Page 1
Degree Thesis
Master’s level (Second cycle)
COVID-19 and structural breaks
The case of the Swedish Housing Market Author: Olle Rönningsberg and Sander ten Hove
Supervisor: Charlie Lindgren
Examiner: Moudud Alam
Subject/main field of study: Micro data analysis
Course code: MI4002
Credits: 15
Date of examination: 2021-06-09
At Dalarna University it is possible to publish the student thesis in full text in DiVA. The
publishing is Open Access, which means the work will be freely accessible to read and
download on the internet. This will significantly increase the dissemination and visibility
of the student thesis.
Open Access is becoming the standard route for spreading scientific and academic
information on the internet. Dalarna University recommends that both researchers as well
as students publish their work Open Access.
I give my/we give our consent for full text publishing (freely accessible on the internet,
Open Access):
Yes ☒ No ☐
Dalarna University – SE-791 88 Falun – Phone +4623-77 8
Page 2
COVID-19 and structural breaks:
The case of the Swedish Housing Market
Olle Rönningberga,*, Sander ten Hovea, Charlie Lindgrena
aDalarna University, 791 88 Falun, Sweden
ABSTRACT: This paper analyzes how the COVID-19 pandemic has affected the Swedish
housing market, and in particular prices and shifts in trends. Different classes of housing objects
in various counties are investigated. Combining web scraped housing data for the entirety of
Sweden between 2016-01-01 and 2021-03-31, including economic, demographic,
socioeconomic and locational data, a hedonic regression model is used to estimate how different
variables influence the housing price. The base model is subsequently used to investigate if
statistically significant structural breaks exist in relation to the COVID-19 pandemic for the
different object types in the entire Swedish market and in certain specific counties. Structural
breaks are found for the housing object types ‘Fritidshus’, ‘Lägenhet’ and ‘Radhus’ in the entire
Swedish market and for “Villa” in Stockholm county shortly after the pandemic outbreak,
suggesting there is evidence for a pandemic infused shift in housing price regime on the
Swedish housing market for these object types in stated areas. Splitting the hedonic regression
model into three, one pooled regression, one before and one after the identified breaks, and
comparing the shifts in impact of the housing price determinants suggests different pandemic
effects on different object types. The result indicates that for the object types ‘Lägenhet’ in the
entire country and for ‘Villa’ in Stockholm county, living area has an increased impact on the
price while the locational variable population density has a decreased impact after the
breakpoint date compared to before. This could suggest that for permanent housing objects in
these regions, living area might have become increasingly valued over location during the
pandemic. The results further indicate the direct opposite effect on the shifted impact in living
area and the population density for the price of the temporary housing type Fritidshus in entire
Sweden. However, an indication for increased impact of the areas socioeconomic level is noted
for all these three object types. These results hold as a ground for further research in the subject.
Keywords: COVID-19, structural breaks, hedonic regression, housing market, Chow test.
*Corresponding author: [email protected]
Page 3
1. Introduction
The Swedish housing market has been a hot topic to debate, especially since the financial crises
of 1990 – 1994 and more recently 2008 – 2009. Since then, the Swedish household debt in
relation to disposable income has increased yearly, while the household interest expenditure
continuously decreased, following prolonged low interest rates (Swedish Central Bank, 2020).
This has led to soaring housing prices and subsequent speculations of an imminent bubble
(Claussen, 2013).
Entering 2020, the COVID-19 pandemic struck. The worldwide spread has had prolonged
effects on both social life and different market dynamics. To prevent an expected negative effect
with respects to the housing market, Sweden made a temporary easing on amortization on new
and old mortgages (Finansinspektionen, 2020). Now, over one year into the pandemic, it has
become possible to evaluate potential pandemic effects on the Swedish housing market. How
have trading volumes and prices affected different phases and regions? Do the effects differ
depending on type of property and location within regions?
This paper will focus on the interactions between housing prices and economic,
demographic, socioeconomic and locational variables. We will investigate how the COVID-19
outbreak has influenced the importance of different variables related to the housing price, and
if there is evidence for structural breaks related to the pandemic and the Swedish housing
market. This is made possible using the hedonic regression and the structural break test.
Our results show that for different object types, breakpoints are found at different dates,
varying between counties. Relating this to COVID-19, many breakpoints reside in the summer
of 2020 short after the pandemic outbreak. Comparing the shifted impact of housing price
determinants in split and pooled hedonic regressions indicates that the impact of certain house
price determinants has changed in different ways for different object types.
The paper is structured as follows: Section 2 will give a theoretical background on housing
price determinants as well as COVID-19 in relation to the housing market. Following this,
Section 3 accounts for the method used, i.e., primarily the hedonic regression model and
structural breaks, with statistical tests for the latter. Then, Section 4 proceeds with the empirical
analysis. Finally, Section 5 concludes with a discussion regarding the results of the study, as
well as suggestions for future research.
Page 4
2. Theoretical background
2.1 Housing price determinants
Variables from both the supply and demand side influence prices on the housing market. These
variables come from micro as well as macro-level. At the micro-level, individual property
characteristics and locational attributes are important, while at the macro-level economic and
demographic indicators like interest rates and population growth come into play with potential
effects on both the supply and demand side of the housing market (Galati et al., 2012;
Olanrewaju et al., 2016).
