses. So, you might have to run multiple models if you have different types of hypotheses.
3. Yes. It means the difference at the model-level is originating from those two paths.
4. Even when the difference is not significant, if the estimates are very different, or significant vs non-significant, this is also a form of moderation. For example, in your case, you can say that X is a good predictor for Y for two of the groups, but not for the other group. So, there is moderation, despite the lack of statistically significant difference.
Thanks Sir for your reply. I have two more questions:
1- how can I do t-statistic after using multi group analysis to make comparisons of path coefficients between the three groups? (Note: I have download your excel Stat-tools but it just calculate t-statistic between 2 samples not 3, how can I use this tool to make comparisons between 3 samples?)
2- which one of tables I should use it to write the path coefficient in t-statistic formula, regression weights or standardized regression weights?
1. Group comparisons must be done in pairs. So, you can do this as A:B, A:C, B:C or you can do it as A:BC, B:AC, C:AB.
2. The formulas always use unstandardized coefficients.
Many thanks Sir.
1-I want to know what is the better method to test multi-group analysis and to examine the significant differences of each path across my groups? is it multi-group analysis by using t-statistic or multi-group analysis by using chi-square difference test? (Note: I have 3 groups).
2- when I did chi-square differences test for each path. I found Δ ꭓ2 of one of my paths was insignificant, indicting that there is no difference of this path across the groups, while the path coefficients unstandardized estimates differs across groups (0.586, 0.371, 0.457), what I should report about that?
3- Also, there is another path that was insignificant in each model and its Δ ꭓ2 was insignificant too across groups, does it mean that this path is invariant across groups and there is no moderation effect for this path or not (because it was already insignificant in each model)? what I should report about that?
1. The chi-square difference test is the most accepted approach.
2. Differences in paths can exist, even if they are not statistically significant differences. These coefficients have sufficient levels of error to not be distinguishably different.
3. If there is no difference between paths, then the moderation is not significant for this path.
Hi James, the video was great!
My exact model is this: the question is how satisfied are parents with children walking (y1 in the SEM1), School bus (y2 in the SEM2) and family car (y3 in the SEM3). SEM1, SEM2 and SEM3 all have the same structure and the same variables, but different weights of different significance. So I have three separate models.
So this is not actually a multi-group analysis. But is there any solution to combine these three models in the same model? It is important as I like to compare the regression weights between the three models (to see if the weights are significantly different at a given confidence interval, when each two models are compared together). What do you suggest?
Thanks in advance
If you just want to know which leads to greater satisfaction, then you should create a satisfaction score (maybe average) from the satisfaction items, and then do an ANOVA where the transportation mode is the factoring variable.
Dear Professor, thank you so much for these extremely helpful videos. It could be that I am missing something obvious. However, I was wondering whether it would be possible to keep the previous moderation indicators in the model and then proceed with the multi-group analysis. In other words, I did not understand why we would exclude Experience as a moderator from a model when testing for the male/female differences. I would appreciate any help from the community as well. Thank you in advance.
You can definitely keep the moderators in there. It just makes it more complex. You would essentially be testing moderated moderation.
Dear Professor Gaskin,
Many thanks for your superb videos! When running Multigroup Analysis, your demonstration compared Males vs. Females per path. Is it possible to compare more than two groups, e.g. Males (High) vs. Females (High); Males (Low) vs. Females (Low)? In other words, the Multigroup Analysis accommodates four groups as opposed to two. Many thanks in advance.
You would need to do the comparison in pairs, or one group versus all. So, just create a new variable in your dataset to indicate each of the four groups. Then in AMOS, compare across two groups at a time.
@@Gaskination That's excellent! Many thanks for your extremely helpful advice, which will be implemented.
Dear sir, your videos are simply awesome. You are explaining simply and understandably. Expecting more videos. Sir, I have one doubt too. When performing model analysis, I got an error like the model is unidentified and adds 9 constraints. How to resolve this issue sir? Could you please help me?
Sounds like some of your latent variables are missing a constraint. Each latent variable must have at least one constraint. This constraint is placed on the path of the first indicator by default, but it can be moved around (doubleclick a path and go to the parameters tab and type 1 in the box), or it can be placed on the latent variable as a variance constraint. Either is fine and will result in very similar (usually identical) estimates. Thi might also happen if you added error variables manually. Make sure each of the error variables has a constraint on the path to the observed variable.
@@Gaskination Thank you very much for this information sir.
Hello Dr. Gaskin. Thank you so much for these life-saver videos. I also watched your measurement invariance videos. And I am just curious if we can conduct metric invariance test using this cool Multi-group analyses feature in AMOS?
