nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Hospital LOS model per physican's group

sweetycha

we have like 20 physician group( clusters of physicians)in the hospital. We would like to focus which physician groups should we target in order to reduce patients LOS(length of stay). Firstly we need to identify which physician groups have higher contribution ( also to what extend) on average LOS like finding out which physician groups are more likely to have higher LOS regardless of patients insurance, severity and disposition ( final destination (like home or rehab or transferred to another hospital). we need some kind of stat of this information to present to executives.

Also want to understand if there are certain physician group who have higher LOS only when patient's insurance is Medicare.. if other insurance then they do not have higher LOS.so we can target just certain insurance company of those physician group and see why insurance is taking longer time for these particular physician group.

I have moderate understanding in stat .. wanted to ask an expert opinion.

Please advise

Thank you

Best,

Accepted answers

dale_lehman

The coefficient of .24 would say that that group has .24 extra LOS compared with the average of all the groups (and is statistically significant). Each of the groups has a coefficient indicating its LOS compared with the average of all other groups - one group is held out from the report (you need to click on Expanded Estimates to see all of the groups). For an interaction term, you would put both variables in the model as factors, then highlight one of them in the factor box and the other in the potential factors (holding down the control key) and clicking the box that says "cross."

All comments

dale_lehman

You will have to provide more detail to get a more detailed answer, and you should also review some of the documentation on modeling in general. I can provide some initial guidance. Your response variable appears to be LOS. You want to include factors in your model for payer (Medicare, other,....), physician group, and other appropriate variables you have data on (type of diagnosis would seem to be important). I would suggest starting with the Fit Model platform and just doing a linear multiple regression to see whether any or all of the variables seem to be related to LOS. There are more sophisticated approaches, but I'd start simple and work you way up.

sweetycha

Thank you for your prompt reply

I have started with simple GLM model like this:

model 1.PNG

where as my dependent variable is opportunity days ( we have national benchmark LOS called GMLOS and our hospital los, so its basically LOS - GMLOS = opportunity days)

Here I just want to get the following answers:

1. If patients financial class (insurance) and disposition is constant, which group has higher OPPORTUNITY DAYS.

I think individual beta estimate on this model will give this answer. M model result1.PNG

MY result :

Am I on right track? Please advise... Also how do I interpretate the result to executives..They are basically wanted us to tell them which group they need to focus in order to reduce the average LOS.

oppdays.PNG

dale_lehman

You are on the right track. I'm not sure you need to use GLM - I'd start with a simple regression. Also, I can't tell how many different financial and physician groups you have - if there are many and a number have very few observations, you might consider regrouping them into a smaller number of categories. I supposed it might also be important to investigate potential interactions between financial categories and groups - the way you have it modeled, the financial categories will have the same impact for all groups. But, the general point is that looking at the estimate and its confidence interval is the main focus of your attention. And that is what you would want to communicate - but the issue of what and how to communicate is too complex to just outline here. It depends on the audience, the results, and the use it will be put to. Simplifying to communicate is good - oversimplifying is dangerous.

sweetycha

Thank you so much. for shedding light to my analysis. Far way to go but I am positive to start with something. I really appreciate your guidance and support.

In Jmp I am not sure where to add the interaction term. You brought a very good point as we are exactly testing out which group to what insurance group we need to start investigating. When we do not have interaction and we pick the group that have higher beta estimate plus p value < 0.005.

It is important to learn if there is particular group that has higher LOS when insurance is ABC or something .

Can you please help me who do I interpretate the B estimate of highlighted group on my previous post?

P_Bartell

I'm going to make two suggestions, one for you @sweetycha and the other for the consumers of the information you are conveying.

For yourself, I suggest you take a look at JMP's Statistical Thinking for Industrial Problem Solving online, free course. Upon successful completion of the course you'll have a sound grasp of the principles of Statistical Thinking, of which @statman has touched upon as well as basic analysis methods consistent with the course's learning objectives. Which are all up all in are two fold; 1. Enable the practice of Statistical Thinking and 2. "How to in JMP.". Granted the examples are not in a clinical setting, but, having done some pro bono work in a hospital setting, all the statistical thinking concepts and practices can be extended to a hospital setting.

For your consumers; First ask yourself, across the group, 'How literate are they wrt to Statistical Thinking'? If the answer is illiterate (or weak at best), then give each person a copy of Don Wheeler's "Understanding Variation, The Key to Managing Chaos" and demand each reads from cover to cover before you meet with them. Then keep your information dump to them exclusively graphical, with easy to read and interpret charts and graphs only. P - values, parameter estimates, models, and tabular output will only serve to muddy the message. My advice wrt to your graphics is "If it does not add, it subtracts." Keep them simple and uncluttered. One message per graph. Test your audiences understanding for each.