Dear Professor James Gaskin, I am really thankful for this. This session provided meaning to some of the many separate readings and books I consulted. I loved the tips and your style to see statistics in a fun and interesting way. This is key to learn and acquire knowledge. So kudos to you! The fact that it is live and there is interaction is even more interesting.
Great video James thanks a lot. These are so helpful. One question if you would be so kind to respond. I am trying to follow along using the SPSS file in the link provided. However this appears to be different to the one in the above video (there are more items). I have tried to manually select the same items as you do in your video (I get 47) but im still slightly out (KMO 0.901 and df 1128). Is there a link the original data set used in this video. many thanks
Dear Professor James Gaskin, thanks for the informational video! I was wondering if it's possible to do EFA with a questionnaire with more than 200 items? I'm developing a personality questionnaire that taps into over 30 constructs based on the Five-Factor Theory. However, in the EFA examples I have read or seen, it's usually done with less than 100 items (i.e. the example you provided in your video). Due to the complex nature of my questionnaire, should I split the EFA in 5 sections, each investigating the items relation with each other within their respective big five? OR should I do it as a whole?
Wow! That's huge! If you get enough data (like n>800), then you could try it all together. It will probably have lots of discriminant validity issues though. In this case, you can separate it into a sort of sensitivity analysis by doing an EFA with the items from two personality types at a time. Then alternate to make sure you have good discriminant validity between them all.
Hi Professor Gaskin, Thanks for the reply! So just to clarify if I understand you correctly, what I could do is look at the items for each big five-factor (i.e. OCEAN) separately or in pairs (e.g. look at Extraversion and Agreeableness items together in one EFA)? If I had a conceptual idea about which items group together, and do EFA on a scale rather than item level, would that be considered valid?
@@chuwainicholaschan190 I agree with the first half of what you said. The last question is probably not advisable since these would each then represent distinct dimensions through a single computed variable, and would therefore not load onto the same factor.
Hi - For your example of performing factor analysis with a formative factor, which you mention around 1:50:00... You suggest you can give each item a score (e.g., 1) and sum them up to create scores/frequencies/etc for your factors. However, I thought you needed at least 3 items per factor to perform factor analysis, and creating a score will essentially turn this into one item. You also wouldn't be able to measure reliability of the factor using Cronbach's alpha. Am I missing something?
The score would create a single item that you would then use in a causal model, not in a factor analysis. Cronbach's alpha is not relevant to formative factors anyway, because it assumes indicators are correlated (which is not an assumption of formative factors).
Dear Prof. Gaskin, thanks for your vedio. I am wondering is it right to do EFA and CFA using the same data set. Or should I use a small sample to do EFA and then use the bigger data set to do CFA? P
It is considered more rigorous to do them on separate samples, or on random samples of the same data. However, most researchers are often struggling to get enough data. So, they do not have this luxury. If you have enough data, then it is recommended to do it on separate samples.
You can do a CFA in a PLS software, such as SmartPLS, but do not do any EFA, as the EFA is built upon the assumption of inter-item correlations (which is not a requirement of formative measurement). Here is a useful video: ruclips.net/video/Qx_7pN7eAes/видео.html
Dear Professor, I did 5 interviews and converted those interviews to frequencies. Now I need to run a factor analysis to see whether some codes relate to each other. My number or variables are greater than 5 (cases/participants). What should I do with this? Could you please give me your opinion? Thank you
At 32:06 you mention that the optimal number of manifest variables to describe a latent construct is 4. If you could please provide a citation for this it would be most appreciated (possibly Hair?).
I wish I could remember. It wasn't Hair though. I just did another search for it and couldn't find it. Sorry about that. If I come across it again, I'll make sure to post it on my references page on the statwiki.
Do you know where in the book Bollen mentions this? I searched the whole book just now and couldn't find it. If not, no worries. But, if you have a page number or quote, that would be awesome!
Hi Sir Hope you are fine. I have a query. If i am using all borrowed scales in my research. Do I still need EFA? As people have different views on that. Kindly suggest Thanks and Regards Kavya
I always conduct an EFA, even when using scales borrowed from others. The EFA identifies discriminant validity issues better than a CFA. Nevertheless, if your CFA shows good validities and model fit, then you are probably fine to skip the EFA.
