Major milestone: Code matching

Throughout the school year, we conducted a school survey in Austria, Germany and Switzerland every three months.

After we finished the data collection, we moved on to cleaning and matching the three waves. To ensure the anonymity of the young people throughout the survey, we always started with a self-generated questionnaire. Code collected.

After the survey was completed, the next step was to match the three waves of data. Sophie and Dora from the Swiss team devised a matching strategy for this: they wanted to automate the data matching in R to ensure a more organized process. However, it quickly became apparent that the young people had often entered the code incorrectly in all three surveys. The question then became: how could they still find their matches?

A structure had to be created that would generate as many matches as possible without potentially creating false matches. Over the summer, possible strategies were considered with the help of the AWARE team. Based on this, Sophie and Dora created a decision tree and coded it in R. This tree was designed to allow for errors such as transposing initials in the code (including checks to ensure that a young person shared the same sociodemographic characteristics). This allowed for the creation of a longitudinal dataset of 626 young people who participated in all three waves of the survey.

The decision tree progresses from strict to tolerant : First, all six code parts must match (1: Exact) . Then, exactly one code part may differ (2: Fuzzy). Next, typical swaps are allowed – e.g., the first two initials (3: Swap) . Afterward, the remaining lines are automatically evaluated: A model trained on the already secured matches decides whether a line matches a person ( 4: AutoRec). However, this only applies if certain additional criteria are met – same school, essentially the same name, plausible date of birth, and the same sociodemographic characteristics. Everything else is checked manually (5).