Extensive research has previously evaluated numerous economic and demographic
variables that could affect prices and transaction volumes on the housing market. In Sweden,
rising disposable income and falling mortgage rates correlated strongly with the rising housing
prices by increasing demand for residential properties (Claussen, 2013). A comprehensive study
of 15 Organization for Economic Co-operation and Development (OECD) countries concluded
gross domestic product (GDP) growth rate and real interest rates affect changes in housing
prices (Englund & Ioannides, 1997). Although real interest rates provide prediction power in
the long run, nominal interest rates are suggested to influence housing prices in the short run
(Turk, 2015). Another economic variable consistently related to housing prices is the household
debt to income ratio. A study on Swedish housing prices during the years 1970 – 1997 verifies
the long-term relationship between housing prices and debt to income ratio (Barot, 2001). The
same study shows the demand side of the housing market to have a short-term relationship with
employment rates and population growth.
While real economic variables are affecting housing prices, the household’s expectations
are also influential. For example, the expectation of future interest rates could affect the amount
a household will borrow, which would affect the demand side of the housing market (Oikarinen,
2009). Looking at the macro-perspective beyond the economic factors, population growth both
in general and in certain age cohorts, was related to development in housing prices both
internationally and in the European market (Girouard, N., 2005; X. R. Wang et al., 2017)
It is also well established that inherent characteristics, socioeconomic and locational
attributes are major determinants of the pricing of a residential property. Not surprisingly,
characteristics like the size of the property correlate positively with property prices (Garrod &
Willis, 1992; Li & Brown, 1980; Linneman, 1980). However, different inherent characteristics
aside, a residential property is immobile as such, leading the location to influence the valuation
Page 5
heavily (Kiel & Zabel, 2008). Previous studies have found locational demographic and
socioeconomic factors like unemployment, household income and educational level to correlate
with housing prices (García & Hernández, 2004; Jumadi, 2010).
2.2 COVID – 19 and the housing market
History shows that pandemics affect housing prices. For example, during the SARS pandemic
in 2003, where a virus was transmitted from Mainland China to Hong Kong, the number of
house sales decreased during the pandemic, March 15, 2003 – June 23, 2003. At the same time,
the average price declined by 1 – 3 percent. However, the market recovered quickly afterwards
(Wong, 2008). The current COVID – 19 pandemic is not yet over, despite current vaccination
efforts, and it has had a far heavier impact on social lives than previous pandemics (Helsingen
et al., 2020).
On March 13,2020, the United States of America declared a COVID-19 related national
emergency. After a consecutive sales increase of 10 – 20% in 2019, the number of houses sold
decreased at the start of the national emergency with 45% (Nuredini, 2020). However, in the
same paper states that until April 2020, the value of houses sold remained steady, which
contrasts from the most recent economic crisis. Wang (2021), who analyzed the impact of
housing prices in four different U.S. cities, found that only one city faced a decrease in average
housing prices. The prices in the other cities were actually increasing. The author concludes
that COVID-19 has a severe impact on local housing markets where the local economy counts
on tourism, service, and aviation.
Elaborating on the U.S. housing market, Zhao (2020) concludes that between April and
August 2020, the growth rate of the median price per square foot has increased quicker than
any other four-month period before the global financial crisis 2008 – 2009. This shows that the
prices are inclining further in the pandemic’s timeline. Furthermore, Liu and Su (2020) mention
that in 2020 the demand for houses near central cities and highly populated areas has, compared
to 2019, decreased in the U.S.. They suggest one cause is the raised adaption of teleworking
and reduced appraisal of consumption amenities.
Furthermore, Hu et al. (2021) analyzed the Australian housing market regarding COVID-
19 on two levels. Their analysis focused on two clusters of factors, epidemiological and policy
interventions. Whereas epidemiological factors include current and new COVID-19 cases,
embraces policy factors measurements like lockdowns, quarantines and social distancing. The
Page 6
researchers unveiled that the policy interventions show a minor impact on the housing market.
On the other hand, the number of positive cases significantly affects the Australian market.
Ahonemie and Putkuri (2020) analyzed survey data regarding Nordic banking markets and
the COVID-19 pandemic and concluded that the customer’s confidence in the economy has
weakened, and intents to start a loan or buy a property fell down during the pandemic.
Moreover, respondents assumed house prices will decrease over the period May 2020 until May
2021.
Recent research states that shutdown orders affect consumer behavior and have a negative
effect on businesses that rely on human interactions (Koren & Pető, 2020). Since Sweden did
not apply strict shutdown orders, government restrictions could have less effect. Still, Google’s
mobility report (Aktay et al., 2020), which measured how the visit rate for specific places has
changed during the pandemic, shows significant changes in people’s mobility behavior. As seen
in Table 1, the number of people around public transport has decreased with 44% compared to
a baseline non-pandemic times, which is determined as January 3 until February 6 2020
(Google, 2021). Visits to retail and recreation by 28% and to workplaces a significant decrease
by 31%. So, even without the strict shutdown polices, it is noticeable that the altered mobility
behavior in Sweden is similar to the U.S. in certain regards.
