SPSS - Double Mediation using PROCESS (Model 6)

hi everybody I would have to say that the number one most recommended thing or requested thing on my channel has been model 6mm model 7 for process in SPSS so this video is going to cover model 6 and then assume to come video will cover model 7 so I always recommend opening the templates to PDF that comes with process when you install it because I think it really helps me see what the models trying to achieve especially as you add more mediators more moderators or covariance so this particular model covers a double mediation where we have one X that we think goes through two mediators before it hits Y you can do up to 4 mediators with model 6 I have no desire to look at even half of that so we're going to do just two mediators because 3 & 4 mediators is just variations on the theme so we're work through the whole thing from start to finish and showing you data screening and then how the model works how to interpret the output and just a quick kind of small guide on APA all right so that's temple stuff PDF it comes with part comes with process alright so what this data set has is um submit up data that I'm using um so it doesn't work perfectly but it would be a good example of what how this works so 1x variable two M's and a Y and what I've done is I have taken that um example picture from Hayes's guide to write on as we go may get a little bigger here and so I just kind of made up this example from some grading data so what we have is our course evals that we all know a notoriously problematic and what we did was we looked at we didn't do this it's made up but the gray that the students are making in the course so they think they're making predicting the fact that they like the instructor so this question is like this instructor is in the top 20% of all courses that I've had at the University but we think that the grade that they're making really predicts whether or not they like the instructor but maybe it's moderate I'm sorry mediated by how fair they think the grading is and how much they think the exams and assignments match the lecture so maybe that grade is in part due the fact that they think the instructors fair and that the items are easy to study for which then predicts that they're the best instructor ever right and we had another question on here that this is a course I wanted to take that also does some funny things to your instructor ratings especially if you teach statistics so let's look at the data set here this data set will be available on stats tools after I'm done if you haven't checked out our website it is really awesome we have tons of tons of materials if you're looking for spss or our and what you do when you get here is you click on the sidebar to see all kinds of different courses that I have taught and all of their corn and corn materials you can look for from it for here you can read about our Facebook page you can look up at our application so all of this is on here or will be when I remember how to upload things and what you do is you click for this one it's usually event stats examples and then here is my other couple of ones for mediation moderation as well one thing we're working on this semester is adding links to all the materials on YouTube so that you can find them a little easier all that aside let's talk about double mediation so first we're going to is practice data screening so most of this data screening is taking from either the Andy field book or the tabachnikov Fadel book I do use the rule where P is less than zero zero one before we decide that anything is problematic which does not match the field that matches the tabachnikov okay so first thing I always want to do is make sure that these averages which is what they are are within range so we're going to check for accuracy first somebody analyzed descriptives and then I can do frequencies or descriptives okay I'm going to move everything over but I'm going to turn off display frequency tables because these are decimals and that will make me a giant giant tables under statistics I can get the men in the max just to make sure that they look right I also usually like to ask for mean and standard deviation just to make sure that nothing seems crazy out of the normal size continue hit okay see we got here alright so for all of them I don't have any missing data so that's fortunate I don't have to look at that step I have 3500 cases the means for the questions don't seem too irregular they run from 1 to 5 so I don't have any they're out of range that I need to fix and the standard deviations are about half a point so now this looks too wild it's within the range I would expect for these 1 to 5 Likert scales so the next thing you always check after accuracy is bore missing data I don't have any missing data so I don't have to worry about that on this example which is good the next thing we're do is check for outliers because this is work Russian style data we could actually test this using all of our predictors predicting Y which is part of the steps for double mediation but you can't actually do it to the process plug in so there's a couple of ways to approach this I'm not saying my way is the best way but what I'm going to do is just use Mahalanobis distance to see if the pattern of answers across all four variables is unusual you could also do traditional regression style outlier analysis but I would do that on the step with all of the variables included so X and 1 in them too predicting Y and use either Mahalanobis cooks or leverage so either