What do you mean by SEM?

What do you mean by SEM?

What do you mean by SEM?

welcome everyone in the last session we  discussed about structural equation  modelling we're just briefed about  structural equation modelling and what  is the role of structural equation  modelling right so in this case when you  talk about structural equation modelling  which is nowadays highly used in all  kind of researches structural equation  modelling is build up built upon  basically three kind of statistical  techniques you can say right

basically  the three are regression or if I say it  is multiple regression right so multiple  regression factor analysis right factor  analysis and the third I say is the  chi-square right the three well the  three statistical measures that are  highly used in structural equation  modelling are these three right I  squared so the chi-square talks about  the goodness of fit right so goodness of  fit index so now it tells about the  whether to the obs

erved and the expected  model the object model and the expected  model are they more or less same or  different and if they are same then it  is a good thing right if their clothes  are same that is good but if there is a  lot of difference between the observed  and expected then it is not good okay so  what in the in the first part we talked  about the measurement model right so we  said that the entire structural equation  modelling can be said as

a combination  of the measurement model and then the  structural model structural model okay  so so the measurement model is what we  say is a confirmatory factor analysis  right so apart if you if you can go  understand the word so there are  exploratory factor analysis and other  was confirmatory factor analysis okay so  this part confirmatory factor analysis  part is a measurement model okay so  confirmatory factor analysis is used as  a measure

to reduce the measurement  error of a study right that means what  when you have measured certain things  and that those things could have been  measured and you know the researcher  with the the the respondent might not  have taken seriousness serious interest  or something and something has happened  right so in a cfa case what happens is  here it tries to take build a latent  construct basically a latent construct  which is made up of certain va

riables  right so this variables together will  represent the the construct let's say  okay so this construct in the  measurement models we say what we do is  we basically create a covariance matrix  or a covariance structure right in a  covariance structure a covariance  structure if you remember we have also  discussed that it is like a correlation  structure only but a correlation is a  special case of covariance right so here  we try to see whet

her the model there is  a covariant relationship covalence  relationship between the constructs  right and finally what we do is we try  to see in the CFA in the CFA case we  first check the construct and for its  reliability and validity okay so you  checked for validity through different  the construct you know we have the  construct validity so so to do the  construct validity we use certain ways  for example we convergent validity  discriminant

validity right and  nomological validity right so convergent  validity I am sure you are very clear  because in convergent validity only  thing you want to see is that whether  the items are converging towards the  main construct okay so if the items are  sufficiently having a larger  you write the correlation between the  item and the construct is high or  interitum correlation you can say is  high then we say this is a high  reliability so that he

lps us also to  understand the convergent validity there  is a convergent convergence but  discriminability occurs when we try to  measure key whether two constructs are  sufficiently separate from each other  okay so construct a and B should be  sufficiently different from each other  so to do that what we use is the average  variance extraction right value and  compare it with the let us say the  correlation right correlation and we  always say th

at the average variance  extraction extraction should be higher  than the squared correlation value right  the squared correlation value whatever  comes it should be the a ve should be  higher than that then there shows a  discriminant validity okay but let us  say now comes a case once we have done  the measurement we have it's a  measurement model so we have measured  and we found that the construct is  justified and it is explaining what it  shou

ld explain okay it is doing its job  with the help of the model fit indices  we have said now the model fitting  business are basically we use the Chi  square by Chi square by degree of  freedom which we always say should be as  less as possible so in between let us  say anything less than 3 is good we say  right we have taken a magic value of 3  right but some people some researchers  also say anything up to 5 is still good  right so what does it m

ean the Chi  square by degree of freedom basically  talks about that it compares against the  baseline model that means there is a  model there is a model and it compares  give what happened what is the saying is  the observed and the expected model the  absurd model in the expected model the  difference should not be very high right  so when we are when we are measuring the  Chi square by degree of freedom and if  it is sufficiently threes around l

ess  than 3 then we say that the model is a  fit model okay there are other fit  indices also for example  the goodness effect with GFI IFI  incremental fit in this right and you  have CF I write confirmatory fit indices  so these are some of the absolute fit  indices which are used and you have a  magical number we say of 0.9 so 0.9 so  if it is if these values are above 0.9  and you can you might not take all the  values you might take only one of

