Table 2.
Health district–level median absolute error (MdAE) outcomes by forecasting method, look-back length, and forecast length.
| Methods | Look-back length (days), n | Look-ahead length (days), n | dfa | Forecasting method with lowest MdAE | Median (95% CI) (cumulative cases) | P valueb | Statistically significantly lower MdAE than other methodsc, n (%) |
| All (N=60) | All | All | 59 | MA (3, 1)d | 3.07 (2.41-3.74) | <.001 | 53 (90) |
| G1e (n=5) | 1 | 1 | 4 | Growth rate | 4.03 (3.01-5.04) | .31 | 1 (25) |
| G1 (n=5) | 1 | 3 | 4 | Growth rate | 8.96 (6.56-11.36) | .93 | 0 (0) |
| G1 (n=5) | 1 | 7 | 4 | Growth rate | 16.48 (11.67-21.28) | .96 | 0 (0) |
| G2f (n=5) | 3 | 1 | 4 | MA | 3.07 (2.41-3.74) | .01 | 1 (25) |
| G2 (n=5) | 3 | 3 | 4 | MA | 3.20 (2.50-3.90) | <.001 | 4 (100)g |
| G2 (n=5) | 3 | 7 | 4 | MA | 7.88 (5.71-10.05) | <.001 | 1 (25) |
| G2 (n=5) | 7 | 1 | 4 | AR | 3.57 (2.67-4.47) | .01 | 0 (0) |
| G2 (n=5) | 7 | 3 | 4 | MA | 5.52 (3.96-7.08) | <.001 | 2 (50) |
| G2 (n=5) | 7 | 7 | 4 | AR | 11.83 (8.16-15.49) | <.001 | 1 (25) |
| G2 (n=5) | All prior | 1 | 4 | MA | 3.14 (2.47-3.80) | .04 | 1 (25) |
| G2 (n=5) | All prior | 3 | 4 | MA | 3.16 (2.54-3.78) | <.001 | 3 (75) |
| G2 (n=5) | All prior | 7 | 4 | MA | 7.68 (6.22-9.14) | <.001 | 3 (75) |
aDegrees of freedom represent the number of forecasting combinations minus one.
bP values were calculated for statistically significant differences in medians across groups.
cThis was based on comparisons of notch ranges. MdAE was interpreted within geographic levels.
dMA: moving average; (3, 1) represents a 3-day look-back and a single-day forecast length.
eG1 includes naïve, MA, autoregressive (AR), growth rate, and Holt-Winters (HW) methods.
fG2 includes MA, AR, growth rate, HW, autoregressive moving average, and autoregressive integrated moving average methods.
gMA (3, 3) achieved statistically significantly smaller MdAE than all four alternatives.