This channel is dedicated to understanding Structural Equation Modeling. The videos for this channel focus on step by step instructions in order for you to run an analysis with your data.
Thank you Prof. Please my little quesrion is, would I be right to say that this approach suits a structural model as you have here and not CFA measurement model? (Just for clarity for those like me who have struggled through several vdeos to understand the difderences in approaches). Thank you 🙏🏽
Thank you a lot! May I ask: is control variable age numeric or what else? If I have income (ordinal scale) or education level (ordinal scale) as control variables, do I need to transfer these variables? Thanks again
Hi! Thank you for this, it's very helpful! For a moderated mediation model, I've been told Process is a better fit when the moderator is a multilevel categorical variable. SEM is better suited for continuous moderator. Is this true?
This not taken from an article. This is my template based on my own publishing history. You can cite my book Applied Structural Equation Modeling using AMOS if you need a citation.
With invariance testing, you are constraining the measurement properties to be equal across the groups. You are then examining to see if a significant difference exists. Unconstrained means the two groups are assessed independently and no relationships is constrained to be equal across the groups.
Hello Mr. Collier, I have an amos result where the GoF is fit. However, the CMIN/DF value = 2.980. I read in your book on page 66 about measurement model analysis in AMOS, that CMIN/DF < 3 can be said to be fit. What is your basis, saying that it can be <3? While others say <2. Please explain. Thank you
That is based on the work of Kline 2011 (referenced at the end of chapter 4). You will also see some say an "acceptable" fit is anything under 5. CMIN (chi-square) can be biased especially with large samples. Understanding the relative chi-square: chi-square/df will present a more rounded approach.
Thanks. The data is this video was made up for the book Applied Structural Equation Modeling using AMOS. I needed a clean data set to work with. Saying that, the original research using these variables was published in the Journal of Business Research. The title was: Idiosyncratic service experiences: When customers desire the extraordinary in a service encounter
@@joelcollier9387 I apologize for any confusion. I would like to cite your method for testing the moderation effect in my paper, so I need a scientific reference for your method. I don't require a reference for the data. Thank you! :)
@@yl3355 You can cite the book but you can also cite Marsh, Herbert W. Zhongline Wen, and Kit-Tai Hau. (2008). "Structural Equation Models of Latent Interaction and Quadratic Effects", in Gregory Hancock and Ralphy D. Mueller (eds) Structural Equation Modeling: A Second Course. Greenwich, CT: Information Age
1. In the moderated mediation section, you described it as "Moderated Mediation With a Continuous Moderator". What would be the method if there was a model with latent variables with sub-dimensions? 2. When you constructed the model in your book, you were taking the interaction structure as CentreAdapt_X_Friend and the adapt and friend structures as normal. However, in Figure 7.40 I see that you take the structures adapt and friendly as comp_adapt and comp_friendly. Is this a specific exception to moderated mediation?
Trying to test moderated mediation with higher order constructs (with subdimensions) is extremely difficult. The reason why is you need to form an interaction term with the moderator and the independent variable. The most popular was to handle this is to create composite variables. To your point 2, the book talks about comp_friendly and comp_adapt. This is where the latent variables were created into composite scores for each construct. This is not unique to moderated mediation....it is just the most efficient way to handle a complex problem. Can you test moderated mediation with latent variable? Yes. I have another video on RU-vid called how to test moderation with Unobserved variables that explains how to do this. Hope that helps.
Hello, first of all, thank you very much for your reply. I have improved my analysis by following your suggestions and your book. If you have time, I have a few more questions. 1. For example, considering figure 7.19 in the moderation section of your book, if "Adaptive Behaviour" has 3 sub-factors and "Customer Delight" has 2 sub-factors, will there be a change in the analysis direction of the model? 2. For the Modareted Mediation model, I arranged it as Path Model, Full Indicator Model, Mixed Model in accordance with the instructions in the book. Unlike yours, the Interaction variable gave slightly different results in all 3 techniques. Even the path model was meaningless. Other variables in the model were close to each other. Could the sample have an effect on this? Unfortunately, my sample is limited to 205 due to the subject. There are also 49 variables in the model. 3. When I created the "composite variable" for the path model, unlike yours, the variables had sub-dimensions. I first took the average of the sub-dimensions and then calculated the value of the variable by taking the average of the sub-dimensions. Is the path I followed correct? I am grateful for your contribution in advance.