from  the absolute fit indices and one from  the incremental fit indices right so if  you take these values and see that this  is above 0.9 then along with the  chi-square by degree of freedom and  these we say that the model is a fit  model right and we also find something  called the RMS CA this is also an  important thing which takes the root  mean square error basically and RMR  right so these this is the residual the  root mean of the residual

s right so  residuals are the unexplained  so unexplained part the residual the  square of the unexplained right  basically it takes right and the RMS  yield now if these two values are  sufficiently good below 0.8 we say you  know for our MSA is 0.08 8 percent below  8 percent less than 8 percent then it is  a fit model so after doing all this we  have understood that there is a the  measurement model the measurement model  has been used to check f

or the  constructs validity and reliability but  now conscious situation where we have to  check whether there is a relationship  between the constructs or not or what  kind of a relationship exists between  the constructs rates so to do that we  use the path analysis with the path  analysis right so the path analysis  basically is the structural part only  right so what is happening here we are  saying model a construct a is effective  let's say co

nstruct B this is a simple  case like a regression only but suppose  now we are adding a third variable let's  see the third construct now we are  saying that if if in this case  it is becoming more than a simple  multiple regression in a multiple  regression case you had this was the  illicit dependent variable this is the  independent variable but now in this  case this has change its characteristic  and it has become a new independent  variable a

nd this has become the final  dependent variable so in such a  condition which is actually what it is  happening in real life right so one  becomes the father of a house is depend  the family's depend on the father but  the father is may be dependent on his on  the office or the you know is office  ambience of his the climate or is let us  say the boss right or whatever ever he  is working so if he is tomorrow let's  say suspended from his job then

his  family will get affected or if he is  given a promotion then his family will  be productive positively affected so  let's say if I take it although it is a  very crude way of understanding this is  the boss or the employer let us say the  boss is a bad word the employer the the  father and let's say the children so the  children are dependent on the father the  father is dependent on the employer  right so this is the case where you want  to es

tablish such kind of relationships  to do that we use structural equation  modelling so it's a process for stress  testing a structural theory right a  structural theory is a conceptual  representation of the hypothesis the  hypothesis relationships between the  constructs now this relationship that we  have built there hypothesized  relationships we are yet to prove it  right so how do we prove it now we'll  see we'll see so how do we prove it so

after you know the mechanism becomes  like exactly like a multiple equation  only okay so it can be expressed in  terms of a construct more a structural  model that represents the theory because  set of structural equations so  everywhere there is a new equation it is  a case like a simultaneous equation  right so you are creating different  equations right from here to here here  to here so this equation that you are  building now you are testing t

hem with a  represent the structural model repels  the theory with a set of such equations  and is usually depicted with a visual  diagram the visual diagram is this okay  so what are we focusing here  we are focusing on the overall and the  relative model fit right and the size  direction and significance of the  structural parameter estimates now the  estimates the word estimates if you  these are this is some simple the  regression estimates ther

e are the  regression coefficients right now with  the help of the estimates we can say  whether a hypothesis is to be accepted  or rejected the estimates will help us  to say whether a hypothesis is to be  accepted or rejected now this is a  structural model but one thing before I  get into the structural model please  remember that the structural model the  relationship that you have built in this  case was the C right it should always  emerge fro

m an existing theory it is not  that we liked something like you know  the is affecting C we will say ah our C  is getting affected by a now that is  something like out of bases out of  reason right so just pulling an arrow to  make some you know to just prove your  point or prove somebody's point is not  the right thing to do your model should  be built upon very clear structure  theoretical premise okay so there are  two constructs here customer s

hare and  customer commitment customer share has  got four variables customer commitment  has again got four variables right this  lambda if you see the lambda these are  the factor loadings as you saw a factor  loadings in factor analysis expletive  factor analysis similarly they're also  the factor loadings okay so my special  theory hypothesis is that customer  commitment results from customer share  the customer share right so the arrow  moves f

rom that means the arrow is  moving from customers share to customer  commitments because thermo commitment is  a dependent variable right so structural  equation modeling there are five six  stages the one two three four here if  you see defining the construct of  developing the overall measurement model  designing the study to produce empirical  results assessing the measurement model  validity is a part of the measurement  model of the CFA right