Yes. There is another video in this series that shows how to do that. It is one of the two CFA videos.
Hello again. Actually, I checked both CFA videos in 2016 SEM Series but I could not find that you using multiple-group analyses button in AMOS to test metric invariance. (You are doing it with Chi-square difference test by using excel tools) Am I looking at the wrong place? Thank you.
Dang. I think you're right. Looks like I only use it for multiple groups comparisons in the causal model: ruclips.net/video/-j3LkADfgWs/видео.htmlh58m17s Sorry about that. I'm working on a plugin to do it, but it's not ready yet...
Thanks Dr Gaskin!
Please could you recommend a resource for conducting SEM in AMOS with a binary outcome variable? I noted your response to another RUclipsr was to use Bayesian estimation but I also wanted to know how the interpretation of output differs for binary vs Likert scale output.
Thanks again!
I don't have a good video or resource for this. Interpretation is just like for logistic regression, where the regression weights are interpreted in reference to the reference category. In the case of binary variables, it is just the lower value (usually zero) of the binary scale. So, for example, if I have a binary outcome variable that is purchase (one) or not purchase (zero), and a predictor like education. Then, if I observe a positive regression weight, I would interpret this: as education increases, the customer is more likely to purchase. If it is a negative coefficient, then it would be interpretted: as education increases, the customer is less likely to purchase.
Hello doctor James Gaskin.
When I analyze many groups at the same time, AMOS give me one set of fit indices and all are good.
But, when I separately analyze each groups, there are various set of fit indices, some are horrible fit.
Which one should be a criteria to be written on Ph.D. dissertation?
Use the one where all are estimated simultaneously. The ones with horrible fit are probably due to sample size being small in the group.
James Gaskin Thank you very much, Dr.
Dear Prof Gaskin, In your 2013 multi-group video you show how to perform the critical ratios for differences test for analysing differences in beta values between variables. Which approach is better, the critical ratios for differences test or the Chi Square difference test approach as outlined in this video?
chi-square difference test is better and more accepted. The problem with the critical ratios approach is that it suffers from family-wise error.
This is so helpful! I appreciate your effort in creating this video.
One question I have is the reference. I'm trying to do multigroup analysis on gender and include the result in my manuscript. I wasn't able to find papers on reporting. How can I put references when including the result of p-value?
Here are some helpful references for this topic: statwiki.kolobkreations.com/index.php?title=References#Moderation_and_Multigroup
Thank you so much!!
Dear James,
I enjoy you lectures! What if the two groups to be compared differ in size (e.g. the sample of males is twice the size of females), this will effect standard errors of the same variable for each group. The coefficient for a particular effect may then be statistically significant for one group but statistically non-significant for the other group. Perhaps the effect is in the non-observed reality in both groups present but sample size in one group fails to reach that conclusion. I face this problem when investigating whether a particular hypothetical causal model fits to populations that are similar (same type of study population) but live in different countries. When pooling country records the model is confirmed, but when fit to country-specific situations (multi-group) than country-specific configurations emerge. Apart from substantive explanations for this, the problem is that there are major differences in country-sample size ranging from about 250-500, while the pooled sample is about 2000 cases. Anyway to control for differences in sample size in multigroup sem?
Thank you for replying
The only way (that I can think of) to control for these differences is to take a random sample of the larger groups to match the lower sample size of the smaller groups. So, if you have 100 females and 200 males, just take a random sample of 100 males when doing your analysis.
Thank you James for your quick reply. It will give me a good idea about the impact of sample size on model fit, sampling errors and statistical significance of model parameters.
Hi James,
thank you for providing new insights into multigroup analysis. Can you please state a reference for this approach? Thank you :)
It's just a chi-square difference test. I like this article:
Floh, A., & Treiblmaier, H. (2006). What Keeps the E-Banking Customer Loyal? A Multigroup Analysis of the Moderating Role of Consumer Characteristics on E-Loyalty in the Financial Service Industry. A Multigroup Analysis of the Moderating Role of Consumer Characteristics on E-Loyalty in the Financial Service Industry.(March 26, 2006).
Dear Dr. James Gaskin, thank you for your video. I am running a multi group analysis in AMOS. For a particular path in my model, the path coefficient is highly significant in one group and the same path coefficient is highly insignificant in the other group. However, the chi-square difference test for this path in the two groups comes out insignificant. Please advise!
This is because there is too much error in the non-significant path. So, all you can say is that the path is meaningful for one group, but not the other.
