Linear Regression + Mediation + Moderation

okay this tutorial will cover linear regression mediation and moderation analysis and it's pretty much the same what I showed in our workshop class it's basically I recorded it so that you could have it so you could look at it in your own time and because linear regression mediation and moderation will form part of your assignment it's very important that you get it and you understand the steps to take in SPSS to analyze your data so what are you gonna do is just what we did in the workshop um and this is the file for multiple linear regression so going to the website will allow you to download the files as usually so for the first exercise so we need Sol seg all groups so going to our website s so groups download save and here we go our file should download in a second okay here we go so we have our file where we have H financial stability security don't remember oh we don't even have labels to it okay so we have H financial let me pick on it what we have their financial security health quantity of social interactions and life satisfaction and also we have groups to which groups of participants belong to i don't precisely remember what are the different groups but it doesn't really matter we know there are three groups and on some variable they define them in three groups so what were interested in is in how age financial security health quantity of social interaction can predict life satisfaction when we talk about the relationship and prediction we need to run a correlation and regression analysis so because in contrast to an oval and t-test regression and correlations study the relationships an over T test study mean differences here we're not interested in the differences of life satisfaction depending on which group you belong to here we interested how other variables that are only scale can predict life satisfaction so depending what age you are can we predict how satisfied you are with your life so to do that we'll go to analyze and the first analysis we want to run is the correlation because if you have no relationship between the variables it's true it's meaningless to put their prediction line because without the relationship you can't form the prediction so we just go to bivariate and we've put all our variables interest in the variable table here you have Pearson correlation chosen by default and it's run on a parametric it's a parametric test Spearman is the correlation as well but it's a nonparametric test so here we use Pearson options well you as always choose means and standard deviation continue and nothing really else you need to choose everything else is chosen for you press ok and here we have our descriptive statistics with means for age financial security health quantity of social interaction interaction and life satisfaction and when you look at the means it's always helpful to know on what scale award the variables measured you have the scales of the measured variable in your lecture slides and I'm just trying to find them unsuccessfully but yeah check in your lecture slides that should be quite clear okay there we go financial security was measured to on a scale from 1 to 45 so and we have mean of 23 health was measured on scale from 1 to 30 and we have mean of 30 who wore quantity of social interaction was measured on scale from 1 to 14 we have 16 analyzed satisfaction was also measured in 1 to 14 we have 16 for instance one of the groups on the Thursday group you guys had a variable we're in a task with ANOVA you had repeated measures and the variable was measured only different scales and you have to Center the variable in regression you know it doesn't really matter if they are measured on the same scale or not so let's have a look at our correlation and here we go with them life satisfaction we have h correlating positively financial security haw relating positively health correlating positively and quantity of social interactions correlated positively so we know that they are all correlated and their financial security life satisfaction showed to be the strongest showed the strongest correlation coefficient out of all those variables so in order for us to see if we can predict life satisfaction knowing age financial security health and quantity of social interaction we need to run the regression so go to analyze regression linear here we go so our dependent variable is life satisfaction if you perform multiple regression you just put all your independent variables we live group out for now so we'll put all our independent variables in one block so block one of one all variables are there we use enter method because here you have stepwise removal backwards forward but as we covered it in the lecture they're not really a wise choice when running the regression so what we're interested in here probability and read 0.5 nothing really we need to change in here plots you could plot your regression but we won't look at it now you could also look at the standardized residuals histogram and normal probability plots when you test one of your assumptions so we won't cover it at the moment our options already did statistics here you'll get the estimate you can ask for the confidence interval in linear regression confidence interval will represent as with what with 90 95 percent confidence interval with present that with 95% confidence you can say that your regression line will fit between the lower and and their upper bounds you can run the case wise diagnostic from here which is one of them and which is needed for testing one of your assumptions you can also request descriptive statistics colinearity diagnostic part and partial correlations something who touched on is when you exclude one variable from either another predictor and kriti and the dependent variable or whether you exclude only from the predictor and not from the dependent variable so that will turn part and partial correlation R squared changeable we're not really gonna get anything from choosing it in a multiple regression so we'll just continue okay so this is our regression output we get our descriptive statistics we also get the correlation so but the correlation table looks slightly different it's actually the same tables just separates in a correlation coefficient significance and number of participants where before it would give you correlation significance and number of participants in one square so that's the