Table 1. Google Mobility Report for Sweden and the U.S., 02-02-2021.
Category Sweden U.S
Retail & recreation -28% -26%
Grocery & pharmacy -12% -15%
Parks +5% -25%
Transit stations -44% -45%
Workplaces -31% -36%
Residential +11% +13%
Page 7
3. Method
3.1 Hedonic pricing model
The hedonic price model has a long history in science as the method to examine various factors
affecting price in the housing market. The theoretical background of the model, that constitutes
of a linear regression model, aims at ascribing prices to attributes of a good by deriving the
attributes marginal contribution to the observed final price, stem from consumer theory by
Lancaster (1966) and Rosen (1974). According to the model it is therefore the attributes that
the consumer values and derives utility from, rather than the good itself. While the model
postulated by Lancaster views goods as members of a group that can be consumed together,
Rosen presumes a range of goods that are not consumed together. With the durable nature of
the housing market, Rosen’s interpretation implying housing prices as determined by a
collection of hedonic values is the more applicable. Estimating the marginal price contribution
of attributes is possible by regressing the sales price against the various variables assumed to
affect the housing price. The hedonic price model further assumes homogeneity of the housing
goods. The homogeneity of housings are however questionable, as housing typically differs in
terms of type of object and locational attributes (Chau & Chin, 2003).
3.2 Structural break model
When studying observation for one or more variables over a certain time span, it is common to
look at time series for trend analysis or to detect certain patterns or shift in the variables. On a
macro-economic level, one can speak of a structural break when the time series data shifts
unpredictably (Luitel & Mahar, 2015). To detect structural breaks is important in research and
to assess whether a structural break exists, the Chow test is used often, which is a method that
determines if identical variables in different datasets are related enough to be grouped as one
(Chow, 1960). Looking at a structural break at a standard linear regression, considering the
model
𝑌𝑖 = 𝑥𝑖𝛽𝑖 + 𝑢𝑖 (𝑖 = 1, 2, . . . , 𝑛) (1)
Page 8
where 𝑖 represents time, 𝑌𝑖 is the outcome of the dependent variable (in this situation the sold
price), 𝑥𝑖 represents a vector of regressors, the independent variables, and 𝛽𝑖 is a vector of
coefficients. Finding a structural break means that the null hypothesis,
𝐻0 ∶ 𝛽𝑖 = 𝛽0 (𝑖 = 1, 2, … , 𝑛)
(2)
that the regression coefficients remain constant, can be rejected. In various cases it is assumable
there are 𝑚 breakpoints, which means that the coefficients shift to another regression. If 𝑅𝑆𝑆
denotes the residual sum of squares of a regression over a full period and 𝑅𝑆𝑆1 the sum of
squared residuals for a regression over 𝑡 = 1, … , 𝑇’ and 𝑅𝑆𝑆2 the sum of squared residuals for
a regression over 𝑡 = 𝑇’ + 1, … , 𝑇, where 𝑇’ denotes the number of observations before the
breakpoint. The Chow test statistic is then given by
𝐹 =
𝑅𝑆𝑆 − (𝑅𝑆𝑆1 + 𝑅𝑆𝑆2)
(𝑅𝑆𝑆1 + 𝑅𝑆𝑆2)∗
𝑇 − 2𝑘
𝑘
(3)
and follows F-distribution with k and T-2k degrees of freedom. An obvious limitation to the
original Chow test is that the structural break point has to be known in advance to be tested.
To solve the problem of testing for structural break without knowing a breakpoint in advance,
different test methods have been put forward. The Quandt-Andrews (1993; 1960) Sup-𝐹
statistic test for parameter stability can test against, without providing timing, a single-shift
alternative by getting the largest F statistic over all potential breakpoint dates. The test is built
on a series of 𝐹 statistics for a change at 𝑖. The Ordinary Least Squares residuals (𝑂𝐿𝑆𝑖) from
a regression with breakpoint, one regression before and after the breakpoint, are compared to
the residuals from a full regression without breakpoints (𝑂𝐿𝑆), e.g., unsegmented regression
(Andrews & Ploberger, 1994). This yield 𝐹 statistics for i = nh, . . . , n − nh(nh ≥ k) and rejects
𝐻0 if Sup-𝐹 gets too large by
𝐹𝑖 =
𝑂𝐿𝑆𝑇𝑂𝐿𝑆 − 𝑂𝐿𝑆𝑖𝑇𝑂𝐿𝑆𝑖
𝑂𝐿𝑆𝑖𝑇𝑂𝐿𝑆𝑖 (𝑛 − 2𝑘)⁄
(4)
The approach suggested by Zeileis et al. (2003) provides a way to determine 𝐹 statistics,
rejecting 𝐻0 if the supreme functional is too large. By applying a regression model of the time
Page 9
series of OLS against the sold price, the 𝐹 statistics will be computed. In case there is proof for
a structural change in the model, the breakpoint is added to the model.
