We use panel data to analyze the determinants of speaking fluency and wages of immigrants. Our model takes account of two problems that may bias OLS estimates of the impact of speaking fluency on earnings. First, subjective variables on an ordinal discrete scale, such as self-reported language ability, can suffer from misclassification errors. The model decomposes misclassification errors into a time-persistent and a time-varying component. Second, the model accounts for correlated unobserved heterogeneity in language and earnings equation. The main finding is that these two generalizations of the standard model both lead to substantial changes in the estimated effect of speaking fluency on earnings.