only difference it's exactly the same table this one is our model summary table which has quantity of social interaction health financial security in age entered all together and it also tells us how a dependent variable is life satisfaction model summary in here we have our correlation of our variables with our predictors with the dependent variable and we have this value squared which gives us R Square and if you put this value in the percentages it would be forty five point four percent of variance our four variables accounting in life satisfaction however it's better to use adjusted r-square because it's just that value and it's always well maybe no but usually I only saw lower then I just are just adjusted r-square low with an R squared because it takes them into account the number of variables entered in the model as well as the number of participants i knowa will basically tell you if your model is significant so it's like testing the main effect of the model and in here we have our coefficients which were interested in and standardized coefficient better this one gives you values the way you look at them is for instance one standard deviation increase in age with the result at point three four nine increase in our life satisfaction however unstandardized coefficient is hard to interpret because as we know they were all on a different scale and we didn't do to the scale we didn't do anything to the scale who did not standardize it because we automatically get standardized coefficients pattern and its transfers it its transfers all the variables into standardized variables so now we can have a comparison between them so we can see that the age and even stronger than age is health predicting life satisfaction but all of our predictors are significant so this one is the basic multiple regression if you want to do a hierarchical regression you follow the same procedure is just you enter the variables in two or more blocks so if we will reset everything at the moment and let's say our life satisfaction still a dependent variable our age financial security and health we'll put in the first block because we possibly let's say we want to control for age financial security and health and C home which variants quantity of social interaction accounts for in life satisfaction in pendant lis of age financial security and health so we're pulling out those variables from quantity of social interaction so by pressing next we can enter quantity of social interaction into the next block now in here you are interested not in this one in statistics and you are interested in our square change this is choosing R square change will give you information how much variance does the second step of the model accounts for additionally to the first step of the model so continue and okay there we go so regression table looks similar but now it shows us that we have two models so we have in the first step our three variables and the quantity of social interaction was entered as a second block this is the model summary where we can see our R squared change so what does that give you is that R squared in the second block represents a sum of our square change for model 1 and a square change for model 2 so point 4 1 5 plus point zero three nine will be point five four or five for M and in here it will tell you if you are addition on the second step accounts for significant portion of variance so our first three variables account for significant portion of variance in life satisfaction account for 41 point point five percent and our quantity of social interaction account for additional three point nine percent of variance in life satisfaction that was a significant so ANOVA gives you the results for the two models and here their coefficients is what you're interested in so again standardized coefficients and we have model one age financial security and health all of them significant now quantity of social interaction when we add it out it's also significant predictor of life satisfaction and when you report your regression if you perform hierarchical regression you have to report standardized coefficients for both for both blocks for model 1 and 4 for block 1 and for block 2 so if we look at our exercises so this is the table for your correlation and this would be the table for your regression where you have and standardized coefficient my standard error of the coefficient and then you have standardized beta coefficient and you put the stars where they are significant so this is the output you would get now let's move on and look at our grouping variable so going back to our data file we see that we have groups and we did not record them as anything we have three groups in total as I covered in the lecture when you have a nominal variable which would be a group belongingness you have to down the code the variable now how you dama code the variable go to transport in transport menu you actually do everything you need to do to variable you can compute the variable you can record them to same variable in two different variable so all the variable transformation will be done from Sport box now we need to we want to record it into the different variable so here we go this is the record in two different variable box we'll put our group as the input variable and outcome variable let's say oh I know remembered what the group was it was never married single and or with the partner without the partner that's now I remember what is responded to let me just stick it out okay there we go so the first group was with the partner always single previously with the partner now single let's just add it to our group variable so that will remember so value one was with partner oops but now yeah - was single and three was previously with the partner now single previously partner now single okay so now we have our variable coded and it will be easier for us to remember and understand what we are doing so transformed variable recording two different variable to do the Dharma coding which is group and let's say our first group we want to put dummy one and it would be partner we can call it the link that say we can know that their dummy one would respond to people having a partner now all the new values so previously one was with the partner and because we have a dummy variable for partner we'll leave one as the value now in group two singles they are not with the partner so they are zero and as well