3.3 Software
In order to test for structural break, the package in the R statistical software strucchange is
applied (Zeileis et al., 2002). A sequence of 𝐹 statistics is computed by the function 𝐹𝑠𝑡𝑎𝑡𝑠
𝐹𝑠𝑡𝑎𝑡𝑠(𝑓𝑜𝑟𝑚𝑢𝑙𝑎, 𝑑𝑎𝑡𝑎, 𝑐𝑜𝑣. 𝑡𝑦𝑝𝑒, 𝑓𝑟𝑜𝑚 = 0.15, . . . )
where the resulting object is a sequence of 𝐹 statistics. The sequence time series is plotted with
boundaries and tested for significance by sup-𝐹. With the maximum 𝐹 statistics of the sequence
corresponding to the minimum OLS estimator of a 2-segmented breakpoint partition, the
breakpoint is extracted from the resulting 𝐹𝑠𝑡𝑎𝑡𝑠 object by
𝑏𝑟𝑒𝑎𝑘𝑝𝑜𝑖𝑛𝑡𝑠(𝑓𝑜𝑟𝑚𝑢𝑙𝑎, 𝑑𝑎𝑡𝑎, 𝑏𝑟𝑒𝑎𝑘𝑠, ℎ = 0.15, . . . )
that calculates the optimal segmentation from a triangular 𝑅𝑆𝑆 matrix and yields the number of
breakpoints set by the parameter break. Where evidence for structural change in the regression
exists, the breakpoint is added to the 𝐹𝑠𝑡𝑎𝑡𝑠 sequence and plotted for display. To confirm that
found the breakpoint is optimal we use the breakpoint and run the original Chow test by
𝑠𝑐𝑡𝑒𝑠𝑡(𝑚𝑜𝑑𝑒𝑙, 𝑡𝑦𝑝𝑒 = 𝐶ℎ𝑜𝑤 , 𝑝𝑜𝑖𝑛𝑡)
4. Empirical analysis
4.1 Data collection
The property sales data for Sweden is collected by web-scraping Booli.se. This website holds
the largest record of housing sales prices in Sweden, and stores information like property type,
size, number of rooms, asked and sold price, which comes directly from the Swedish brokers’
administrative systems (Booli.se, n.d.). The initial collected dataset contained data points from
May 2000 until March 2021. However, the data used in this paper is sampled by using the years
2016 and onwards. Although somewhat arbitrary, the intent is to cover a time period with more
Page 10
comparable market conditions regarding the pandemic period, which had its inception as of
March 2020, while omitting more turbulent events such as the 2008-2009 financial crisis.
Furthermore, only the Swedish object types ‘Villa’, ‘Lägenhet’, ‘Fritidshus’, ‘Radhus’ and
‘Kedjehus’ were used, which still represent the vast majority of common living
accommodations in Sweden. Translations for these object types into English, including
translations for other Swedish terminology used later in the paper, can be found in Appendix
A2.
Although the address was included in the scraped data, the city was missing. Therefore,
the longitude and latitude coordinates were used to retrieve the city for each record. These
coordinates were fetched in a function from the R package “Revgeo”. This library is able to
return specifics as house numbers, postal codes, streets, cities, countries based on reverse
geocoding. Each sold house is attached to a particular DeSo (Demographic Statistics Areas)
region. Using geographical information system QGIS, the location coordinates were utilized to
retrieve what DeSo area a property belongs to. This region can be linked to a range of
demographical datasets registered by Statistics Sweden and therefore the housing data is
enriched with demographical and socioeconomic variables derived from the corresponding
DeSo area. Each DeSo area has a unique code of 9 characters. The first four represent the län
(county) and kommun (municipality). The fifth position is a letter: A, B, C. These letters groups
statistical areas in three different categories. Category A symbolizes zones outside urban areas,
also known as countryside. B represents urban areas outside the municipality’s central city.
Category C is in the municipality’s central city. Category A represent 18% of the DeSo areas,
B 10% and C 72% (Statistics Sweden, n.d.). The sixth, seventh and eighth position are for
sorting geographically, from south to north. The ninth and last position is a placeholder, in case
of a future area split. DeSo data includes, for each individual area; the percentage of population
with higher education, percentage with risk of poverty, percentage of dwellings that are rentals,
percentage of household with children, percentage of population gainfully employed and lastly
the square meter per person in the area.
Statistics Sweden (SCB) was used to retrieve variables such as population,
unemployment, interest rates, income, GDP and consumer price index with fixed interest rates
(CPIF). Data concerning household’s expectation on the economy is gathered from the National
Institute of Economic Research (NIER) which is also used as a measure to calculate households’
average financial situation. Table A1 in the appendix describes the complete list of collected
housing data and variables with acronyms later used in the paper.
Page 11
4.2 Data cleaning, visualization and descriptive statistics
The full dataset contained 865 438 observations on sold properties of 5 different object types.
After cleaning the data, removing observations with missing values for sold price and the
number of rooms, the remaining dataset contained 859 732 observations.
