Establishments in the Covid-19-Crisis (BeCovid): A High-Frequency Establishment Survey to Monitor the Impact of the Covid-19 Pandemic

2.1 Design of the Survey and Data Collection

The survey’s target population comprises all establishments with at least one employee subject to social security contributions, but excludes the public sector, private households, and extraterritorial organisations. Hence, it covers the entire private sector in Germany. Our sampling frame is the establishment file of the Federal Employment Agency, which contains all establishments (without the aforementioned entities) that have to submit employee notifications to the in social security system. In this context, an establishment is defined as a regionally and economically separate unit with employees that is assigned its own unique establishment identifier. In practice, this entails, e.g., that plants of the same company at different locations are treated as their own establishments. Moreover, this establishment definition is identical to other data from the IAB, like the Establishment Panel (Ellguth et al. 2014) or the administrative Establishment History Panel (Ganzer et al. 2021). The sample is drawn via disproportionate sampling, stratified by establishment size (1–9, 10–49, 50–249, and 250+ employees) interacted with five broad economic sectors. Large establishments with more than 250 employees are defined to be one stratum, irrespective of their sector, resulting in a total of 16 strata.

For each individual wave, the sample is divided into two parts: A refreshment sample consisting of establishments that have not yet been surveyed at any time during the study, and a repeat sample of establishments that had, in a previous wave, given their consent to be surveyed again (or those that had not yet been successfully contacted). Initially, an establishment was rotated out of the sample after a maximum of six participations. Other reasons for dropping out of the (repeat) sample were no participation for seven straight waves after the first-time interview and an establishment’s last participation being more than four waves ago. These measures are applied to prevent biases due to cluster effects of establishments that are particularly willing to participate, but also to survey a larger variety of establishments throughout the study to keep the burden low for the participants. From wave 17 onwards, however, those restrictions were lifted for large establishments (250+ employees) in order to ensure an ample supply of possible participants, such that the goal of at least 150 large establishments per wave can be realized up until the end of the study.

Table 1 provides an overview of how many interviews were conducted successfully during each wave, which share of these was repeat respondents, and the wave’s overall response rate. After varying between 1500 and 2000 for the first seven waves due to budget restrictions, the targeted sample size of roughly 2000 was consistently met in the subsequent waves. In total, across the first 23 waves, 17,772 establishments participated in the survey for a total of 43,973 observations. Starting with the third wave, the proportion of repeat respondents hovered around an envisaged level of roughly 60%. The only exception is the reduced share in wave 8 (46%), caused by the fact that, for the first time, a large part of the repeat sample was rotated out due the aforementioned rules. There is a great deal of variation in the wave-specific response rates, ranging from 12.3% to 23.6%. A reason for this lies in the confluence of short field periods and seasonal effects. Contacting respondents which are authorized to respond to surveys for their establishment was less successful during vacation periods and resulted more often in appointments for participation in a subsequent wave or in refusals, examples include wave 11 during Easter 2021 or wave 16 in that year’s summer vacation period. Similarly, during lockdown periods, particularly affected economic sectors such as accommodation or food services were more difficult to reach. Another, mechanical, reason for differing response rates is the varying number of repeat respondents that were available to contact in each wave, since prior respondents are more likely to participate than those establishments that have not been interviewed. An alternative way to calculate and interpret response rates in a high-frequency panel survey is based on the percentage of establishments that participated at least once throughout the entire study period. This preliminary total response rate is 22% through the first 22 waves, although this value has to be interpreted with caution as recently drawn establishments were contacted fewer times than other ones.

The main instrument of the survey is a questionnaire designed for a standardized CATI interview,^[1] containing questions that are only addressed to establishments surveyed for the first time, panel questions that are asked in each wave and questions on one or more thematic subjects (see Section 3 below for more details on the survey’s contents). The questionnaires were developed by IAB researchers and various project partners. A particular challenge concerns the very short time intervals between the waves, so that the time for developing and changing questionnaires is limited, while the pandemic situation itself further required swift adjustments to some questions (e.g. concerning lockdown).

Still, many efforts are made to ensure that the instrument undergoes rigorous testing prior to the field phase. A cognitive pretest (∼10–15 interviews) is conducted by IAB staff and focuses on the new questions of the wave in question. In this pretest, emphasis is placed on whether and how respondents understand and react to the questions. A second pretest (∼100 interviews) checks the entire instrument and is conducted in the telephone studio of the survey institute Kantar Public under the usual field conditions. The reporting mainly includes analyses of item non-response, interviewer notes, and the duration of the questions. The results of both pretests are jointly considered in the revision of the questionnaire.

Before the first participation, all establishments receive a postal cover letter to invite participation, signed by the head of the Federal Employment Agency and the director of the IAB. The cover letter points out the nature of a short telephone interview which is designed to address the economic problems of the Corona crisis. A second page of the cover letter includes information on data protection.

2.2 The Weighting Scheme

Given the targeted establishment population of the private sector, weights are available to obtain representative results. Whether weights should be applied, however, depends on the research interest at hand (Solon et al. 2015). In BeCovid, weights are constructed to make the sample of every wave representative of the German private sector economy; no panel weights are generated. This entails that the same establishment taking part in two different waves will not have the exact same weight as the sample composition varies across waves. Creation of the weights comprises a procedure in three broad steps. The first step utilizes administrative data from the establishment file to calculate the probability of inclusion that is solely based on the design. The second step updates these design-based weights, building on additional field-, contact, and administrative data to correct for potential selection effects that arise due to unit-nonresponse. Using generalized regression, the third step calibrates the weights. Final, calibrated weights allow for representative analyses by East and West Germany and by the aggregated sectors and establishment size classes which define the strata.

During the survey’s first year, another set of weights was generated through adjustments in the second step which are connected to short-time work (STW), a leading indicator of crisis severity during the pandemic. Estimated amounts of the total workers and establishments in STW from BeCovid turned out to be significantly higher than benchmarks from the BA that have a time lag of six months. With the possibility of linking the survey to administrative data on the establishments’ STW usage from the BA (resulting from the survey’s definition of an establishment), comparisons show that establishments receiving support in the form of STW are more likely to participate in the survey. This is probably explained by the fact that the BA pays out the STW support and is also a sponsor of the survey as evident to the establishments from the cover letter inviting participation. Hence, for example, establishments receiving support from the BA in the form of STW might participate in the BA-sponsored survey as a token of gratitude. As a consequence, the results on STW were no longer representative of the private-sector economy. Given the data linkage, however, the increased likelihood of participation for STW establishments can be adjusted for in the weights construction’s by correcting for this particular instance of non-response bias. Correspondingly, the data come with a refined set of weights that restores representativeness with respect to STW. Results that are not connected to STW specifically, however, are mostly unaffected when comparing the original and refined weights. Further details on this are presented in Fitzenberger et al. (2021) who also show that other firm surveys similarly overestimated STW numbers. Intrinsically, due to the institutional set-up of STW in Germany with employers having up to three months to settle their claims, the STW numbers elicited during the survey that asked for a tally of a not yet completed month should be read as an estimate from the establishment’s side (Kagerl et al. 2022). Exact figures for the surveyed establishments can be obtained ex post from administrative data, also see Section 4.