previously with the partner which is Group three they are not with the partner now so they also get to zero and continue and change okay here looking at our data fall there we go so dummy partner everyone who had one have one now they still remained one and everyone with the twos and threes have zeros now we need to record a second dummy variable because the way you record the dummy variable you take the number of the groups in your grouping variable minus one so we have three groups - one two so we need two dummy variables transform recording to the different variables so reset will put Group steel and we have it as dummy - and let's say it's single that's what we dummy food to we'll represent groups of singles all the new values say group number one with the partner they are not single that group number two are single and therefore they single so we give them that one to identify that they in a single group and now group three are previously with the partner single so they still get zero continued change okay so there we go our next dummy variable so what we have to represent belonging to group one they'll have to score on a dummy one-one and dummy to zero and in group two should score zero and uh me one and one on dummy two now the third group has zero zero and it's also refers to as reference group when we'll run our regression and look at the output we'll know what why we need the reference group as well because dummy variables that correspond to one variable group you can't enter them in a different steps of hierarchical regression it just doesn't make sense even though we separated one grouping variable into two dummy variables the dummy variables will go together so let's say we want to run our regression and we want to see if belonging to or the status of the relationship has any effect on life satisfaction in addition to other variables because in previous some regression that's just resulted in the previous study analysis we found that all of our predictor variables had a significant accounted for significant portion of variance in life satisfaction so let's see if the group accounts for any additional variance so we'll put all our psychological predictor variables into independent block and in the next block so putting them in the first block we control for their virulence we put this to other variables now going to statistics our square change because we want to see the percentage of variance our group accounts for we can also ask for descriptives but let's not clutter our output okay there we go so our again output quantity of social interaction health financial security nation block 1 and our group variable which is relationship status is in model 2 so when we look at the model summary would we see that again just as before our psychological variables together accounted for four forty five point four percent of parents and it was significant now having a partner or being single or now having a fun in being single accounts for additional 7.3 percent of variance which is significant ANOVA says everything is significant so we're pretty happy now standardized coefficients better we'll see that this is just what we obtained before from our multiple regression and the last step of a hierarchical regression with the data now having relationship status in the block two it tells us how much its accounts for in the block two and we see that dummy 1 and dummy to account for signal if s is accounted for significant portion of ours but now the way you can interpret this batter is whether dummy one with the partner is different from our reference group and our reference group had zero zero on both of the dummy variables so the reference group is pretty with the partner now single so they are different and singles with previously with the partner now single are also different so but I mean it's one of the ways of mobile it's probably the only way of reporting their dummy coded variables when you look at the group membership in a regression but the most what we're interested is that the relationship state is accounts for 7.3 percent of variance okay so this is the multiple regression now let's move to radiation and moderation so no we don't want to save anything now mediating moderator let's start with a mediator Ms mediator so we need to download that file ms mediator download okay it's loading but my my SPSS does not behave particularly well so the example we looked at the lecture was less fun is that [Applause] oh that's why is that Terry are we now that's a motor right oh no no no this is not them lecture this is your one of your exercises sorry no I don't want only yes I want to close it I don't want to save anything we need howl howl howl howl will mediate them there we go now I recognize the data set okay so there we go we the tusky well the question is does maternal care predict self-efficacy and if the relationship between maternal care and self-efficacy is mediated by self-esteem if you read the examples from the lecture it will refresh your memory so let's have a look first before we run any mediation moderation regression as usually we just want to see their relationship provide correlation from here choose our self esteem self-efficacy nothing else to choose correlation we see that with a 5-month self-efficacy self-esteem correlated strongly maternal care slightly weaker low esteem and maternal care correlated quite strongly so following come barrelling Kenny conditions for testing mediation moderation what we need is IV and DV our predictor and our dependent variable to have a significant relationship and also our IV which is maternal care should have a significant relationship with self-esteem and that's what we actually see in here yeah and our mediator which is self-esteem should have a significant relationship with self-efficacy so we satisfy to all the criterion and now let's run the mediation to run mediation and moderation that we'll cover in a second is the same you go to linear regression you put your myself I think the Zener dependent variable maternal care you put in the independent variable click Next and you food self-esteem we covered it in more details at the lecture why we put them in this particular order the purpose of this video is just to show you the steps now always click options statistics are square change continue okay what we get is our variables the stages at which they were entered we get summary models so we