you may use multilevel analysis.

See this http://www.ssicentral.com/hlm/help7/faq/FAQ_Advantages_of_multivariate_multilevel_models.pdf

I can use multi level modeling to compare variables across sites, but I actually want to quantitate the amount of variability across multiple choices within a single institution and then compare that variability to an outcome. Multi level modeling will just account for the different sites, and in the end I am more interested in the variability at each site, not account for it. 

Dear friend

you may use some statistical or AI methods.

But AI is always beneficial over such kind of stochastic problems.

you may look at some of my papers (heart paper too):

Mandal, I., Sairam, N. New machine-learning algorithms for prediction of Parkinson's disease (2014) International Journal of Systems Science, 45 (3), pp. 647-666. DOI: 10.1080/00207721.2012.724114

Mandal, I., Sairam, N. Accurate telemonitoring of Parkinson's disease diagnosis using robust inference system (2013) International Journal of Medical Informatics, 82 (5), pp. 359-377. DOI: 10.1016/j.ijmedinf.2012.10.006

Mandal, I., Sairam, N. Accurate prediction of coronary artery disease using reliable diagnosis system (2012) Journal of Medical Systems, 36 (5), pp. 3353-3373. DOI: 10.1007/s10916-012-9828-0

Mandal, I., Sairam, N. Enhanced classification performance using computational intelligence (2011) Communications in Computer and Information Science, 204 CCIS, pp. 384-391. DOI: 10.1007/978-3-642-24043-0_39

Mandal, I. Software reliability assessment using artificial neural network (2010) ICWET 2010 - International Conference and Workshop on Emerging Trends in Technology 2010, Conference Proceedings, pp. 698-699. DOI: 10.1145/1741906.1742067

Mandal, I. A low-power content-addressable memory (CAM) using pipelined search scheme (2010) ICWET 2010 - International Conference and Workshop on Emerging Trends in Technology 2010, Conference Proceedings, pp. 853-858. DOI: 10.1145/1741906.1742103

Mandal, I. A novel approach for accurate identification of splice junctions based on hybrid algorithms (2014) Journal of Biomolecular Structure and Dynamics

pp. 1-10 | DOI: 10.1080/07391102.2014.944218 PMID: 25203504 http://www.tandfonline.com/eprint/TzMeXxpEXxujtEATHwqY/full

I hope this will help you.

Thank you.

Best regards

Dr.Indrajit Mandal

You need multilevel modeling. If you are interested in quantifying variation between institutions (assuming treatment effect is fixed across institutions), you can use a random intercept model. If the treatment effect vary between institutions (not constant over institutions), a model with random intercept and random treatment-by-institution effects (random slope) is needed.

Thanks for everyone's replies so far.

I currently know what medications are given at each institution and how often each of those medications is given. So I can see that there is variation at each step within the system. It is that variation I want to quantify, I think just using a multilevel model will take into account the variation in medications as well as all the other ways that the institutions differ. So I would like to figure out a way to pull that variation out from the rest of the model and quantify it separately. We know that there is quite a bit of regional variation that exists between institutions and we think that it is multifactorial. The idea of this study is to decide how much of that variation is from the choices that the physicians make in treatments. 

It depends on your response variable. If you use a linear mixed model or a generalized linear mixed model, ICC (intra class correlation) can be computed. 

Austin,

From where I sit, my reading of your description of intra-institutional variation in medication treatment patterns for the same clinical condition already suggests to me why you would not want to use multi-level modeling (MLM). MLM is most commonly used when there is a basis to assume there is an institutional homogeneity of pattern practice.  It (MLM) is then used to account for, and analyze cross institution variation, as if there is a common practice pattern within each institution to begin with, and your description of things counters that assumption.  In my experience, your description of intra-institutional variation is the more common reality of clinical practice, stemming from patient to patient variation in clinical circumstances, and is the main reason why MLM is not always appropriate for comparative variation analyses.  In my view, many advocates to use MLM, especially without considering essential assumptions, has taken on a form of method-idolatry, and it therefore does not deserve regular use for the above key reason - there commonly exists intra-institutional variation in practice patterns.

I think the model you choose needs to flow from your hypotheses about planned comparative differences that you do want to assess. If, for example, you want compare physician to physician differences, both within and across institutions, you will need sub groups of similar patient cases, with markers for treatment eligibility candidacy for each patient to determine ideal candidates for particular medications/procedures, so that you will be comparing ideal eligibles for each physician. And you will also need reliable markers for treating physicians to sub-group their "panel" of ideal eligible patients treated/not treated. I am assuming that you have a preferred treatment protocol that is guideline based, and that allows for determining patient specific eligibility for each and all treatments for which you want to observe variation, i.e., aspirin use, beta blockade, etc. Additionally, I am assuming you have abstraction case data to derive denominator eligibility status for each patient as well. 

Given these things, you should be able to determine for each physician, for their pool of ideal eligibles, for any treatment (medical/procedural), what their compliance rate relative to the guideline based - preferred practice would be, and what the range of rates of best practice is for each institution.  At this level of comparative analysis there is first, simple, descriptive statistics that let you see, without additional adjustments, what the variation in clinical practices may be relative to preferred guideline practice, which implicitly expects 100% compliance for all ideal eligibles. 

With those simple figures you can construct institutional comparisons, aggregating the pool of physician results, to see what variation exists across institutions relative to preferred best guideline practices, for each treatment.  

Then, with those findings you can pursue additional adjustment for physician, institutional and other community factors that might account for variation in practice, or patient outcomes, using some form of multivariate logistic analyses.  

Not knowing what details you have to work with in data or key hypotheses makes substantive suggestions challenging to offer, but I hope by making my assumptions somewhat clear, you may be able to make some headway on your variation analyses.  Best of luck.

Hi,

I think, your study is related to Multi Level Modeling (MLM) method.  So, you go through this link and do your research.

http://books.google.com.sa/books?id=IAamvxsVlGAC&pg=PR12&lpg=PR12&dq=compare+system+variability+across+many+institutions&source=bl&ots=4WqwHZa4N6&sig=wPF7yFldeKDUbvsIqIUqemImL1Q&hl=en&sa=X&ei=KP9IVNWiOc_PggSppoLoDQ&redir_esc=y#v=onepage&q=compare%20system%20variability%20across%20many%20institutions&f=false