Housing affordability sets us apart: The effect of rising housing prices on relocation behaviour

Method

The study focuses on two questions: First, how do rising local housing prices affect the probability of relocation? Second, how do rising prices affect the decision of which neighbourhoods households relocate to, regarding distance to the city centre and local unemployment rates? A difference-in-differences (DID) design with household fixed-effects is used to assess the behaviour of households over time. Local rent price increases resemble the continuous treatment dose. ⁵ Pre-treatment matching is applied to make treatment and control units more comparable with regard to the household and neighbourhood structure. In line with previous research, the cut-off points for low- and medium-incomes are the 40th and 80th income percentiles (see Pinkster et al., 2014). ⁶

For the first part, two situations are examined that usually put pressure on households to relocate. In particular, excessive housing cost burden, which occurs when a household is required to pay more than 40% of income on rent; and excessive space burden, which occurs when a household has fewer rooms than appropriate for the size of the household. ⁷ The difference in relocation probability of households in steady markets before and after these excessive housing burdens is compared with that of households in rising markets.

For the second part, the distance to the city centre and the local unemployment rate are compared before and after relocation. In a cross-sectional analysis, the correlation between local prices and outcomes might be spurious, as changes due to social mobility of existing residents and changes due to migration cannot be distinguished. Household fixed-effects incorporate all unobserved time-invariant changes at the household level that potentially alter both the variable of interest and the outcome.

The regression model can be written as follows:

Local unemployment rat e i , t , l = Relocatio n i , t + Housing market chang e i , l first + [ Relocatio n i , t * Housing market chang e i , l first ] + LifecourseControl s i , t + Household F E i + CountyF E i , l + ∈ i , t , l

(1)

with i indicating a household observed at year t and location l. The location l_first indicates the original location of a household, that is, the first housing and neighbourhood setting before any relocation. The set of variables LifecourseControls captures changes in the life course of a household over time, such as becoming widowed or having children. NUTS3 region fixed-effects account for different average levels of the outcome variable when moving between 402 counties.

Burden indicators are evaluated with respect the housing costs and space of the initial dwelling. In this way, the direction of the effect from outcome to cause is ruled out. Following Hainmueller et al. (2019) for applying multiplicative interaction models, marginal effect plots are shown to interpret the estimated results over the relevant range of local price changes. Histograms are attached to the plots to evaluate the common data. Estimates for low-, medium- and high-price changes (binning estimator) make it possible to assess the linearity assumption of the interaction effects.

Data

The SOEP is the largest household panel in Germany. It started in 1985 in West Germany and was extended to East Germany by 1990 (Wagner et al., 2007). It provides comprehensive information about the economic, residential and social situation of households and their members. For the analysis, 56,210 observations of 12,837 households are considered for the years 2007 to 2016. ⁸ Observations located in the top and bottom 1% of local price changes are removed to improve common data support and limit the influence of extreme values (2,229 observations). As home ownership requires extensive financial security and is often considered a once-in-a-lifetime decision in Germany, the analysis focuses on initial tenant households.

The postcode information of each household is used to link the SOEP to the regional housing market. ImmobilienScout24.de is the largest online market platform in Germany. As a market leader with around 1.2 million listings per month, ImmobilienScout24 covers the vast majority of housing offers in Germany. ⁹ For each listing, information on rental price, number of rooms, contact request, social housing status and location is aggregated at the postcode level.

Additional neighbourhood indicators such as the local unemployment rate are provided by MICROM GmbH (Goebel et al., 2014). By consolidating several public and private data sources, MICROM offers fine-grained information on neighbourhood structures. In cooperation with the SOEP, yearly MICROM variables are linked to SOEP household-year observations. Several variables are employed in the pre-treatment matching to better capture the neighbourhood contexts such as the share of migrants, neighbourhood status and the relocation volume in the fine-grained postcode-8 area. ¹⁰

Pre-treatment matching

In an ideal natural experiment, treatment is assigned randomly and does not depend on any pre-treatment variables nor the outcome. This is the case if the control and treatment groups differ only in their treatment assignment but are otherwise similar (Imbens and Wooldridge, 2009). However, this is often not the case with observational data. Matching is one method to reduce model dependency by balancing treatment and control groups on observed covariates (Rubin, 1973).

