Parameter Estimation
Tuesday 3 June, 10:20--12:10
- Moderator: Bill FOX
- Bristol-Myers Squibb Co.
- Xiaorong PAN, Mai ZHOU
- University of Kentucky, Lexington, KY, USA
Recently, Owen(1988 Biometrika 237-249), building on an earlier
suggestion of Thomas and Grunkemier, showed that empirical likelihood
ratio statistics for various parameters $\theta(F)$ of an unknown
distribution F have limiting chi-square distributions and may be used
to obtain tests or confidence intervals in a way that is completely
analogous to that used with parametric likelihoods. Here we
investigated the methodology of empirical likelihood ratio with
censored data.
First of all we considered the problem of making inference for the
parameter $\theta(\Lambda)$ for iid data and the parameter $\beta$ in
Box proportional hazard model with right censored data and obtained
the explicit solutions similar to Owen's, which is quite easy to
implement with software. We also investigated the approach with doubly
censored data. In some complicated models, to fine the maximum under
constrain is difficult. Motivated by Owen(1988) and our own results,
we suggest to use one particular method to search the maximum under
constraints which is quite successful and easier to implement.
[Xiaorong Pan, Department of Statistics, University of Kentucky, Lexington, KY 40506, pan@ms.uky.edu]
- M. AHSANULLAH
- Rider University, NJ, USA
In this paper some distributional properties of the
generalized order statistics from power function distribution are given.
The minimum variance linear unbiased as well as best ( in the sense on
minimum mean squared error ) invariant estimators of the parameters based
on the first m generalized order statistics are given.It is shown that
many of the known results based on ordinary order statistics and record
values can be obtained as special cases of the results of this paper
[M. Ahsanullah, Department of Management Sciences, Rider University,
Lawrenceville, NJ, 08648-3099, ahsan@enigma.rider.edu]
- Mikhail B. MALIOUTOV, Rostislav PROTASOV
- Northeastern University, Boston MA, USA
We consider usual mixed model $ y=X\beta+\sum_{a=1}^tU(a)\phi(a)$,
random IID normal $n(a)$ components of $\phi(a)$ have means $0$ and
unknown variances $d(a), a=1,\dots,t,$ whereas
$\beta\in R^d$ is unknown nonrandom parameter and $X$ is a nonrandom matrix
of size $N$ times $d$.
Under essentially the same conditions as in Miller
[Ann.\thinspace
Statist.\thinspace 5\thinspace
(1977):746--762]
we prove Local
Asymptotic Normality of the family of measures $P_{\theta}$ on Borel
$\sigma$- field of $R^N$ depending on $\theta=d(1),\dots,d(t), \beta$,
with an appropriate
normalizing diagonal matrix indicating
rates of convergence of different components of parameter vector to asymptotic
normality.
Under some additional uniformity conditions the above result implies local
asymptotic minimaxity of MLE and of the first step of the iterative procedure
for finding MLE which is essentially the Gauss-Newton iterative reweighted
procedure for the extended regression model where the components of $ y,
y\cdot y^T$ are considered as measurements and the initial guess for the
parameter vector is assumed to be sufficiently close to its true value
to provide convergence of iterations in probability.
The above results on minimaxity justify study of optimality of certain designs
for the simplest mixed models, namely, one-factor models when applying MLE.
The results obtained show that the asymptotically optimal designs differ
drastically from
those for ANOVA with fixed parameters.
[Prof. M. Malioutov, Mathematics Dept., 5676 Lake Hall. Northeastern University, Boston MA 02115 , MLTV@neu.edu]
- Mohamed DARIF, Ph.D.
- Kuwait University, Kuwait
- James D. NEATON, Ph.D., Thomas A. LOUIS, Ph.D.
- University of Minnesota, Minneapolis, MN, USA
Logistic regression has been used for binary data modeling in many areas
including econometrics and epidemiology. The maximum likelihood estimation
would produce coefficient estimates with minimum variance. However, this
might not be the case if the variables included in the model are measured
with error. Inferences based on the estimation with error-in-variables may
be misleading. This is a serious problem particularly in epidemiological
studies. Simulations were used to study the effect of measurement error
on the coefficient estimates when we have two variables in the model, one
of which is measured without error. This study was run using different
coefficient values, various values of the correlation between the two
variables, two rates of event probability, different sample sizes, and
two levels of measurement reliability.
The results of this study showed that the coefficient estimates can be
attenuated or overestimated. In a situation where the variable of
interest is measured with no error and there is a confounding factor measured
with error, one is likely to see an attenuation of the coefficient estimate
for the confounder. More importantly the coefficient for the variable of
interest can be biased in either direction. This is also likely to occur
when an important factor is not accounted for. In general, measurement
error lead to coefficient estimates that are less variable, and have lower
coverage probability than the ordinary estimates.
[Mohamed Darif, Ph.D., Department Of Community Medicine,
Faculty of Medicine, Kuwait University, PO Box
24923, Safat 13110, Kuwait,, darifm@hsccmail.kuniv.edu.kw]
Back to JRC program
Back to JRC Home Page