SEM Series Part 6: Multivariate Assumptions

alright time for some multivariate assumptions that is the next step in the process you'll see here we have three different multivariate sumption linearity multicollinearity and homoscedasticity I actually not going to show you homoscedasticity why well because a lot of people say it's not even an issue another reason is because you often find heteroscedasticity if you have bimodal data or or data that is theoretically moderated by different groups so in our case we have this low and high experience low and high frequency we expect there to be a heteroscedasticity relationship between the residuals and the values for each variable and so to say we need homoscedasticity is kind of I don't it's not relevant so I'm actually going to drop this if you really want to do homoscedasticity test to watch my videos they actually say that they're not the best way to do it there's also no great way to do this there's no easy way to do this what I would say in my paper if I had to talk about homoscedasticity is say we expect heteroscedasticity because we are moderating our model with group multiple multi group moderators so I didn't run a homoscedasticity this because we expected heteroscedasticity and then say it 10 times fast and you'll get a medal alright let's go to linearity and multicollinearity for linearity this is something we need to do let me go get that model we want to do this for sadly each relationship in our model to all how painful is that I'm just going to do these four right here just to keep it simple in fact I might just do two of them and you can do the rest on your own so we want to do a linearity test between usefulness and information acquisition let's go do that we need our composite data let's grab that file open data what a terrible interface let's just go open it directly there it is alright and do a linear no we're not going to do linear regression test we're going to do a curvilinear a curve estimation go down to the bottom here are all of our variables let's throw one of our dependent variables in there information acquisition and throw usefulness in there as our independent variable and I want to estimate logarithmic inverse linear quadratic cubic power just throw them all in there these are all the different possible relationships you could have between these two variables and we hit OK and then I look at this model summary table or I go down to this fancy thing but basically I'm checking for a couple things I want to see is the linear which is on top the linear relationship strong look at that F value that's so powerful it is also significant okay does it have the strongest F highest F it does that means this is definitely a linear relationship let's try a different one caressed amasian takeout usefulness throw in joy I haven't tried this so hopefully it'll give us something more fancy there we go you see some of these are missing it tells you why up here somewhere right here ok and we see linear is significant or its strong R square is significant right here and it's F value is actually almost the highest it's fairly close its high enough and it is significant so we can say it is sufficiently linear to be tested in a a structural equation model so what if this was not significant well then that's a limitation we're just kind of to deal with because we're using Amos and Amos just calculates linear relationships so that's a limitation of our paper what if it was significant but some other variable some other relationships like cubic or quadratic were much stronger like they had F values that were triple the F value for the linear well again I'd say it is sufficiently linear because it is significant and so we're going to say that's not a limitation okay that's linearity you do that for each relationship go ah so painful so for playfulness usefulness playfulness enjoyment etc etc and what would you report you'd report one line hopefully that says we did a curve estimation for all the relationships in our model and determined that all relationships were sufficiently linear to be tested using a covariance based structural equation modeling algorithm that such as the one used in Amos and that's all you'd need if there were any that were not linear or sufficiently linear you just say what I said earlier which is its limitation we're going to deal with it was more quadratic or more inverse or something like that than linear okay the next thing was multicollinearity you do this when you have more than two variables predicting another variable so in our case the only things that require a multicollinearity tests are playfulness comprehensiveness abuse and a typical use because they're on the same level we're not going to include these interaction terms with it so go to SPSS run a linear regression regression linear and what you're going to do is go to the bottom and grab those three so what was that that was confused a typical use and playfulness throw those independent take one of them out and put it up independent and go to statistics and give me code linearity diagnostic it I'm gonna check the model fit so it doesn't mess up the area or muddy the area run and we're going to look at this to efficient stable it's the VI half that we're looking for the vif should be less than three hopefully if it's less than ten we're pretty happy but less than three is the ideal so that looks good let me swap it out on your regression again stick a typical use up there just do this a few times see if these change very much one point one eight nine the last time it was 1.1 1.2 to four so not very different we can do it one last time just make sure you feel good as to come with copies for playfulness playfulness up there hey okay and one point two six four so we're awesome no multicollinearity issue what if we did have multicollinearity issue oh well you could say that's fine it's limitation were willing to accept or you could remove one of the independent variables that's kind of painful but according to the statistics you're already capturing enough of the effect that each of those variables has on the dependent variable because they're the same effect so you're not actually losing anything yeah I would just say it's a limitation okay and that does it for multivariate assumptions

Channel: James Gaskin

Views: 43,792

Rating: 4.9183674 out of 5

Keywords: Screencast-O-Matic.com, multicollinearity, linearity, SEM, multivariate assumptions, Statistics, SPSS

Id: Gkp1DKbU-es

Length: 8min 15sec (495 seconds)

Published: Thu May 02 2013