People often ask me about using different climate data sources for social science research. Reanalyses.org is a new wiki site that will be a useful resource for people looking to use reanalysis data (see this RealClimate post for a short explanation of what reanalysis is and why this site is new/useful).
For social scientists looking at climate impacts, it's probably worth considering using reanalysis data.
First, if you work on problems in low income countries, there may be a relatively thin observational record in your region of interest (see here and here). Using reanalysis helps interpolate between existing observations using a physical model, as opposed to using only statistical techniques.
Second, the record of weather observations (again, particularly in low income countries) is influenced by social events (again, see here). If you are a researcher concerned about endogeniety, this may be a serious issue. For example, Vecchi and Knutson (2008) document the endogeneity of hurricane observations (and I can say with certainty that endogenous weather observations influence climate impact estimates because I have a paper showing it). My guess is that reanalyses can help mitigate this issue somewhat, since human observations are balanced against model estimates. But, as Gavin and others have pointed out, reanalyses are still strongly influenced by the observational record so it is unlikely that using them can completely fix this issue.