Who Gets to Work from Home? Evidence from an Employer Survey Experiment in Germany

2026-03-09

2026-05-31

In each of three allocation choices, respondents assign 0–4 remote-work days per week (half-days allowed) across two candidates. The primary outcome is the within-pair allocation difference, coded so that positive values indicate an advantage for the focal attribute in that choice (mother vs. father in Choices 1–2; migrant vs. non-migrant or parent vs. childless in Choice 3, depending on arm). A key secondary outcome is the total number of days granted within a pair (the sum across the two candidates), capturing overall generosity and enabling analysis of performance effects on the extensive margin of granting remote work. After Choice 2, respondents also report (i) a third-person benchmark allocation for the same pair, (ii) which candidate would be more productive when working from home, and (iii) which candidate is more likely to decline if no remote work is offered; the module concludes with a short battery of 0–10 attitude items on perceived consequences of remote work. For more details, see attached pre-analysis plan.

The experiment is embedded in the SOEP-LEE2-Compare establishment survey (wave 2026) and administered in the “Home Office/Remote Work” module via CATI and, where applicable, CAWI. Respondents are establishment representatives (typically managers, owners, or HR staff) with responsibility for staffing or workplace-flexibility decisions. The experimental block is designed to be brief and standardized: interviewer prompts and vignette wording are scripted, and respondents complete three independent allocation tasks in which they distribute a fixed weekly budget of up to four remote-work days between two hypothetical candidates (half-days allowed; a zero allocation is permitted).

Candidate profiles are one-line vignettes with a fixed job context and randomly varied attributes: gender, migration background, parenthood, and a performance signal. Choices 1–2 are gender pairs contrasting an otherwise identical mother and father, with migration toggled across the two choices and the performance signal held constant within respondent. Choice 3 is randomly assigned to be either a migration pair (migrant vs. non-migrant of the same gender) or a parenthood pair (parent vs. childless with gender and migration held constant within the pair). Across all choices, the two profiles are identical in non-focal attributes and left–right placement is balanced by design; randomization is implemented ex ante at the firm/interview level and merged into the survey instrument.

Not available

Randomization is implemented ex ante in Stata at the firm/interview level prior to fieldwork and merged into the CATI/CAWI instrument. Each sampled establishment is pre-assigned a complete set of six candidate profiles (A–F) that fully determines the left and right profiles in each of the three choices. Randomization is stratified by region and uses blocked procedures within strata to achieve balanced 50/50 splits for all binary attributes and to enforce the intended 50:50 share of the two Choice 3 pair types. A fixed random seed and a reproducible generation log are retained.

All randomized attributes are binary: gender (woman vs. man), migration background (born abroad vs. born in Germany), parenthood (has children vs. childless), and a performance signal (excellent vs. average). The assignment follows simple design restrictions: within each choice, the two candidates are identical in all non-focal attributes, and only the focal attribute differs across left and right; left–right placement is balanced by construction. In Choices 1–2, the performance signal is drawn once per respondent and held constant, migration is constant within a choice but toggles across the two choices, and gender is mirrored to balance side presentation. In Choice 3, respondents are assigned to either a migration pair (contrast in migration background) or a parenthood pair (contrast in parenthood), with remaining attributes held equal within the pair and randomized across respondents.

Firm

No

None

1000

1000