Unintended Consequences of the 2015 Refugee Surge on Residential Building Permits in Germany

2.1 Data

Our data set covers a maximum of 390 out of 401 counties and cities with county status in Germany.^[2] For simplicity, we refer to counties in the remaining text, which include both counties and cities with county status. The panel data include yearly observations from 2011 to 2019. Our panel setup is unbalanced, as there are missing observations for a number of variables. We exclude the year 2020 and all following years, as the Corona/Covid pandemic caused a delay in new housing construction (Mohammed et al. 2021). Our main data source is the INKAR data base,^[3] with some additional variables drawn from the German regional data bank.^[4] Variable definitions and source codes are provided in Table 1 and basic statistics in Table 2. For some sensitivity checks, we also draw on alternative refugee data and some other variables from Lange and Sommerfeld (2024). These data are available only for the years from 2013 to 2018, which reduces the number of observations by about a third. The variables associated with this alternative data set are covered in the last four rows of Tables 1 and 2.

Our variable to explain is the residential building permit rate. Building permits are a good indicator of the current state of the real estate market, as they capture current decision making (Lerbs 2014). Therefore, they should be sensitive to temporary processing problems of overwhelmed building authorities on the one hand and possible sentiment changes by prospective home builders. Our permit rates are defined as the number of residential housing units in new buildings per 1,000 residents.^[5] We differentiate between two permit series, those for all housing units, regardless of what type of building they are in, and one of its two components, the permits for housing units in one- and two-family houses. The latter we will call traditional permits. We omit permits for housing units in buildings with 3 or more units, which includes high-rise apartment buildings, student and senior dormitories, and refugee shelters.

Our focus is on the traditional permit rate. In contrast to building permits for units in multi-family buildings, the traditional permit rate is less likely driven by public policy objectives, such as providing housing for refugees or the decision making of commercial housing developers. In that sense, traditional permits more closely capture the individual-family decision making about home building and the extent to which they may be influenced by the influx of refugees. If the building authorities were indeed overwhelmed in 2015/16 in their efforts to provide for refugee housing, their attention to permits involving traditional permits would likely have suffered more than their attention to housing units in multi-family buildings, which includes refugee centers. If, by contrast, the refugee surge is associated with fewer traditional permits not only in the wake of the refugee surge but for years thereafter, it becomes difficult to blame overwhelmed building authorities. A more likely reason for such a longer lasting development would be a sentiment change among prospective home builders.

For our refugee data, we rely on a very general measure, the share of the population seeking protection from war or persecution (refg in Table 1). The 2015/16 refugee wave increased that share significantly. To distribute the burden of refugee inflows, refugees have been registered and dispersed into all German states and counties according to the Königstein Key since 1993. However, the 2015/16 surge overwhelmed many local authorities leading to numerous local adjustments to avoid homelessness. In addition, refugees often took things into their own hand and moved on from their designated localities. These issues are discussed in detail in the literature on the impact of the 2015/16 refugee surge into Germany (e.g., Berbée et al. 2022; Gehrsitz and Ungerer 2022; Lange and Sommerfeld 2024). As they may induce endogeneity of the general refugee numbers, we also employ in a sensitivity check an alternative refugee series from Lange and Sommerfeld (2024) along with an instrumental variable measure devised by these authors and tied to the Königstein Key.

We apply economic, demographic, and housing-related controls. First, we control for economic variation across counties through the inclusion of unemployment and GDP variables. Demographic changes are captured by the level of population, net migration, and the share of foreigners in the population. A larger share of foreigners, in particular, may moderate the response of building permits similar to how it moderated the response of hate crimes to foreign-born residents in a recent study by Entorf and Lange (2023) of the 2015/16 refugee crisis. In terms of housing-related controls, we rely on two variables: the area of land sold that is ready for housing construction and the average price of that land. We expect that building permits increase when land transactions increase as new owners are more likely to build on previously vacant land.

2.2 Methodology

Our base regression model takes the form

where the dependent variable is the logarithm of the building permit rate (prm) of county i in year t and where ɛ _it is the equation’s error term. The term ϑ _i denotes county fixed effects, which control for all observed and unobserved time-invariant county-specific factors. These include locational characteristics that do not change with time. By contrast, the term involving cntrl captures J county-specific, time-varying control variables, such as unemployment.

