Это видео недоступно.
Сожалеем об этом.
SEM Series (2016) 6. Multivariate Assumptions
HTML-код
- Опубликовано: 21 апр 2016
- Multivariate normality, outliers, influentials in SPSS using cook's distance. BEWARE, there will ALWAYS be multivariate outliers, even after you have removed some. I don't even address multivariate outliers anymore through the Mahalanobis D. Instead, I use Cook's D: • SEM Series (2016) 6. M...
i like all the videos, and are very helpful
Hi James, thank you very much for the video and the series. It's super helpful! I was wondering if a few paths in the path analysis do NOT have linear relationships, but most of the other paths do have linear relationships, can I still run a path analysis that assumes a linear relationship? The paths that do not have linear relationships are between a predictor and two mediators. I am interested to test the indirect effects that involve these variables. Thank you for your help in advance.
You certainly can, but just make note of it (the curved effects estimated as linear) when you write it up.
Hi James,
You're pretty cool! It's very useful for me!
But I have a question. How did you creat another 8 new variables in the new dataset? In the video at the bottom SPSS data freme of 1:04.
Thank you so much!
Here is a video showing how: ruclips.net/video/dsOS9tQjxW8/видео.html
Hi James, thanks for this series so far. I am just wondering if I missed any part of the video as I do not know how come the items useful-1 to useful-7 have been converted to 1 single construct Useful (same as other items)? Thanks again.
Here is the full playlist: ruclips.net/p/PLnMJlbz3sefJaVv8rBL2_G85HoUko5I--
Look at the end of the CFA part 2 video for factor imputation.
James Gaskin thanks for your reply, I always support your channel and recommend my friends to subscribe
Hi Prof. James,
Thanks for the wonderful video. But I have some doubt related to the multivariate normality.
Q1. What is the best way to detect an outlier and test for multivariate normality? mahalanobis distance or Cook's D ?
Q2. When you have multiple antecedents (constructs) and multiple consequences for a second-order construct, then what is the way out for testing multivariate normality of the whole data?
I think Mahalanobis D is too strict. I would definitely go with Cook's D for multivariate data.
Hi James, How do I get the values for Latent variables if I did not do the Data Imputation in CFA while checking for Common Method Bias ?
You help is really appreciated
You can do data imputation without CLF as well. Or you can do it during an EFA in SPSS by checking the box in the Save options (in EFA) to "save factor scores as variables).
Mr. Gaskin, you are doing an amazing job.
I have two questions.
(1) Some researchers say that there is need to find separate run of EFA test for each single variable? If no, please share some authentic references
(2) What about univariate and multivariate normality tests? Is there any difference? Please elaborate.
1. You don't need a reference to justify running the EFA with all reflective latent first order factors. It is common practice and the most appropriate action. Running them separately is useless, because it prevents you from being able to determine convergent and discriminant validity, which is the primary purpose of the EFA.
2. Univariate is for individual variables. Multivariate is for relationships between variables.
James, say I have 5 mediators, is it okey to correlate them with one another? Theoretically, it makes sense that they are correlated. Or, should I just covary the errors as recommended by the modification indices? Thanks!
You won't be able to correlate them because they are endogenous. But you could covary their errors, or draw regressions between them if that also makes sense.
Hello Dr. Gaskin! So helpful videos,thank you so much! I would like to ask, what is exactly the ID variable that you use in the graphs with the Cokk's distance ? I donot understand which variable from my dataset I must use. Please, if it is possible, make a video to show mediation analysis with only latent variables. Another big problem, that I wasn't able to solve, is how do I perform SEM with non-normal data (sample size N=400, 45 observed variables representing 10 constructs)? Bootstrapping is recommended?
The ID variable is just an artificial incremented number. You can just create a variable for the row number and use that. I do have a video for mediation with latent variables: ruclips.net/video/-j3LkADfgWs/видео.html It's toward the end of this video. 400 is a pretty large, and it will mitigate most issues due to non-normal data. To further mitigate, you can bootstrap. If you want to go even further, you could use Bayesian estimation, instead of maximum likelihood estimation.