Dear James, thanks for your very informative videos! I'm a beginner at SEM and trying to do multigroup latent growth curve analysis using the "Multi-group analysis"-tool in AMOS. When attempting to do so, it appears to me as if the tool removes some of the constraints that I had imposed on my original model (equal factor loadings and intercepts over time) - is this true? If so, should I put them back manually to be able to interpret my models? Also, under Model comparison in the output I get 8 different tables, and I’m unsure which to use. The tool created 9 different models for me (named unconstrained -> measurement residuals).
I've never done a latent growth curve model in AMOS, so I'm not sure. Sorry about that.
Hi, Dr. Gaskin, I've recently came across a question about SEM model fit.
What's the difference between Multi group (male120/female140) SEM model fit and overall (all260) SEM model fit? Why they are different? Can I use the good multigroup model fit as an argument to continue following multigroup analysis, even though my Measurement Invariance testing result is very significant?
When split into groups, it is considered a configural invariance test. When not split, then it is the goodness of fit for the model. When reporting model fit, you should report it unsplit. When doing invariance tests, you should report it split. If your invariance test is significant (i.e., the groups are different), then you need to identify a solution for those differences to achieve at least partial metric invariance (if you plan to do multigroup moderation moving forward).
Hello, Prof. Gaskin! Is it possible that the chi-square difference test results in a significant p value, meaning that there is difference between males and females at the global model level, but when testing for each path at a time, none of the single paths results in a significant p value? I mean, I found out that the model is different for males and females, but I couldn't find the path where this difference is...What could be the explanation for that? Does it mean that the answers of males and females were different for a certain latent construct, but this difference did not effect the influence of this factor in the other constructs (for example, females had stronger negative attitudes than males, but these attitudes did not effect the purchase intention differently for males and females)? Thank you very much for the videos, I am learning a lot from your channel!!
It is possible, though not common, for this to happen. Are you using a latent model or a composite model? (latent would be like CFA, but with regression lines also between latent factors, composite would be using factor scores and regression lines only between the boxes). If using a latent model, make sure you first do a measurement invariance test. This might be the reason you are finding model-level differences, but no path differences.
@@Gaskination Hi James, may I know why the model-level could be invariant, but the individual paths could be different? I have found in my results, indeed, the unconstrained and constrained model are not significantly different, but when I check individual paths, two paths are significantly different across groups. I would be grateful to know how to interpret this and if there is any reference for it. Huge thanks!!!
@@junyixie1008 When taken together, the individual paths can exhibit small enough differences that the whole model does not appear different. However, when observed in isolation (focusing only on that path), the difference can be significant. I'm really bad at remembering references... Here is a link to some potentially useful references: statwiki.gaskination.com/index.php?title=References#Moderation_and_Multigroup
Hi James,
I may have missed this, but for a basic mediation model (Model 4), what is the recommended sample size for each group when doing a multigroup analysis in AMOS?
Thanks!
There are lots of ways to calculate it. I prefer a basic formula: 50+5x, where x is the number of observed variables. This applies to all groups individually.
Dear James, thank you for your video. I have a question, this analysis is the invariance analysis of model?
This video is the structural invariance. There is also measurement invariance which refers to something else.
@@Gaskination thank you very mucho!
Hi dear James Gaskin,
Firstly i should tell you thanks for videos. I 'd like to ask you a question. How to analyze formative constructs in AMOS if our scale's items were treated as a formative measure of the construct (in contrast to the Likert statements which are reflective measures (true or false etc.))? Thank you in advance
+ibrahim arısal AMOS won't allow for formative measures. Instead you will need to use a PLS software like SmartPLS. An alternative is to turn the formative measures into a score (sum or average), and then use the resultant new variables in a path analysis in AMOS.
Thanks, a lot Dr. James. You did a great work.
Could you please tell me how to write research questions and hypothesizes if I have multigroup (i.e. High achievement and low)
Usually the RQ would be about how some grouping variable will affect the relationship of interest. The hypotheses will explicitly state the direction and strength of the effect for one group versus the other. Here is an example: statwiki.kolobkreations.com/index.php?title=Structural_Equation_Modeling#Multi-group_effects
Dear Prof. James Gaskin, thank you very much for your helpful videos. I'm having some doubts as to how to test indirect effects, using your MyindirectEffects estimand, in a multigroup analysis. I tried to use it but the result was the same for the two groups...is there a way to test it separately for each group? Thank you in advance.
You would have to uncheck the box for "all groups" in the object properties window, parameters tab. Uncheck it before naming the paths. Then name the paths for one group. Then run it and then switch to the other group (also make sure to remove the names for the first group). Or you can use my moderated mediation estimand. This allows you to name the paths for both groups and run them together.
