From welcome culture to welcome limits? Uncovering preference changes over time for sheltering refugees in Germany

Europe recently experienced a large influx of refugees, spurring much public debate about the admission and integration of refugees and migrants into society. Previous research based on cross-sectional data found that European citizens generally favour asylum seekers with high employability, severe vulnerabilities, and Christians over Muslims. These preferences and attitudes were found to be homogeneous across countries and socio-demographic groups. Here, we do not study the general acceptance of asylum seekers, but the acceptance of refugee and migrant homes in citizens’ vicinity and how it changes over time. Based on a repeated stated choice experiment on preferences for refugee and migrant homes, we show that the initially promoted “welcome culture” towards refugees in Germany was not reflected in the views of a majority of a sample of German citizens who rather disapproved refugee homes in their vicinity. Their preferences have not changed between November 2015, the peak of “welcome culture,” and November 2016, after political debates, media reporting and public discourse had shifted towards limiting admission of immigrants. A minority of one fifth of the sample population, who were initially rather approving of refugee and migrant homes being established in their vicinity, were more likely to change their preferences towards a rather disapproving position in 2016. Experience of contact with refugees and migrants, higher education, and general pro-immigration attitudes explain acceptance of refugee and migrant homes as well as preference stability over time. Country of origin and religion of refugees and migrants are considered less important than decent housing conditions and whether refugee and migrants arrive as families or single persons. In this respect our results highlight the importance of humanitarian aspects of sheltering and integration of refugees and other migrants into society.

provides an overview of sample characteristics which are used in our empirical analyses to explain class membership in the latent class choice model and a 2-state Markov model. We consider all respondents who did not have any missing values for the variables considered in this study. As shown in Table B, an analysis of selection bias regarding retest participation using a binary logit model reveals a statistically significant and positive effect for age only. Older respondents have a slightly higher likelihood to participate in the retest survey. Most notably we do not find any significant effect of the attitude towards immigration on the likelihood of retest participation.

Experimental design
In the stated choice experiment respondents were asked to choose between three alternatives regarding migrant and refugee homes in their vicinity the alternative they prefer most. The choice alternatives were labeled "Refugee Home," "Migrant Home," and "Refugee and Migrant Home" representing whether refugees, migrants or a mix of both live in the planned homes. The order of these alternatives was randomized across choice sets. The choice tasks included five choice attributes (see below) with eight levels (country of origin and religion), six levels (number of persons), two levels (families or single persons), and four levels (type of home and distance to respondent's house/flat). With respect to country of origin and religion we selected two countries (Syria and Serbia) that are within the top-ten countries in terms of numbers of asylum seekers being registered Germany in 2015, and two countries (Nigeria and India) that are less common as countries of origin of asylum seekers in Germany [1]. Another important selection criterion was that all of these countries have a sizeable population of Christians, which forms a common reference regarding religion.
The full factorial of all attribute-level combinations is 3,623,878,656 = (8 x 6 x 2 x 4 x 4) x (8 x 6 x 2 x 4 x 4) x (8 x 6 x 2 x 4 x 4). We employed an optimal orthogonal in the differenced (OOD) design [2]. Orthogonality ensures that the influence of a single attribute can be determined independently from the influences of the others. Besides orthogonality, the choice design was constructed to minimize the overlap between attribute levels across alternatives in a choice set, thus forcing respondents to make trade-offs between the single attributes. We obtained 36 choice sets which were blocked into six groups of six sets each, and each respondent answered one such group. The order of choice-sets within each group was randomized.

Instruction to the stated choice experiment in November 2015 and November 2016
In English

[Screen 1]
"New homes have to be established for refugees and migrants who come to Germany. We would like to know which homes you would prefer. On the following choice sets you can choose between different plans for the establishment of a new home. Please imagine that these homes will be built in your place of residence and will be there for at least 3 years.
The home can be for the accommodation of  Refugees only or  Migrants only or  Refugees and Migrants.
Refugees are persons who were forced to flee (e.g. due to membership of a specific social group or their political beliefs). Migrants are persons who left their country voluntarily (e.g., because they hope for a better life). Both groups -refugees and migrants -can ask for asylum in Germany. Then they are asylum seekers. It is more likely for refugees than migrants that their asylum application will be confirmed and that they can stay in Germany. In principle there can come refugees and migrants from any country to Germany." [Screen 2] "In the following each home for refugees and/or migrants will be described with the following attributes.
Mainly families or single persons: The persons can be mainly families with children or single persons.

