SmartPLS 4: FIMIX (Finite Mixture Analysis)

[@demiluigi]

Hi Dr. Gasking. Thank you so much for the video. Could I use the FIMIX method to assess that there are no significant differences in the outcomes, due to a variation in the experimental treatment applied to a portion of the sample in the study?

[@Gaskination]

ANOVA would be simpler and better for such a test.

[@demiluigi]

@@Gaskination thanks James. But my doubt is I already have the model run in SmartPLS. Even though, do you recommend better running a separate ANOVA?

[@urielaldavapardave3313]

Buenas noches profesor James. Dos consultas por favor. A) PLS-SEM se puede utilizar para validar instrumentos de medicion; b) en caso de tener un modelo de medicion de segundo orden con algunos constructos de primer orden medido formativamente y otros medidos reflexivamente, como se debe proceder, es decir, en el constructo de segundo orden se debe colocar todos los ítems de forma reflexiva o formativa?. Gracias su amable respuesta.

[@Gaskination]

colocar todos los indicadores para el factor de segundo orden reflexivamente

[@tazyeenalam]

My generate group option is showing an error in a loop which says "WrappedValue(numericValue=2.0, stringValue=null)....can someone tell me how to fix it? Thanks

[@Gaskination]

Sounds like there is a problem with your data. Check it outside of SmartPLS. Specifically, look at the grouping variable values. You can also look at them inside SmartPLS to see how smart PLS has interpreted them.

[@tazyeenalam]

@@Gaskination okay. How can i fix it? How do I identify the problem?

[@Gaskination]

@@tazyeenalam You would have to look at it as I've indicated and then discover for yourself if there is an encoding error (for example, invalid characters such as mathematical symbols).

[@anisalshargabi1244]

Thanks Dr. James for excellent explanation! - When using two-stage disjoint approach for higher order constructs where latent variables from 1st stage LOC are used for the 2nd HOC, on what level should the FIMIX analysis be conducted? on the LOC or the HOC? or should it be on the original model as a whole including LOCs and HOCs in the model using the repeated indicators approach? - Should the control variables (e.g. gender, age, region, etc.) be included in the model when doing the FIMIX analysis? Thanks!

[@Gaskination]

-I think it should be the lower order model, since it is assessing latent groups in the items. -Those can certainly help identify latent groups, though they are not the focus of your theory. So, if we found latent groups associated with them, this might not be concerning. I'm not sure what the right answer here is, but if it were me, I would probably conduct it without the control variables (although, if it were me, I probably wouldn't do FIMIX at all unless the findings were irrational or weak).

[@farahfarhana4014]

Hi Prof., I wonder if AIC3 - segment 5 and CAIC point towards the fourth segment, there are 5 segments does it mean there is unobserved hetero? segment 4 is 0.073 (sample size 21), segment 5 is 16, min sample size is 55.

[@Gaskination]

I don't think I understand your question. You must be referring to your data and model, since the video shows only 4 segments.

[@farahfarhana4014]

@@Gaskination yup correct Prof. My data

[@Gaskination]

I would recommend looking at what seems to be driving those differences, as I do in 6:26 in this video. If there is nothing clear, then perhaps it is just randomness manifesting as a group. I would not worry too much about latent heterogeneity unless it is strongly affecting important variables.

[@Nbl-369]

Dear Sir My research model is look like this, adding two moderators. One With Participation and Autnomny, another with learning and feedback. questionnaire and interview method are used separately to collect dat. I am eager to know this model would be the final model? As a whole. ?

[@Gaskination]

I'm not sure if I understand your question. The model in the video is likely the final model in this example.

[@Nbl-369]

What if this model add 2.more variables i.e 2. moderators itself. One moderator with participation and autonomy and another moderate learning and feedback. Then what could be the possible way to run it. Plz make a video using Amos.

[@Gaskination]

@@Nbl-369 Like moderated mediation? Here is a video of it in AMOS: kzbin.info/www/bejne/nZ2tmZqVm911l8k

[@roshanpanditharathna9142]

Hi professor, where should I find the literature that confirms the lower value of ∑ (AIC3: CAIC) on 1st segmentation decides the no segmentation needs. I keep searching, but I did find none for far. Could you please give me a clue about literature? appreciated as always professor.

[@Gaskination]

hmm... Don't know if I saw that in another video or in a journal article or book... The references on the smartpls website are pretty useful. Sorry to not be much help on that... Remembering references are definitely my weak spot.

