MEHDI RAZZAGHI
331
chemical hormesis we refer to Calabrese and
Baldwin (2000). However, as shown in Razzaghi
and Loomis (2001), in developmental toxicology,
more than a single replication of an experiment
must be considered before a chemical can be
declared as being hormetic. For the present data,
therefore, in order to fit a monotonic dose-
response function, one might consider replacing
the observed incidence of zero by an estimate of it.
In such a situation, it would seem unreasonable to
estimate the probability of response in the 5 mg/kg
dose group as 0, as given by the maximum
likelihood method. In this case, because n = 98,
from (2), (6), (14) and (20),
0.046 p
~
,021.p ,005.p .007, p
ˆ
*
ni 1,
*
ni
====
are four different point estimates for the
probability of response at the first nonzero dose
level.
In order to further investigate the
properties of these estimates, a probit model was
used to fit the response probability p as a function
of the natural logarithm of dose, i.e.
d) log b a(p +Φ= (23)
Using PROC PROBIT in SAS (1996), it was
found that the maximum likelihood estimates of
the model parameters are
0.987.b
ˆ
and 03.601 a
ˆ
== Using these
parameter estimates, it is found that the point
estimate of p when d = 5 mg/kg is .022.
Furthermore, the standard deviation of
5 log b
ˆ
a
ˆ
+ is 0.163. Based on these quantities, if
the 95% confidence interval is evaluated for the
predicted proportion, one finds that this range is
(.010, .046). Interestingly, although the minimax
estimator
p
is equal to the upper bound in this
range, both the Bailey estimator
p
ˆ
and the
Bayesian estimator
p
*
ni
are outside this range and
far too small to be plausible. Therefore, in this
instance,
and p
*
ni1,
the minimax procedure appear
to produce more realistic estimates of p compared
to other methods.
Discussion
Lack of occurrence of rare events in biological and
physical experiments is not uncommon. In such
situations, the maximum likelihood estimate
becomes unusable and one needs to resort to
alternative statistical methods. Here, I have
considered this problem and investigated the use
of several other statistical techniques and the
minimax estimator.
It is immediately noted from (2) that for
the Bailey estimator,
.
n
1
0 p
ˆ
= This property
also holds for the Bayesian estimator considered
by Basu et al. (1996). However, for the minimax
estimator, from (18)
.
n
1
0 p
~
=
This means that
for relatively large values of n, both
p
ˆ
and the
Bayes estimate lead to numerically smaller values
than the minimax estimator. Actually, it can be
shown (Roussas, 1997) that the Bayes estimate for
the family of beta prior and SEL has the same
asymptotic distribution as the maximum likelihood
estimate for arbitrary fixed values of α and β,
while the asymptotic distribution of
p)- p
(n is
normal with mean
p
2
1
− and variance p(1-p).
Thus, I can say that the minimax estimator is
comparatively more conservative.
However, as discussed by Carlin and
Louis (1996), although informative priors enable
more precise estimation, extreme care must be
taken in their use because they also carry the risk
of disastrous performance when their informative
content is in error. Although using a non-
informative prior leads to a more conservative
Bayes estimate, there may be situations when
Bayes and other methods underestimate the value
of this rare event. This result is demonstrated
through an example in developmental toxicology.
The conclusion of this paper is not
necessary to recommend the minimax or any other
estimator in all situations when there is a zero
response. Rather, the goal is to increase awareness
and recommend that more caution should be taken
when any single method is used to estimate the
success probability when sample shows zero
occurrence. The choice of the estimate should to a
large extent depend on which kind of optimality is
judged to be most appropriate for the case in
question.