δ _t represents country-wide time fixed effects, which capture time-varying factors at the federal level that affect all counties. As an example, δ _t controls for the impact of changes over time in country-wide price levels, interest rates, environmental standards, and a host of other variables that may affect building permits. The term involving cnty in Equation (1) represents interactions between our time fixed effects variable year and H time-invariant county attributes. These relate to the variables west, big7, s_county, and growth in Table 1. The term allows for time fixed effects to differ across (a) the western and eastern counties of Germany, (b) the cities belonging to the big7 group and all other counties, (c) the four county categories based on differences in population density, and (d) the five county categories based on different levels of growth momentum. This specification makes it possible for important time-invariant county characteristics to influence not only the level of the dependent variable but also how it evolves over time. It also acknowledges the fact that country-wide time fixed effects, as given by δ _t in Equation (1), are unlikely to capture the heterogeneity in observed trends across counties. With our specification, each county is influenced by a sizable number of potentially different time trends above and beyond any country-wide trend.

The term involving variable refg in Equation (1) captures the impact of refugees on the building permit rate. By interacting refg with the time fixed effects variable year, we allow the data to determine whether and in which year a change in the refugee rate impacted the building permit rate. Since the dependent variable and the refugee variable are both expressed in logarithms, β _t in Equation (1) represents a series of elasticities of the building permit rate with respect to the refugee share. If there was no appreciable response of the permit rate to changes in the refugee share before the refugee surge of 2015/16, we expect the β _t to be zero before 2015. Similarly, if the 2015/16 refugee surge made building permits more responsive to the refugee share, we expect to see some or all of the β coefficients to be different from zero starting in or after 2015. The sequence of elasticity values β _t over time also allow us to differentiate between the two key competing hypotheses of the refugee surge. An immediate but temporary rise in the responsiveness of building permits to the refugee surge would point toward a lack of land for development or overwhelmed building authorities. A more gradual but longer lasting response would more likely be associated with potential home owners reassessing their decision to build in the aftermath of the refugee surge.

We investigate the short-run versus longer run impact of the 2015/16 refugee surge also by Equation (2),

where treat stands for a 0/1 treatment indicator variable that turns from 0 to 1 either in 2015 or later and that remains at 1 for the remainder of the sample. For a short-run, temporary impact, we would expect the coefficient β ₁ to be smaller in absolute terms and less statistically significant the longer after the refugee surge variable treat turns from 0 to 1, e.g., 2017 as opposed to 2015. Just the opposite is to be expected if the refugee impact is gaining momentum in the longer run. We note that β ₀ represents the elasticity of the building permit rate with respect to the refugee share prior to the refugee surge. By contrast, the sum of the coefficients β ₀ and β ₁ represents the elasticity after the surge.

We conduct two sensitivity checks. The first one is intended to rule out reverse causality concerns for our control variables by replacing the J contemporaneous controls with interactions between our time fixed effects variable year and the beginning-of-sample (2011) values of our control variables,

This is an approach applied, among others, by Jeanty, Partridge, and Irwin (2010).

In a second sensitivity check, we draw on alternative refugee data, an associated instrumental variable, and some additional control variables (Lange and Sommerfeld 2024). The refugee data for this check come from the Federal Office of Migration and Refugees (BAMF) and cover the time period 2013–18, instead of the years 2011–19. The alternative refugee variable (refg_infl) represents inflows of refugees per 100,000 residents per year,^[6] in contrast to the end-of-year stock variable refg that is being used for the baseline estimates. Since the refugee numbers induce a potential endogeneity bias due to endogenous sorting by refugees or last-minute refugee reallocations by local authorities, Lange and Sommerfeld (2024, Section 3.2) construct an instrumental variable (IV) for the refugee inflow variable. This IV predicts the number of refugees per county on the basis of fixed ex ante allocation quotas that are not influenced by the actions of refugees or local authorities. For our checks with the alternative refugee data and the IV estimates, we rely on simplified versions of Equations (1) and (2) over the period from 2013 to 2018.

3.1 Baseline Regression Results

The regression results from estimating Equations (1)–(3) on the baseline refugee data (refg in Table 1) over the period 2011–2019 are collected in Table 3, with the dynamic elasticity estimates pictured in Figures 1–3.^[7] All models in the table contain a full set of control variables, county fixed effects (CFE), country-wide fixed effects (year), and a large number of interactions between time fixed effects (year) and time-invariant county characteristics, such as a location in the western as opposed to the eastern part of Germany.

Figure 1:

Refugee variable interacted with year time fixed effects (Table 3, Mod. 1, prm_all).

Figure 2:

Refugee variable interacted with year time fixed effects (Table 3, Mod. 2, prm_12).