To get a better understanding of the dataset, summary and descriptive statistics are
presented in Table 2. Table 3 further provides overview of the housing data by summarizing
the dataset by object type. As table 2 shows, the average property is offered for 30.9 days on
Booli.se, has 3 rooms and has 93.4 square meters as living area. The mean list price on Booli.se
is 2 734 030 Swedish kronor (SEK), while the actual paid price is 2 926 288 SEK, with an
average difference of 7,1%. The mean price per square meter is 35 027 SEK. Note that all prices
are adjusted according to the CPIF defined by Statistics Sweden. The original sold price per
square meters originally came from Booli.se. However, evaluating the data unveiled some
irregularities, and was therefore recalculated by dividing the price over the property size. This
resulted in higher data quality.
What stands out in table 2 is that there are some irregularities in the numerical data. For
example, the lowest recorded price per square meter is 3 SEK, and also the highest price per
square meter (10 800 000 SEK) seems ambiguous. Before heading into the breakpoint analysis
and hedonic regression modeling later, the observations were studied more closely, and outlier
removal was conducted. This eliminated the top and bottom 10% of the sales prices and living
area resulting in a better, while still not optimal, model fit. First, the outlier rate was set at 5%
(Chen et al., 2017). However, after applying trial and error, the rate was set at 10% to obtain a
better model fit. Although a higher rate could have resulted in a better model fit, 10% was set
as the threshold to keep the dataset large and representable to the market. Since the data on
prices are highly skewed, to be able to use the price in the linear hedonic model, the sale prices
are transformed into the logarithm of the sale price, resulting in a Gaussian distribution shown
in the right graph of Figure 1 (Olofsson & Andersson, 2012).
As Figure 2 shows, the dataset contains property data from the entire country. Due to
the fact that the northern part is less populated, a smaller amount of property data is available
from this area. Displaying the Borlänge and Falun centers and their vicinity, it becomes visible
how properties are labeled into the different DeSo areas. Borlänge and Falun are both red (label
‘C’) other urban areas in both municipalities are in ‘B’ (yellow) and the rest in category ‘A’
marked in blue. Figure 3 shows the number of properties sold over the years. It is apparent that
object types ‘Lägenhet’ and ‘Villa’ are the most popular properties as well as there are more
Page 12
apartments sold for every year studied. Please note that 2021 only contains data until the 31st
of March.
Table 2. Summary statistics of web scraped housing data from Booli.se.
Variable Mean Min Max 1st Qu 3rd Qu
Days active 30.9 0 3 307 13 29
Living area (m2) 93.4 1 2 852 59 120
Rooms 3 1 56.5 2 4
List price (SEK) 2 734 030 2 65 629 800 1 348 484 3 470 422
Sold Price (SEK) 2 926 288 1011 145 692 000 1350750 3 744 400
Difference list/sold 195 368 -23 330 160 31 600 516 0 332 630
Price per m2 35027 3 10 800 000 16017 46 064
Table 3. Summary statistics per object type.
Type Houses
sold
Average list price
(SEK)
Average sold price
(SEK)
Average price
per M₂
Fritidshus 17 448 1 452 307 1 598 683 27 691.13
Kedjehus 35 348 3 309 115 3 386 318 27 878.63
Lägenhet 397 823 2 555 815 2 783 674 44 882.92
Radhus 49 439 3 222 074 3 372 894 30 061.49
Villa 359 674 2 992 901 3 039 636 25 936.93
Figure 1. CPI adjusted price and logarithm CPI adjusted sales price
Page 13
Figure 2. Property locations, including DeSo areas, for Sweden as well as Borlänge and
Falun municipalities.
Figure 3. Sold properties per type and year.
Page 14
Figure 4 provides a more detailed view on the volume change per object type, by displaying the
amount of sold properties by month compared to corresponding month in previous year. The
figure includes a break line for the day (11th of March 2020) the World Health Organization
declared the pandemic to provide perspective on where a potential change on the market is
hypothesized to have occurred. The volumes for object type ‘Fritidshus’ show a steep increase
shortly after the pandemic outbreak, potentially indicating an increasing demand for this object
type. This could stem from people wanting an extra home outside the urban areas, as remote
working seems here to stay (Bartik et al., 2020). The other object types also show some
volatility regarding volumes after the pandemic outbreak, but not as apparent as object type
‘Fritidshus’. When looking at the median square meter price per object type in Figure 5, every
object type already showed an increase in price before the pandemic. However, prices are rising
quickly after the 11th of March in 2020, especially for object type ‘Fritidshus’, indicating that
the increased trading volumes might be reflected in prices.
Figure 4. Volume change in %, from previous year and month, in amount of sold objects per
month by type.
Page 15
Figure 5. Average price per square meter per object type.
4.3 Structural break analysis
As the hedonic regression model assumes homogeneity of the properties analyzed, the data is
split on object types when applying the regression model and estimating for structural breaks.
