Treatment of unit and item nonresponse in surveys   

Nonresponse inevitably occurs in most, if not all, surveys. Essentially, survey statisticians distinguish between two types of nonresponse: unit nonresponse and item nonresponse. Unit nonresponse occurs when all the survey variables are missing or not enough information is available. Item nonresponse occurs when some but not all survey variables have missing values. Weight adjustment procedures are typically used to treat unit nonresponse, whereas imputation is generally used to treat item nonresponse. The main objective when treating (unit or item) nonresponse is to reduce the nonresponse bias, which occurs if respondents and nonrespondents are different with respect to the survey variables.

The workshop will consist of three modules, with time for questions and discussion after each.

In the first module, causes and prevention of nonresponse will be briefly discussed. Also, point estimation in the complete data case will be presented. The concept of bias, variance and mean square error in the design based framework will be explained. The Horvitz-Thompson estimator and calibration estimators (that include the post-stratified estimator as a special case) will be covered. Finally, we will explain the link between two-phase sampling and the nonresponse situation, which will bring us to speak about the nonresponse mechanism (including the concepts of ignorable vs. nonigorable nonresponse mechanisms).

The second module will focus on the treatment of unit nonresponse. Two weight adjustment procedures will be presented: (i) nonresponse propensity weighting that consists of adjusting the weights of the respondents using the inverse of the estimated response probabilities and (ii) nonresponse calibration weighting including generalized calibration. The concept of reweighting classes and their construction will be presented.

The final module will concentrate on treatment of item nonresponse using single imputation: commonly used imputation methods will be presented and their properties will be discussed. The construction of imputation classes will be presented. Variance estimation in the presence of nonresponse will be covered including techniques such as the jackknife and the bootstrap. Finally, multiple imputation and its properties in the context of surveys will be presented.