How to assess prevalence of grossly inaccurate data in recorded body weights in medical records? I have around 5 million observations of weights of patients over the course of a year.  Some patients have one measurement during that year and others have 30-40 (or more).  I've noticed a great deal of variability.  For example, on one day a patient will have a recorded weight of 120 lbs and the next a recorded weight of 220 pounds - obviously, one of those is incorrect.
In this data set, we also have the locations (multiple) that are entering the weights, so it would be helpful to know which of the locations are entering the inaccurate data.  
Eventually, we'd like to use this weight data in various studies but would first like to get an idea of how often these bogus weights appear in this collection.
Any suggestions? 
 A: From an epidemiological perspective, weight data is often of intensely poor quality. There's variation by time of day, by measurer, by scale, by subject (subjects perceived by study staff as being "overweight" often have over-estimated weights)...it's really rather horrendous.
Two potential ideas for how to deal with the problem:


*

*Some sort of regression calibration from a validation study. If you have access to the study sites themselves, and they're not, you know, massively far flung, you could run a validation study with a small number of individuals at each site who have been weighed by a "gold standard" scale and staff member - preferably a very good, very careful one, or using a more reliable method. From there, you can quantify the measurement error at each site and adjust for it.

*If you have other measurements that use weight - such as BMI - you can try to reverse calculate weight to see where the variation isn't due to measurement error, but entry error (for example, hitting 2 instead of 1 when entering a weight of 120 pounds).