On the subset, we estimate the cointegrated residuals by OLS from the multiple linear
regression model. In this model, the logarithm of the sale price acts as the dependent variable
and the house price determinant variables as regressors. To test for structural change, the
cointegrated residuals are used as regressors for the logarithmic sale price. Treating the
breakpoint date as an unknown parameter, all potential breakpoint dates are evaluated
individually for this model by sequential 𝐹-tests. As the 𝐹 statistic crosses the boundary
threshold, there is evidence for a structural break at 5% significance level. The time series of 𝐹
statistics is utilized to estimate breakpoints above the boundary. As several breakpoints exist
above the boundary and we are interested in when the single most optimal breakpoint is located,
this point is dated to where the 𝐹 statistics is the highest. The estimated breakpoint is displayed,
together with 𝐹- and 𝑝-values for the Chow test on the full regression model evaluated with a
break on this suggested point in Table 4. Looking at this table, the breakpoint analysis suggests
Page 16
that a structural break for object type ‘Fritidshus’, with minor deviation in dates, took place
during the summer of 2020 for Sweden as a whole, and for every county studied. We also note
structural break for object types ‘Lägenhet’ and ‘Radhus’ in the entire Swedish market and for
object type ‘Villa’ in Stockholm county. As object types ‘Lägenhet’ and ‘Villa’ are prominent
object types and the ones with most observations in our data, we will look investigate these
breakpoints further in section 4.3.1 together with the breakpoints for object type ‘Fritidshus’ as
this object types shows structural breaks shortly after the pandemic outbreak in every county
studied.
Table 4: Breakpoints per object type and county.
Object
type
Län Single
breakpoint
F-stat Chow
test on BP
P-stat Chow
test on BP
‘Fritidshus’ Sweden 2020-08-01 498,82 2.2e-16
Stockholm 2020-08-27 92,06 2.2e-16
Dalarna 2020-08-28 41,48 2.2e-16
Skåne 2020-07-31 47,40 2.2e-16
Västra Götaland 2020-06-30 60,45 2.2e-16
Lägenhet Sweden 2020-08-12 1211 2.2e-16
Stockholm 2019-08-22 730,12 2.2e-16
Dalarna 2017-08-22 51,36 2.2e-16
Skåne 2018-08-31 942,49 2.2e-16
Västra Götaland 2017-02-14 488,79 2.2e-16
Radhus Sweden 2020-06-01 449,85 2.2e-16
Stockholm 2020-08-20 247,87 2.2e-16
Dalarnas 2018-10-08 8,18 3.188e-4
Skåne 2020-05-15 104,45 2.2e-16
Västra Götaland 2017-04-25 98,28 2.2e-16
Villa Sweden 2019-12-02 3023,8 2.2e-16
Stockholm 2020-07-31 425,65 2.2e-16
Dalarnas 2018-11-30 224,13 2.2e-16
Skåne 2019-01-11 776,64 2.2e-16
Västra Götaland 2018-12-21 617,94 2.2e-16
Kedjehus Sweden 2019-03-01 333,55 2.2e-16
Page 17
Stockholm 2019-11-29 117,75 2.2e-16
Dalarna 2018-06-15 32,54 2.365e-14
Skåne 2019-04-12 146,86 2.2e-16
Västra Götaland 2019-06-05 48,45 2.2e-16
4.3.1 Pandemic related structural breaks
Evaluating the breakpoints for object type ‘Fritidshus’ properties in the entire country, we look
at the 𝐹-statistics plot, which is depicted in Figure A. In order to assume a breakpoint, the 𝐹-
statistics must be above the threshold, which is represented by the red line. Since most
observations have 𝐹-statistic above this level (approximate from observation number 2 200 and
beyond), it can be assumed that a breakpoint exists at almost every record. In order to find the
optimal breakpoint, the highest sup-𝐹-statistic is extracted. This is at observation number 12
781, which corresponds to 2020-08-01.
Plotting the monthly median logarithm sales price of object type ‘Fritidshus’ in Figure
6, the suggested breakpoint date represented by the dotted line is right at the steep increase of
prices in august 2020. This further emphasize the suggested breakpoint.
Figure 6. Time series for the log of monthly median price of object type ‘Fritidshus’ in
Sweden
A similar claim can be made for the price of object type ‘Lägenhet’ in all Sweden. The
breakpoint is detected at observation number 298 646, as displayed in 𝐹-statistics plot in
Figure A2, which corresponds to 2020-08-12. Looking at the median monthly logarithm of
Page 18
sales prices over time, in Figure 7, the suggested breakpoint is again located right in a steep
price increase.
Figure 7. Time series for the log of monthly median price of object type ‘Lägenhet’ in
Sweden
Looking at the log of monthly median price of object type ‘Villa’ throughout Sweden did not
result in a breakpoint in the month post COVID-19. As listed in Table 3, the breakpoint is
determined at 2019-12-02. However, looking at Stockholm county individually there is a
breakpoint detected at 2020-07-31, as indicated by observation point 37 210 as in Figure A3.
As Figure 8 shows, the median price for object type ‘Villa’ in Stockholm county steeply
increases around this point, reaching values not registered before.
Figure 8. Time series for the log of monthly median price of object type ‘Villa’ in Stockholm
county
Page 19
4.4 Hedonic regression analysis
Before constructing the hedonic regression model, highly correlated variables to the sales price
and to each other were removed. An example of a variable highly correlated to the sales price
that had to be removed is the square meter sales price. An example of a variable correlated to
another variable that also had to be removed is the risk of poverty in an area, that is highly
correlated to the employment level in the area.