approach appropriate and then there were other ways to do this but let's say we just want to look at Mahalanobis values and we want to do this for all four variables because we just want to see if the pattern is unusual across all four of them since we're going to use all of them in one way or another to do that in SPSS one thing you can do is create a random variable this is from the tabash Nick butcher transform compute variable I'm going to call it random because then I remember what it is over here in our function group I'm going to click on a Type R till I get two random numbers I'm gonna use RV Chi square Chi square because the residuals and Mahalanobis values and a lot of these as assumption tests our Chi square distributed so we want to create a random variable in the same distribution if you've watched a lot of my videos you'll see that I picked the number seven it works for me you really need anything bigger than two otherwise you don't get enough variance for this to give you a picture so pick something a little bigger than two but you don't have and there's no reason for seven other than it works alright so that created me this random variable now to get all the asari the assumption tests and Mahalanobis what we're going to do is analyze a Russian and then linear I'm going to predict my random variable so none of these variables should be very good at predicting a random variable which is why the scatter plots of our residuals should look random as well as the the residuals should also hover around zero because someone behind so I'm gonna be low but they shouldn't create a predictable pattern so I move all four of these over because I have four of them that makes my degrees of freedom for for this next analysis so it's the number of things that you put in this independence box under nope not statistics sorry this has been a while under plots doozy pred for Y Z residual and X that's going to create us a residual statistics plot for checking ingenuity and homoscedasticity click on the histogram button for normality and the probability plot for linearity continue under save click Mahalanobis now you could do all three here as well but cooks and leverages really depend on you predicting why so I wouldn't recommend it okay if you want to use the all three methods that I've shown in some of my regression videos you should do X m1 m2 pretty damn why not this random thing so click continue and then click okay we're going to get some output but I'm going to ignore it for right now I'm just going to look at Mahalanobis so for that I need a chi-square table so the cutoff for this is Chi square distributed which is why we made up a random variable with chi-square and I'm going to use fork set for variables x + 1 into NY I'm P less than point 0 1 because we want things to be really nuts before we correct them so have 1847 would be considered an outlier so I'm going to sort Mahalanobis now a max sometimes a right-click and sort doesn't work so what I'm going to do is Duke data sort cases Mahalanobis and descending it's always positive that helps us out a lot so I clearly have some wild answers here where a student hated a professor but then they rated all the other ratings pretty high so that might so looks kind of like a typo so what we're going to do is first move the cat he's decided to use the computer as a scratching post all right is we're going to get rid of everybody who's over eighteen point four seven so I could filter those out or I can delete them I highly recommend saving datasets as you go so that you don't actually lose participants so what I would normally do is to file save as I call this no outliers and then delete anybody over 18 for seven so hopefully this doesn't make my effect go away I still see the same results because otherwise I checked whether or not this worked before I did a screen so we have a no outliers data set great now all the output I have here is now incorrect because these charts are based on having those outliers so what I'm going to do is rerun that regression so analyze regress linear and just click OK now I will get another Mahalanobis column out here but I'm gonna ignore that you don't want to double delete so these people were outliers given the data I had I don't want to delete them again so let me close this thing so my computer doesn't freak out and now let's look at those charts I'm not gonna look at the first one because the first one includes the outliers so this first thing here is normality I would say this is all right we have a huge data set so even with deleting a couple hundred people or not even 100 20 or 30 I still have 3500 cases I'm well past the N equals 30 for normality so I'm not too worried about it but when you're looking at these charts what you want to look for is between 2 & 2 are most of the the bar centered over 0 and between 2 & 2 this one does have a slight skew to it although most of the data is between 2 & 2 it is slightly higher so it has a slight positive skew then we'll check out linearity oh it's questionable I don't tend to get too concerned one that bows away from the bars a little bit unless it starts to make a big S shape or it looks like a bow and arrow so I'd probably say this was okay remember to use the squint test so if he's going to add it kinda looks like the line is probably fine so this is not great but it's alright now this is actually quite bad if you want to see a little better you can double-click on it and tell it to add a line of best fit add fit line at total that'll work can't ever remember yeah