[@PedroRezagado]

Hi James, thank you for the explanation. I would appreciate it if you let us know about the literature that you mentioned when you were getting rid of some indexes, considering useful only AIC3, CAIC and EN. Thank you.

[@Gaskination]

Lots of good literature here: smartpls.com/documentation/algorithms-and-techniques/fimix-pls

[@arifashraf8569]

Hello James thank you for your wonderful work . We are waiting for SMART PLS 4 and Plugins for AMOS 28.

[@Gaskination]

Hopefully both are coming in June.

[@BEKatowice]

Hi James, I have followed your very informative video coming up with 2 pretty stable fimix segments, and then I have partitioned the sample based on discrete segment assignments. However, when I initiate IPMA for any of the 2 dependent constructs I receive information that there is a singular matrix problem. Could you please suggest some remedies?

[@Gaskination]

Singularity matrix occurs when you do MGA if you have also included the same variable in the model. For example, if I performed MGA based on gender, but I also included gender in the model, then for each group, the gender variable would not vary. This creates a singularity matrix. All variables must vary. Hope this helps.

[@BEKatowice]

@@Gaskination thanks for your reply. I understand but the variable I have used for MGA or IPMA (the same problem) is just discrete segment assignment (2 segments) coming from FIMIX with a pretty stable solution. When I then try to compare the results of the model (without of course segment assignments as the antecedent) the singular matrix problem comes.

[@Gaskination]

@@BEKatowice Yes, so if that variable is in the model, but also used to segment, then it will have zero variance within each segment. This will cause the singularity matrix.

[@BEKatowice]

@@Gaskination Dear James :) thank you. Sorry but I am terribly lost here ;) To give you a bit better picture: I have a pls-sem model with 3 clear antecedents, one mediator (partial) and one dependent variable + 2 control variables. For this very model the fimix procedure provides me with some latent segments partition which I keep in new data file (created by smartpls 4.0). Then, I compare the same model in the MGA using discrete assignments (segments) for inter-group comparisons. This is exactly when the singularity matrix comes. Thus, I do not use for model the same variable like for MGA-segment comparison, because in the model I DO NOT USE CATEGORICAL VARIABLE ESTABLISHED FOR SEGMENT ASSIGMENTS. Additionally, I have checked carefully through the indicator correlation matrix and it appears that the only "1" correlation appears between the same variables, which is all fine. Thus, I am totally lost with your comment, when you write "variable in the model, but also used to segment". Actually all variables from the model are used in fimix procedure to segment initially, right? This is then about potential threat of including the same variable that COMES OUT FROM THE FIMIX for MGA comparison, which is not the case in my iterations.

[@Gaskination]

​@@BEKatowice Sorry for the confusion. If possible, export the new dataset and then examine it in Excel or SPSS. Sort it by segment. Then look to see if there is any segment that has zero variance for one of the variables in the model. For example, if for one of the segments, all gender values are 1, then this will create a singularity matrix. In this case, that variable would need to be removed from the model, or omitted when conducting the FIMIX, and then added back in afterwards. This might also be caused if one of the segments is very small (e.g., n

[@hoanganhvien]

Thanks you!

[@ragwsjuve856]

Thanks for the video James! I have just a little question on the difference of FIMIX and MICOM/MGA (you covered that in this video: kzbin.info/www/bejne/gmOvlXSdgrCgeNE). Is it correct that MICOM/MGA is done earlier in the analysis than this FIMIX, but we should do both indepent of the results in MICOM/MGA? Cheers!

[@Gaskination]

FIMIX detects latent, non-theorized and non-specified groups within our data by examining heterogeneity between subsets of the data. MICOM analyzes heterogeneity among determined groups in the data. So, if I did not have groups in my data already specified (such as single/married or leader/subordinate), I would run FIMIX first and then possibly run MICOM if I decided to use the detected latent groups as a grouping moderator. If I had groups already identified a-priori, I would run FIMIX just as due diligence and then proceed with MICOM before MGA.

[@ragwsjuve856]

@@Gaskination Can I maybe ask you another thing :) If the data points to 4 segments and with that 4 discrete segment assignments. Would you group 1 as low and 4 as high in your case? Or what would be the way there? Cheers!

[@Gaskination]

@@ragwsjuve856 I would recommend naming the segments based on the values of the variables that were used to form the segments. In some cases, that might mean low vs high. In other cases, they might be more like cluster groups with profiles that are more nuanced than low/high on a single spectrum.