Article dans une revue: Empirical analysis of social networks is often based on self-reported links from survey data. How we interpret such data is crucial for drawing correct inference on network effects. We propose a method for testing whether survey responses can safely be interpreted as a link and, if so, whether links are generated by a unilateral or bilateral link formation process. We present two empirical illustrations of the test on risk-sharing links in Tanzania and on communication among Indian farmers, respectively, demonstrating the ability of the methodology to discriminate between competing data-generating processes.