This paper demonstrates that consequences of errors in trade classification (buy/sell indicator) are possibly worse than suggested by previous studies. An errors-in-variables model of classification errors shows that the bias in regression-type microstructure models depends in a complicated way on the probability of trade-reversal in addition to the probability of error. This has implications for empirical research using regression type estimation involving trade-indicators. I propose instrumental variable (IV) estimation as an alternative to OLS, and show that the bias of the IV estimator is typically much smaller and only depends on the probability of classification error. The method is illustrated using 51 stocks from the TORQ database. The empirical analysis shows that even a crude estimator of classification error renders the bias statistically insignificant.