it usually adds a line at zero for you it also gives me the regression line which I didn't really want but that's gray ah can you tell I've been making our videos for quite some time now I don't remember how to do any of this goes this all right well I'll have the fit line but I didn't really need it so here across 0 for homogeneity what we want to see is a cross 0 in both directions we want to see the data is evenly spread so from 2 to 2 normally it's a common theme because these are z-score you want most of the dots to fall here we're running 2 to negative 4 that's not so hot and then here on this axis the horizontal axis oh sorry that's vertical it's running 2 to 6 so that's also not good so I would say this data is probably not home with genetic right this considered raining for almost get a sisse T what you want to see is an even spread of the dots regardless of the zeros so you have to kind of ignore the zeros and draw a line around the dots and this was actually not too bad the problem is that generally most of the dots here are concentrated right in the middle and so the issue becomes the fact that there is more of a spread this way so we have this many dots it's kind of hard to see but I would say this one's actually probably okay because the spread of the dots is roughly even across most of where the dots are these are the problem dots out here then we're gonna have trouble predicting because they are likely not falling within the same range as everybody else so I always say this datasets kind of yeah and a better option might be for me to try this an AR and use some nonlinear regression some rank regression that would deal with these we're going to be difficult residuals but for the purposes of this video we're talking about process which is part of SPSS so that's how you would check for accuracy issues but didn't have any missing data we talked about outliers and assumptions for regression test so the next thing I'm going to do is actually try process so keeping let's close this thing keeping this picture in mind we're going to run this model its allies regression we go down here the process plug-in Hayes's website explains how to install so I plug in so we'll click on that and then what we're going to do is oh I'm hit reset because I had already done this but we're just the X into the X variable spot Y into the outcome variable and then both moderators go into M I'm sorry mediators like you say in moderation my fault this is all about mediation ignore me if I say moderation I'm also dealing with a very needy cat that dislike warning all of the attention so Andy feel would be excited because there's cats not in the video but at least as part of the video all right so we've got both M's in there now the order that M 1 and M 2 will show up in is in this order so if you decide you want to switch the order of them you can click on one of them and drag them into a different order so that will control which one comes up first as M 1 which one is M 2 so they're not separate boxes anymore we want to change model number to 6 over here or we will not get the right model under options here what we want to do is pick effect size for models 4 & 6 so we can see the effect sizes although I would recommend looking at what they are because they're not what I'm used to seeing we won't get the Sobel test because that's model for only so I also want the total effect model 4 & 6 and compare indirect effects for 4 & 6 so click continue conditioning is for moderation so I don't need that I don't have any categorical variables but I'm really excited to see that he's added these multi categorical options so if you have categorical variables this is how you can change which one is the indicator variable for dummy coding and then the long names option which is the long time coming that will let you continue without changing the variable names if you have more than 8 characters whew so great updates to process here so I'm gonna click OK now let's look at this output or after runs I do have a very big data set so this does take a minute to run there it goes so let's go through what's in this output sort of one piece at a time although all the most of the action the part where you're going to be interested in whether or not mediation really happened is here down here at the bottom total direct and indirect effects and a total effect model a little easier if I just kind of start at the top and label them all and then I can think about the picture so the first thing we get is X predicting m1 and its coefficient here I'm standardized coefficient is 0.5 1 so I've got this X predicting m1 here which is the path a 1 and so we're going to add a little text box here oops what happened okay well I thought I could add a little text box here there it goes and that was 0.5 1 so the grade I'm making predicts whether or not I think the professor's grading scheme is fair so um it may be that I think the grading scheme is fair and therefore I do better in the class but if I perceive that I'm doing really well in the class I think the grading is fair because I'm doing a good job so clearly all you people who are making FS are just complaining because you are done something like that ok so that's a significant effect and that's good we need X to predict m1 I'm tell us ignite over here by looking at P and if I wanted to write this up I would use T here at 3000 555 so this is APA style I can work a word here come down here thank you slide your T don't auto capitalize 3 5 5 5 equals I pull that T values just 65 90s a that P