Our model relies on the assumption that men born in birth year 1959 are similar to those in the adjacent years.
Such a distribution was observed in the adjacent years of 1996 and appeared also in the P-state of the 1970s.
We observe that males born in 1959 more often complete tertiary education, especially university education, than males born in the adjacent years (columns (2), (3) and (4)).
We identify the causal effect of military service by comparing the long term outcomes of this exempted birth cohort with the outcomes of those born in the adjacent years.
This effect can be attributed to the system of conscription if we assume that the schooling and earnings capabilities of males born in 1959 are not different from the capabilities of males born in the adjacent years before and after 1959.
Statistically significant excess deaths were computed by detecting the data points at which the all-cause deaths exceeded the mean of the adjacent years +2 SDs (6, 10 ).
The percentages of all-cause deaths by age group were computed for 1918 and 1920 and compared with the respective averages of percentages for the adjacent years (1917, 1919, and 1921).
Using data from the 1895 1945 Statistical Abstract of Taiwan (9 ), we compared monthly deaths during the 2 waves of epidemics in 1918 and 1920 with deaths during corresponding nonpandemic periods of the adjacent years.
The excess percentages of deaths for 1918 and 1920 in age groups 10 19, 20 29, and 30 39 years were computed by subtracting the average percentages of deaths in these age groups during the adjacent years from the respective true percentages of deaths in these age groups during 1918 and 1920 (Table 2).
Moreover, the percentages of excess yearly number of deaths reported by 12 large public hospitals and public medics for each year during 1918-1920 over the averaged yearly number of deaths of the adjacent years of 1916, 1917, 1921, and 1922, are given in Table 2.
Consistent with the theory that this transient drop in compliance was driven by a backlog of ongoing studies being registered en masse, the average delay to registration for studies registered in 2005 was several hundred days longer than in either of the two adjacent years (table 2).
