16.11.2014 Views

McGraw-Hill

McGraw-Hill

McGraw-Hill

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Earning Estimates, Predictor Specification, and Measurement Error 223<br />

assuming that the measurement error, U, and residual model error,<br />

e, are independent and that measurement error is not a function of<br />

the flash predictor.<br />

Alternatively, one can write the preceding equation as:<br />

I)<br />

b<br />

b=<br />

l+ [variance(u)/ variance (flash predictor)]'<br />

This suggests that the estimated coefficient, 6, has a bias toward zero<br />

that depends upon the variance in the measurement error relative to<br />

the variance in the flash predictor. Other things equal, the greater<br />

the variance in the measurement error, the more biased the est<br />

of 6. Why? Because the consensus provides a more noisy estimate of<br />

flash earnings as measurement error increases.<br />

For models with more than one explanatory variable, effect the<br />

of measurement error becomes a bit more complex. It will depend<br />

on, among other things, the number of predictors with measurement<br />

error, the correlations between predictors, the correlations<br />

between measurement errors, and the signs of the regression coefficients.<br />

For the special case where predictors are not correlated and<br />

not related to measurement errors, and measurement errors across<br />

predictors are independent, the regression coefficients will be biased<br />

toward zero. As predictors become more highly correlated,<br />

however, any bias will depend on the signs of the regression coefficients<br />

and the variance of the measurement errors relative to the<br />

variance of the predictors, other things equal.=<br />

ENDNOTES<br />

The authors thank Judith Kimball for her editorial assistance.<br />

1. One can choose from a variety of data vendors as well. For this study<br />

we use Institutional Brokers Estimates System (IBES) data.<br />

2. There are a variety of options regarding the estimates to include as<br />

well as their weights. A dynamic weighting strategy, for example,<br />

would weight more recent estimates more heavily than older<br />

estimates. .<br />

3. Trimmed means remove a proportion of the most extreme<br />

observations from a data set and compute the mean of the remaining<br />

observations. This procedure reduces the influence of outliers.

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

Saved successfully!

Ooh no, something went wrong!