Precision, time, and cost: a comparison of three sampling designs in an emergency setting

Appropriate response to a nutritional emergency requires reliable and timely data about the health, nutrition, and food security status of the affected population. The assessment method commonly used in emergency settings is a 30 × 30 cluster survey [1]-a method that provides statistically reliable results if administered and analyzed correctly, but that can be time-consuming and expensive to administer. Studies have been conducted or are currently underway to investigate alternative methods to collect reliable data in emergency settings [2-4].Deitchler et al [3] recently compared the performance of two alternative sampling designs, a 33 × 6 cluster design (33 clusters, 6 observations per cluster) and a 67 × 3 cluster design (67 clusters, 3 observations per cluster), to a conventional 30 × 30 cluster survey for estimation of child-level indicators. Data for the 33 × 6 design, 67 × 3 design, and a standard 30 × 30 design were collected concurrently in the Siraro woreda in Ethiopia during the 2003 nutritional emergency. The study showed encouraging results with respect to the statistical reliability and time savings offered by the 33 × 6 and 67 × 3 designs. However, that study did not sample for each design independently; rather, data were shared among the designs when the same primary sampling unit (PSU) was selected for sampling by multiple designs. The investigators recommended that a second study, using independent samples for each design, be carried out to validate the 33 × 6 and 67 × 3 designs [3]. This paper addresses that recommendation.In the study described here, we use independent samples, representative of the same area, and collected concurrently, to compare the performance of the 33 × 6, 67 × 3, and 30 × 30 designs. Data were collected in the Administrative Units (AUs) of Fur Baranga and Habila in West Darfur, Sudan in September-October 2005 in order to compare: 1) child- and household-level indicator results (point estimates and 95% confidence intervals