Building on the estimated structural breaks, three hedonic regression models are created and
compared for object type ‘Fritidshus’ throughout Sweden, ‘Lägenhet’ throughout Sweden and
‘Villa’ in Stockholm county. The first model comprises observations before the suggested
breakpoint date, the second model thereafter and the third for the full (‘pooled’) set of
observations. Applying this comparison for object type ‘Fritidshus’ throughout Sweden in
Table 5 we see the adjusted 𝑅2 for all three models are around 0.35.
From the results in Table 5 we note that DACT has a significant negative impact on the selling
price, and also that the impact differs between the regression before and after the breakpoint
date for object type ‘Fritidshus’, see Table 4. For every additional day the object is active for
selling, the price before the break decreases with 0,03%, since (1 − exp( −0,0003274) =
0,03%, while after the break the price decreases by 0,08%. LIVA, the living area, shows as
expected a significant positive impact on the price for object type ‘Fritidshus’ Sweden. For the
pooled regression, every square meter increase in living area yields a 1,29%, with same
calculation as above from exponential, increase in price. Interestingly, as while the overall
prices have increased after the breakpoint, the effect of living area on the price is larger before
the break with 1,29% price increase per m2 compared to 1,27% after the break.
Before the breakpoint, for every percentage unit increment of DPEM, share of gainfully
employed in the area of the property, the sales price is estimated to increase by 11,6% while
after the break the corresponding figure was 219,1%, and while not being a significant
coefficient before the breakpoint date it is deemed as highly significant after the break. A similar
claim could be made for DPSE, that represent share of people with higher educational level in
the property area, although DPSE shows significance both before and after the break. That both
DPEM and DPSE increase after the break could indicate that the price for object type
‘Fritidshus’ have increased substantially in socioeconomic stable areas, or that more objects
have been traded in these areas that has higher prices compared to before the break. DM2P, the
Page 20
measurement for population density, is having both an increased and more significant impact
on the price for object type ‘Fritidshus’ after the break, further suggesting increasing prices in
more populated areas after the breakpoint date. Combining this with the noted change in
influence of the living area in relation to the price, this could indicate that smaller object type
‘Fritidshus’ in more densely populated areas have been traded more often and / or has
experienced an increase in price.
Traditional strong housing price determinant reflecting economic development, UNEM, and
GDPL displays a negative relationship with the prices for object type ‘Fritidshus’ after the
breakpoint. This is an interesting development that begs the question if the price increase post
pandemic is sustainable, as these variables in the long term usually has a positive relationship
with the price.
Comparing the split regressions for apartments in the entire Sweden in Table 6, we note
the same relationship, as for Fritidshus, in a decreased impact of DACT on the price after the
break, but also that DACT after the break is no longer deemed as a significant coefficient. This
seems reasonable as for the regression after the breakpoint date, merely 0,003% increase in
price is estimated from every day an object is active. For apartments, the living area (LIVA)
experiences an increased effect, estimating the selling price as every M2 prior to the breakpoint
estimated an increase in price of 0,95% while after the break the effect is 1,03%. As the impact
shifts from 0,95% to 1,03% equaling an increase of 8%, this indicates that the prices for larger
apartments have increased compared to smaller ones. As DPEM, DPSE and DDIM are variables
that represent the socioeconomic level of the area of the object, these three variables are as
expected estimated to have a significant positive effect on the selling price for apartments.
Notably, we see that the effect of these variables is increasing from the regression before the
breakpoint date to the regression after, suggesting an added price premium for apartments in
socioeconomic stable areas after the break. We further note that DM2P is, as expected, highly
significant for apartments, but that this coefficient’s effect on the selling price actually is
reduced after the break. This indicates that the price after the break is not increasing as much
as before in relation to how densely populated the area is, potentially reflecting increased
demand in less populated areas. UNEM again shows a negative relationship with the sales price,
reflecting the inverse relationship between soaring housing prices and increasing
unemployment during the pandemic.
Page 21
Table 5: Hedonic model comparison for object type ‘Fritidshus’ in Sweden by breakpoint date
2020-08-01.
Dependent variable:
Log of sale price ‘Fritidshus’
Before After Pooled
DACT -0.0003274*** -0.0007591*** -0.0003895*** (0.0000603) (0.0002024) (0.0000577)
LIVA 0.01277*** 0.01262*** 0.01275*** (0.000337) (0.0008201) (0.0003121)
DPEM 0.11 1.16*** 0.26* (0.16) (0.41) (0.15)
DPSE 0.40*** 0.47** 0.41*** (0.08) (0.20) (0.08)
DDIM 0.003397*** 0.002946*** 0.003339*** (0.0001648) (0.0004271) (0.0001538)
DM2P 0.00000001206 0.00000002744* 0.00000001513** (0.00000000790) (0.00000001613) (0.000000007087)
UNEM 2.79** -4.34 7.16*** (1.19) (6.87) (0.91)
GDPL 0.004 -0.01 0.01** (0.005) (0.04) (0.004)
HAFS 0.001 0.06* -0.002 (0.004) (0.03) (0.004)
EI12 0.03* -0.29 0.06*** (0.01) (0.22) (0.01)
ER12 0.17** 0.78 -0.005 (0.08) (2.78) (0.07)
Constant 11.39*** 10.56 11.23*** (0.22) (6.46) (0.21)
Observations 9,332 1,589 10,921
R2 0.35 0.37 0.35
Adjusted R2 0.34 0.35 0.35
Residual Std. Error 0.48 (df = 9284) 0.48 (df = 1551) 0.48 (df = 10873)
F Statistic 104.14*** (df = 47; 9284) 24.39*** (df = 37; 1551) 125.74*** (df = 47; 10873)
Note: Dummy values have been used for Month and Counties. Full table in Table A3 in Appendix. *p**p***p<0.01
Page 22
Table 6: Hedonic model comparison for object type ‘Lägenhet’ in Sweden by breakpoint date
2020-08-12.