the five and six  specifying the structural model and then  the  assessing the structural model validity  is the five and six are the part of the  structural equation right so the  structural model so remember when you  have let's say after you have done the  measurement model right so whatever  you're the Fitness in descended right  the structural model should be better  than the measurement model the  structural model should always be better  or h

ave a richer value or higher value  than the measurement model if it is not  happening then there is not much of a  significant improvement right so we need  to understand that the measurement model  was only measuring the constructs but  structural model you are putting in some  relationships right so the structural  model has to be more powerful and you  know robust then the measurement model  okay okay so how do you convert a  measurement model i

nto a structural  model so now you see the a this is a  measurement model so ABCD are these are  all there's a covariance relationships  right so there is a double-headed arrow  to form from one construct to the other  right so this double-headed arrow is  basically nothing but still talks about  the covariance relationship right but  here if you look at the structural model  it is very clear in saying that a and B  FX c c then FX d okay now these t

hings  that you see that for example this this  symbols right these are called the error  terms so if you these are the sometimes  we use it simple we use e 1 e 2 right  kind of error terms now what is the  error terms time and again I've been  explaining the error terms are nothing  but the unexplained part right so in  this case there are 4 items or variables  related with each Latin constant and 4 4  4 4 and there are 2 error terms here  right be

cause every dependent variable  will have a error term right the depend  obviously the independent variables do  not have an inner term right so why then  this is also an independent variable  then why it is having additive now  because it is also independent variable  for these  okay so that is why now I'm not getting  into the recursive and non recursive  model but just to if you want to  understand because a recursive model is  something where th

ere is a if you look  back this is a non recursive model and  the recursive model is only the only  difference is that when there is a  one-sided when the arrow moves from one  direction towards one direction only we  say it is a recursive model on the other  hand when there is a dyadic relationship  or a double the path is moving both  sides right so in that case this is  called a non recursive SEO model right  so now let us look at this case this

is  a it's back from the book of the thumb  and black had an understand so from  there we have brought so these are some  of the constructs and their  relationships so these are the  endogenous variable this is the  endogenous variable these two are the  exogenous variables exogenous means  independent and endogenous means at the  dependent okay now hypothesis is there  are several hypothesis if you can see  the h1 is saying EP now what is EB if we

want to see then I think it is B it will  be here  EP is environmental perception AC is  attitude towards coworkers yes I is  staying in tension OC is organizational  commitment and Jas is job satisfaction  okay  so environmental and what this  co-workers I think attitude towards the  coworkers are affecting the job  satisfaction and the organization  climate finally affecting the staying in  tension so the hypothesis is that  environmental effects

jjs job  satisfaction elemental also positively  affects OC AC effects AC effects  Jas AC effects OC right  and then Jessie J s sorry effects si and  OCFS si so one two two here four five  six seven right and Jase you know  see there must be one so if you see job  satisfaction leads to organizational  commitment this is one more right kind  of a relationship so this is a very  complex relationship right so so I'm not  getting into back into CFA now

so yes so  so let's look at this right so if you  are having such a condition now you have  established relationship now what should  you do after you have got the structural  model intact then what is what you need  to do is you need to find out the first  the model fit right so as you did in the  measurement model similarly you need to  find the model fit for this case also  and I said that the model fit should be  better than the measurement mod

el right  so once you have the measurement model  then what you do is after the you check  for the model fit suppose it is the  model is fit then you go to the next  thing that is to check for the fitness  industry and assess also you have  suppose your check and they are all  coming point 9 and above so that means  the model is now clear it is a fit fit  model right there is no problem with it  but in case in case there is a problem  of let's say t

here's a poor model  fitness is showing then a most has a  facility to improve the you know to  improve the mo see the software which  you use for little and a most a most  being a very popular software they are  to improve the model you can do one  thing you can use the modification  indices so modification indices indices  are one such value which is used to  improve the model fit now if you have  the modification indices then this  indices what i

t does is basically try to  find out which are the error terms which  have got lot of you know there's a which  there's a lot of unexplained part into  it so that those two errors are combined  so that the unnecessary are as  unexplained part will get reduced and a  model will improve better right so  so uh okay so once you've done this then  what happens is let's go back to the  hypothesis now what do you do is the you  need to go for the estimates

