Resolved flow statistics extracted from large eddy simulations (LES) do not directly compare to unfiltered flow statistics determined, for example, from experiments. Scalar statistics in reacting flows are investigated here from a priori tests based on experimental data. Mean scalar fields are found to be quite well estimated from LES, taking into account mass-weighted statistics, but resolved variances are strongly affected by the filtering operation and decrease roughly linearly when the filter size ? increases. Accordingly, subgrid scale scalar variances should be modelled to compare LES and experimental results. A simple gradient model is found to provide very good estimates of the mean subgrid scale variance of the progress variable.