Can we do exploratory and confirmatory factor analysis in the same data set? In all your EFA and CFA you used the same data to build the model. If it is correct please can you send reference thank you
Most studies do EFA and CFA on the same dataset, although it is considered more rigorous to do it on separate datasets, or separate samples from the same dataset. However, most scholars don't have enough sample size to split the data. As for a reference, I'm not sure there is one. It is just common practice.
The KMO and Bartlett’s test whether the set of variables are adequately correlated, but not whether the variables are normally distributed. For univariate normality, you can just check skewness and kurtosis.
@@Gaskination thanks. What if the skewness and kurtosis don't meet the criteria as well? Is it okay to use Satorra-Bentler or Yuan-Bentler adjustments? (My items are Likert rating)
@@sarahoseingholizade9155 I'm not sure of any effective way to transform Likert scale ordinal variables. If the skewness and kurtosis is less than 10, then there may still be reason to use them. I think it was Sposito (or maybe it was Kline...) who said up to 10 is probably okay under certain circumstances... Here are both citations: Sposito, V. A., Hand, M. L., & Skarpness, B. (1983). On the efficiency of using the sample kurtosis in selecting optimal lpestimators. Communications in Statistics-simulation and Computation, 12(3), 265-272. Kline, R. B. (2011). Principles and practice of structural equation modeling (3rd ed.). New York, NY: Guilford publications.
I host one every May in Utah, but sometimes also do them at other universities throughout the year. Often it is less expensive for a university to fly me out and compensate me than to fly out all their interested students and faculty.
Hello sir.....i need ur opinion...in my study, i measure women empowerment in terms of pre and post analysis....i asked same questions to respondents twice...One for after intervention and another for before intervention....do i need to conduct EFA/CFA for Both set of responses or just one /post scenario in enough???
It would be best to do EFA/CFA for both separately, just to make sure they are both valid sets. The other option is to do it for the change from pre to post.
James Gaskin thanks for ur respond... actually i intend to measure empowerment from pre to post using SEM, in which latent variables are treated as observed variables.....do i still need to conduct EFA and CFA???
factor analysis should be done for all reflective latent factors. If you don't use them as latent, then you can just create averages. But if you use them as latent in any causal analysis, then you should do factor analysis.
Dear Prof. Gaskin, Please help me with a query. In my research work, I have used questions from 5 existing univariate scales to measure perceived discrimination at work. I did not modify, adapt or adopt the questions, I have used them as they were in entirety which led to a total of 21 questions. I'm confused now in terms of data analysis. Q1. Should I use the total score on the 21 questions as the combined score on perceived discrimination. When I did some preliminary analysis, there was obviously a difference when I took the total score as compared to looking at any of the 5 scales' scores alone. Q2. Should I be doing an EFA first? Because these questions have not been asked together before. Your response shall be greatly appreciated. Hope you and everyone around is keeping safe :) Cheers.
If these are part of reflective scales, then yes, you should do an EFA first. If prior literature has recommended summing them, then that is a valid approach as well.
I cannot believe you have put all these insightful videos for free James. I learned a lot from you. Thank you so much.
Dear Professor James Gaskin, I am really thankful for this. This session provided meaning to some of the many separate readings and books I consulted. I loved the tips and your style to see statistics in a fun and interesting way. This is key to learn and acquire knowledge. So kudos to you! The fact that it is live and there is interaction is even more interesting.
Thanks james I learn a lot from you
Many thanks, James, I have learn a lot from your video, they are really helpful.
Thank you for this and your other videos.
Thank you, James.
Great video James thanks a lot. These are so helpful. One question if you would be so kind to respond. I am trying to follow along using the SPSS file in the link provided. However this appears to be different to the one in the above video (there are more items). I have tried to manually select the same items as you do in your video (I get 47) but im still slightly out (KMO 0.901 and df 1128). Is there a link the original data set used in this video. many thanks
The dataset linked should be identical, or nearly identical. The exercise will be the same either way.
Thanks I got the file.
