This study uses a nationally representative survey to analyze a key survey design decision in child labor measurement: self-reporting versus proxy interviewing. The child/proxy disagreement affects 20 percent of the sample, which translates into a 17.1 percentage point difference in the national rate of child labor by type of respondent. As a result, marginal effects from standard child labor supply functions show important child/proxy differences, particularly when the household experienced some adverse weather and income shocks. Moreover, we find that attitudes and social perceptions toward child labor are not related to the likelihood of disagreement, while proxy respondent's past experience as child laborers emerges as an important predictor of the disagreement. A modified bivariate choice model reports statistically significant probabilities of misclassification that ranges between 9 and 30 percent according to alternative definitions of child labor.