A strategy for smoothing the distribution of the binomial random variable X is to base our inference on X + Y, where Y
is a comparatively small random noise, using X + Y instead of X in the formula for our chosen confidence interval. Having
a smoother distribution leads to a better normal approximation, which in turn reduces the coverage fluctuations of the
interval. From a purely probabilistic perspective, the split sample method can be seen to be a special case of this strategy.
Let Z be a random variable which, conditioned on X, follows a Hypergeometric(n, X, n1) distribution. Then it follows from
(1) that
˜ X = n˜p d=
n
2n1
Z −
n
2n2
(X − Z),
so that
˜ X d=
X + Y with Y =
n
2n1
Z −
n
2n2
Z +
n1 − n2
2n2
X.
The conditional distribution of Y when n = 11 is shown in Fig. 1.
From the above distributional identity it is clear that the split sample method relies on the sufficient statistic X as well
as an additional random variable Y, and that it therefore can be considered to amount to adding discrete noise to binomial
data. We note however that the noise term Y is a deterministic function of the sequence X1, . . . , Xn. Conditioned on the
sequence, it is therefore more natural to think of ˜ X as resulting from splitting the sample rather than adding noise to X.