Alternative measures for evaluating population forecasts: a comparison of state county and sub-county geographic areas.

Forecasting error strategies have included the average error rate and the variability mean absolute percentage (MAPE) median absolute percentage error (MEDAPE) algebraic percentage and root mean square errors (RMSE). Alternative measures of accuracy and "utility" that are compared include box and steam-and-leaf plots of the absolute percentage errors the index of dissimilarity (IOD) and the proportionate reduction in error (PRE). Data was obtained from 1980 census data on the nation states counties and townships or census tracts for a variety of populations by size and rates of decline. Forecasts were made for 1990 from the 1980 census and compared with the latest census results. Plots were found to be useful for illustrating variation in error: symmetry central location and outliers. The IOD measured the misallocation of the forecast across geographic areas. The PRE evaluated the estimates reduction of error found by using the cohort component method instead of forecasting from census data. Forecasts from 1980 census figures and explicit forecasting of 1990 population were compared using a variety of the aforementioned tests of accuracy. Findings show that the 1980 census base for forecasting area population was just as accurate as explicit forecasting. Explicit forecasting did not improve accuracy of estimates for cities. 1980 census-based forecasts and explicit forecasts were too high in smaller census tracts and too low in higher census tracts. Census-based forecasts were considered more accurate for small tracts. Explicit forecasts were better for larger sized tracts. Growth rates were forecasted higher in areas with changing population with 1980 census-based forecasts. Explicit forecasts were too high in declining census tracts and too low in high growth tracts. Explicit forecasts were more accurate for tracts that grew by 20-49%. 1980 census forecasts were more accurate for declining tracts. MEDAPE was a very useful measure for detecting errors due to skewed distributions. IOD was a useful measure when forecasting small geographic units. PRE was useful for all geographic areas and for standardizing comparisons. Explicit forecasts were more accurate for estimates of counties and states. In 32% of the cases examined explicit forecasts increased error.