Figure 3:

Refugee variable interacted with year time fixed effects (Table 3, Mod. 5, prm_12).

Models (1) and (2) differ only in terms of the dependent variable. Model (1) and Figure 1 provide the results for the building permit rate for residential housing units in all types of buildings, from single- and multi-family homes to senior homes and refugee centers. We can observe in Figure 1 a movement away from zero in the elasticity of permits with respect to refugees in 2017 and 2018, with only the 2018 deviation being statistically significant. Overall, there appears to be no systematic reaction in the elasticity of building permits to the refugee crisis. This is not entirely surprising as permits for all types of housing, including those for refugee centers, are lumped together. By contrast, Model (2) and Figure 2 focus on what we call traditional building permits, those for housing in one- or two-family homes, on which we will focus in all remaining regressions.

In Figure 2, we can observe a distinct downward trend in the elasticity estimates for the traditional permit rate. This trend entails an increasing (negative) sensitivity of the permit rate to the refugee rate. It sets in 2016 and keeps the elasticity of building permits with respect to refugees in the range between −0.08 and −0.10 for the remainder of the sample. Values in that range imply that a doubling of the refugee rate, a common occurrence in the wake of the refugee surge, would drop the traditional permit rate by 8–10 percent.

The shape of the downturn in the elasticity values and the evolution of their absolute size suggest that the 2015/16 refugee surge had a more sustained, longer-lasting impact than what can be expected from temporary administrative issues by the building permit authorities.

That the refugee surge had more than a short-run impact on the elasticity of traditional building permits is underscored by Models 3 and 4 of Table 3. These models correspond to Equation (2) and capture in condensed format the change that occurred in the building permit elasticity after the refugee surge. Model (3) puts the building permit elasticity at 0 before 2015 and at −0.08 from 2015 to the end of the sample. Model (4) replicates the estimates with 2016 replacing 2015 as the treatment year. The elasticity value before 2016 remains at zero, while that from 2016 onward reaches −0.09. Unreported results show that the estimated coefficients do not materially change if 2017 replaces 2016 as the treatment year. These estimates confirm the message of Figure 2: the sensitivity of building permits to the share of refugees in the resident population tends to rise in absolute terms after the refugee surge, instead of declining again after an initial uptick. This speaks against a temporary impact of the refugee surge on the building permit elasticity and in favor of a longer lasting impact, one that gains momentum over time.

3.2 Sensitivity Checks

Model 5 of Table 3 identifies the structure of a first sensitivity check of the results contained in Models 2 through 4. It implements Equation (3) where all contemporaneous values of the control variables are replaced with interactions between their 2011 values and the time fixed effects variable year. These interactions put the counties’ building permits on different time paths conditional on the 2011 values of the counties’ control variables.^[8] It rules out endogeneity issues that may arise from the control variables.

Figure 3 illustrates the estimation results of Model 5 that are of interest to us: the responsiveness of the traditional building permit rate to the share of refugees in the population over time. The figure strongly confirms the results of Models 2–4 of Table 3. Two results stand out. First, the refugee share at the county level has no statistically significant impact on the building permit rate prior to 2016. The elasticity of traditional building permits with respect to the refugee share is effectively zero. Second, building permits become responsive to the refugee share in a statistically significant way in 2016. Importantly, the impact of the refugee share is not of the short-run, temporary type, but its impact gains momentum over time. This points toward a material, longer run increase in the absolute value of the elasticity of building permits with respect to the share of refugees in the population. In particular, Figure 3 suggests that a 100 % increase in the population’s refugee share, say from 1 % to 2 %, lowers the permit rate by approximately 15 % by 2018 or 2019.

Table 4 collects a number of additional sensitivity checks. These cover the estimation results for alternative refugee data, explore the influence of additional control variables adapted from related work on the impact of refugees by Lange and Sommerfeld (2024), and check to what extent endogeneity in the refugee data may impact the results through the application of an innovative instrumental variable developed by Lange and Sommerfeld (2024). All variables used for the models of Table 4 that are different from those employed for the models of Table 3 are described at the end of Table 1, with basic statistics provided at the end of Table 2. All models collected in Table 4 use the traditional building permit rate (prm_12) as the dependent variable. They cover the time period 2013–18, instead of the years 2011–19, as do the models reported in Table 3.

