This paper develops asymptotic distribution theory for single-equation
instrumental variables regression when the partial correlations betwe
en the instruments and the endogenous variables are weak, here modeled
as local to zero. Asymptotic representations are provided for various
statistics, including mio-stage least squares (TSLS) and limited info
rmation maximum likelihood (LIML) estimators, Wald statistics, and sta
tistics testing overidentification and endogeneity. The asymptotic dis
tributions are found to provide good approximations to sampling distri
butions with 10-20 observations per instrument. The theory suggests co
ncrete guidelines for applied work, including using nonstandard method
s for construction of confidence regions. These results are used to in
terpret Angrist and Krueger's (1991) estimates of the returns to educa
tion: whereas TSLS estimates with many instruments approach the OLS es
timate of 6%, the more reliable LIML estimates with fewer instruments
fall between 8% and 10%, with a typical 95% confidence interval of (5%
, 15%).