Dear Dr. Gaskin,
I want to know whether the technique you use in the video is applicable for structural model with latent construct? What i mean is, my model have multiple factors with multiple items.
If this technique is applicable for latent construct, what is the procedure to analyze specific path? Which path in the structural weight model that I need to deleted? Since for latent construct, there also path for each items (factor loadings). Do i need to delete that too?
Thank you in advance.
Yes, this can work for latent models as well. Just focus on the causal paths between latent factors and don't worry about the indicator paths.
Hi James, just a question, do we have to do both interactions and multigroup analysis or we can do one of them? I am confused with what are they. If not wrong interaction is a control or moderator with no categories, while, multigroup has various categories, e.g. age, gender etc.?
zzz just one or the other is fine. Interaction for continuous moderators and multi group for categorical moderators.
thanks James, blessings from UK :-)
Professor Gaskin,
Thank you for the tutorial. I am using AMOS 24 and when I do multigroup analysis the program "creates" 6 models (unconstrained, measurement weights, structural weights, structural covariances, structural residuals and measurement residuals). Why do we only need the unconstrained and the structural weights models in the multigroup analysis?
Thank you
The other models are for measurement invariance tests. You can use them if you like.
Hi James,
Surprisingly, I have found different significant results using this technique as demonstrated in this video and by using the plugin from statwiki page. Although the effect size is quite similar yet there are some different results in terms of significant results. Could you please advise on this problem?
Many thanks - James. I highly appreciate it.
The plugin uses a chi-square difference test as well. However, we have noticed that in version 26 and 27 of AMOS that it has trouble saving the chi-square and df of the second model. This results in a zero difference, which is wrong. If this is the issue you are experiencing, then I recommend using the approach in this video instead of the plugin.
@@Gaskination Heaps of thanks, James, for this clarification. I highly appreciate it. Looking forward to meeting you on May 11 as I will be one of the participants from RMIT University, Australia, attending the SEM online course. Have a great weekend.
Hi Dr Gaskin,
I was wondering what is an orthodox way to approach the EFA and CFA prior to multigroup analysis. Should I run two distinct EFAs and CFAs for each group?
It is customary to run them together. You can conduct invariance tests as well if you suspect there will be measurement differences. You can search my channel for "invariance": ruclips.net/user/Gaskinationsearch?query=invariance
Hi James, thank you for your great videos. Unfortunately, I could not quite find the solution to my current problem. My model is based on a 2x2 mixed design. One factor (scenario) splits my sample into two unbalanced groups (288/312) and the second factor (stimuli) is measured with two conditions within-subjects. I can only compute a variable assigning the subjects in one of the two conditions of the first factor (scenario). So for testing the difference between the two conditions of the second factor (stimuli), I would need an according dummy-variable which I don't have since each subject was measured on both conditions. Does AMOS provide any similar solution comparable to a mixed ANOVA that is able to split the within-subject factor? Would be great to get some help as I've been search the web for a solution for quite a while. I am fairly new to AMOS and I would very much appreciate your help.
AMOS does not, but you can just create a variable that splits into four (to indicate each group). Then you can do comparisons between pairs at a time: A:B, A:C, A:D, B:C, etc.
Hi James,
very nice and informative video about the model comparison. I loaded 4 Groups into my Amos Dashboard, but Amos still gives me only 1 "Nested Model Comparsion" Output like you did with 2 Models. Shouldn't i get more Information because i used 4 models? I would like to see if there are any significant differents between the 4 models in one of the variables. Maybe you can give me some information in how to handle that. Cheers and thanks in advance.
We can only calculate the differences between two at a time. The human brain (and consequently our tools) can only conceive of pairwise comparisons. This can be A:B, A:C, A:D, B:C, B:D, C:D, or it can be one vs all: A:BCD, B:ACD, C:ABD, D:ABC.
Dear Prof James Gaskin
Thank you so much for your informative video. I have a question about the confusing results of the structural invariance in my project. When I tested the full model by following your video, the p-value was insignificant. Thus, all paths are invariance across the groups. However, when I checked the estimated loadings in each group, one path was significant in group A and NOT significant in group B. I wonder if something wrong with my data. Thank you so much in advance.
That's totally possible. It can also be due to low sample size in one group.
Thank you so much for your reply. In my study, I have 250 in sample A and 242 in sample B. It is not a significant difference in sample size. Could you please give me some advice on how I could solve the problem?