Our first results (Figure 4) pertain to a model that follows in its structure Model 2 of Table 3 and is a counterpart to Model 1 of Table 4, with the difference that variable log(regf_infl) is replaced by interactions with the time fixed effects variable year. Figure 4 corresponds to Figure 2. The difference between Figures 2 and 4 lies in the fact that Figure 4 is based on (a) a different refugee variable (refg_infl), (b) a shorter time frame (2013–2018), and (c) the addition of two control variables. The first control variable is a time-varying 0/1 indicator for the presence of initial refugee reception facilities (EAE) within a county. The presence of such a facility is expected to draw a larger number of refugees to a county. The second variable (military) is time-invariant and provides the number (0–5) of empty military installations (typically barracks) that could be used to house refugees. Again, counties with such unused military facilities can be expected to draw more refugees. We employ the variable military for the construction of 5 additional interaction series with our time fixed effects variable year.

Figure 4:

Refugee variable refg_infl interacted with year time fixed effects; analogous to Table 3, Mod. 2 (alternative to Table 4, Mod. 1, with log(refg_infl) replaced by year interactions, prm_12).

The response pattern of the elasticity values in Figure 4 is noticeably similar to that of Figure 2. The key take-away is that the percentage response of building permits to a given percentage rise in refugee inflows is negative after 2015 and increases in absolute terms over time. The response pattern of Figure 4 confirms the earlier result that building permits have become more sensitive to refugees after the refugee surge and that this sensitivity gained momentum over time instead of being of a temporary nature.

Model (1) of Table 4 gives the point estimate of β ₀ in Equation (2) in the absence of a coefficient β ₁. At −0.03, it represents the average elasticity of building permits with respect to refugees before, during, and after the refugee surge. It is not statistically different from zero at standard levels of confidence. Model (2) of Table 4 allows the building permit elasticity to vary before and after the refugee surge according to Equation (2). This results in two elasticity estimates, one for the time period from 2013 to 2015 and one for the time period from 2016 to 2018. The first of these elasticities is zero, and the second one has a value of −0.045 and is statistically significant at the 10 percent level. This value is only about half the elasticity estimate of the corresponding Model (4) in Table 3. One reason for this discrepancy is the shorter time horizon that is employed for the estimation of the models in Table 4; it does not allow for capturing the full momentum of the sensitivity of building permits to refugees shares.

Model (3) in Table 4 does not use any 0/1 indicator variables to separate between before and after the refugee shock of 2015/16. Instead, it instruments variable refg_infl with the number of ex ante predicted refugees, a variable designed for and implemented in another context by Lange and Sommerfeld (2024). The resulting elasticity estimate more than doubles in absolute size to −0.07 from a value of −0.03 in Model (1). It also switches from being statistically insignificant to being highly significant. The IV estimate of the refugee variable means that a doubling of the refugee inflow lowers standard building permits by about 7 % over the estimation period from 2013 to 2018. Comparing the estimate to those shown in Table 3, we see that the instrumentation in Model (3) makes the building permit elasticity very similar to the ones predicted by Models (3) and (4) of Table 3 for the time period after the refugee surge. The first-stage estimates of Model (4) in Table 4 confirm that the IV estimates are statistically acceptable in the sense that we do not have a weak instrument problem.^[9] The instrumental variable is statistically significant as reflected in the corresponding F-test statistic.

Models (4)–(6) of Table 4 replicate Models (1)–(3) for the refugee variable used for the estimates presented in Table 3. This makes it possible to assess what difference the choice of the refugee variable makes for the results.

The results of Model (4) confirm those of Model (1) that the average building permit elasticity with respect to refugees over the years 2013–2018 is not statistically different from zero at common levels of confidence. If one allows the elasticity to be different before and after 2016, as in Model (5), the elasticity value for the years 2016–2018 turns out to be −0.086, which is statistically highly significant and almost twice the value derived for Model (2). It is in fact close to the elasticity value generated in Model (3) by IV estimation. Model (6) shows the IV estimates that correspond to Model (3). The estimated elasticity with respect to the refugee share log(refg) is not different from zero statistically. From the first-stage results, we can see that this is the result of a week instrument problem. The corresponding F-statistic is not statistically significant at any common level.

Comparing the estimates from Models (1)–(3) to those of Models (4)–(6), we note that the definition of the refugee variable matters for the IV estimators. Outside of those, the results are rather similar. In addition, the IV estimates of Model (3) in Table 4 tend to be fully consistent with the OLS estimates shown in Table 3 if one is willing to assume that the refugee surge of 2015/16 played a critical role in the increased sensitivity of building permits to refugees after 2015.