Combining Data from Different Sources: Different Techniques, Different Worlds

Combining data from different sources means to achieve a complete data file from different sources which do not (e.g. data fusion) or do (e.g. record linkage) contain the same units or may have an overlap. In any case, statistical or probabilistic matching or imputation techniques might be used.

On Day 1, after an introduction on all the practical and organisational aspects of the course, definitions are given trying to overcome the “Babylonian confusion” of many terms like e.g. statistical matching, data fusion, fusion, data integration, propensity score matching, single imputation and multiple imputation. Then it is discussed and shown that many real world research problems can be seen as missing data problems, likewise holds for statistical matching situations. The main focus of the course is on data fusion. Traditionally, data fusion is done on the basis of variables common to all files. It is well known that those approaches establish conditional independence of the (specific) variables not jointly observed given the common variables, although they may be conditionally dependent in reality. This implicit assumption will be illustrated by a little simulation study and there from four levels of validity are derived. These validity levels a data fusion procedure may or may not achieve will be discussed in greater detail.

On Day 2, the identification problem that is inherent in data fusion is shown and bounds for the inestimable correlations between the variables not jointly observed will be proposed and carefully explained. Since we treat the data fusion situation as a problem of missing data by design, we suggest imputation approaches to multiply impute the specific variables using informative prior information to account for violations of the conditional independence assumption. A short introduction to the multiple imputation theory will be given, too. Regarding the third level of the levels of validity a quality measure is proposed and simulation studies will illustrate the usefulness of these new ideas. Another shorter part of the course is devoted to propensity score matching and Rubin’s Causal Model (RCM). It will be shown the similarity of the potential outcome approach of the RCM to the data fusion situation. However there is a small but important difference of these two situations which makes propensity score matching a perfect choice for the RCM but not a good choice for data fusion. Guidance will be given to the participants which matching or imputation technique should be used in which situation. Finally, a “virtual”, i.e. anonymised data fusion project will be discussed with the participants pointing out many practical problems that may occur while undertaking data fusion projects. On Saturday afternoon the participants will follow the virtual project advice by themselves with hands on real world data and in the computer lab. After having finished the matching project in the lab, another real world project regarding the empirical application of the RCM is presented as a showcase.

Little specific prior knowledge is to be expected. Participants should only have:

• Experience with inferential statistics
• Basic knowledge of the linear regression model
• Basic understanding of matrix algebra

Furthermore, the participants should simply be willing to reflect and talk about their research questions.