Although the DID research design with household fixed-effects mitigates the effect of group differences, they could still differ in more substantial terms. Nearest neighbour matching is performed to make treatment and control units more comparable with regard to household structure and initial neighbourhood context. Each household enters the matching sample in their first year of observation. A regression-based propensity score is estimated to select the best control match for each observation in the treatment group. Observations outside the support of the distance measure are discarded (15 households). The 70th percentile of the price change distribution (or a price increase of 3.9% per year) distinguishes rising and steady housing markets. For the main analysis, local price change is entered as a continuous variable. ¹¹

The results show an overall improvement in the comparability of control and treatment units. ¹² Before matching, households in rising (versus steady) housing markets experience higher average levels of housing cost burden and room stress. They are less often retired and widowed, and reside in areas closer to city centres with a higher share of migrants and purchasing power. After matching, the means and standard deviation of all life-course-related variables are not significantly different according to a two-sided t-test. Pre-treatment matching also removes control units in less expensive and less popular regions.

Descriptive results

Table 1 provides summary statistics of variables by income group and local housing demand.

OPEN IN VIEWER

Table 1.

Descriptive statistics.

	Min	Max	Mean	Low-income		Middle-income
				Steady market	Rising market	Steady market	Rising market
Excessive cost burden	0	1	0.16	0.25	0.3***	0.08***	0.1***
Excessive space burden	0	1	0.07	0.07	0.1**	0.06*	0.05***
Move	0	1	0.09	0.09	0.08***	0.09	0.09*
Distance to closest city (km)	0.9	74.3	13.65	15.2	12.39***	13.86***	11.58
Local unemployment rate	0	0.39	0.09	0.1	0.1	0.09***	0.09***
Income-to-rent	0	4.33	0.29	0.34	0.37***	0.25***	0.27***
Household income (Euro)	100	7500	1663.65	1221.24	1227.18	1961.51***	1963.54***
Housing costs (Euro)	1	3800	422.86	368.46	400.13***	443.32***	475.98***
Household size	1	13	2.05	2.1	2.07	2.04	1.98
Living space (m2)	8	400	71.46	65.43	65.33	76.1***	74.35***
Rooms per person	0.25	11	1.7	1.6	1.56	1.78***	1.78***
Without partner	0	1	0.56	0.63	0.67*	0.49***	0.53***
Widowed	0	1	0.12	0.11	0.09	0.14**	0.13**
Number of children	0	9	0.45	0.56	0.56	0.39***	0.38***
Number of relatives	0	9	0.02	0.03	0.04	0.02	0.01***
Unemployed	0	1	0.46	0.34	0.37	0.52***	0.55***
New employment	0	1	0.15	0.14	0.14	0.15*	0.15
New educational degree	0	1	0.01	0.01	0.01	0***	0.01
Handicapped	0	1	0.19	0.2	0.19	0.18**	0.18
Retired	0	1	0.28	0.3	0.27	0.28	0.26
Years since moved in	0	109	11.03	10.2	10	11.45***	12.04***
Homeowner	0	1	0.03	0.01	0.01	0.05***	0.04***
Social housing	0	1	0.06	0.11	0.1	0.03***	0.05***
Urban	0	1	0.7	0.68	0.78***	0.67	0.76
Median local rent (Euro)	138	1950	439.63	399.84	479.35***	426.03***	506.02***
Contact requests per offer	0	28.67	5.47	4.35	7.44***	4.83***	7.36
Offers – overall	0	1,302.6	20.32	18.18	23.69***	19.27	23.82
Offers – affordable	0	1176.62	8.07	5.36	3.99***	10.17***	10.23***
Offers – social housing	0	32.38	0.46	0.45	0.56**	0.43	0.49
Observations	56,210	56,210	56,210	18,344	6745	21,998	9123
Households				4703	1908	4295	1931

Notes: Low income is defined as the lowest 40% percentile in income distribution; middle income is defined as 40% to 80% percentile in income distribution. Housing market variables are based on the first observed postcode area a household has resided in. Affordable housing offers denote rental listings that require less than 30% of a respective household’s income on rent and provide sufficient rooms for household members. Housing offers are normalised to offers per 1,000 inhabitants in the respective area.

Adjusted Wald test with *p < 0.10, **p < 0.05, ***p < 0.01 comparing against the respective lower income group.

Data source: SOEP v33 2007–2016.

In rising housing markets, low-income tenants are more likely to suffer excessive housing costs and space burdens than middle-income tenants by 20 and 5 percentage points, respectively. At the same time, they face greater room stress as on average they reside in smaller dwellings and with more household members. Regarding neighbourhood context, they live in neighbourhoods with a higher average unemployment rate. The differences between income groups still exist for households in steady housing markets but are much less pronounced.