There is no way to solve it because it is not a problem. It just means that it is significant for one group and not the other. This is perfectly possible. Just means more error in the non-significant group. You can try bootstrapping to reduce that error, but I doubt it will make a lot of difference.
Thank you so much for your useful recommendation. I really appreciate it.
Thanks for the video James, I am wondering about testing multigroup analysis and moderation using critical ratio in ur stat tools, is it still valid? and what difference between this new way and the old one (critical ration "z-score")?
Many thanks
+Riad Cheikh The critical ratios approach suffers from family-wise error. In order to correct for this bias, you have to include a Bonferroni's correction, which is to divide the desired p-value (e.g., 0.05) by the number of hypotheses being simultaneously tested. The test p-values must then be less than this new corrected p-value in order to be significant at the original level of confidence (e.g., 95%). The chi-square difference approach is far more accepted. I actually just came up with a new way that is very direct, but before make a video about it I'm going to get it approved through some experts.
Dear Dr. James Gaskin. Because there are many types in testing invariance across groups (e.g. metric variance, configural variance, scalar invariance...), Is the "Structural Weights" model in video above called "congfigural invariance" (same structure across groups)? Or what is it called, when we want to imposed equal structural paths for groups? Thanks and best regards
Structural invariance refers to multigroup comparisons. In this case, it is not a measurement invariance. And, whereas in measurement invariance we hope for invariance (no difference between groups), in structural invariance we hope for differences between groups.
Thanks again for your enthusiastic response. However, I still have several problems with my analysis. the CFA and SEM model are fine with model fit indices (e.g. TLI, GFI, CFI >0.9). Nevertheless, when it comes to multi-group analysis, these indices drop down poorly ( just around 0.87). Furthermore, according to your recommendation, global test must take place first, then P value. In my case, the model fit indices are underrated although the p-value somehow indicates significant difference between groups. I test with totally 3 groups: age, income and gender; and all of them revealed underrated indices. So, what should I report these outcomes in my paper? Do I necessarily get rid of this multi-group analysis due to poor fit? Thanks and best wishes
Keep the analysis if you passed measurement invariance. Report the drop in model fit. You might also look at the regressions table in the modification indices. This table will change based on which group is showing. Maybe address any regression it suggests if these will have a major impact on your model and can be theoretically justified.
Dear Sir
Thanks very much for sharing such great tutorial. I was wondering if you please guide me about co-varying the error terms. In my model, I have two exogenous variables (one 2nd order factor and one simple reflective construct), one endogenous/mediator ( 2nd order factor) and one endogenous/dependent (2nd order factor). My question is whether I can co-vary the error terms of the exogenous variables (2nd order factor) with other mediator and endogenous variables (2nd order factor). Thanking you very much for continuing such wonderful job.
It is generally considered best practice to not covary error terms. However, error term scan be covaried in special circumstances, such as for variables at the same level (e.g., two mediators, or two DVs). Dave Kenny has a website that explains this in more detail: davidakenny.net/cm/respec.htm
Good afternoon Professor Gaskin. I did a multi group analysis via spss Amos. One of my paths showed to be significantly different for men and women. However, both betas are insignificant. Did moderation still occur? Kind regards
Good question. This happens occasionally. You can say there is moderation, but that you cannot determine with confidence what the individual effects are, since the individual effects cannot be distinguished from zero (i.e., they are non-significant). However, you can at least say that the effects (although undetermined) are different for men and women.
Dear James, thank you. What sample size do we need to conduct path analysis in Amos? and what about the sample size of the groups we compare? Should they be equal? Any assumptions?
The minimum that I like to follow is 50+3x, where x is the number of included variables. Group sample sizes need to meet this minimum as well, but they do not need to be equal. If they are vastly different, like 65 and 500, then it may be best to randomly sample from the larger group to more closely match the smaller group.
@@Gaskination Thank you, it helps a lot. I read something everywhere saying min sample size needs to be 200 and you need 10 participants per parameter. It s confusing if you have just begun to learn something!
Btw, I agree with everyone, you and your videos are just amazing. I need to take more time for your channel.
Hello Dr Gaskin, I noticed that you did not include your moderating variable in your EFA or CFA, is this always the case ?
The moderator in this case is gender. As this is a single observed variable, rather than a latent factor, it does not belong in the measurement model. Only latent factors belong in the measurement models (EFA/CFA).
Dear Dr. James, thank you very much for this helpful video. My question is whether multigroup analysis with the binary outcome variable can be performed in Amos? If the Bayesian and MCMC approaches are used, no chi-square is reported. How can we compare the model difference? Thanks.