Dear Professor James Gaskin, thanks for the informational video! I was wondering if it's possible to do EFA with a questionnaire with more than 200 items? I'm developing a personality questionnaire that taps into over 30 constructs based on the Five-Factor Theory. However, in the EFA examples I have read or seen, it's usually done with less than 100 items (i.e. the example you provided in your video). Due to the complex nature of my questionnaire, should I split the EFA in 5 sections, each investigating the items relation with each other within their respective big five? OR should I do it as a whole?
Wow! That's huge! If you get enough data (like n>800), then you could try it all together. It will probably have lots of discriminant validity issues though. In this case, you can separate it into a sort of sensitivity analysis by doing an EFA with the items from two personality types at a time. Then alternate to make sure you have good discriminant validity between them all.
Hi Professor Gaskin, Thanks for the reply! So just to clarify if I understand you correctly, what I could do is look at the items for each big five-factor (i.e. OCEAN) separately or in pairs (e.g. look at Extraversion and Agreeableness items together in one EFA)?
If I had a conceptual idea about which items group together, and do EFA on a scale rather than item level, would that be considered valid?
@@chuwainicholaschan190 I agree with the first half of what you said. The last question is probably not advisable since these would each then represent distinct dimensions through a single computed variable, and would therefore not load onto the same factor.
Thanks a lot Sir
Hi - For your example of performing factor analysis with a formative factor, which you mention around 1:50:00... You suggest you can give each item a score (e.g., 1) and sum them up to create scores/frequencies/etc for your factors. However, I thought you needed at least 3 items per factor to perform factor analysis, and creating a score will essentially turn this into one item. You also wouldn't be able to measure reliability of the factor using Cronbach's alpha. Am I missing something?
The score would create a single item that you would then use in a causal model, not in a factor analysis. Cronbach's alpha is not relevant to formative factors anyway, because it assumes indicators are correlated (which is not an assumption of formative factors).
@@Gaskination Thank you! Loving your videos!
Dear Prof. Gaskin, thanks for your vedio. I am wondering is it right to do EFA and CFA using the same data set. Or should I use a small sample to do EFA and then use the bigger data set to do CFA? P
It is considered more rigorous to do them on separate samples, or on random samples of the same data. However, most researchers are often struggling to get enough data. So, they do not have this luxury. If you have enough data, then it is recommended to do it on separate samples.
Dear Prof Gaskin, Do we apply EFA, CFA for construct validity in case of a formative scale as well ? Thank you so much in advance
You can do a CFA in a PLS software, such as SmartPLS, but do not do any EFA, as the EFA is built upon the assumption of inter-item correlations (which is not a requirement of formative measurement). Here is a useful video: ruclips.net/video/Qx_7pN7eAes/видео.html
I could use some of this. Thx
Thanks James. Can we get the data please?
drive.google.com/open?id=0B3T1TGdHG9aEN05aRGdIUGhDMDg
Dear Professor, I did 5 interviews and converted those interviews to frequencies. Now I need to run a factor analysis to see whether some codes relate to each other. My number or variables are greater than 5 (cases/participants). What should I do with this? Could you please give me your opinion? Thank you
A sample size of five is not sufficient to run EFA. I would recommend just doing it by hand. Just look for patterns in the frequencies.
@@Gaskination Thanks a lot for your feedback professor
At 32:06 you mention that the optimal number of manifest variables to describe a latent construct is 4. If you could please provide a citation for this it would be most appreciated (possibly Hair?).
I wish I could remember. It wasn't Hair though. I just did another search for it and couldn't find it. Sorry about that. If I come across it again, I'll make sure to post it on my references page on the statwiki.
Bollen (1989) suggests that 3 to 4 indicators are the best number of indicators for each latent construct.
Great and thanks!
THANKS!
Do you know where in the book Bollen mentions this? I searched the whole book just now and couldn't find it. If not, no worries. But, if you have a page number or quote, that would be awesome!
Hi Sir
Hope you are fine. I have a query. If i am using all borrowed scales in my research. Do I still need EFA? As people have different views on that.
