摘要：Survey statistics has always been a somewhat specialised area due in part to its view of the world. In the rest of the statistical world data are random and we model their distributions. In traditional survey statistics the population data are fixed and only the sampling is random. The advantage of this ‘designbased’ approach is that the sampling, unlike the population, is under the researcher’s control. The basic question of survey statistics is If we did exactly this analysis on the whole population, what result would we get? If individuals were sampled independently with equal probability the answer would be very straightforward. They aren’t, and it isn’t. To simplify the operations of a large survey it is routine to sample in clusters. For example, a random sample of towns or counties can be chosen and then individuals or families sampled from within those areas. There can be multiple levels of cluster sampling, eg, individuals within families within towns. Cluster sampling often results in people having an unequal chance of being included, for example, the same number of individuals might be surveyed regardless of the size of the town.
