Hi,
I am new to the concept of SEM and I am considering its use within my research. As a very brief descriptor I am looking at how elite athletes’ participation in post-athletic career development activities (e.g., educational courses or work experience) may be related to their sporting performances during their careers.
Essentially I will be using some validated psychometrics as predictor variables, one that looks at self concept and another that looks at conscientiousness as a personality trait. I also intend however to include two other predictor variables, one that measures positive experiences in the athlete’s career development activities and the other measuring the support they receive from their sporting organisation. The questions for these variables have been developed out of the findings of a prior qualitative study and would represent two more latent constructs in the model.
Now, here is my problem. I have two major constraints, 1. Response burden (e.g., the overall survey package is becoming very large) and 2. Time constraints of when I have to get the project finished. The questions I have developed for each of these new dimensions are relatively brief and simple and contain two questions for each of the main themes pertaining to both a positive experience in a career development activity and support provided by the organisation (16 and ten questions respectively). Although they have been confirmed for face and content validity I wonder about the potential for adequate factor analysis and reliability testing as there are only two questions per theme. There is also limited time for reiterations prior to the major study and consequently any factor analysis on these items (confirmatory I guess considering each pair of questions represents a theme) would occur after collection of all the necessary cross-sectional data in the major study.
Considering the time and total length of survey pressures I am under, to try and develop two constructs with credible construct validity may be out of the question (unless I get lucky in the SEM process). I have also sought to find existing scales/questionnaires that will measure what is required for these predictor variables however for these two concepts, the findings in my first (qualitative) study suggest that the environments are specific enough to warrant tailor made questions. So my question is,
Can an SEM include some ‘latent constructs’ that have been measured by observed variables that are not validated psychometric measures yet have been pragmatically produced and confirmed for face and content validity?
Additionally, how important does reliability testing become for those specific constructs if the observed variable is just a simple questionnaire? How credible would the latent variable be?
So there it is and apologies for taking so long to arrive to the point. As a final question, is it possible to include a dependent variable in an SEM that is a simple rating of athletic performance on a 5 point likert scale? Essentially the latent construct would be ‘performance’ however it is measured by only one observed variable (the 5 point scale).
Cheers,
Matt
Printer-friendly version
I'm a little confused as to all that you're doing, so I have a few questions. I'm not quite sure what you plan to do with the items from these two new scales you're constructing. It sounds like you have one 16 and one 10 item scale, but that the items on one scale relate in some way to items on the other. Is that correct?
Getting to your big question, the answer to "Can an SEM include some ‘latent constructs’ that have been measured by observed variables..." is yes. This is, at some level, the whole point of SEM. I'd probably add a few dimensionality tests to further support that your theorized structure is reasonable, and otherwise rest on standard SEM fit indices to support my final model fit.
On your last question, yes you can include ordinal variables, though you should treat them as ordinal rather than as a continuous variable. There's more info in the documentation and elsewhere on the forums. This variable won't really be a latent variable: you'll just be modeling the observed variable. There is nothing wrong with that.
-Ryne
Hi Ryne,
Thank you very much for your reply and apologies if I was a little confusing. The items on the two new questionnaires I am developing are to create two separate latent constructs in the SEM. The SEM will also include other latent constructs that use items from previously validated psychometrics. All the latent constructs will be linked in a path diagram.
I guess what was going to be the deal breaker between using SEM or not for me was the fact that the items I have developed for the two questionnaires will not have gone under the same rigorous process used to develop and validate psychometrics. They essentially will just be simple questionnaires (so really just observed variables related to the same topic). If it meant that using an SEM would not be defendable (providing the model fits) then I would look to other statistical methods.
From your answer however it appears that occasionally in SEM there is no choice but to model a latent construct on the measures that you have available (or have the time to create) and in some cases they are 'pseudo latent constructs' with your/my example of modelling a simple 5-1 rating of player performance and modelling it as a latent construct being testimony to that. My thought would be as long as I mention the limitations of my measures in the methods and discussion then SEM would still be a viable statistic for me to use, especially considering I wish to see how these constructs behave in relation to each other and the DV of performance?
Please feel under no pressure to respond. I understand you are probably quite busy. Using SEM appears quite interesting to me at this stage.
Cheers,
Matt
Dear Matt
The process you are describing could be seen as itself a validation of the constructs you are hoping to develop and the items you have chosen. If all your items load on your hypothesised construct then all is fine. If not you may have some more investigation to do.
Michael
Thank you Ryne and Michael!
I can see that as long as the model fit indices suggest that the items load on the hypothesised construct then I will be good to go and that it is ok to do the 'validation' work in the first use of SEM as essentially this is implicit in the SEM process. I could also adjust the model to a point where there is appropriate fit by removing items etc.
Thanks again for your help and suggestions.
I agree with everything Michael Dewey said. What you're doing is validating the new measure concurrently with its first empirical use. Using SEM should provide sufficient model fit stats that you can support this model over other alternative models with different factor structures. You can also use just the items for your new measures to carry out whatever other methods you see fit to evaluate reliability and validity in separate steps from your model.
Good luck, and have fun with this!