Households in rising housing markets are primarily located in urban regions and in neighbourhoods with a higher average unemployment rate and that are closer to the city centre. Around 70% of households in rising markets live in urban regions. As 67% of households in steady markets live in urban areas, rising prices neither apply to all neighbourhoods of a city nor to all cities in general. Regarding the tenure structure in those markets, 10% of low-income households reside in subsidised housing while this is the case for only 4% of middle-income households.

Differences in life course contexts exist mainly between but not within income groups. For example, independent of whether households reside in steady or rising housing markets, the number of children, cohabitation, level of handicap or retirement status do not show differences within low-income samples.

Rising housing markets and the probability to relocate

Figure 1 visualises the marginal effects of excessive cost and space burden on the probability of relocation, moderated by local rent price increases. ¹³ The underlying regression models include household and county fixed-effects as well as life course control variables. The 95% confidence intervals are displayed in grey. The histograms at the bottom of Figure 1 indicate common data support for price changes between −5% and +15% for movers (grey histogram) and stayers (white histogram).

OPEN IN VIEWER

Figure 1.

Marginal effects plot – excessive housing burden.

In the case of zero price changes, excessive costs and space burdens relate to a similarly higher probability of relocation for both income groups. In other words, households move if they face too little space or too high costs. For an excessive cost burden, rising local prices dampen the probability of relocation for both income groups. However, low-income households are much more strongly affected by this dampening effect: according to the estimates, for them a 10% local price increase per year makes relocation less likely by −28.5 percentage points. For middle-income households, this effect is much lower at around 10 percentage points but not significant in the full model. Binning estimates for the lowest-, middle- and highest-third of price increases indicate a U-shaped rather than a linear effect (white points and 95% confidence interval, Figure 1). Households in markets with very little or even negative price changes might be eager to move away. The marginal effect becomes stronger again in the top rising markets, particularly among middle-income households.

Excessive space burden has a positive effect on relocation for both income groups. For low-income households, the regional housing market barely alters the strength of the effect. Middle-income households are much more likely to move when prices increase. As described above, household units are selected based on propensity score matching. Regression results for the unmatched sample (see Appendix A2, available online) show much stronger effect sizes. This suggests structural differences in the treatment and control group that are mitigated by propensity score matching.

Rising housing markets and relocation outcomes

Figure 2 visualises the marginal effects of moving on distance to the closest city centre (top) and on the local unemployment rate (bottom) for low- and middle-income households. There is strong evidence from the regression estimates that low-income movers relocate further away from the city centre. If local prices increase by 10% per year, low-income movers end up in neighbourhoods that are on average 1.5 km further away. The effect for middle-income households is not significantly moderated by local rental price changes.

OPEN IN VIEWER

Figure 2.

Marginal effects plot – moving.

The estimates support the argument that rising housing markets foster social segregation through selective mobility. Accordingly, there is significant indication that low-income households end up in neighbourhoods with higher unemployment rates when local housing markets are rising. If prices increase by 10% per year, low-income households move to neighbourhoods where the local unemployment rate is 60 percentage points higher than before.

Again, estimation based on the unmatched sample revealed a much stronger effect size (see Appendix A2, available online). Comparing affected households with all other households seems to overestimate the effect as it does not account for structurally different household contexts. Propensity score matching helped to mitigate selection effects.

The hidden costs of rising housing markets

Access to affordable housing defines in what kind of dwelling and neighbourhood a household can reside. At the same time, where people live has lasting effects on their life chances (Chetty et al., 2016; Stephens and Leishman, 2017). Therefore, dynamics in the housing market might enhance spatial inequalities (Pawson and Herath, 2015). The present study shows that rising housing markets affect the residential choice of low- and middle-income households differently. In times of rising local housing prices, low-income households become less likely to relocate when faced with excessive housing burdens.

This has several implications. First, lower mobility means a potential misallocation of housing. Households are likely to remain in housing that is too small or too large if moving is costly. The aggregated effect would be to restrict supply and raise prices even further. Reduced mobility would affect not only vulnerable households but would also extend to other income groups. There are early signs that this is the case in Germany. Mobility rates in the 20 largest cities have declined from 15% to 8.6% in the last 7 years. ¹⁴ Second, the lack of affordable housing can change where households search for new jobs. Low-skilled labour might be prevented from moving into regions with high labour demand as housing costs are too high. Past research shows that limitations in labour mobility have substantial adverse effects on aggregated productivity (Glaeser and Gyourko, 2018; Hsieh and Moretti, 2017). Third, residing in low-price housing segments on the outskirts of the city imposes additional costs on their residents. Longer commuting times have implications for fuel and car maintenance costs as well as intangible costs such as less time for recreation or time spent with the family (Belsky et al., 2005). If such neighbourhoods correlate with high crime rates and levels of pollution, residents are also exposed to additional health risks (Galster, 2012). Fourth, a social-spatial concentration at the periphery creates costs for the municipality (Galster et al., 2008). Direct public costs occur to maintain local infrastructure and transport systems that allow access to city centres. Potential indirect costs arise if educational and career opportunities of residents are limited. This would require additional social welfare transfers and lower income tax in the long run.