Type of home:
The home can be an empty, renovated house; a container; an empty large building (e.g., building center or hospital) or an existing multi-purpose hall (e.g., gymnasium) can be used.

Statistical model
To understand citizens' preferences towards new homes for refugees and migrants and to model transitions in these preferences over time, various models were estimated. In the following we discuss the modeling strategy and the results of selected models. All models were estimated using Latent Gold Choice 5.1 [3].
To obtain an initial impression of the preferences in the sample as a whole we started with the standard Multinomial Logit Model (MNL), in which we included respondents from both waves (n=418) and all attribute levels as described in the previous section. This model yielded a (pseudo) R 2 of 0.1538, indicating that the attributes could reasonably well predict individuals' choices.  Since our interest was not in revealing individuals' preferences towards a range of countries and religions, but was actually focused on preferences towards Muslim and Syrian migrants/refugees, we estimated a second MNL model in which we only considered these attribute levels. Therefore, new dummy variables were created for Muslims (from either Syria or Nigeria) and for Syria (including Muslims and Christian). The other religions and countries were defined as the reference group. Reestimation of this -considerably more parsimonious -MNL model yielded a pseudo R 2 of 0.1474, indicating that the explanatory power of the model decreased only marginally. It should be noted, however, that the fit of the model (in terms of the Log-likelihood) was significantly lower.  We applied latent class choice modelling in order to reveal and explain transitions in preferences over time. Typically, the decision to consider a certain number of latent classes is based on model fit and model parsimony (i.e. the number of parameters). Various Information Criteria (weighing both) are available to this end [4]. In the context of latent class modelling, the Bayesian Information Criterion has been shown to perform particularly well and is therefore considered in this analysis [5].
Modelling individuals' transitions between the latent classes over time requires a low error rate in the modal assignment of individuals, i.e. the assignment to a latent class based on the highest probability. Otherwise, observed transitions between the latent classes over time could not be (solely) attributed to real preferences shifts, but would instead also reflect measurement errors. Therefore, as a second criterion in deciding upon the number of latent classes, the proportion of classification errors was considered [6].
Using the model considering only Syria and Muslims as country/religion ( In terms of the proportion of classification errors the 2-class model, with an error rate of around 8%, performed much better than models with 3-5 classes, with an error rate of around 16-17%. It follows that based on this criterion the 2-class model was clearly preferable for the purpose of our analysis. The decision was therefore made to opt for the parsimonious 2-class model. Selection of this model also meant that the complexity of the transition model would be kept at a relatively low level, and hence, that this model could be interpreted in a straightforward way.   (Table D). However, across all estimates there is a consistent difference towards a stronger rather disapproving preference. For example, there is a stronger dislike of (exclusive) migrant homes and Muslims. In addition, individuals' choices in class 1 are more strongly influenced by the number of people being sheltered and the distance between their home and the shelter location.
Citizens in class 2, representing the remaining 20%, have a more positive preference for refugee homes. They are indifferent as to whether refugees, migrants of a combination of both are being sheltered near their home and are also indifferent with regard to the number of refugees being sheltered. Similar to class 1 they prefer Syrian people and families (over single persons), but they have no particular dislike of Muslims. Individuals in class 2 have a strong preference for the two higher quality accommodations -an empty large building and renovated house -over the two less quality accommodations -a multi-purpose hall and container. Finally, the sign of the variable 'distance to you house / flat' is opposite to the sign in the first class, indicating that individuals in class 2 actually prefer the sheltering location to be closer to their home.
The (pseudo) R-square of the 2-class model (R²=0.2262) is substantially higher than the 1-class model (R²=0.1474). Interestingly, this increase can mainly be attributed to the second class, in which the choices can be predicted very well by the included attributes (class 2 R²=0.4286).