@@cihanuyank9749 So lets just clean up a little terminology first and that should help things. Adaptive behavior does not have "sub factors" those are the observed indicators or measurement items for the construct. To form a composite variable, one usually takes an average of all the observed variable scores and then you have a single score that represents to the construct. You will most likely have different results if you run this moderated mediation model as a path model and then as a full indicator or mixed model method. The reason why is you are accounting for measurement error in the full indicator and mixed model. You are not accounting for measurement error in the path model. Your sample is a little small for that much analysis and that might have something to do with it. With that small of a sample, I would encourage you to run it as a path model and have all your constructs as composite variables. It will make things easier for you.
Hi Joel- thanks for the informative video! Question- is this bootstrapping required for the final CFA model and final SEM model? or is it just for the final SEM model? (currently writing a master's thesis and it's greatly appreciated) :)
@@joelcollier9387 Thank you Joel! Final question- the p value for the Bollen-Stine bootstrap is non-significant (0.642), but it states that the model fit better in 1788 sample, and worse in 3212 samples. Can it still be claimed that the model fit is adequate with the bootstrap samples?
AMOS can not account for Sample weights. You need to weight the data in SPSS (new data column) and then pull that variable into AMOS. Little bit of a work around
Sir, please tell how to do Invariance of Formative Measures (Structure invariance, Slope invariance, and Residual invariance)? Please share a link of the video if u have one...
If the measures are formative to an unobservable construct, then you do not have to do an invariance test. Invariance testing is for unobserved constructs with reflective measures
Sir, please tell how to do Invariance of Formative Measures (Structure invariance, Slope invariance, and Residual invariance)? Please share a link of the video if u have one...
Hi Dr Collier, if i'm going to run a study to see the mediation effect of iq on male and stress, and the mediation effect of exam scores on stress, i am only interested to see the efffect on male sample x=gender (male 'x1'/ female 'x2' ) y=stress m1=iq m2=exam scores 1. compare stress between male and female (t-test) 2. male predict iq 3. iq predict exam scores 4. iq mediates male and exam scores 5. male predict exam scores 6. male predict stress 7. exam scores predict stress *1 is t-test *2-7 are mediation analysis i'm going to split the study into 2 parts, for the first part is the t-test comparing the gender on stress, then in the second part is the mediation, so is it possible to run the mediation just with 1 category (male) as x?
Short answer...no. You can not run a categorical variable with just one value. You will have no variance in the data. If the value is categorical you are going to need to assess at least two categories
My structural model fits (with all the data). However, when I run a group analysis the Amos output message is "iterations limit reached, the results that follow therefore are incorrect". is there anyway I can fix this? . Please a second question, what happens when one of my constrains ends up in a model that does not fit , Amos wont stimate the chi-square results, how can I assess if the difference is significant or not for those cases?
There are a lot of reason for an iteration limit....if I am taking my best guess, you have multicollinearity in your model. Look for the two unobserved variables that are highly correlated with one another. You may have so much overlap that AMOS can not get a unique solution
Thanks prof. for your kind reply. Please tell should i go for measurement invariance testing(MICOM) of my groups, if i have created them from the same collected data using cluster analysis on 4 of my observed items (assessing attitude). Although all my respondents are tourists, but they differ in their attitude, so later on after clustering, i want to know differences in them on "experience and satisfaction relationship". Shall i do measurement invariance, or how shall i proceed...?
If your survey items are the exact same and you do not believe that your groups are fundamentally different, then you can proceed without doing an invariance test. Saying that, I almost always see reviewers asking for an invariance test in multi-group analysis even if there is no fundamental difference between the groups. I would say go ahead and perform it because it makes your study look all the more rigorous.
If all the groups to be compared are different, then their responses to the questions would also vary, as per their attitude & perception. Then why do we want to establish " Measurement Invariance' before conducting Multigroup analysis?. Although the questions asked are same, but the responses are different for the different groups, so how can they possess measurement Invariance.... Please explain.
With a two group analysis you are looking for differences in the structural relationships across constructs. You are not looking for differences in the measurement properties of the construct across different groups. If you think managers and salespeople are going to interpret a question differently (even if it is worded similarly) then your differences have more to do with how the construct is measured as opposed to group differences. Thus, you would need to perform a measurement model invariance test.
@@joelcollier9387 Thanks prof. for your kind reply. Please tell should i go for measurement invariance testing(MICOM) of my groups, if i have created them from the same collected data using cluster analysis on 4 of my observed items (assessing attitude). Although all my respondents are tourists, but they differ in their attitude, so later on after clustering, i want to know differences in them on "experience and satisfaction relationship". Shall i do measurement invariance, or how shall i proceed...?