Amos should be able to run an MGA with a binary outcome, as long as the grouping variable is not also the outcome variable, and as long as all variables included in the model have variance within each of the groups. If no Chi-square is produced, then maybe it is because the model is saturated (i.e., every possible relationship is drawn). You need degrees of freedom to produce the chi-square.
@@GaskinationDear Dr. James, thank you very much for the prompt reply. I think my model has qualified for your mentioned conditions. The problem is by using the Bayesian approach, only "posterior predictive p", "DIC", and "Effective number of parameters" are reported in the fit measures. It looks like this approach doesn't produce Chi-square. Therefore, there is no way to do model comparison.
The Maximum Likelihood Estimate can be used to run an MGA with a binary outcome and get a result, from technologically speaking. However, the AMOS tutorial says that "In Amos, models involving categorical
responses require the use of the Bayesian and MCMC option" (No. 10, www.structuralequations.org/). Is that true? If it is true, how to do MGA analysis with a binary outcome? Thanks again?
@@jinxinliu7822 In this case, you would have to compare the outcomes without the chi-square. This can be done by comparing the regression weights and also by using an estimand that subtracts the regression weight of one group from another group. This will produce a confidence interval around the difference and estimate a p-value. I have an estimand that could be used for this. It can be found here: statwiki.kolobkreations.com/index.php?title=Plugins#LIST_OF_ESTIMANDS: It is called "MyGroupDifferences". There are instructions there on how to use it. I've never tried running an estimand with Bayesian estimation though, so I don't know for sure if it will work... Good luck!
@@Gaskination Dear Dr. Gaskin, thanks for this helpful suggestion. I'm wondering whether the maximum likelihood estimation can be used for a binary outcome in Amos. In another word, is the Bayesian approach necessary for a binary outcome? Thanks so much!
@@jinxinliu7822 I think Bayesian is preferred but not required. ML can be used. Just interpret your regression weights as 'more associated with the higher binary value' if positive and 'more associated with the lower binary value' if negative.
Hi Dr Gaskin,
Thanks for your highly informative videos!
Please could you clarify if it is okay to use the "groups" feature in AMOS when I have multiple categorical moderators?
For example, I have gender (male,female) and residency (resident, non-resident). Can I create those two groups as moderators in the same model or do I have to run separate models, one with gender as moderator, and a second with residency as moderator?
If I can use one model, then can I interpret results for only one (of four) group at a time or is it possible to interpret for 2 groups each, e.g. male AND resident?
Thanks a lot!
Good questions. AMOS can handle as many groups as you create. However, for simplicity of interpretation, I'd recommend only having two groups at a time.
@@Gaskination Thanks!!
I preassume that hostile sexism positively strengthen the relationship between low self-esteem and organizations, following your lesson if HS dampens the negative relationship, means do not support my hypothesis
correct. That would be opposite to your theorized position.
Hi Dr. Gaskin, can we test the multigroup effect if there is a big difference in sample size such as male n=250) and female (n = 70).
You can, but you are likely to find no differences (that are statistically significant), due to the amount of error in a small sample size. I would recommend assessing differences with less emphasis on p-values.
Dear Prof. Gaskin, I'm trying to develop two models (one for local students and one for foreign students) based on a structural model. Examples of my hypotheses are
H1: Peer influence has a direct effect on local student achievement
H2: Peer influence has a direct effect on foreign student achievement.
I using multigroup analysis to test the hypothesis and to determine the sig. path. Am I doing the right thing?
The baseline hypothesis would be something like:
H1. Peer influence has a positive effect on student achievement.
And then the moderating hypothesis would be something like:
H1a. The positive effect of peer influence on student achievement is stronger for foreign students.
Or maybe it is for local students; I don't know... And then you would test one model, with peer influence having an effect on student achievement, but you would test it across this grouping variable (local/foreign).
Hello again Prof. Gaskin, thank you so much for the prompt reply. I have two more questions:
1. If I have H1 and H1a. as mentioned. Do I need to find additional theory/studies to support H1a? What if my study objective is making a comparative path model between these two groups but not testing the moderation effect?
2. If I'm testing H1 and H1a. kind of hypotheses using multigroup analysis, is it right to present a path model for overall and for each group in the end? Thank you
@@noraasikinabubakar9514
1. You would need to supply some abductive reasoning and some supporting literature to justify the moderation effect. Comparing a path or model based on group membership is the definition of multi-group moderation.
2. Yes, this is a good approach.
Hi Prof. Gaskin, many thanks. I'm very clear now. Thank you so much
Inspiring.....Thank you How to get your excel program?