Dear Dr.Gaskin, thank you for your reply! I have some additional questions..If regression weights are greater than 1 in a full SEM model how can i fix this? I mean that some paths from the IV'S to the DV's are greater than 1 specifically. Is it really a problem? I think this indicates multicollinearity issues, which I don't understand why we should check multicollinearity using the factor scores that emerged from the CFA. Model fit is quite good although...Unfortunetely, I also can't download the plugin AxB estimation so as to run a Mediation analysis. Do you know what might be a problem with the plugins?
1. Regression weights can be greater than 1. But STANDARDIZED regression weights should not be greater than 1. If they are, then there could be normality issues with your data. Make sure to check for that right from the beginning (even before EFA).
2. The estimand is called "MyIndirectEffects" now. It should be available from the 'Plugins and Estimands' link on the left navigation panel of the statwiki.
Thank you for your videos! Do you have a citation for the argument that large sample sizes mitigate issues of non-normality? I have data that have high values for Kurtosis and also have a sample size of 400.
Hi Dr.@james Gaskin, i have a small sample of 120 with non-normal data (5/9 variables), watching your SEM series, do i perform bootstrap at every stage? if so do u use bootstrap ML? videos on bootstrap ruclips.net/video/DktTdTnFkCI/видео.html didnt clarify this. Also, i read that u mentioned imputing data without CLF for some error i faced. what will the difference be and do i need to account for anything? would imputing data as Bayesian instead of regression help with non-normality of data? seeing that you suggested bayesian estimation for non-normality data.
hello dr Gaskin. i have problem with multivariate, which is i have unnormality data thats about 14.000. even thought , to be normality is needed about -2.00 - 2.58. how to fix that without using bootstraping?
If you mean the skewness and/or kurtosis values are 14, then this is very high. Perhaps it is due to the wording of the question. For example, if you ask people if they think their children are smarter than other children, they will all respond very positively.
Hi James,
Thank you for tutoring! I love all your videos! However, I have a question about master validity: if two important factors (cannot delete either one) are highly correlated, and the square root of the AVE for factor is less than the absolute value of the correlations with another factor. Do you have any way to fix it? (I tried the ways you showed in previous video already, but it did't work)
Another option is to combine them into a second order factor (if that is theoretically reasonable). Another option is to conduct an EFA with the items from just those two factors, and use that EFA to try to separate them. Then return to the CFA using the EFA as guidance for which measures (items) to retain. Another option is to first assess and extract any specific biases (such as method bias), which will tend to raise the correlation between factors, and then check validity.
@@Gaskination I tried all of these already. EFA shows good and the problem factors can't be combined into a second order factor due to the theory. If you want, I can send you the outputs through email. : )
Nice. What if there are two factors of a second order reflective construct that are showing VIF way higher than the threshold of 10? What to do about that? I want to keep the variable as its important for my study but you said one way to treat this problem is to create a 2nd order factor but they are already the first order factors of a 2nd order construct...Any hint?
In this case they are expected to be strongly correlated, and you won't be predicting anything with them. Instead you'll be using their parent factor to predict other factors. So, it is only a concern if the parent factor has a high VIF.
Hello Sir, I have a question again. I have a dependent variable which has single item construct measured with 5 point likert-scale. So it has ordinal data. How can I do these cook's distance and Multivariate normality tests. Because you used linear regression. I did my analysis EFA and CFA up-to this point without this DV. Can I do these tests only for remaining IVs and DVs and add my ordinal DV in the last causal model. Is there any tests I need to do with my ordinal DV? Please help. Thank you :)
Since your DV is a single observed variable, you can do this cook's distance test just during the causal model.
How to conduct Mardia's multivariate skewness and Kurtosis? Thanks please
I'm not sure. I've never done it. I know some SEM software calculates multivariate normality. AMOS does (using Mahalanobis D).
Hi James, I have an amos version with no plugings at all, my question is how to get the plugins?
Plugins don't work on versions earlier than 20 I think.
Hi sir! I am actually doing research study for my degree and I basically just following your videos and statwiki website for guidance. How can I cite you as one of the reference of my study? Thanks!
