Tid: 5 februari 2020, kl. 13-14


Within the theory of survey sampling, being design-based means that the only randomness in the estimation process of some population quantity stems from the random mechanism generating the sample. How the values of the study variable are actually created is completely ignored. This approach for making statistical inference soon developed its own culture of notation which differs somewhat from what is commonly used within other areas. For instance capital letters stand for population totals instead of random variables. This may have made some statisticians sceptical and/or confused about the design-based concept.

The other approach to survey inference means that one relies on a stochastic model for explaining the outcomes of the study variable and not much attention, if any, is paid to the actual sampling mechanism. Model-assisted estimators have been suggested as a sort of compromise, where there should be some robustness in terms of coping with a misspecified model.

For some years now the problem of nonresponse has increased and this offers new challenges in terms of finding good estimates of population quantities. Probability sampling, which is the core of design-based thinking, has by some been proclaimed doomed and it is said that the sample should be obtained in other ways combined with some modelling.