is less than 0.01 because it is here you could also include the model statistics for this so our r-squared is 0.55 since this is only one variable that variable predicts in one quite a lot because all of that R squared is due to this one variable but the real focus with mediation is really to look at the individual variables predicting and I don't know the overall model statistics while they're great I don't know that it's necessarily the most interesting part of this analysis so now we're going to put up to M 2 so I've got X predicting M - so I got 0.6 once is considered a as well but it's like sort of a 2 if you will oops so we said this is 0.6 1 over here and so that means that the grid I'm making predicts that I think that the grading matches what's in the lecture now for this one what I would do is have T but the degrees of freedom we're going to change because it's a different model so 3 5 5 4 and that is equal to C here 2391 so that's a big effect and it is significant it's significant because it's less than 0.05 0.05 single zero I'm writing them as 0.001 because APA five and six indicated Michigan's exact p-values when reporting and that's the closest that's sort of the closest thing we can get given our outputs generally people stop at three decimals alright I'm not reporting B because I'm putting into my picture here so now we need m1 predicting m2 and this is the one where I said it doesn't really work but that's okay and so in this model it would be kind of difficult to argue that this is double mediation because this path is not significant but for example sake it's 0.04 okay and that's not significant here what did I do I am terrible working word today get gracious there it is so it's point zero four so the fairness that I feel um doesn't seem to predict whether or not I I think it matches the lecture but that's probably because fairness is in there with the grade I'm making and that is such a strong predictor so this may be a little bit of suppression here because these might all be so highly correlated that's difficult to tell but what would be really better as a fairness predict match lecture as well and so we really need in one to predict him too but it doesn't in this case this would also be three five five for our T value is less than one so it's clearly not significant point oh nine seven and our p value equals 0.33 right so now we're going to predict why and in this particular model of this last model I have all of them predicting Y so this is the sort of model with the indirect effects so I've got X is point zero three now so that's good it's it's a small number which is going to probably be what we're looking for so that's C prime which generally is written on the inside and the pictures I've seen and now we have let's see m1 predicting why well I stopped early on here here it is so let's see how much does in one product why so I've got 0.3 one so what I think is happening is the fairness predicts whether or not they think you're a good instructor which is a good thing if they think that you're fair they think you're a good instructor you always want to be fair and m2 hear 0.07 is also predictor so if they think that your tests match the lectures that you're giving them so there's a match between class and exams they also like you but it seems to me like fairness is a little bit stronger these are all in the same scales so while they're not beta I can probably I can compare them a little bit I can't say it's significantly different without doing some tests Robin but it seems to me like this one's a little bit stronger so now I have all of these filled in except C I really need to see that's usually the first step okay so I have that everything is predictive except M 1 M 2 which is kind of a problem but let's look at X 2 y so right here underneath this last one now I didn't write these in so there would be 3 5 5 3 let's do that real quick so we've got T 3 5 5 3 equals we're going to have to do X to point to one P's point O three okay so it's significant which may not be the end of the world remember that your your mediation does not have to totally eliminate the relationship that's sort of like the original view of mediation was that X to Y I went totally away well one this is a huge data set so we've got some some issues with sample size might be pushing significance here as well but I it doesn't have to totally go away it could be partially mediating it or it's just about moving the relationship between x and y through m instead of all the way all through x and y so sort about redirecting that relationship that's there we haven't tested that yet we're almost there we've got em - so we got 1578 and 7.4 - so three five five three 1578 and that one is significant and then t3 3 is 7 point 4 - if I can remember correctly key lesson point zero one so that would be all the predictors for that step last but not least we need to see so we've got X predicting Y all by itself with nothing else and that's point 2 3 and it is significant so we dropped point 2 points okay let's consider the indirect effect which is the direct effect - the total effect or total effect - the direct effect either way so the indirect effect the total amount of it is point 2 oh so we're going to get down here to get this out here in a second but let's write this one up so it's 3 5 5 5 24 52 and it is significant so at this point what can I say do I have mediation well adding em1 and em2 has eliminated the relationship between x and y quite a bit not totally it's still significant but quite a litt