There are two sources of bias in the existing gravity equations used to assess the impact of non-reciprocal preferential trade policies (NRPTPs). The main inconsistency comes from the use of aggregate export flows at country level to analyse the effects of trade preferences which, by contrast, apply at product level. The second source of bias is that the literature does not deal with the main econometric issues which are likely to be present when a gravity equation is estimated. This paper discusses the first problem using evidence based on three levels of data aggregation (total exports, total agricultural exports and 2-digit). Furthermore, the estimation methods take into account the unobservable country heterogeneity as well as the endogeneity of trade preferences and the potential selection bias which zero-trade values pose. We consider all NRPTPs granted by 8 major OECD countries to exports from developing countries over the period 1995-2003. We find two key results. First of all we show that the impact of NRPTPs on total exports is positive, whatever the estimator. This means that, other things being equal, the national exports to the preference-giving country of a preferred country are higher than those of a non-preferred country. Secondly, when the analysis is conducted at 2-digit level, it emerges that the preference premium is very high in many 2-digit sectors, whatever the preferential treatment (GSP and/or other preferences). This finding stands in contrast with the result obtained when total exports are considered, which places the preference gain at lower values.