right the  estimates so the estimates we'll assume  in a most at least helps you in  providing this estimates this estimates  as I told you as nothing but the  coefficient regression coefficients for  each value now it will also give you  with it the estimates as well as its  significance right so when it gives you  the significance right from the  significance if suppose something is  significant right as the significance is  told shown in triple

star or double star  look depending on point zero one or  point zero five or point zero zero one  right kind of a significance level so  when when you are going through that  estimates you need to check that  relationship suppose J s ep2 Jas  now if that relationship is it will give  you an you know significant or non  significant relationship it will explain  that so if there if your significant  relationships significant then we say  that yes Jas

is really getting affected  by EP and it is not due to chance that  it has happened only this time right  similarly suppose out of it suppose you  found only one of them that's a Jas and  OC j s job satisfaction leading to  organizational climate leading to  ordinal climate this is not coming  significant that means what in this  model we have theoretically assumed that  job satisfaction also needs to organize  the climate and thus it affects stayin

g  in tension right so this is not coming  true that means this hypothesis is a  false hypothesis we cannot we cannot say  that your J is actually effect so C and  then that effects si right so this is  one thing that is that you are given  right so after the end after the end you  will get all the estimates and with  their significant no value significance  values and through that you can  accept or reject your hypothesis right  so in this case the

re are seven  hypotheses and we said tentatively let's  say only one of the hypotheses is not  accepted or it is rejected let's say so  it is rejected so we say in this study  there are six seven hypotheses out of  which the first six hypothesis we're  significant and they were they it showed  a positive effect as the relationship  you have to explain through the  relationship but only one of them did  not show a significant relationship that  means

it shows that there is no effect  okay okay so when you do this when you  do this you need to understand you need  to understand that there is something  also for example let's say there are two  constructs let's say two constructs this  is a third construct let's say this is a  fourth construct let's say right now  there is something we need to understand  ABCD for example okay okay  so now if you see in this case a is  affecting D also through C

okay B is  also affecting B through C and as well  as directly in this case if you shall  look at it right so B has a direct  effect also and B has an indirect effect  also so let us say these two are point  the factor loadings are point five and  point five right and this is let us say  point four right and this is let us say  point three these are the factor  loadings I'm giving now what will you  understand from here there are two types  of effec

ts one called the direct effect  and the other called the indirect effect  now in this case you have to understand  that the indirect effect from A to D  through C an indirect effect means when  the the variable passes through  intervening variable that means a  mediating variable a intervening you can  understand a mediating variable right  which is behaving like a mediator in  between right so 0.5 0.5 in to 0.4 is  equal to 0.2 so the indirect eff

ect is  0.24 a to do a 2 D and similarly from b  to  it is also it's a this could have been  point food also  so 0.4 into let's say point four so 0.16  let's say so this is from and what is  the direct effort from B to D now point  three so here we say that the direct  effect from B to D is higher than the  indirect effect and in this case but the  total indirect effect the total indirect  effect is 0.2 plus 0.16 okay so many a  times it happens it

is of interest to a  researcher to know give whether if we if  mediating variable comes into the  picture  whether it is advantageous or it is it  is not required sometimes it's like a  catalyst we have to understand like in a  catalyst the presence of a catalyst will  change the reaction right so direct  effect might be let's say the direct  effect might be significant might not be  significant right so if it is whatever  and the value could be les

s or more  whatever but suppose if it goes through  a mediating effect the mediator then  suppose we see that it is significantly  improving the the value is improving  then we say that there is a presence of  a mediation effect right so let us see  what is this mediation as you can see  here right so k goes through to e  through m m being the mediator right so  M mediates the relationship between K  and E if the diet effect of K on E is  diminished

when M is also a predictor of  E so we will if suppose you have a  direct effect  nobody wants a mediator until unless the  direct effect is not so good right so if  your mediator only adds to the  explanation or adds to the strength of  the relationship then only the mediated  is a justification otherwise there is no  justification for mid-age mediation okay  now adding direct effects so how do you  check for mediation  although this will not expl

ain much I  will do it directly here let's say there  are two types of mediation so one is a  partial mediation and we said the other  is a full mediation that means we will  Hey the mediator has a partial effect or  the mediator has a full effect now what  is this full and partial let us say a M  let us say B okay so this is something  like this so when a has a significant  effect on the mediator and M has a  significant effect on the dependent  va

