1.
Point Estimation and Principles
Q: What is the principle of data reduction in estimation?
A: Data reduction refers to summarizing data using statistics like sufficient, complete, and
ancillary statistics without losing information relevant to parameter estimation. It helps in
deriving efficient estimators.
Q: What is the difference between Bayes and Empirical Bayes estimation?
A: Bayes estimation incorporates a known prior distribution on parameters. Empirical
Bayes estimation uses data to estimate the prior distribution, often from repeated or
hierarchical structures.
Q: What is minimaxity in estimation?
A: An estimator is minimax if it minimizes the maximum possible risk (loss) over all
parameter values. It's useful when the true parameter is unknown and worst-case
performance is a concern.
Q: What is admissibility?
A: An estimator is admissible if no other estimator performs better in terms of lower risk for
all parameter values. Inadmissible estimators can be improved upon.
2. Bayesian Estimation
Q: What are the advantages of Bayesian estimation in the linear model?
A: It allows incorporation of prior knowledge, provides a full posterior distribution, and
leads to more precise inference especially with small samples or prior information.
Q: What is predictive inference in Bayesian analysis?
A: It refers to predicting future observations by integrating the predictive distribution over
the posterior of the parameters.
Q: What is the James-Stein estimator?
A: A shrinkage estimator that dominates the MLE in estimating multivariate normal means
(dimension ≥ 3), reducing total risk by shrinking estimates towards a central point.
Q: What is the role of the EM algorithm in estimation?
A: The EM algorithm is used for maximum likelihood estimation when data is incomplete or
has latent variables, by iteratively applying Expectation and Maximization steps.
3. Robust Statistics
Q: What are M-estimators?
A: Generalizations of MLE that are robust to outliers. They minimize a chosen loss function
and provide resistance to model deviations.
Q: What is an influence function?
A: A tool to measure the effect of a small contamination at a point on an estimator. It helps
to assess the robustness of an estimator.
Q: What are L- and R-estimators?
A: L-estimators are based on linear combinations of order statistics (e.g., trimmed means).
R-estimators are based on rank statistics and are robust to outliers.
4. Evaluation of Estimators
Q: What are U-statistics?
A: U-statistics are unbiased estimators constructed from i.i.d. samples that have minimum
variance among unbiased estimators for symmetric functions.
Q: What defines a best unbiased estimator?
A: One that is unbiased and has the minimum variance among all unbiased estimators
(UMVUE).
5. Confidence Sets
Q: How do you find the shortest length confidence interval?
A: By selecting an interval with the minimum expected length for a given confidence level,
often using pivotal quantities or likelihood ratios.
Q: What are UMA and UMAU confidence sets?
A: UMA: Uniformly Most Accurate; UMAU: Uniformly Most Accurate Unbiased confidence
sets that achieve the best coverage probability for given significance.
Q: What are randomized confidence sets?
A: These involve randomization to achieve desired coverage probabilities, particularly when
exact confidence limits are not achievable deterministically.
6. Bootstrap Methods
Q: What is a bootstrap confidence interval?
A: An interval derived by resampling the data repeatedly and calculating the statistic’s
distribution empirically.
Q: What is meant by accurate bootstrap confidence sets?
A: These intervals have better finite-sample performance, often using bias-correction or
percentile methods, and converge asymptotically to the true distribution.
7. Simultaneous Confidence Intervals
Q: What is Bonferroni's method?
A: It adjusts confidence levels to control the family-wise error rate when making multiple
comparisons, by dividing alpha by the number of intervals.
Q: What is Schaffer’s method in linear models?
A: An improved method over Bonferroni, using logical constraints among tests to reduce
conservativeness in multiple testing.
Q: What is Tukey’s method used for?
A: It is used in ANOVA for pairwise comparisons while controlling the family-wise error rate.
Q: What are confidence bands for CDFs?
A: Simultaneous confidence intervals for all points on the cumulative distribution function,
used in nonparametric inference (e.g., Kolmogorov-Smirnov bands).
