A generalization of maximum entropy weighting (Maxent) is studied to allow for solutions when the data are not full rank and, more importantly, to find the best solution possible when the population moments cannot be perfectly matched or when a perfect match would involve a large degree of extrapolation. Maxent generalization is used to reweight a sample to match observed population moments. It is noted that by freeing the unit weights to vary smoothly across units greatly increases the space of potentially good matching solutions and maximazes the use of available information. Propensity score weighting can be combined with MaxEnt When the weights are highly variables, alternative methods might be used, such as PSM or a doubly robust method—-i.e., first weight and then estimate a model on the data.

Starting from the above mentioned Maxent approach, different weighting schemes are computed on a household panel, the LISS panel. From this quite exhaustive panel a representative sample was drawn for survey purposes. Survey data are merged to the panel population information. Then, different weighting schemes and methods are applied for estimation purposes and comparison discussed.