riable right then and suppose a and B  also this is a barren and Kanis method  is a very although some people say is  over the old method but still the it is  a very basic method and the very clear  method suppose the presence of a  mediator and the absence of a mediator  if there is a presence of the mediator  and the absence of a redditor if the  significance level changes let's say the  direct effect the direct effect let's  say when you bring in

a mediator let  said the relationship becomes strongly  significant right and the diet effect  shows not so or trust strong then we say  it is a case of a full mediation right  that means when suppose a to be was  easily now was earlier is significant  now also is not significant let us say  and through this this is becoming  significant right then we say there's a  powerful mediation suppose both are  significant but this has got a lesser  value t

hen this then we'll say it's a  partial mediation right so these are the  ways of just justifying or you know  identifying okay this is ridden so  mediation involves the comparison of a  direct effect with the two constructs  while also including an indirect way  through the third conserve full  mediation is found when the direct  effect becomes non significant as I said  so when it is non significant and in the  presence of the indirect effect wher

eas  full meditation is when it is non  significant whereas partial mediation  occurs when the direct effect is be  reduced but still significant that means  what if I go directly also I will get  some money for example but if I go  through somebody I will get more money  that is a partial mediation but suppose  if I go directly I will get no money but  if I go through somebody I will get  money then it is a full mediation as  good as that right you

have to  understand so so well it says some of  the things okay there is one more  important thing this is called a after a  mediation weeks is there is something  called a moderation also now what is a  moderator a moderator this is always a  you know something like there are some  things which are used as a moderator now  some things automatic moderate the  relationship okay now what is the  moderating relationship now for example  many a cases g

enerally what is used is  gender gender we say is a moderating  variable now what do you mean by  moderating but that means whatever the  relationship you studied in this case  for example or the seven hypothesis that  might behave differently for male and  differently for female okay so if you  have a few exercises the same thing for  male and you exercise the same  relationships and test the relationships  for female that means it is different so

if you take only gender is one that  means there is you are not able to get  into deep and you are not explaining it  well so if you do a moderation test a  moderation test then we can we can  easily explain that whether gender also  moderate is gender is moderating the  relationship between let us say let's  say I am saying here I am in putting in  a moderator okay moderator now I am  saying the relationship between a and B  is moderated this ther

e is a mediator  also but this is a mediating and  moderating effect both are there right  so this if the moderator is affecting  the relationship right is affecting the  relationship so how do you check for  moderation it is very simple there's  nothing very complicated in case you  have continuous variables also you can  use a moderator because generally there  is a confusion between researchers that  moderator should be a categorical  variable or

not right if you have a  continuous variable also what you can do  is you can use some logic and create  some categories in it for example let's  say income from 0 to or let's say any 0  to 10 lakhs is let's say  1:10 to let's say 20 lakhs is to 22  let's say above is 3 now why doing this  what I have done is I have categorized  the variables right so moderation can be  tested with multi group SDM right so  what you do is simple in a moderation  te

st you first define the groups and  then you run it by taking each part of  the each group individually so by taking  it will have two values will generate  one suppose in this case the all the  seven hypothesis that seven you know Jas  and all these things so if you have for  male you have will have a particulars  the seven hypothesis the relationships  right seven hyper similarly for female  of so you will get it right now then how  do you know ki

that moderator what  should you do to look at the effect of  the moderation the best is to see is  there any variable or any relationship  that is significant in male and baby C  not significant in female or vice versa  only pick up those those relationships  which are significant in both the cases  the ones which are not significant in  though in either one of them they are  items for deletion  okay so by doing this basically what you  are doing i

s you are you are using a  moderation technique hi this is another  way also which I was saying the bod  resting variable can be collapsed into  groups which I explained or cluster  analysis also can be used to identify  the groups right that is also by the  hierarchical clustering you can identify  the groups also that is one more method  so whatever it is so you understood that  structural model helps in identifying  the relationships how they go

and there  is if there is an immediate ringing  effect we can check and finally if there  is a moderator how does the moderator  affecting the relationship that also can  be checked right so this is all we have  for SVM now thank you 

Post a Comment