Many epidemiological studies have used house characteristics associated with indoor sources as simplified proxies for personal nitrogen dioxide (NO2) exposure. Stove type and presence of a pilot light often been used as the two key characteristics, but significant overlaps have remained in the NO2 concentrations in the exposed and unexposed groups. This has contributed to inconsistencies in epidemiological findings, due to potential misclassification of exposure. In this study, other possible proxies were analyzed by cross-table analyses and were investigated in terms of improvements in both classification and predictive power. Adding building type to the above two proxies resulted in 0-5% of households with concentrations overlapping the observed range for the opposing stratum, compared with 22-42% for the two-proxy model. In spite of this performance, the predictive power of regression models for indoor NO2 was not improved by the addition of the third proxy, and the potential sample population was significantly limited. Using these analytical methods to choose descriptive proxies and evaluate the tradeoffs in their implementation can help epidemiological studies improve their designs and therefore optimize the robustness of their conclusions.