There is a lot of debate around this. If you use PROCESS, it does not even give you model fit statistics. If your model fit is bad in SEM, then it points to a model that is not representing the data. In essence, the model does a poor job explaining the data collected. Saying that, it does not mean it poorly explains certain relationships. It means that the totality of your model does not explain the data very well.
@@joelcollier9387 thank you so much for your response. Please would you recommend any references I can check to dig into this discussion ? I am interested in a certain relationship between two constructs of my structural model. The regression weight is 0.24- which is low, but it exists.However, my CFI is 0,.86, Is there any way I can address how reliable this relationship is since the model doesn't fit?
Thank you for the video, what time is the best to do the bootstrapping, my model is not fitting so I am working on my modification indices, I have reach a point were I dont think I can improve more my fitting indicators with the MI. Should I transform non-normal variables to (near) normality before running the model?
My best guess is if you are having model fit issues...you most likely have measurement issues with your latent constructs. You might want to find the troublesome items and see if it warrants deletion if it is not contributing to the validity of the construct.
@@joelcollier9387 thank you for your response, I will look at my latent constructs. My data does not follow a normal distribution, some literature suggests to transform the non-normal variables before uploading them to Amos, if necessary (Niels.J , 2013)., would this approach be far different from bootstrapping?
Thank you for the video! I selected the test for normality, but the output does not include this analyisis; my data set has missing data - this may be the reason?. I was planning to use the data imputation option; however, I am unsure how this will affect the normality test results. Do you have any suggestions?
I always encourage you to impute data in SPSS/SAS/Excel before starting to analyze the data. It is the easiest way to address missing data. You can do it in AMOS but it is not that easy. Try imputing your missing data and then try it again.
Thanks for the video! What would be the difference between analyzing a control variable (for example, Gender Female /male) vs analyzing the model with groups (group 1: female, group 2 male) when to use one or the other process?
Control variables are trying to account for variance in your model in explaining the dependent variable. Two group analysis is examining if relationships are significantly different across the groups. Control variables are necessary if you think that a variable will change the outcome of your dependent variable....for instance Age and technology acceptance.
Yes, you need to assess model fit before you even start addressing the two group analysis. You will have a model fit statistic for each group that needs to be presented
Great! I have a question regarding the measurement and structural models with control variables. When evaluating these models, should we include the control variables in the model and assess the model indices, or do we need to remove the control variables to assess the measurement and structural models?
I assess model fit before adding control variables. Many control variables will have relatively no influence but will cause unexplained variance in the model. In my opinion, assess the model fit with the initial structural model and then add you control variables after that. Is it wrong to assess model fit with the control variables included? No but it will usually produce more unexplained variance which could hurt model fit...especially if the control variables are not significant.
Professor, this is a very great explanation. How can we get this Customer Delight Data. Sav file ? Professor, I got your book. Please help me to find the dataset to practice myself.
Can we apply the principles of PROCESS for constructing structural equation models (SEMs) in simple models that involve one independent variable (IV) and one dependent variable (DV) in the context of moderated mediation models?
Yes, you can. With one DV the analysis will be exactly the same. PROCESS is quicker to use because of the preformed models. If the preformed models do not line up with your model, you are better off using SEM
Great! Can we apply the same approach to examine mediation (not for the independent variable) and determine whether the moderator moderates the effects of the mediator on the outcome?
Yes, you absolutely can do that. You will just have an interaction term with the mediator and moderator to the DV along with a moderator with a relationship to the DV.
Thanks! What would be the indirect effect of ID on DV? I mean do I need to have a direct effect from ID to the moderator and interaction term as well?@@joelcollier9387
If you are talking about the degrees of freedom calculation, it is : Rigdon, Edward E. (1994), "Calculating Degrees of Freedom for a Structural Equation Model", Structural Equation Modeling, 1 (3), 274-278
what about multivariate kurtosis value? Your data is not normal in this example. What about CR value. That has to be between +-1.96 to asssess if data is normal
Thanks for making it easier for us to understand, can you please share link to reference for those ranges as i have tried search Arbuckle, James L. (2017) but unable to find kurtosis -+10 range. I need to quote in my paper, please
No. One of the assumptions of SEM is that the DVs are continuous. You might want to try the PROCESS macro otherwise you will need to do logistic regression to find your answers
Thanks for the great explanation. I am attempting to determine how to find in AMOS or calculate the t-value you have in noted parentheses on your template table. Would you have a quick explanation? Thanks!
In the output in the estimates tab, you will see a column called "C.R". That stands for Critical ratio which is also the T-value for that specific relationship you are trying to test. Hope that helps.