Kindly suggest
Thanks and Regards
Kavya
I always conduct an EFA, even when using scales borrowed from others. The EFA identifies discriminant validity issues better than a CFA. Nevertheless, if your CFA shows good validities and model fit, then you are probably fine to skip the EFA.
Can we do exploratory and confirmatory factor analysis in the same data set?
In all your EFA and CFA you used the same data to build the model. If it is correct please can you send reference
thank you
Most studies do EFA and CFA on the same dataset, although it is considered more rigorous to do it on separate datasets, or separate samples from the same dataset. However, most scholars don't have enough sample size to split the data. As for a reference, I'm not sure there is one. It is just common practice.
May I know the literature that served as basis for your decision to retain items with factor loadings of .50 up. Thank you!
Hair et al 2010 (Multivariate data analysis) is probably a good one.
Do the Bartlet and KMO suffice for testing normality for EFA or shall we also use Kolmogorov-Smirnov and Shapiro-Wilk test?
The KMO and Bartlett’s test whether the set of variables are adequately correlated, but not whether the variables are normally distributed. For univariate normality, you can just check skewness and kurtosis.
@@Gaskination thanks. What if the skewness and kurtosis don't meet the criteria as well? Is it okay to use Satorra-Bentler or Yuan-Bentler adjustments? (My items are Likert rating)
@@sarahoseingholizade9155 I'm not sure of any effective way to transform Likert scale ordinal variables. If the skewness and kurtosis is less than 10, then there may still be reason to use them. I think it was Sposito (or maybe it was Kline...) who said up to 10 is probably okay under certain circumstances... Here are both citations:
Sposito, V. A., Hand, M. L., & Skarpness, B. (1983). On the efficiency of using the sample kurtosis in selecting optimal lpestimators. Communications in Statistics-simulation and Computation, 12(3), 265-272.
Kline, R. B. (2011). Principles and practice of structural equation modeling (3rd ed.). New York, NY: Guilford publications.
@@Gaskination Ahan, thanks so much
Sir How we can get the working file
drive.google.com/open?id=0B3T1TGdHG9aEN05aRGdIUGhDMDg
When is your next SEM boot camp? Can I sign In? How can I sign in?
I host one every May in Utah, but sometimes also do them at other universities throughout the year. Often it is less expensive for a university to fly me out and compensate me than to fly out all their interested students and faculty.
Hello sir.....i need ur opinion...in my study, i measure women empowerment in terms of pre and post analysis....i asked same questions to respondents twice...One for after intervention and another for before intervention....do i need to conduct EFA/CFA for Both set of responses or just one /post scenario in enough???
It would be best to do EFA/CFA for both separately, just to make sure they are both valid sets. The other option is to do it for the change from pre to post.
James Gaskin thanks for ur respond... actually i intend to measure empowerment from pre to post using SEM, in which latent variables are treated as observed variables.....do i still need to conduct EFA and CFA???
If you aren't using latent variables, then you don't need to do factor analysis.
James Gaskin sir, im using latent variables but for particular analysis these are treated as observed..
factor analysis should be done for all reflective latent factors. If you don't use them as latent, then you can just create averages. But if you use them as latent in any causal analysis, then you should do factor analysis.
Sir great video. Can I have the spss file
The file is available on the homepage of the statwiki.
Sir what is the file name
drive.google.com/open?id=0B3T1TGdHG9aEN05aRGdIUGhDMDg
Dear Prof. Gaskin,
Please help me with a query. In my research work, I have used questions from 5 existing univariate scales to measure perceived discrimination at work. I did not modify, adapt or adopt the questions, I have used them as they were in entirety which led to a total of 21 questions. I'm confused now in terms of data analysis. Q1. Should I use the total score on the 21 questions as the combined score on perceived discrimination. When I did some preliminary analysis, there was obviously a difference when I took the total score as compared to looking at any of the 5 scales' scores alone. Q2. Should I be doing an EFA first? Because these questions have not been asked together before. Your response shall be greatly appreciated. Hope you and everyone around is keeping safe :) Cheers.
If these are part of reflective scales, then yes, you should do an EFA first. If prior literature has recommended summing them, then that is a valid approach as well.
Thank you so much for your quick revert Prof. Gaskin. You've been really helpful. Have a great day :)