6.6. The EM Algorithm 407
By expression (6.6.1), the conditional distributionZgivenXis the ratio of (6.6.12)
to (6.6.11); that is,
k(z|θ,x)=
∏n 1
i=1f(xi−θ)
∏n 2
i=1f(zi−θ)
[1−F(a−θ)]n^2
∏n 1
i=1f(xi−θ)
=[1−F(a−θ)]−n^2
∏n^2
i=1
f(zi−θ),a<zi,i=1,...,n 2 .(6.6.13)
Thus,ZandXare independent, andZ 1 ,...,Zn 2 are iid with the common pdf
f(z−θ)/[1−F(a−θ)], forz>a. Based on these observations and expression
(6.6.13), we have the following derivation:
Q(θ|θ 0 ,x)=Eθ 0 [logLc(θ|x,Z)]
= Eθ 0
[n 1
∑
i=1
logf(xi−θ)+
∑n^2
i=1
logf(Zi−θ)
]
=
∑n^1
i=1
logf(xi−θ)+n 2 Eθ 0 [logf(Z−θ)]
=
∑n^1
i=1
logf(xi−θ)
+n 2
∫∞
a
logf(z−θ)
f(z−θ 0 )
1 −F(a−θ 0 )
dz. (6.6.14)
This last result is the E step of the EM algorithm. For the M step, we need the
partial derivative ofQ(θ|θ 0 ,x) with respect toθ. This is easily found to be
∂Q
∂θ
=−
{n
∑^1
i=1
f′(xi−θ)
f(xi−θ)
+n 2
∫∞
a
f′(z−θ)
f(z−θ)
f(z−θ 0 )
1 −F(a−θ 0 )
dz
}
. (6.6.15)
Assuming thatθ 0 =θ̂ 0 , the first-step EM estimate would be the value ofθ,sayθ̂(1),
which solves∂Q∂θ = 0. In the next example, we obtain the solution for a normal
model.
Example 6.6.1.Assume the censoring model given above, but now assume that
Xhas aN(θ,1) distribution. Thenf(x)=φ(x)=(2π)−^1 /^2 exp{−x^2 / 2 }.Itiseasy
to show thatf′(x)/f(x)=−x. Letting Φ(z) denote, as usual, the cdf of a standard
normal random variable, by (6.6.15) the partial derivative ofQ(θ|θ 0 ,x) with respect
toθfor this model simplifies to
∂Q
∂θ
=
∑n^1
i=1
(xi−θ)+n 2
∫∞
a
(z−θ)
1
√
2 π
exp{−(z−θ 0 )^2 / 2 }
1 −Φ(a−θ 0 )
dz
= n 1 (x−θ)+n 2
∫∞
a
(z−θ 0 )
1
√
2 π
exp{−(z−θ 0 )^2 / 2 }
1 −Φ(a−θ 0 )
dz−n 2 (θ−θ 0 )
= n 1 (x−θ)+
n 2
1 −Φ(a−θ 0 )
φ(a−θ 0 )−n 2 (θ−θ 0 ).