Difficulty of learning structural equation modeling

[loveoforganic2]

I'm completing a masters thesis on a patient communication education intervention designed to facilitate physician-patient communication. I'm proposing that changing constructs within the health belief model and social cognitive theory will impact patient communication (several constructs within this), which in turn will impact outcomes of the medical encounter (e.g. patient-centeredness of the encounter, but there are also several constructs within this), which in turn will modify treatment adherence and ultimately biomedical health outcomes.

This seems like it would be amenable to structural equation analysis (if this doesn't seem to be the case, please say so). However, for the thesis, I'm supposed to conduct the statistical analyses myself, and will need to justify myself to statisticians. After some preliminary reading, I think I'm going to need to look into alternative analytic methods, but I figured I'd seek input here. How difficult would it be to self-learn structural analysis, coming from a background of having taken only introductory stats?

Thanks in advance for any input

---

Members don't see this ad.

 

[MCParent]

I'm completing a masters thesis on a patient communication education intervention designed to facilitate physician-patient communication. I'm proposing that changing constructs within the health belief model and social cognitive theory will impact patient communication (several constructs within this), which in turn will impact outcomes of the medical encounter (e.g. patient-centeredness of the encounter, but there are also several constructs within this), which in turn will modify treatment adherence and ultimately biomedical health outcomes.

This seems like it would be amenable to structural equation analysis (if this doesn't seem to be the case, please say so). However, for the thesis, I'm supposed to conduct the statistical analyses myself, and will need to justify myself to statisticians. After some preliminary reading, I think I'm going to need to look into alternative analytic methods, but I figured I'd seek input here. How difficult would it be to self-learn structural analysis, coming from a background of having taken only introductory stats?

Thanks in advance for any input

IMO, learning to /do/ it (i.e., draw the right squares and lines in amos) is pretty easy. Learning what's actually happening can be harder, but it's not that rough.

If you're using only manifest variables (i.e., not creating latent constructs), then it's not much more than more complex regression and shouldn't be too hard.

If you're using latent constructs AND you have to explain things, it could also be reasonably easy still, or you could run into things that make it harder (Heywood cases, nonpositive definite matrices, etc). But it's still way easier than HLM or something.

http://www.statisticshell.com/docs/sem.pdf
http://tcp.sagepub.com/content/34/5/719.abstract
http://tcp.sagepub.com/content/33/3/269.full.pdf

 

[loveoforganic2]

IMO, learning to /do/ it (i.e., draw the right squares and lines in amos) is pretty easy. Learning what's actually happening can be harder, but it's not that rough.

If you're using only manifest variables (i.e., not creating latent constructs), then you're it's not much more than more complex regression and shouldn't be too hard.

If you're using latent constructs AND you have to explain things, it could also be reasonably easy still, or you could run into things that make it harder (Heywood cases, nonpositive definite matrices, etc). But it's still way easier than HLM or something.

http://www.statisticshell.com/docs/sem.pdf
http://tcp.sagepub.com/content/34/5/719.abstract
http://tcp.sagepub.com/content/33/3/269.full.pdf

Thanks for the input. I'm glad it's not a "no way in hell" situation - I'd much prefer to do my analyses this way. However, some (half? 😛) of your vocabulary was over my head, but those links look digestible. I'll give them a read over and see where I feel I stand

 

[Member6523]

If you're using latent constructs AND you have to explain things, it could also be reasonably easy still, or you could run into things that make it harder (Heywood cases, nonpositive definite matrices, etc). But it's still way easier than HLM or something.
\]

I don't know about that - HLM is pretty straightforward in terms of design. SEM can get pretty muddled when you have to factor in model identification, fitness, trimming, etc. Not to mention the headache of constraining/loosening parameters to fit theory and the model itself. I think there's plenty of opportunity to screw up SEM if you don't know what you're doing. HLM by comparison allows for more lax assumptions and is easier to interpret overall IMO.

 

[AcronymAllergy]

I agree with both of the above posts--SEM is fairly easy (perhaps almost too easy, thanks to AMOS) to learn how to do, but like many other stats-related topics, can be much more difficult to fully understand if you want to get into all the nuts and bolts of it...

...however, my SEM professor (an I/O psych person who was big into SEM) said he very, very much regretted doing SEM rather than something more straightforward for his masters thesis.

 

[MCParent]

...however, my SEM professor (an I/O psych person who was big into SEM) said he very, very much regretted doing SEM rather than something more straightforward for his masters thesis.

Agreeeeee. I've had requests to do SEM in papers where it really wasn't necessary. If you're ever asked this, you can cite the APA itself as a reason not to though! (http://www.apa.org/science/leadership/bsa/statistical/tfsi-initial-report.pdf)

 

[Lisa44201]

SEM itself is not too terrible to learn if you've got a decent grasp on factor analysis & regression. SEM software is where things can get complicated, depending on the specific software and your ability to learn it. What programs are available to you?

 

[Sanman]

Agreed that learning to code the equation into the program is the much harder than learning the theoretical constructs. When I looked at SEM for my dissertation, I ended up going with multiple hierarchical regressions as the number of variables meant a lower error rate with multiple regressions vs SEM. It really depends on what you are looking at.

 

[loveoforganic2]

Thanks all for the input. I'm going to run the idea by one of my program's stats professors after I get through those links and feel I can speak with a good effort put forward.

What programs are available to you?

SAS, with which I have only a small bit of experience unfortunately. Worked a lot more with SPSS in undergrad

 

[coldsweat]

You need a large sample size to obtain trustworthy parameter estimates with structural equation modeling - how many participants do you plan to recruit?

Based on your description of your study, it's not immediately evident to me why you need to use SEM. Maybe I'm not understanding your study fully, or I don't have enough details, but it appears that standard analyses used in treatment studies combined with mediation analyses could answer your questions.

I use structural equation modeling when I have multiple measures of the same construct. You mention multiple constructs, but you don't mention how many measures of each construct you have, which will determine whether SEM is feasible for your study. Before you use decide to use SEM, you need to construct the model and determine that it is not underidentified (resources on SEM will explain what this means). I recommend Rex Kline's book Principles and Practice of Structural Equation Modeling as the best introduction to SEM.

 

[Lisa44201]

SAS, with which I have only a small bit of experience unfortunately. Worked a lot more with SPSS in undergrad

If you're going to try SEM, you need SEM software; SAS/SPSS won't build the model you've described. AMOS? EQS? Lisrel? Mplus? (Not ideal for beginners unless you're not afraid of code.)

I'm going to respectfully disagree with Coldsweat - I think you could use SEM to model this.

However.

That's a hell of a model for a Master's Thesis.

I'd simplify it. A lot. The way I read your opening paragraph, you've got four measurement models and three structural models. I've seen PhD dissertations with less than that.

 

[MCParent]

Given that OP is new to SEM (so doesn't know what words to write to describe building a model), and provided few details, I don't think anyone can comment on the feasibility, or lack thereof, of using SEM or any other method ATM. OP should take a look at the links I posted and the Kline book (and any other basic summaries) and see if that helps her/his approach.