
#1




Performing a CaseMix Adjustment
I don't have much knowledge on this, so I'd really appreciate it if someone could critique my method. I'm given data on a clinic that currently has a program for prematurely born babies. I have the babies' birth weight, length of stay, and health issues. I'm also given benchmark data which contains all of the above categories as well but from other clinics. I'm asked to find the average length of stay by source (benchmark or program), and perform a casemix adjustment on that. What I did was I plotted the average length of stay by birth weight of all the data and found that as birth weight increases, average length of stay decreases (duh). Thus, for any data points that have a higher than normal length of stay for their respective birth weights, I'm looking at the health issues and deciding whether the issues may have caused length of stay to increase despite their higher birth weight. For example, if the baby had surgery or something serious done, I would expect that they stay longer in the clinic. Can anyone critique my method or give me some advice? My biggest problem is that because I plotted the data before performing the casemix adjustment, I don't know if I can give credibility to the trend of the plot. Also, I'm doing all the stuff in Excel.

#2




Interesting. My first thought was that you don't really have enough demographic information on your population to really adjust. But upon further thought, I think you might. Here is how I'd approach this problem, assuming that you can't get any additional data elements.
The first step is to see if you have enough data to feel comfortable that you are able to tell if the populations are actually different. For example, if there are only 20 births in the program clinic but 12,000 births in the benchmark clinics, it is unlikely that the program clinic births are credible enough to warrant adjusting. Personally, this doesn't have to be anything real formal, just a smell test. If you don't feel that the data is credible enough and you need to convince your management of that fact, then do some fancy statistical tests that you learned in school. Or bluff your way through it. I use the second approach because management finds my efforts to avoid work charming The second step is to convince yourself and those who are requesting the analysis that there is actually a difference in the relative morbidity of the two populations (program clinic versus benchmark clinics). You only have two data elements that you can use to verify this, the birth weight and the health issues. What I would do is look at the distribution of these two elements in the two populations. Look at the distribution of birth weights, the mean birth weight, median birth weight, etc. for both populations. Do they look different? If all of the statistics come back looking similar then there is no need to make an adjustment because there is no real difference to adjust for. Repeat the comparison process for the health issues. See how easy I made that part sound? Actually, I suspect this is going to be problematic. I don't know how your "health issues" are represented in the data. I'll take a couple of guesses and explain what I'd do. You may have bunch of different fields with each representing a different type of condition. You can do things then. First, you can count the number of different conditions and use that for a demographic comparison. Second, you can compare the number of conditions in each block and see if the incident rate is similar between the two groups. If there are differences between the two populations, you can adjust based upon the conditions that vary. However, I'm guessing that isn't how your data is set up. It is more likely that you have a field that has a bunch of typed in notes. This is a problem for you but I've been down this road before too. My advice is to create some additional columns that track two things, the number of conditions the baby has and a severity level. You will need to go through every line and assign the two numbers. For example, a normal vaginal delivery would have 0 conditions and a severity of 1. A csection would be 0 conditions with a severity 2. A child born prematurely with a congenital heart defect and down syndrome would have 3 conditions and a severity of 10. Now that you have given both populations some comparable statistics you compare them just like you did with the birth weight. Do they look different? If not, it is once again time to bluff your way out of some meaningless work and endear yourself to management. The upside here is you have some cool data, maybe even some impressive full color graphs, to tell your story of why you are too lazy, whoops I mean smart, to continue with the analysis Honestly, I suspect that when you get to this point you will find that the distributions of the populations are similar and no adjusts will actually be needed to compare the ALOS on an apples to apples basis. If you find that the populations are different you need to move to step three. Step three is two determine what the impact of your indicator variables (birth weight and health conditions) on ALOS. It sounds like you went directly to this step when you started your analysis. You made a good start. You found a correlation between birth weight and ALOS. Now you need to figure out how to apply differences. Is the relationship linear? In other words, if weight increases by 10% then ALOS decreases by 1%. Is it exponential? Is it something else? You'll be able to tell roughly from the shape of the graph. Best case it is linear and you can use excel to get the formula for the line and use that formula to adjust the program raw results. Basically you need to come up with a function that approximates the shape of your data plot and use that function to adjust your program results to the same basis as the benchmark program. Using my example from above, if the ALOS for the benchmark program is 3.5 days and the average birth weight is 7.5 pounds while the the ALOS for the program is 4 days and the average birth weight is 6.75 pounds then the adjusted ALOS for the program will be 3.96 days. This is obtained by determining that the birth weight is 10% lower in the program population thus ALOS is 1% higher due to difference in birth weights. I get 3.96 by diving 4 by 1.01 (decreasing the ALOS by 1% to get the two populations on an apples to apples basis). Repeat for the health conditions. No tricks this time. One thing to be aware of is that there may be some comorbidity between birth weight and the health issues. If fact, it is pretty likely. So instead of adjusting separately, you will want to combine the two fields together in some manner and then plot them out. Or you can replace step three with a regression analysis. I've only used SAS to do this so someone else will have to give advice on an excel method. Step four, adjust the program results as described above using your function (or coefficients of correlation if you did the regression analysis). Step five, present your findings and collect all the glory. Sorry for the wall of text and I hope it helps. 
#3




I would do the following:
Create a regression model based on the benchmark data. This will lead to some choices. (a) Does the relationship between birthweight and length of stay appear to be linear? First option could be to create a regression model with just a birth weight predictor and see how well it fits. (b) How many health issues are in the data? (c) Does adding health issues improve the model fit significantly? I would use AIC/BIC to determine this (maybe adjusted Rsquared if that's all Excel will give you). Just keep in mind Rsquared will always increase with more predictors so this won't help you. (d) Also think about interaction. Multiple health issues in the same person may affect the length of stay differently than the individual issues alone. Once you create a reasonable model on the benchmark, then use this model to predict the length of stay for the newborns in the program. Compare the expected length of stay to the actual, then that can help to see how different they are. You can also create a prediction interval at a certain level to determine the uncertainty in the prediction estimate. I don't know how to do this in Excel unfortunately. I believe it should be possible but you may have to manually calculate some of the items by looking up the formulas. Last edited by AMedActuary; 11162017 at 11:29 AM.. 
#5




Hmm I've been working at this for a while now, but I still have some questions. First of all, since I'm supposed to find an adjustment for the benchmark data as well, how can I use the benchmark data as a model? Wouldn't there be no need to adjust the benchmark data in that case? Also, how do I assign severity values to these conditions? My biology education stopped in high school, so I really don't know which conditions are extreme or minor. As of right now, I've divided the conditions into five groups based on average length of stay percentiles (020%, 2040%,...) using the benchmark data. Is this fine?

#10




Anyways, the problem specifies that I only use conditions to adjust LOS, so I guess I can leave birth weight out of it. My final method is as follows: since I have four classes of severity (1,2,3,4), the average of that is 2.5, and I will multiply or divide the LOS of each observation by x such that the observation's severity multiplied or divided by x is 2.5. What do you think?

Thread Tools  Search this Thread 
Display Modes  