It's on the homepage of the StatWiki
If the relationship between two variables is moderated by three moderators and there are interaction effects among moderators, should we break the models so as to study two moderators and their interaction and then drop on moderator, introduce the third one and then study the interaction between the first and third moderator?
Also is there a reference that supports this?
It depends on your theory. If you are theorizing simultaneous moderated effects, then do them all at the same time. If these are not confounding, then you can do them separately. As for a reference,
Hayes, Andrew. F., 2013. Introduction to Mediation, Moderation, and Conditional Process Analysis: A Regression-Based Approach, New York: The Guilford Press
Can we do Multigroup analysis for other factors like Income, Education level, etc?
Yes. You can split income into low and high, or education into less and more, etc.
Dear James, just another question, what if my hypothesis is: "Female consumers have more positive effect of satisfaction on loyalty, when the role model is female, while male consumers have more positive effect of satisfaction on loyalty, when the role model is male"... Could you kindly suggest, what shall I do in this case, please? Do I have to create 4 multigroups, male-like-male, male-like-female, female-like-male and female-like-female or is there any other way to resolve this issue. Kindly suggest please.
Correct. Four groups. Two separate hypotheses (or a two part hypothesis).
Thanks for the reply, very kind of you. If I create four groups, it means I will be doing changes in my data set according to male-likes-male, male-likes-female, female-likes-male and female-likes-female and will be creating one construct (with scores from 1 to 4), right?
And thankyou so much for all the movies and information, you have provided. You are a real Angel and God Bless You for that :-)
You are correct. That is the right approach.
Thankyou so much James... very kind of you. I will surely add your name in my dissertation and publications, blessings xxx
Hi, thank you very much for this video. I do have two data set one from France and the other from Pakistan. i want to compare the results but the measurement models are different e.g. only 4 out of 8 observable loads on affective commitment from that data from Pakistan while from french data 7 variables of affective commitment out of 8 loads. so how to compare the results?
One way to do this would be to just use the variables that load on both. Another way is to run the model with all data (from both countries) and see what loads with all data.
Dear James, many thanks for this (and all your other) very helpful videos! One question remains, could you please help me with the following?
I compared two groups with the X2 test but am wondering how to report on it. As far as I know, a X2 test would normally be reported the following way: Χ2(df, n) = ..., p < ...
What would be the n in this case? As I have two, that of group 1 and that of group 2.
Thank you very much!
Report it as Group 1 n and Group 2 n, and then the X2 and the df are for the difference between the constrained and unconstrained models (rather than the actual X2 and df). Then the p-value is for the test.
How can I made a working model using of MIF ( Multi Influence Factor) method...please help
I've never worked with MIF. Sorry about that. Good luck!
Greetings.
I am conducting multigroup moderation analysis with three categorical groups.
Now my question is since i need to do comparison in Pairs A:B, A:C, & B:C.
Do i need to add all the three groups at once in the Model, run it and then do pairwise comparison.
Or
I need to run the model three times with only two groups in my Structural Model each time.
You can add all three groups and then just make paired comparisons. As for model fit, it should not change between these models since the models are the same. If there are large measurement differences between groups (i.e., measurement non-invariance), then perhaps you'll observe very small model fit differences. So, you only need to report model fit for the model once.
@@Gaskination by making paired comparison do you mean that we will end up having same
chi square difference for all the groups(in amos Model output Comparison option) and then look for path level difference to make inference ?
In other words we will have only one chi square difference value for a single path in model comparison for all the three groups.?
@@basimhussain7520 No, the chi-square difference will probably change depending on which groups you are comparing.
@@Gaskination if we will include all the three groups in amos in a single model and run it for moderation then in output we have only single model comparison chi square for a particular path for all the three groups.
Having said that will that one model comparison chi square difference suffice for making comparison for all the three groups ( by comparing their regression coefficient and significance level in output option- "estimates" in Amos.
( Here i am running model only once with three different categorical groups. Afterwards i check model comparison in amos to see the chi square which is common for all the three groups- since amos shows only one cmin/df for a single path for all the three groups. Thereafter i see the regression estimate of the three groups to make pairwise comparison)
Is this right way to do it.
Or there is any issue with this approach?
Sir,
Greetings. I have enjoyed your video very much. However,
I have collected data on a study entitled "Relationship between employees Satisfaction and Loyalty: Comparative study of Public vs Private bank". 4 factors in satisfaction (Data were collected in Likert scale). Data were collected from two group of respondents (Private bank; Public Bank)? I had also run Structural Equation Model in AMOS.
