Natural gas is a ubiquitous fuel, obtained from a variety of source deposits that present an inherent variation in composition. As newer sources of natural gas become available (such as Liquefied Natural Gas and shale gas) the compositional variation is expected to increase, which can affect emissions during combustion in appliances, including criteria pollutants. Unfortunately, experimental observations of the effect of natural gas composition on combustion products are sparse due to the wide range of burner designs and high cost of experimentation. The current work develops a rigorous methodology for statistical inference on available data that accounts for the limited nature of experimental observations. The goal is to overcome data size and quality limitations and provide best estimates of emission response to fuel composition change by identifying a continuous probability distribution with a high likelihood of representing the data and high correlation to the experimental observations. Quantitative measures of agreement between the data and a set of candidate distributions form the basis of the evaluation. In addition, qualitative assessment of the reliability of distribution identification is derived from a quantitative rating system for desired features of the data set and chosen distribution. Finally, this methodology is applied to sample data from the Lawrence Berkeley National Laboratory to develop a comprehensive and self-consistent set of emission factor estimates applicable to investigations of modeling the effect of natural gas interchangeability on urban air quality. By following the developed process, representative distributions, ranges of estimates, and evaluations of the estimate reliability are obtained for changes in CO, NOx, NO2, and HCHO emissions as a function of change in fuel Wobbe Number for six classifications of residential appliances. (C) 2013 Elsevier Ltd. All rights reserved.