A note on multi-stage sampling with list frames: the PARC/CIMMYT survey of wheat harvest technology

Abstract

In collecting farm-level data for the purpose of initiating social science research on selected cropping systems, many researchers in Pakistan use random sampling of villages with nonprobability sampling of farmers within villages. Given typical timing and logistical constraints, this relatively low-cost method often provides sufficient data to meet information needs. Studies generated through non-probability sampling provide useful background information, and often serve to identify critical issues for further research. When the purpose of a survey is to generate estimates of population characteristics for policy-makers, the reliability of the estimates becomes more important. The major shortcoming of non-probability methods is the researcher's inability to calculate sampling errors and confidence intervals because the probability of selecting the elements of the sample is unknown. For this reason, such sampling methods are often termed "nonmeasurable" or "judgment" designs (Hansen, Hurwitz, and Madow, pp. 8-9). In many research settings, however, non-sampling errors constitute a large part of total error of estimates. Nonsampling errors are common to both probability and nonprobability surveys, and their magnitude and direction can rarely be measured (Casley and Lury, pp. 86-88). During March and April of 1987, the PARC/CIMMYT Collaborative Program in Economics sponsored a pilot survey on wheat harvest technology in the rice-wheat production area of the Punjab. In the harvest technology survey, the team experimented with a form of multi-stage probability sampling using list frames. While the sample design described in this note reflects specifically the type of information needs identified for this survey, the lessons learned in the implementation of the method may have general implications for similar sample surveys undertaken by Pakistan's social science researchers.