A Note on the Dangers of Using Regressions to Analyze Earnings Forecasts

When data exhibit cross-sectional variation in scale and regression parameters, pooled regression parameters can exhibit severe biases. It is commonly assumed that normalizing per-share earnings data by a firm's stock price eliminates cross-sectional variation in scale. This study shows that this is not always true. The hierarchical model explicitly allows for both types of variations in earnings data, and therefore provides more reliable parameter estimates. This study examines analyst exaggeration and herding behavior using empirical Bayes estimates from a hierarchical model. There are significant differences between hierarchical and conventional estimates, which sometimes substantially alter one's conclusions about analyst behavior