Page 221 - Contributed Paper Session (CPS) - Volume 7

CPS2068 Jan-Philipp Kolb et al.
from the German population register by a random sample. Every two years a
panel refresher is carried out. The German general social survey (ALLBUS) is
used for the data restocking. Except for an oversampling in Eastern Germany,
the ALLBUS has the same sample design as the GESIS Panel. Participants in the
ALLBUS are asked if they are willing to participate in the GESIS Panel. A
restocking takes place every two years. So the first cohort is made up of the
panellists who have been participating since the beginning of the panel. In the
second cohort are people who have participated in the survey since 2016, and
the panellists who have just been recruited are part of the third cohort.
Mixed mode means that there are two groups of participants-- the
panellists have the choice between an offline questionnaire and an online
version. These two modes ensure for example that people without an internet
connection are also represented in the panel. It does not matter if they do not
have the internet because of conviction or age.
Access panel means that researchers can submit study proposals. These
proposals are then peer-reviewed, and if the review is positive, the submitters
have a slot of about five minutes in the panel at their disposal. The proposed
study can be either cross-sectional or longitudinal. The GESIS panel started in
2013 with more than 32 waves and over 40 studies running in the panel since
then. The data is available as scientific use file, and it is also possible to use an
extended version in the GESIS secure data center.
Our target is to predict unit nonresponse in the GESIS Panel. More
specifically, we want to predict whether a person has a high likelihood to fall
into one of the following AAPOR categories (Smith and others 2004).
• 211 Refusal
• 212 Break-off: Questionnaire too incomplete to process
• 319 Nothing ever returned
• 2112 Explicit refusal
• 211221 Logged on to survey, no item complete
On this basis, we e.g. flag a unit nonresponse if a person breaks off at the
start so that the questionnaire is too incomplete to process. The result is a
binary variable, which has value one for non-respondents and zeroes for
respondents.
Potentially we could use more than 4000 variables as predictors. As a first
block of potential predictors, we have substantive survey information from
studies like “Lifestyles in everyday life” with study token aa or “threat
perception and political trust” (study token bi). Examples for these variables
are age, gender or educational attainment of the panelists, as well as distance
from place of residence to a large city (Kolb and Weyandt 2018). Some of the
variables are collected regularly if the study is longitudinal. Other variables are
only available at one point in time.

208 | I S I W S C 2 0 1 9