Two-step estimation in linear regressions with adaptive learning