Portfolio credit risk models are very often constructed with correlation matrices serving as proxies for interrelations in the creditworthiness of each company. In addition to the size of the matrix, estimation of correlation is also complicated by the fact that defaults are rare and credit-sensitive securities such as stocks, bonds and credit default swaps (CDS) are noisy. Therefore, we present in this paper an estimation approach based on credit-sensitive instruments that is both statistically consistent and highly parallelizable. A simulation study shows that the method is reliable and has better statistical properties when benchmarked against other correlation estimators. In an empirical study based on the CDS premiums and stock prices of 225 firms listed on the CDX North American indices, we analyze the correlations computed using numerous approaches. Overall, we find that ignoring noise severely underestimates correlations, whereas equity correlation is poorly related to the best correlation estimates inferred from the CDS market.