I have a data set of medical drug stock levels at health centres and I want to forecast monthly consumption over the following 3-6 months. However about 30%-40% of the data is missing and some of the reporting intervals are irregular. Several other variables can potentially be used to further predict consumption. I'm looking for suggestions for how to best approach this problem.
Some additional information about the data set:
- Total of ~4000 facilities
- Data spans 1 year for most facilities. A small subset has ~3 years of data.
- Facilities submit monthly reports for current stock on hand for up to 20 products.
- Stock receipts occur roughly every 3 months but only ~30% are explicitly reported - the others can be detected by an increase in the stock level
- Less than 1% of all product/facility combinations have 1 year of continuously reported data that can be used for testing
- Occasionally the data contains data points mid-month. These data points are particularly important when a facility reports 0 stock and consequently its consumption is 0.
- A stock-out frequently goes unreported until end of the month. This is hard to detect if combined with an unreported receipt of stock that same month.
- The consumption rate of different drugs is likely proportionally co-related for one facility: if a facility consumes 2 times the amount an average facility consumes for drug X, that facility will likely consume 2 times the average consumed of drug Y.
- The consumption rate is likely correlated to the size of the community it's serving (catchment population). This data is available for a subset of facilities.
- The consumption is likely correlated to the consumption of geographically proximate facilities. GPS coordinates for ~60% of facilities is available. All facilities are grouped into districts of 10-50 facilities of varying geographic size.
I'm assuming it's best to start with a simple model and slowly build up its complexity to include additional factors. What models and methods are most appropriate?