08.03.2014 Views

Variance Estimation for the General Regression Estimator

Variance Estimation for the General Regression Estimator

Variance Estimation for the General Regression Estimator

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

16<br />

<strong>the</strong> average over <strong>the</strong> samples of (<br />

ˆ<br />

) 2<br />

T − T<br />

, are also shown. In <strong>the</strong> Hospitals population, both<br />

estimators have negligible bias at ei<strong>the</strong>r sample size. The GREG is considerably more efficient<br />

in Hospitals than <strong>the</strong> π -estimator because of a strong relationship of Y to x. In <strong>the</strong> two Labor<br />

Force populations, both <strong>the</strong> π -estimator and <strong>the</strong> GREG are nearly unbiased while <strong>the</strong> GREG is<br />

somewhat more efficient as measured by <strong>the</strong> rmse <strong>for</strong> all sample sizes and selection methods.<br />

Table 2 lists <strong>the</strong> empirical relative biases (relbiases) of <strong>the</strong> nine variance estimators,<br />

defined as 100( v − mse)<br />

mse, where v is <strong>the</strong> average of a variance estimator over <strong>the</strong> 3000<br />

samples and mse is <strong>the</strong> empirical mean square error of <strong>the</strong> GREG. The rows of <strong>the</strong> table are<br />

sorted by <strong>the</strong> size of <strong>the</strong> relbias in LF(mod) <strong>for</strong> srswor’s of size 50, although <strong>the</strong> ordering would<br />

be similar <strong>for</strong> <strong>the</strong> o<strong>the</strong>r populations, sample sizes, and selection methods. In <strong>the</strong> Hospitals<br />

population, <strong>the</strong> sampling fraction is substantial, especially when n = 100. As might be expected,<br />

this results in <strong>the</strong> estimators that omit any type of finite population correction (fpc)— v R2<br />

, v D ,<br />

vJ<br />

∗ , and v J —being severe over-estimates in ei<strong>the</strong>r srswor or pps samples. Because v R1<br />

lacks a<br />

term to reflect <strong>the</strong> model-variance of <strong>the</strong> nonsample sum, it under-estimates <strong>the</strong> mse badly when<br />

<strong>the</strong> sampling fraction is large.<br />

In <strong>the</strong> Labor Force and LF(mod) populations, increasing sample size leads to decreasing<br />

bias. The estimators v π , R1<br />

v , v R2<br />

, and v SSW have negative biases that tend to be less severe as<br />

<strong>the</strong> sample size increases. The jackknife v J and its variants, v ∗ J , v ∗<br />

JP , are over-estimates,<br />

especially at n = 50. The estimators, v D and v DP , are more nearly unbiased at each of <strong>the</strong><br />

sample sizes than most of <strong>the</strong> o<strong>the</strong>r estimators.<br />

The empirical coverages of 95% confidence intervals across <strong>the</strong> 3000 samples in each set<br />

are shown in Table 3 <strong>for</strong> <strong>the</strong> Hospitals population. The three choices of variance estimator that

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!