When i inputted data in SPSS; Type of bank was considered as nominal Variable (1 for private; 2 for public bank). All data from the respondent ( employees of public bank and private bank) were recorded in a one data set (that means, i have created one data set). I would like to test the hypothesis: there is no significant difference in employee satisfaction and loyalty in case of private bank and public bank.
Can I run multiple Group analysis in AMOS? Does require two set of Data or Single Set of data (where both employees from public and private)? How can i prepare two set of data from single set of data? Could you please help me?
How i can do this? please suggest me sir.
Thanks in Advance
Just use one dataset and do a t-test or ANOVA where the factoring variable is BankType and the dependent variables are satisfaction and loyalty. This will show you whether there is a difference between private and public bank employees in terms of satisfaction and loyalty.
Respected Sir
Thank you so much for your valuable guidelines. Just one more question
can I run Multi group analysis in amos for testing the hypothesis whether there is a significant difference between private bank and public bank in terms of satisfaction and loyalty?
Thanks
Multigroup in amos will let you determine if the relationship between satisfaction and loyalty is different when comparing public to private banks. Again, the multigroup is for testing differences in the *relationship* between variables, but will not tell you if public or private banks have more or less loyalty or satisfaction.
Dear James, thanks for the helpful instructions! I would highly appreciate if you could help me with the following problem. I did a normal path analysis which run smoothly and then I conducted a group analysis with the same model with 4 different groups. The notifications "sample moment matrix is not positive definite" pops up and suggests 4 reasons for this problem. How is this possible that only by adding the groups to the path analysis this problem comes up? I was desperately looking for a solution and would be really happy if you could help me - Thanks a lot!!!
This might be because one or more of the groups has a small sample size.
My path model includes 1 independent, 3 mediator and 5 dependent variables. The groups have a size of 48, 72, 75 and 90 participants - is this already too small? Thanks for your support!
48 is pretty small. See what happens when you exclude that group. The problem with low sample size is that there is insufficient data to reduce error and thus minimize (execute) the model.
I tried that after your first message and it was still not working! Then I need to find a different solution.. Thanks for your quick support!!
You might also try looking for clues in any results it gives (if it gives any results) - look for high standard error. Try removing a single variable at a time until you find the one that is giving you the trouble.
hello dear James. i would like to know that if we have one model and we aimed to comparison study between two states , how we can compare our results ? could we run SEM separately for states ?
Multigroup analysis makes more sense. Just follow the video above. This is a good way to test differences between two groups (such as states).
Thank you so much
Dear Professor Can you please make a whole video lecture about multilevel analysis with smart PLS, as in my thesis there are 3 models two based on individual level and one is on team level please guide please
Funny story... I've never done multilevel analysis. I've just never been involved in such a study. So, I'm not sure how to do it. It's on my list of things to learn eventually.
Thank you professor for your reply please suggest any other channel or something i am stuck in this@@Gaskination
@@qaisermohiuddin413 Having never done it, I'm not sure who to point you toward. I checked another channel I follow, and they don't have it either. I guess you'll just need to google it or search on RUclips.
Professor Thank you for your response @@Gaskination
Hi James, i'm wondering if i should use this one or the another one in this link ruclips.net/video/mirI5ETQRTA/видео.html that you post earlier.
- This one is using the fully constrained model,
- The one in the link as i see used the partly constrained model
Some of my friend thesis are using partly constrained model, but i based mine on fully constrained. is there many different between two way? or i can use both?
oops! I should have left the whole model constrained EXCEPT the path I was testing... Sorry about that. Good catch. Sorry also about the late reply. Somehow this comment didn't make it to my inbox. I only just stumbled upon it while responding to another comment.
@@Gaskination last thing, at the end of the clip, when we check if the relation is stronger from male to female, you said the p-value < 0,05 means the regression p-value, right? since the chi-squared test's value already < 0,05 mean they are different.
@@phuongnguyenoanviet9594 Sorry, that wasn't very clear. It would be good to check the p-value for the regressions and for the chi-square difference.
is multigroup analysis equal in meaning to multilevel modelling?
No. These are different. MGA is to compare multiple groups. Multilevel modeling is usually referring to modeling that includes group levels, such as country, state, city.
@@Gaskination would you please make a video about multilevel modelling?
@@hishonline I've actually never done it. I haven't come across the need to. So, I don't have the expertise to make the video for it.
If we want to compare 3 groups how should we do? Thanks in advance
Two at a time. A:B, A:C, B:C or A:BC, B:AC, C:AB
first seems good, thanks