In optical scatterometry, a proper measurement configuration has a significant impact on the precision of the reconstructed profile parameters beyond the quality of the measured signatures. In this paper, we propose to determine an optimal measurement configuration for optical scatterometry with the application of global sensitivity analysis (GSA). For each measurement configuration, we define a metric called the uncertainty index to evaluate the impact of random noise in measured signatures on measurement precision by combining the corresponding noise level with the main effect defined in GSA. Experiments performed on a one-dimensional silicon grating with its true dimensions close to its nominal values have revealed a trend that the lower the uncertainty index, the better the precision of the reconstructed profile parameters. This trend shows an agreement between the theoretically predicted and experimentally obtained optimal measurement configurations. The uncertainty index also predicts an optimal measurement configuration for a set of grating samples with various dimensions, which shows a similar trend in agreement with that by numerical simulations. In contrast, the optimal configuration predicted using the local sensitivity analysis method is significantly dependent on the nominal dimensions of the samples, and consequently it is difficult to achieve a proper configuration for all the investigated samples. The results suggest that the defined uncertainty index by the GSA method is suitable to determine an optimal measurement configuration, especially for a set of samples with relatively large dimensional variation. (C) 2014 Elsevier B.V. All rights reserved.