Finally, the (significant) influences of the six variables in the class membership model are in line with the preferences of both classes: individuals who are higher educated, have a stronger general proimmigrant attitude, who have been in contact with refugees, and already have a sheltering location near one's home (significant at 10%) have a significantly higher probability to belong to the second class. Gender and age do not significantly influence class membership. The model results reported in Table F are used to investigate how changes in the specification of a refugee or migrant home affect the probability of choosing this home. Changes in the choice probability are calculated based on a choice between three homes, using a sequential process where one characteristic of one out of the three refugee/migrant homes is altered at a time, while the remaining two homes retain a reference specification. This reference home is defined by the following characteristics: joint accommodation of refugees and migrants; non-Syrians; Muslims; mainly single persons; container building; accommodation of 138 people (mean of attribute levels 'Number of persons'); distance to respondents' home 1.43km (mean of attribute levels 'Distance to your home/flat'). Multinomial logit choice probabilities are calculated using parameter estimates for each of the two classes. If all three homes share the same (reference) specification, the probability of choosing any one out of the three homes is 1/3. If the specification of one home is altered, for example if the home accommodates Syrians rather than non-Syrians or Muslims rather than non-Muslims, the choice probability of this home will increase or decrease relative to the probability of choosing any of the two reference homes. The magnitude of the difference in choice probability between the evaluated home compared to the reference homes illustrates the relative influence of the different characteristics on respondents' decisions. Table G below reports the probabilities for selected characteristics. 95% confidence intervals for differences in probability between evaluated and reference home are calculated using a Krinsky and Robb [7] procedure. Notes: # indicates that effect was not significantly different from zero in latent class choice model; corresponding effects on probabilities are also not significantly different from zero. Class 1 refers to the "rather disapproving" class and Class 2 to the "rather approving" class.
As stated above, the identification of the two classes allowed modelling and interpreting transitions of individuals between these classes over time. To this end, a 2-state Markov model was estimated [8]. Based on modal assignment, each individual in each wave is (deterministically) assigned to one of the two classes, i.e. the class with highest membership probability for that individual. In the Markov model it is assumed that class membership represents a 'state' of an individual. In line with the firstorder Markov assumption, it is assumed that a person's state membership at the second point in time (2016) is influenced by his/her state membership at the first point in time (2015). In addition, the six explanatory variables previously included in the class membership function are again included in the model and assumed to predict initial state membership as well as state membership at the second point in time.
In order to test for parameter equality we compared constrained (MNL) models where equality constraints were imposed on specific model parameters with the baseline MNL model. As shown below in Tables H and I, likelihood ratio tests indicate that in all cases the models with equality constraints imposed on the parameters (with more degrees of freedom) result in significant reductions in model fit. Hence, the (absolute values of the) parameters associated with "mainly families," "an empty large building" and "multi-purpose hall" are significantly different and greater than the parameters associated with "Syria" and "Muslims."  (Table F). However, it should be noted that, due to the modal assignment of individuals (via which measurement errors are introduced), education level is no longer significant. The probability of being member of the first state in 2016 is positively influenced by state-1 membership in 2015. This means that, irrespective of (i.e., also when controlling for) the influences of the included explanatory variables, there is an 'autonomous' tendency of citizens to stay in or move to the first state. Finally, education level, the pro-immigrant attitude and contact with refugees decrease the probability of staying in / moving to state 1 in 2016. The parameter estimates can be used to compute the predicted transition probabilities reflecting the movement of individuals across the two waves. Table K presents this matrix. Individuals assigned to state 1 (rather disapproving) in 2015 have a probability of 90% of remaining in this state in 2016. In contrast, citizens assigned to state 2 (rather approving) only have a probability of 56% of remaining in their respective state in 2016. They have a substantial probability (44%) of moving from the "rather approving" state to the "rather disapproving" state.