see that model 1 which is fast our maternal care accounted for 7.4% in self-efficacy and the second step accounted for additional 8.7% both of them was significant and that's the same would anova shows us so let's look at the coefficients so in here how we can see that it's a mediating relationship maternal care predicts significant claim self-efficacy what that means at one standard deviation change in maternal care self-efficacy increases by 0.271 yem now if we look at the second step where self-esteem was added maternal care became known significance of one change in matured in maternal once and deviation change in maternal care resulted at 0.14 to increase in self-efficacy and this increase was not significant however self-esteem is significant and it's actually much the increase in self-efficacy is higher that with than from maternal care so one standard deviation change in South efficacy results 0.32 change in life satisfaction when I say when I look at standard coefficients that's assuming that other independent variable or other predictors main constant which is usually not the case because independent variables themselves correlate with each other but this is the estimations were getting so because maternal care became non significant we can conclude that the relationship between maternal care and self efficacy was fully mediated by self esteem in your exercises that you needed to do it's not always the case sometimes you can have partial mediation and you have to check if your mediation is significant using samples tests I will probably give you the link to the video that unfilled shot explaining the process model that the one that I'm mainly using if you are interested you could look at what he is got to say now this is everything there is to mediation let's have a look at the moderation say moderate and we need oops how will hassle moderator how will hassle moderator downloading the file so moderator is slightly different from the mediator so what we have with our moderator the way to look at it is the variable that affects the relationship between our predictor variable and our criterion level our dependent variable so moderator is like interaction of actin and over the and you pretty much test the moderation in the same way as you test interaction in an omen so here I'm just quickly going to give you the steps for that first let's run the correlation for all our variables bivariate variables together run our correlation and this is what we get hustle support symptoms and in here we see that symptoms is our independent variable or through our dependent variable sorry and support and hassle are our independent variables so in order for us to test the moderation the moderator we need to Center our predictor variables and then not Center sorry oh yes Center sorry I've been shooting too many videos today so we need to Center our predictor variables and then have a product of multiply centered variables and that will give us our moderator to Center your variable you need to to calculate means or your two variables so our predictor variables are have hassles and support so let's calculate descriptive statistic descriptives hassle support options we just need means and standard deviation we don't even need standard deviation but by default let's do that so we have min 4 for hassle is one seven zero point one nine six and for support is twenty eight point nine six four to compute a centered variable you go to transform compute the variable and put centroid let's say hassle let's start with hassles centered hassles equals two hassles - hassles mean and we can see that our hassles mean is 170 point one nine six okay and we should get our central variable in our dataset centered hassles so 100 176 minus 170 point one nine one and six will equal to point two five point eight so this will basically have central centered hassles is calculated now let's do the same with the support transform compute variable set central support say we pick support minus mean of the support is 22.96 I think nine six four okay so here we go we'll get there to centered variables now to calculate the moderator or to to calculate the interaction effect we'll need to a product of those two variables so we can compute the variable and we can call it central hassle central support and what we do to calculate their interaction central hassles times central support okay so now if we look at our data set we have our moderator for our interaction effect of hassles times support now to run the regression and to check the moderation between hassle weather support has a moderating effect on the relationship between hassles and symptoms we'll need to run analyze regression linear we have our symptoms as the dependent variable hassle and support will go to the independent variable and that will give us sort of main effect for hassle and main effect for support and next we'll enter our interaction which also give us weather support where the support is the moderator in our relationship between hassle and symptoms so statistics r-square changed because we want to see if our moderator or our interaction account for any significant variance in our dependent variable which is symptoms and we press ok so this is the output we get first to model two variable centered support and hassle so the main effect of support and hassle accounted for 33% or 33.4% of variance in some and it was significant our interaction between hassles and support accounted for additional significant 5.4 percent of variance looking at the coefficients we have hassles and support which in here support was not significant yeah but now when we enter our interaction term of hassle and support our moderators support as the moderating to the model it's become significant so what's happening hassle effects symptoms but demand depending on whether you get support or you don't get support or whether you get high support or low support will determine if your the number of hassles you have will affect the number of symptoms your show so this is initially the moderating effect and because this centered hassle support variable is significant we can conclude that our support is the moderator in the relationship between hassle and symptoms okay that is it for this analysis

Channel: Jekaterina Rogaten

Views: 63,777

Rating: 4.7813506 out of 5

Keywords:

Id: 7KCuumQmsxk

Length: 41min 35sec (2495 seconds)

Published: Mon Nov 25 2013