Before reading this, it is worth skimming my previous post on the topic of forecasting English hospital admissions.

The two graphs above show the results of my initial pokings into the NHS hospital activity data.  The first one shows quite clearly how activity has in average increased every year. (I’m not yet quite sure why it isn’t showing any data for years 0 and 1.  By rights it shouldn’t show 0, but 1…)

The second shows that while almost all trusts have positive mean growth over the period, there is a large minority which have only experienced very modest positive growth, much less than the national means in the first graph would imply.

While mildly interesting on its own, this has implications for the inferential analysis which is going to follow.  One of the key issues in performing econometric analysis with panel data is how you treat your units, in this case hospital trusts.  Under one approach, you assume that that each unit has its own unique effect on the variable you are analysing, but that these effects are random.  The second approach says that they are not random but driven by some systematic differences in the units.

Based on intuition one would have thought that the random approach would not be appropriate for hospital trusts because the growth in activity is going to largely be driven by their local population and the the funding levels of their local Strategic Health Authority, i.e. there are systematic differences.  The second graph doesn’t really help us decide which approach is more appropriate because it shows the trusts as being quite neatly distributed, even if the mean is skewed by some outliers.

This means that we will have to use statistical tests to decide which approach is better, and possibly just see which makes the better forecast.