8. Exploratory Factor Analysis

welcome back to this video looking at the exploratory factor analysis so what is exploratory factor analysis otherwise known as EFA basically the purpose of it is to reduce a larger set of questions into a number of factors or sub dimensions so imagine me getting all the students grades for a particular class based on all their assessments and then using an EF a or exploratory factor analysis to break them down into different factors these factors based upon the data might be excellent students moderate achievers borderline passes and perhaps apathetic sailors so we start with a large bunch of data and we reduce it down into different dimensions based upon the the way in which that data falls so why do we do it we do this as usually we often we have sub dimensions nested within an overarching dimension or a construct so in this data set we're going to look at PP which relates to pet pampers and we've got nine items that we've used to measure pet pampers so we're going to try and see if there are any sub dimensions within those nine items as an example from my PhD a component of this involved discerning between the different types of emotions that consumers experience when they're exposed to an innovation so I started with about sixty emotions and then used exploratory factor analysis to reduce it down to a preliminary amount of about thirty spread across six factors of different emotions that all related to each other so the idea is within each factor there's emotions that it's similar to each other but they're distinct from the emotions of the other factors so that way they're sub dimensions of the overarching broad a construct so we can then use these factors or sub dimensions as part of regression analysis or calculate means using those grouping variables or do a range of things with them so how do we do it to do this we firstly select analyze as we do with all of our stuff dimension reduction towards the bottom and then factor just give me a chance to just the screen a little bit within we've over the variables that we want to conduct the essay on into the window so we've got PP 1 so X PP 9 so I'm now analyzing these 9 pet pamper items within select rotation and I'm selecting direct oblem in' so direct oblem and is an oblique rotation that's used because it's likely that the nine items we've looked at that all all relate to each other so in other words then nine items that are measuring a similar thing that being the phenomenon of being a pet pamper if we were using items that didn't relate to each other for example we've got a whole bunch of items including age income purchase history etc where each of those isn't specifically related to each other we would then use an orthogonal eyes rotation and the most commonly one use for that is what we call very max but because all those items are related to each other we're going to use direct oblem continue we're then going to hit okay and run the analysis so the first output we want to view is the the total variance explained when we have here and what we can see here is we've got initial egan values so it's telling us there's nine different factors that have emerged or principal components each with different eigen values so we want to explain as much variance in the data as possible but only through a number of factors that make sense if we have too many factors then it's not going to make sense or we're going to be reducing them down so narrow that it won't make any sense so we could create nine factors here now we've got nine items so that would basically be one item in each of those factors which would perfectly explain the variance of the data as you would expect it's about finding that middle point or that nice balancing point where there's not too many factors but also that there's a large amount of variance in the data explained the way we usually determine that is if the factors the egan value of that factor is above one so we can see here that we've got three factors one two and three they have egan values above one and together they explained seventy one percent or seventy one point zero four percent of the variance in the data so this is a good thing again if we included the fourth factor we need to get the we've told the computer to only include factors that have eigen values greater than one but if we turn that off we would get a little bit more variance explained in the data seven point seven six more but would actually be at the cost of the strength of each of those factors so we're going to move down to the pattern matrix and as we can see here this is a bit messy we've got items 1 to 9 over these three factors or components one two and three now because we're using direct oblem we use pattern matrix if we were doing the more common very max again doesn't suit this analysis but then we would look at the structure matrix I think it's called so looking at this shows our key results with the number of factors extracted expressed through a number of columns three factors three columns and the items that belong to each of those factors but looking at this this is a mess it's very hard to tell what's going on here it's virtually illegible so we need to make this output easier to read so what I should have done in the first place and what I'll do again now by analyze dimension reduction factor so just back to the same approach is options sorted by size and then we're going to suppress small coefficients so we only want to show those coefficients that are greater than 0.3 0 or then run it again same results in terms of all that but now our pattern matrix looks very different that compared to that so the results are the same but just the way in which it's showing us and the order of those items so what we did first is we sort of them based upon the factors so we can now see that rather than being 1 to 9 it's in the order of the factors so factor 1 the items in our 7 5 & 2 factor to 8 6 & 4 factor 3 3 1 & 9 and it's also suppressed or not shown any factor any associations that are greater that a less than point 3 0 and as a result we can see three very clear factors have emerged here so we can see that the items within a factor in other words these three or these three these three are strongly associated with each other but show almost no association or in this case we can't see any association with the ones outside of the factor which would be if they had another Association at a different point so this is because we'd suppress any ones below point three zero so we could turn this off on not predominance we saw before and there would be a bunch of different ones here but also because we've sorted these based upon then they're late - they're all chunk together there but the general rule of thumb is that you don't want an item to have an association of at least twice the size of its largest association with an item of another factor and on top of that we want all of our associations within factors to be about at least point six zero zero so we can see that here these are sorted via biggest to small so the smallest for this factor is point eight point eight to two point seven one they're all greater than point six good and none of them have any factor associations in other factors that are greater than point three we can tell that because we've suppressed any that are smaller than point three so if there was any greater than point three we'd be able to see them here so here we have a very clear IFA with three factors each with three items in it so it's the first part so what we can see is that these items reveal three sub dimensions but it's up to the researcher to interpret what those sub dimensions are and we do that by closely reading the items so I can make sacrifices to Pat my pets my spare money goes towards my pet I can afford to spend money so we can see how those three items relate to each other conceptually in the same way that these three relate to each other and these three but importantly the items within this one a different to the ones that they all will relate to tech tampering they're a different sort of pet pampering to the other ones so we can actually then they label these factors we'll give them a name at least conceptually and arbitrarily I've for example called pet pamper factor one spare resources for their pet and maybe pet pampering factor three is that you're a crazy pet person so what's important though is we let the the concept of the items and theory dictate what we're doing here so the numbers have come out based upon the EFA but sometimes they happen a little bit strangely or that they just a purely statistical things that have come together rather than actually things that make conceptual theoretical sense so we always need to find that balance between the two so if we were in a situation where was some cross loadings as in ones and other factors that are very high or there's perhaps items within a factor that don't relate to the others we've got a problem we need to run it again but perhaps remove those items from that so all we would do in that case would be go okay that's our starting point with those nine items now for example pet pamper seven didn't make sense or as a spurious Kok Association we remove it and then run it again and we find that those associations actually change in that case so it's really important we do that and with EFA this is an iterative process we need to this one's come out clean because that's sort of how I set the data up but generally doesn't come out clean first time around as I mentioned my PhD we started with sixty items and then used it to reduce it down to thirty approximately thirty based upon theory and the results that were there so it is exploratory there's no fixed way of doing it but we do need to sort of find that balance between theory and of data so later on we're going to look at how we can then use these constructs that we've created by creating new variables based upon them and do some main analysis perhaps a regression

Channel: Luke Butcher

Views: 1,222

Rating: 5 out of 5

Keywords: SPSS, EFA

Id: kchN1_wRKhg

Length: 9min 46sec (586 seconds)

Published: Mon May 08 2017