Pre-Analysis Plan: What drives refugees’ location choices? Evidence from a conjoint experiment among Ukrainian refugees

Each respondent is shown three choices each between two hypothetical destination countries with given characteristics, and asked which country they would choose.

We study Ukrainian refugees across Europe in survey I and in Germany in survey II. The first survey’s participants have been previously recruited using Facebook advertisements by Kantar Public, whereas we recruit the participants in the second survey by contacting a random sample of 30,000 Ukrainian refugees in Germany from administrative postal address data, provided by the Federal Office for Migration and Refugees (BAMF). The two surveys are complementary: Kantar Public survey allows reaching Ukrainian refugees across Europe, but is restricted to Facebook users, while our own survey in Germany uses administrative data for a more representative sample, but is restricted to one country.

The selected attributes can be roughly categorized in three broad categories.
Country dimensions:
– Proximity to Ukraine (within 500km, not within 500 km)
Distance is one of the main determinants of population movements and
also of refugee movements in particular (Suleimenova et al., 2017). We
hypothesize that proximity is also relevant in this setting because of (for
example) travel costs, and ease of visiting family in Ukraine. A small
caveat could be that closer proximity to (the conflict in) Ukraine could be
perceived as less safe.
– Average net wage levels (between 16,000 and 100,000 Hryvna)
Earnings potential in destination countries are a main determinant for
population movements (Grogger and Hanson, 2011). We draw the net
wage level from a uniform distribution between the lowest (Bulgaria: EUR
400, corresponding to 16,000 Hryvna) and highest (Ireland: EUR 2500,
corresponding to 100,000 Hryvna) of the according to the EU SILC in
2019.
– Housing cost of a one bedroom apartment on the private market (between 20 and 40% of the average net wage)
Housing costs are an important determinant for the cost of living and
therefore determine the relative attractiveness of destinations. As usually
stated in Ukraine, we study housing costs including utilities of a common
type of apartment. We choose to operationalize housing costs as amounts
relative to wage levels, in order to not draw completely unrealistic combi-
nations of wage levels and living costs.
Policy dimensions:
– Social benefits conditional on unemployment (between 0 and 30% of average wage)
The generosity of welfare benefits have been studied extensively and is a
key element of the immigration policy debate (see e.g., Agersnap et al.
(2020)). Social benefits for unemployed Ukrainians are 0 for Poland,
about 200 euro in Czech Republic and 502 euro in Germany (for individuals over 25 years of age). Nevertheless, many countries (such as Germany) have income-dependent housing subsidies, which would render this amount somewhat higher. Therefore we vary this amount between 0 and 30%.
– Child benefits (between 0 and 10% of average wage per child)
Unconditional benefits for refugees are another policy dimension with dif-
ferent implications than conditional benefits. This is especially salient
because of the composition of Ukrainian refugee families, as many are
comprised of women and children. In order to study the relevance of this
dimension, we elicit the number of children one is accompanied by in both
surveys.
Individual-country dimensions:
– Personal networks at destination (yes, no)
Networks have been shown to be crucial in refugees’ destination choice
(Di Iasio and Wahba, 2023; Crawley and Hagen-Zanker, 2019; Barthel
and Neumayer, 2015; Beine et al., 2011). As we want to relate elicited
preferences to revealed preferences, our surveys also elicit whether respondents have family members or friends in their current destination prior to moving.
– Labor market prospects (easy/difficult to find a job corresponding to one’s qualifications)
Individual labor-market prospects and perceptions thereof are likely to
affect destination choice. As we want to relate elicited preferences to
revealed preferences, our surveys also elicit whether respondents deem it
easy to find a job corresponding to their qualifications in their current des-
tination as well as the three most popular destination countries (Poland,
Czech Republic and Germany) as well as two additional countries (Italy
with weak and Sweden with strong labor markets).
– Knowledge of destination-country language (yes, no)
Immigrants’ knowledge of destination country languages increases earnings (see e.g. Adsera and Pytlikova (2016)) and linguistic proximity to
host-region languages increases earnings among asylum seekers (see e.g. Wong (2023)). As we want to relate elicited preferences to revealed preferences, our surveys also elicit whether respondents knew the language of their destination country upon arrival.

