Workers Comp Data Carveouts – Are You Guilty of Sampling Errors?
The concept of Workers Comp data carveouts surfaces more now that employers look to enhance their workers’ compensation saving programs.
Workers comp data carveouts start when a company, organization, consultant, etc. decide to use oversampling of a small set of data to justify a larger set. The reverse will work under certain circumstances.
For example, and I have seen this happen often – using California data results on any kind of workers comp statistics to think that the same result will occur in another state.
I often do phone consulting for investment companies that are looking to place the work in the workers’ compensation space. The investors will want to look at two or three different states and then use the results derived from those states to try to expand their investing across the US.
I do not use Wikipedia that often as a source. Two sentences were taken from their page on oversampling/undersampling.
Both oversampling and undersampling involve introducing a bias to select more samples from one class than from another, to compensate for an imbalance that is either already present in the data, or likely to develop if a purely random sample were taken.
Many articles appear on this website concerning statistical madness. Sampling errors can be a form of statistical madness.
The Whole Picture
Many software packages do allow companies to isolate one state when examining a Mod calculation. Isolating one state can be Ok to look at just that one state. The mistake is made when a single state company wishes to expand to other states using the data from their HQ state.
Workers’ Compensation rating contains too many variables to think that how your company is rated in one state will allow your company to know what would happen data-wise in another state. Most E-Mod formulas have 33 calculations to arrive at the final number.
An employer would need to operate in the additional states for a few years before the actual results are known – remember that Workers Comp is a delayed lookback system.
If an employer is operating in multiple states but looks at one or two of their main operational states as the bellwether for other states, they would be oversampling and undersampling at the same time. The main operational states would be oversampled and the remaining states would be undersampled to the point of not being sampled at all.
I left the link above in the Wikipedia statement on bias. Bias in statistics leads to unnatural and inaccurate results. I have actually sliced out sections of data as the employer wanted that section to be analyzed. I did not guarantee the results as being a good forecast.
Workers Comp data carveouts usually occur when analyzing one state, a year or other data out of any analysis – mainly the Experience Modification Factor or Loss Development Factor (LDF). One needs to look at the whole picture.
Bottom line – be very careful when using workers comp data carveouts for decisions.
©J&L Risk Management Inc Copyright Notice