Thanks! Check out the homepage of statwiki for a guide on citing some of these resources.
What if I have only one variable that has an VIF greater than 3? what does that mean? and what can I do in this case? 3.67
That VIF is not too bad. I would not worry at all. I would only worry if it is over 5.0.
Hye Dr james, I am currently analyse my data for thesis. Just to have a confirmation, as i read many sources about the multivariate analysis. If i am using AMOS, is it outliers and multicollinearity is enough for multivariate normality?.... I can achieve the univariate normal distribution through skewness and kurtosis, but not the multivariate normality. I am stuck. The model fitness, CFA and validity are all fine. Hope you can clarify me and thanks in advance
Most do not worry about multivariate normality. However, they do check multivariate outliers (through Cook's D or other distance measures) and multicollinearity (through VIF).
Dear James
Many thanks for this video. I have a question: my model has 9 independent variables and 2 dependent variables.
X1, X2, X3, X4, X5, X6, X7, X8, X9 ->Y
Y-> Z
So, how many COO I have to make? I made COO 1 between X1 and Y already. The 2nd one I intend to make X1 to X9 (independent variables) and Y (dependent variable), right?
I do appreciate your help!
I don't understand what you mean by COO.
It is the COOK's distance. First, I made the COO_1 with X1 as independent variables and Y (dependent variables). Next, I made COO_2 with all of my independent variables from X1 to X9 and Y as dependent variables. Is it enough or what else I have to do to find out the outliers?
Oh. I see. Yes, that should be fine. If you want to be really thorough you can do it separately for each X, but altogether should be fine as well.
Sir, Do you have video on Bollen Stine bootstrap amos while model violating multivariate assumption?
I do not
Dear Professor.I am conducting a research on covis patients. My dependent variable is measured on likert scale. My predictors are also measured on likert scale and some are categirical.
Can I run multiple regression on SPSS
I want to rumregression analysis .
You can run multiple regression if each construct is represented by a single variable. If there are multiple variables per construct, then you'll first want to create averages, sums, or factor scores (during EFA, save factors as variables).
@@Gaskination Thank you very much for your reply
Is it necessary to achieve multivariate normality for the model in AMOS as I have read that large sample size mostly deviates from 'Multivariate normality'. Do I need to consider 'Multivariate normality' or not? If yes, to how much extent?
I almost always ignore mahalanobis distances in AMOS, as it will always report something, but these deviations from multivariate normality are not necessarily biasing or bad data. They may simply represent those who do not conform to the theorized model. Instead, I will often look for Cook's distances just in case there are extremely influential multivariate outliers that are unevenly bearing weight on the model: ruclips.net/video/J2EkjIeK-PE/видео.html
Dear James Gaskin,
I really need your help. Can you please tell help me?
Question: I am measuring the impact of socio-demographic variables (e.g. age, gender, marital status, etc.) on workplace incivility and I have conducted ANOVA & regression analysis. But, the reviewer is suggesting me to conduct ANCOVA / MANOVA. Can you please tell me is this the right process?
I don't want to conduct ANCOVA.
I will be extremely grateful to you.
ANCOVA and MANCOVA are not too difficult (although I don't have any video on them...). They allow the researcher to examine the effects of multinomial variables (such as marital status) on some dependent variables. This works well when there are no mediators or moderators in the model.
James
You used "Cook's distance" to check for outliers. Some people talk about "Mahalanobis distance". What is the difference?
Cook's distance and mahalanobis distance are very similar. Both measure the residual or distance away from the line of least fit (more or less) for a specific case. Either will suffice for testing multivariate normality.
i,ve five questions for each 9 variables (7 IV and 2 DV) ive have tried to eliminate outliers through mahalanobs distances but each time when i did this new outliers show up in data set. can you suggest something on this.
That is correct. There will always be multivariate outliers. I don't even address multivariate outliers anymore through the Mahalanobis D. Instead, I use Cook's D: ruclips.net/video/J2EkjIeK-PE/видео.html
James Gaskin Thanks alot sir