Exploratory Factor Analysis (conceptual)

this virtual lecture is about exploratory factor analysis as a review the rows in our data set are the responses of individual participants the columns in our data set are the survey questions they provided answers to so participant number 251 responded with a value of 5 for the item called grade 3 this means that participant number 251 strongly agrees that she deserves the grade she received the set of items from our survey now entered as columns in our data set together measure the constructs we're interested in for example grade 1 2 & 3 comprise the construct called grade as a final bit of review there are two different theories we are testing when we make causal claims the first is called a measurement theory and is assessed using a measurement model this is what I'm going to explain in this video the second theory is a causal theory and is assessed using a structural model the structural model includes causal claims in the form of hypotheses here is an information-rich slide you're welcome to chew on for a while the short story though is the factor analysis is useful for identifying groups of items which many cases or survey questions that are strongly correlated we assume these groups of strongly correlated items represent some reflective factor or construct because they move together consistently here's a visual way to explain it for those who like that sort of thing let's pretend each of these boxes represents a variable in our data set they are all moving or in other words study participants responded differently to each variable for some items or respondent may have answered hi on the scale of responses while for another variable the same respondent may have answered low on the scale of responses the movement of variables in this illustration represents the collective movements of all responses our job in an is to identify which variables are moving together when a goes up does B go up or down or is there no consistent correlated movement hmm to me it sort of looks like a and C oh no they're not moving together maybe be in a Howie and F look like opposites let's see can we follow them for a sec and then oh they're getting off now sometimes when egos up F goes up and sometimes when egos up goes down it is very difficult for a human to accomplish this sort of task it involves some pretty fancy statistical sweetness to figure it out luckily we have software to do the job for us SPSS will identify the underlying groupings of variables for us in this set of variables we can extract two factors represented here in two different colors once identified it's pretty obvious which variables move together how they move and how strongly their movements are correlated whereas before it was a pretty difficult task here's another way to look at it our variables are correlated with each other to a greater or lesser extent a factor analysis finds consistent correlations among groups of variables this slide shows roughly what SPSS receives as an input and then what it gives us as an output after conducting an EF a when trying to find correlated groups of variables the resulting pattern matrix is far easier for us to interpret than the correlation matrix now how would you factor these items what are their underlying traits how do they move together or how are they correlated perhaps two large groupings could be based on political party affiliation but to what extent does that particular trait influence them perhaps we might have groupings based on the ratio of facial hair to head hair what if we observed groupings based on age could we observe groupings based on gender hmm nope not yet what about race you but one of those extracted factors or groupings would be very small the point is that the variance or movement in variables depends on many things if we find a group of variables that seem to move together we must identify an underlying theme or trait common to all variables in that factor the underlying theme is the construct identifying and labeling these underlying constructs is our measurement theory in other words we theorize certain variables move together for certain reasons where the reasons are the latent constructs here is another information rich slide you can pause for later consumption for now I'll just summarize we use the factor analysis to explore the data for patterns in variable correlations to reduce the number of variable traits into smaller lists of constructs and to confirm our measurement theory that certain items capture a latent trait so why do an EFA well let's say we have our awesome theory that it says playfullness predicts joy efficacy predicts usefulness and usefulness and joy predict performance that's all fine and dandy except that when we collect data will use multiple items to measure each of these constructs the circles here represent those items their proximity to each other represents how strongly correlated items they are now notice that this is not nearly as clean as the box and arrow model we envisioned but in reality this is what we are working with not all measured items capture their underlying construct effectively or consistently for example the blue circles represent the playfullness items notice how some of them are tightly correlated or close together while a couple of them are often right field and are more correlated with items from other constructs without doing an EFA this muddy mess is what we would end up testing can we even say that usefulness is a distinct contra from efficacy or that one of the items from joy doesn't really belong to the playfulness items so we do an e FA to assess the actual rather than just theoretical correlations on their items again the circles represent the items and their correlations are represented by proximity z' what we want to do is to capture a construct through a tight group of strongly correlated items in order to do this we may have to trim away some items which are not strongly correlated notice that I haven't moved any items only trimmed away those that were not grouped closely with any other items unfortunately that meant removing all but one of the efficacy items which leaves our efficacy construct unmeasured this occurs fairly often our data collection instrument often a survey simply does not capture all constructs as well as we had hoped on rare occasions we may also find that we have inadvertently captured an unexpected construct when this happens you can go back to the instrument to see if the wording of the items contains a common theme if so then you have accidentally captured a unique construct this construct might then become part of your new measurement model and the poorly measured a construct would then be dropped I mentioned earlier that a factor analysis allows us to reduce the number of variables we are dealing with here's a way to visualize that feature without a factor analysis you might be tempted to test the model on the Left where every X item is causally related to every Y item however if we identify their underlying factor we can simply relate all of the X items together to all of the Y items together which is a much simpler causal model the next several slides are just here for your reference I'm going to show them for a few seconds and then move on to the next one you're welcome to pause to understand the information better these slides will be available upon request and on blackboard for my class I hope this has been helpful in other videos I go through and explain what each of those slides means

Channel: James Gaskin

Views: 133,674

Rating: 4.9250937 out of 5

Keywords: EFA, Factor Analysis, AMOS, Statistics, Measurement

Id: Q2JBLuQDUvI

Length: 9min 50sec (590 seconds)

Published: Wed Nov 05 2014