Weighting and Analysis of Complex Samples
Weighting
We help you calculate selection probabilities in the context of complex sample designs to determine design weights. Furthermore, we support you in the calculation of so-called calibration or adjustment weights to reduce the bias of the estimate caused by non-response.
Weighting is a measure used in statistical data analysis to adjust the sample to the population under investigation. This is necessary when
- the potential bias due to different selection probabilities can be avoided or reduced (design weighting),
- the reduction of potential bias due to non-response is aimed at (adjustment weighting or calibration), or
- a subsequent stratification and reduction of variance can be achieved by adjusting to parameters of the population.
A "weight" is thus a multiplicative factor that can be used for any analysis of a survey data set with the aim of obtaining an unbiased or less biased estimator for the population.
Imputation of Missing Values
By imputation, missing values can be replaced by existing or estimated values to complete a data set. This makes it possible to work with a complete data set, for example for model analyses or for adjustment weighting. In addition, the imputation of missing values helps to reduce the nonresponse bias and to reproduce the covariance structure of the variables and the respective marginal distribution.
The determination of imputation values should always take into account the underlying failure mechanism and the use of auxiliary variables that are as highly correlated as possible with the missing characteristic. Different approaches and donors are available for this purpose, far removed from the imputation model used. Roughly speaking, a distinction is made between simple and multiple imputation as well as hot and cold deck methods for determining suitable donors. Depending on the method, variance-covariance structure and model used, imputation also has an influence on the variance of the estimator to be determined from the data. This must also be taken into account in the subsequent use of the data.
The imputation of missing values for your data set can be commissioned from GESIS both for your complete data set and for selected variables. In consultation with you, we impute missing values using simple or multiple imputation. The dataset for which this is to be done must be made available to GESIS.
Nonresponse-Bias-Analysis
Nonresponse bias can always arise when participants in a survey systematically differ from non-participants. This can be the case, for example, if certain social groups more often do not receive the invitation to participate/ cannot be contacted, more often decide not to participate or are unable to participate (e.g. because they lack the technical equipment necessary to participate). If the reasons for non-participation are associated with the target variables of the survey, it is not possible to reliably infer the population from the responses. The non-response bias analysis offers a method to assess the bias caused by non-response.
We advise on the possibilities of non-response bias analysis and give recommendations for the type of analysis for a specific question. We distinguish between measures of nonresponse bias at the level of individual survey questions (e.g. benchmark comparisons and subgroup analysis) and measures at the survey level (e.g. R-indicators). We give recommendations for dealing with non-response bias and show possible limitations of analysing surveys under non-response bias. We also discuss the possibility of correcting for non-response bias.