the interventions were not randomized, the potential for selection bias contaminating the results still exists. Using panel data does not completely resolve this issue, but allows us to control for unobserved time–invariant effects. Third, we take advantage of the treatment–comparison design of the original evaluation to undertake three types of comparisons using nearest neighbor matching, comparing (1) early and late adopters of the technology; (2) NGO members with access to the technology and those without access to the technology; and (3) NGO members vs. non–NGO members. Heckman, Ichimura, and Todd (1997, 1998) and Abadie and Imbens (2002) show that under certain conditions on the data, all of which are satisfied in this study, matching estimators provide reliable estimates of program impact. Finally, we use insights drawn from the qualitative work conducted as part of the panel study and earlier work conducted in the study sites to obtain additional perspectives on the role of implementation modalities and collective action.

3. DATA AND SAMPLING

3.1 The initial evaluation sample: an overview of the agricultural technology interventions The initial survey for this study was conducted in 1996–97 by the International Food Policy Research Institute (IFPRI) and Data Analysis and Technical Assistance, Ltd. (DATA) to examine the impacts of improved vegetable and polyculture fish management technologies on household resource allocation, incomes, and nutrition. The initial study also aimed to uncover intrahousehold and gender–differentiated impacts of the new technologies, so extremely detailed information was collected on individuals within households. As mentioned above, households were surveyed in three sites where NGOs and extension programs disseminated the new technologies. These sites were (1) Saturia thana, (referred to below as Saturia); (2) Sadar thana, (referred to below as Jessore); and (3) Gaffargaon thana, district, and Pakundia and Sadar thanas, (referred to below collectively as Mymensingh).3 The agricultural technologies and extension programs at each site are unique, resulting in three case studies that may be compared (see Table 1).4 In all of the sites, the technology had been disseminated prior to the initial survey in 1996, so this survey should not be interpreted as a true baseline.

3 Thana (also ) is a subdistrict in subdistrict in the administrative divisions of . 4 This description draws from Quisumbing and de la Briere (2000), Hallman et al. (2007), and recent field visits by the authors.

5