DETECTING MULTIPLE OUTLIERS WITH AN APPLICATION TO R-AND-D PRODUCTIVITY

Multiple outlying observations are frequently encountered in empirical analyses of economic data. Several multiple outlier tests exist, but little evidence is available on their relative power against alternative causes of outliers and influence. This paper analytically and numerically compares the sensitivity of several outlier diagnostics to different forms of data contamination; it also proposes new statistics. Practical issues associated with these tests are addressed using data on R&D spending and total factor productivity from Griliches and Lichtenberg (1984). The tests confirm their hypothesis that previously low estimates of the return to R&D maybedue to outliers. © 1990.