StATS: Physician Performance Data (January 27, 2000).
Dear Professor Mean, Producing statistics of physician performance or group performance or whatever seems to be one of the great growth industries in medicine. Graphs of performance in just about anything seem to be produced - usually with something that looks at first glance like a normal distribution (and almost never with any statistical addenda). But I would like to know whether we can use them sensibly as anything other than pictures. In particular when I am one of the subjects of the analysis how do I interpret my own performance?
Do like everyone else does. When a graph or table shows that you are the best physician in your group, praise the method as innovative and cutting edge. When it shows that you are the worst physician, pull out your stock complaint about "lies, damned lies, and statistics".
Seriously, this is a difficult area. The bad news is that measures of performance will typically be subject to strenuous disputes, even when they are based on solid statistical methods. But the worse news is that the statistical methods frequently used are, at best, simplistic. It would be impossible to do justice to the complexities of physician performance in this brief web page, but I can make some general comments about the limitations of statistics. I also want to encourage the use of control charts as a good way to view this type of data.
Statistics are great at characterizing the behavior of groups, but they don't do as well when they try to characterize the behavior of individuals. With care, you can use Statistics to characterize individuals, but you want to avoid blindly using the Statistical methods that have been developed for clinical trials. Just as a simple example, you should note that characterizing individual behavior is not an activity that fits in well with the traditional hypothesis driven research. Do you set up a separate hypothesis for each doctor?
There is an additional problem. Most rating systems fail to properly adjust for all sources of uncertainty. Some hospitals, for example, have a smaller case load and these estimates are more unstable. So an outlier for a small hospital may simply be normal variation.
A promising approach is the use of random effects models, empirical bayes approaches, and shrinkage estimates (these are all interrelated). These models, unfortunately, are very complex, and require extensive consultation with a professional statistician.
There's a more fundamental philosophical issue. We have a tendency, especially in the United States, to want to rank and rate everything in sight. We have the top 100 movies of the past century, and the Places Rated Almanac of the best places to live. Many companies are returning to employee evaluation systems that enforce a quota of at least x percent unsatisfactory ratings. These efforts to rank and rate seem innocuous enough on the outside, but do they really serve a useful purpose? What are the hidden costs? It may be worthwhile to read some of the thoughts of W. Edwards Deming, Alfie Kohn, and Peter Scholtes. After looking at their perspective, you may decide that your efforts to identify good and bad PCP's may not be appropriate.
It is tricky to decipher when a deviation is part of the random fluctuations that are an inherent part of the medical system and when a deviation is an indication of a special cause that we might want to investigate and learn from. We all have a tendency to overestimate and overreact to small deviations that may be nothing more than normal variation.
When we see deviations, we tend to attribute them too often to the individual and tend to ignore the environment that the individual works in. If there are unacceptably large variations in performance, your first thought ought to be "how do I change the environment to reduce this variation" but it's human nature instead to say, "who should I retrain or reprimand".
Another good book to look at is Understanding Variation by Donald Wheeler. This is a delightful and very easy to read book that explains many of the problems that businesses have with handling variation in their production lines. Again, you need to extrapolate; a doctor's office is not a production line. If you think about some of the ways that physician performance data has been abused and misused, then you will see that these same types of abuses from a business context in Wheeler's book.
Hospital league tables.
Bamji A, Rao JN.
BMJ 2001; 322: 992.
Understanding Variation. The Key to Managing Chaos.
Knoxvile TN: SPC Press, Inc (1993).
Schools' experience of league tables should make doctors think
Peter Tymms and Andy Wiggins
BMJ 2000; 321: 1467. [Full text]
After you've read Walton and/or Wheeler, you may come to the conclusion that the statistical control chart, a tool widely used in industry, has similar applicability in health care. I would encourage you to apply control chart methods to physician performance data as well as a lot of other data that is not usually examined carefully in the health care context.
I'm working with some nurses at Children's Mercy Hospital to use control charts to track medication errors, patient complaints, employee accidents, unplanned sick leave, employee turnover, and additional measures of organizational safety and effectiveness. It has a lot of potential, in my opinion, to handle both this type of data as well as the type of data you are referring to.
I need to add a disclaimer that health care is different from most other businesses. That doesn't mean that health care can't use control charts, but it does mean that we can't blindly apply a process developed for industries that produce fast cars and fast food. So some type of adaptation will be necessary.
This page was written by Steve Simon while working at Children's Mercy Hospital. Although I do not hold the copyright for this material, I am reproducing it here as a service, as it is no longer available on the Children's Mercy Hospital website. Need more information? I have a page with general help resources. You can also browse for pages similar to this one at Category: Unusual data.