If preventing social residential segregation is a political goal, creating affordable and accessible housing options for mixed-income households is critical. According to the OECD, however, a reverse trend can be observed:

[…] rising socio-economic inequalities are gentrifying and dividing European and North American cities and making adequate housing increasingly unaffordable for low- and middle-income households. A near unanimous belief in homeownership has drastically reduced rental housing stock which is a vital tenure option for many households […] Governments in European and North American countries have largely retreated from providing ‘social’ housing in favour of ‘enabling the market’, yet the market has not provided land and housing that is affordable to low-income households. (UN Habitat, 2011: xii)

While subsidised housing requires proof of eligibility based on income, this is not the case for other types of housing. An increase in private housing supply alone is not necessarily successful in providing affordable housing when demand is high (Fingleton et al., 2018; van den Nouwelant et al., 2016: 161). Better-off households offer more steady employment and financial conditions which make them more attractive for landlords. A more sustainable approach requires the extension of housing supply to be accompanied by additional measures. For example, inclusionary policies link the construction of market-rate housing to the creation of affordable homes for less privileged households (Thurber et al., 2018). In several countries such as Germany and France, it is a requirement of any new development project that a share of social housing should be provided. A relaxation in building requirements and additional rent cap regulations are frequently discussed in German policy debates (Hendricks, 2016).

Robustness checks

Some limitations of this research need to be pointed out. Potential concerns could arise regarding the measurement of excessive housing burden. First, the indicators enter the regression in a binary fashion with respect to official thresholds for excessive housing cost and space burden. For example, the threshold of excessive cost burden is 40% of income spent on rent. Changes in the rent-to-income variable from 39% to 41%, therefore, enter the regression in the same way as changes from 20% to 60%. The perceived burden is arguably much greater in the latter case. In a sensitivity analysis, housing cost burden and room stress enter as continuous variables. ¹⁵ The results are in line with the main findings.

Second, changes in the intensity of housing cost burden are measured similarly across income groups. As pointed out by Nepal et al. (2010), middle-income households might be willing to spend more on amenities without perceiving it as an additional burden. The findings are in line with the criticism and underline that low-income households are a much more vulnerable group particularly in rising housing markets.

Third, the anticipation of a future burden might lead to relocations before the actual event. For example, households who plan to have another child might relocate to a larger dwelling some years in advance. If anticipation effects are randomly distributed across households in steady and rising markets, estimates in the DID design would not be biased. However, if anticipation effects are stronger in rising markets, the estimates would indicate that there would be less relocation resulting from excessive burden though relocation took place before. Figure 1 suggests the opposite as anticipation effects seem more prevalent in steady housing markets. In a sensitivity regression, the indicators for excessive housing burden are lagged by 1 year. The estimates are in line with the results presented here.

Another concern regards the use of price data from the largest property portal in Germany as it does not necessarily capture the entire housing market. A share of housing offers might appear on other platforms or marketing channels. If such marketing channels are used randomly across regions and price segments, effects would still be estimated consistently. If they are more popular among affordable housing offers in rising housing markets, however, the moderator variable would overstate the actual local rental price change. For the years 2017 to 2018, price increases could be computed based on the three major online real estate portals in Germany. However, results from the sensitivity regression are in line with the main findings.

In the analysis, the treatment dose is based on the median price of a postcode region. Changes in the median could originate from price changes of similar apartments or from changes in the housing stock offered. It is argued that the origin of price changes is not crucial here as they reflect the de facto supply of housing. Even if it would be more accurate to extract the conditional price development from hedonic prices, movers still rely on what is offered in the market.

Another concern regards the exogeneity of the local housing market development. Households may relocate to neighbourhoods because of preferences for particular neighbourhood amenities. If the same amenities make future rental price increases more likely, the estimated interaction effect would be biased. Pre-treatment matching is applied to make selected households in steady and rising housing markets more comparable. The development of the housing market is generally beyond the control of single residents. However, the in-movement of households with certain dispositions that trigger price increases can still not be entirely ruled out. In a sensitivity regression, only households who have lived at the same address for 3 years or longer are considered. The results are in line with the main findings.