We combine the New Immigrant Survey (NIS), which contains information on US legal immigrants, with the American Community Survey (ACS), which contains information on legal and illegal immigrants to the U.S. Using econometric methodology proposed by Lancaster and Imbens (1996) we compute the probability for each observation in the ACS data to refer to an illegal immigrant, conditional on observed characteristics. The results for illegal versus legal immigrants are novel, since no other work has quantified the characteristics of illegal immigrants from a random sample. We find that, compared to legal immigrants, illegal immigrants are more likely to be less educated, males, and married with their spouse not present. These results are heterogeneous across education categories, country of origin (Mexico) and whether professional occupations are included or not in the analysis. Forecasts for the distribution of legal and illegal characteristics match aggregate imputations by the Department of Homeland Security. We find that, while illegal immigrants suffer a wage penalty compared to legal immigrants, returns to higher education remain large and positive.