Randomization
• As some of the dimensions contain levels that are strictly ordered (so-called valence attributes in voting conjoints, see e.g. (Franchino and Zucchini, 2015)), our tasks could contain trivial profile pairs. As an example, two otherwise identical destinations only differ in whether or not the respondent speaks the destination country language. A rational respondent would always prefer the option with the country where he speaks the language. Such draws hence do not provide any meaningful variation However, as we include 8 dimensions and the inclusion of multi-valued levels of several attributes, the probability to draw such profiles is small. As we do not want to impose any assumptions on whether or not our dimensions are valence attributes, we keep such profiles.
• Randomization is performed by fully random draws in all cases by computer, using distributions reported in the pdf file of our pre-analysis plan.

2023-11-15

2024-01-31

Three choices of hypothetical destination country with given characteristics per respondent.

Information on country in which the respondent actually resides, in terms of (1) whether the respondent expects it to be easy to find a job correspondig to own qualifications, (2) average net wage level (below or above 20,000 euros per year), (3) generosity of social benefits for unemployed refugees and (4) child benefits, (5) whether the respondent spoke the language, (6) whether the respondent had family or friends in the chosen country at the time or arrival, and (7) distance to Ukraine (For this, we divide the sample in three groups: those hosted in a country whose capital is less than 500 kilometer away from the Ukrainian border, between 500and 1000 kilometers and more than 1000 kilometer.)

Survey experiment (forced-choice conjoint)

Randomization of shown country characteristics is done by survey software, according to the rules specified in the PAP document.

Individual

No

There are no clusters in the typical sense of an RCT. The two hypothetical destinations vary along 8 dimensions. Given that every respondent has to make three hypothetical destination choices, we cluster standard errors on the individual level as in its absence the correlation between the responses of the same individual may lead to overstate precision.

About 1000 respondents in the Kantar Public survey and 3000-4000 respondents in our own survey.

In each survey, each individual has an own randomization of the characteristics of the hypothetical destination countries in three pairs.

To infer the significance of our effect, we always use two-sided tests with α = 0.05. We aim to detect effect sizes of 5 percentage points (AMCE), which is smaller than effect sizes for access to welfare and labor market conditions in comparable studies (Ferwerda and Gest, 2021). We performed a power analysis using the R package cjpowR (Schuessler and Freitag, 2020), calculating the power with a combined sample size of 4,000 (1,000 for Survey I and 3,000 for survey II). For the combined sample, the statistical tests of the main effects for our 8 attributes when the true effect size is 5 percentage points have a power exceeding 99%. The power to detect internal interaction effects of 5 percentage effects is 97%. Subgroup analyses on a subsample of 20% (which is the share of men among Ukrainian refugees) of the full sample have a power of 93%. Hence, our studies are sufficiently powered to test aforementioned hypotheses.

Pre-analysis plan

Understanding refugees’ destination choices is key to designing appropriate policies, but little is known about this beyond correlational patterns and the effects of isolated policy changes. To derive causal evidence on how destination country characteristics affect refugees’ destination choices, we conducted a forced-choice conjoint experiment among 3,348 Ukrainian refugees across Europe. In the survey experiment, refugees repeatedly chose between two hypothetical countries that varied on eight relevant attributes. Ukrainian refugees are uniquely suited to study the relative importance of different host country attributes as they have the right to choose in which member state of the European Union they apply for temporary protection. We find that job opportunities are a much stronger driver of destination choice than social assistance. A one SD increase in average wages makes it 16.4 percentage points more likely that the country is chosen in the survey experiment, while a corresponding increase in social assistance increases the probability of the country being chosen by only 4.5 percentage points. The ease of finding a job matching one’s qualifications is valued even more than having friends or family and knowing the language of the country. We also find strong sorting in that respondents who value knowing the language and job opportunities in our survey experiments are also more likely to have settled in countries where this is the case. Conjoint experiments can be used to assess refugees’ anticipated migratory responses to proposed policy changes.

Adema, Joop, Lasha Chargaziia, Yvonne Giesing, and Panu Poutvaara (2025), Refugees from Ukraine value job opportunities over welfare. Proceedings of the National Academy of Sciences (PNAS), 122 (32): e2502420122.

https://doi.org/10.1073/pnas.2502420122