looks like it has done something because C and C Prime are clearly not the same number all of my variables are predictive except here so it may be that there are two mediators but they don't have this linear path between them so they both mediate the relationship between x and y but there it isn't 1 to 2 to 3 to 4 so 1 2 3 4 here so this model may be better as separate mediations and compare the mediations to each other so two versions of modeled for compared to each other or it may be that we've got the order wrong with a couple different things that maybe you could try if they were theoretically meaningful now how to interpret the rest of this because that's pretty straightforward if you just label all of your variables but down here under total direct and indirect effects what's happening so total effect is see the direct effect C Prime the indirect effect here I get double click to activate go away is here this total is 23 minus 0.03 or 0.23 minus 0.03 so that's what 20 comes in now you're no there's no significant values out here and we didn't get so bail because it actually is two different civil tests so what I really want to do is I want to look at the bhoot confidence interval here and see if it crosses zero okay so it's 20 here I don't want this to cross zero you don't want one to be negative and wanted to be positive and it doesn't so it appears to me that this effect is significantly different from zero because the confidence interval does not cross the zero so that implies to me that total across M 1 and M 2 we are mediating the relationship between x and y because that indirect effect is not zero but then you also get these three indirect to fix okay now what you should do is scroll down and you'll see here what that key is so indirect one is X 2 m 1 2 y and so that's basically a single moderation so part of that total indirect effect is X through M 1 right so it breaks them down you'll see it adds up to the 20 up here so most of that action is happening through M 1 and that is a significantly different from 0 due to the confidence interval here not crossing 0 so most of the total effect is X 2m 1 to Y you'll see indirect to here so indirect to is X to m1 to m2 to Y so that's the the rerouting it is 1 to 2 to 3 to 4 and that one is not significant it's very close to 0 and it doesn't seem to support the notion that it is 1 2 3 4 instead it looks like it's 1 2 3 is where the most of the action is happening and that's really because these two variables don't predict each other or m1 doesn't predict them too and so I probably wouldn't say this was double mediation I would say that m1 is clearly a mediator now in in 2 3 here is X 2 m2 to my to Y and that one is a mediator and it's above zero but it's clearly less than m1 here it's really what it looks like it's happening is I do have two mediators but the two mediators aren't related to each other so I have one mediating two mediating but it they don't go through this path of 1 2 3 4 what it looks like is happening is there's two routes to get to Y but there isn't this third um bypass if you will to use a terrible road example hey the other thing that the output will give you is the different effect sizes so I've got partially standardized if direct effects completely standardized indirect effects ratios of those effects I would tell you to lead what Hayes has to say about these and which one he recommends currently I know that preacher and Kelly option has been taken out of this version I'm using 2.16 in this video so I'm not sure which one I would recommend because I haven't looked at these in a while and which one is best I might recommend completely standardized but I would you probably want to get it from the man himself from which one is the best effect size but either way just telling people which one year reporting would be best and be sure you include the confidence interval so that they can judge for themselves that these are greater than 0 the other fantage get here are these little C's the C's are comparing the indirect effects to each other so these are sort of like little mini t-test for which indirect effect is better than another in a sense so I've got 1 to 2 so the indirect effect of 1 versus 2 here now one remembers X m1 Y so this is sort of testing m1 by itself 2 includes both mediators so 1 2 2 clearly oops sorry looking here 1 is a much better one and I can see that by knowing that that's one is the 15 one so those are significantly different from each other again it doesn't across 0 and then c2 down here is 1 2 3 so we're going to need pairwise so 1 2 3 and it seems that m1 is a much better mediator than m2 because 15 square to the 4 and that is significantly different so one is better than having two of them and then one is better than having m2 so m1 is better than a lot of these other things and then the last thing is c3 which is two versus three and those are different but it's basically saying m2 is better than having both m1 and m2 together so to kind of sum all this up what I tell you to do is to report maybe these indirect effects boxes on Y on all of the different predictors here and then an interpretation of how those predictors lay out and what they mean for your particular theoretical model so feel free to ask questions let me know model seven coming very soon

Channel: Statistics of DOOM

Views: 25,776

Rating: 4.9540229 out of 5

Keywords: statistics, spss, mediation, hayes, process, data screening, apa style, model 6

Id: WFWWHF2zltc

Length: 35min 10sec (2110 seconds)

Published: Mon Aug 08 2016