Dependent variable:
Log of sale price Lägenhet Before After Pooled
DACT 0.0000695*** 0.0000278 0.00005345*** (0.00001081) (0.00001778) (0.00000928)
LIVA 0.009412*** 0.01021*** 0.00954*** (0.00003477) (0.00007851) (0.00003194)
DPEM 0.53*** 0.69*** 0.56*** (0.01) (0.02) (0.01)
DPSE 1.58*** 1.64*** 1.59*** (0.004) (0.01) (0.004)
DDIM 0.0007735*** 0.0007996*** 0.0007803*** (0.000007703) (0.00001861) (0.000007153)
DM2P 0.000000335300*** 0.00000022310*** 0.0000003113*** (0.000000007411) (0.00000001083) (0.000000005997)
UNEM 1.61*** -2.31*** 4.06*** (0.11) (0.60) (0.07)
GDPL 0.003*** 0.02*** 0.01*** (0.0004) (0.004) (0.0004)
HAFS 0.002*** 0.01** 0.0003 (0.0003) (0.003) (0.0003)
EI12 0.02*** -0.003 0.05*** (0.001) (0.02) (0.001)
ER12 0.13*** -0.20 0.02*** (0.01) (0.25) (0.01)
Constant 11.58*** 12.82*** 11.58*** (0.02) (0.60) (0.02)
Observations 214,283 42,296 256,579
R2 0.75 0.74 0.75
Adjusted R2 0.75 0.74 0.75
Residual Std.
Error 0.22 (df = 214235) 0.22 (df = 42258) 0.22 (df = 256531)
F Statistic 13,932.63*** (df = 47;
214235)
3,243.31*** (df = 37;
42258)
16,340.89*** (df = 47;
256531)
Note: Dummy values have been used for Month and Counties. Full table in Table A4 in Appendix. *p**p***p<0.01
Page 23
Looking into Table 7 for object type ‘Villa’ in Stockholm, we find a similar but smaller
shift in the impact of living area (LIVA) as an estimator of price from the regression model
after the break compared to the one before, as we did for object type ‘Lägenhet’. Before the
break, every extra M2 estimated an increase in price of 0,308 % while after the break this figure
is 0,314, equaling an increase of 2% compared to 8% for object type ‘Lägenhet’.
For object type ‘Villa’, in contrary to object type ‘Fritidshus’ but in the same way as for
object type ‘Lägenhet’, the variable for population density, DM2P, has decreased in impact on
the selling price. Still a highly significant determinant for the selling price, the decreased impact
of DM2P after the break suggests that object type ‘Villa’ outside densely populated areas have
increased more in price compared to object type ‘Villa’ inside these areas. The variables
reflecting the socioeconomic level of the object’s location however, do not show any clear
indication for object type ‘Villa’ as they did for object type ‘Lägenhet’. DPEM and DDIM show
an increased impact in the same way as for object type ‘Lägenhet’, while DPSE shows a
decreasing effect on the sales price after the break.
Table 7: Hedonic model comparison for object type ‘Villa’ in Stockholm county by breakpoint
2020-07-31.
Dependent variable:
Log of sale price Villa Stockholm county Before After Pooled
DACT -0.0004131*** -0.0002868*** -0.0004113*** (0.00004663) (0.00008211) (0.00004117)
LIVA 0.003075*** 0.003141*** 0.003091*** (0.00004761) (0.000106) (0.00004379)
DPEM -0.31*** -0.02 -0.26*** (0.06) (0.12) (0.05)
DPSE 1.64*** 1.53*** 1.62*** (0.02) (0.04) (0.02)
DDIM 0.0002896*** 0.0003993*** 0.0003026*** (0.00002234) (0.00005581) (0.00002076)
DM2P 0.0000005051*** 0.000000323*** 0.0000004826*** (0.00000002223) (0.000000057460) (0.00000002071)
UNEM 0.85** -4.33** 5.10*** (0.36) (1.78) (0.24)
GDPL 0.0002 0.05*** 0.01*** (0.001) (0.01) (0.001)
Page 24
HAFS 0.002* -0.01 -0.001 (0.001) (0.01) (0.001)
EI12 0.01*** -0.05 0.06*** (0.004) (0.05) (0.004)
ER12 0.16*** 0.46 -0.02 (0.02) (0.72) (0.02)
Constant 13.95*** 13.78*** 13.90*** (0.07) (1.72) (0.06)
Observations 26,469 4,723 31,192
R2 0.60 0.63 0.61
Adjusted R2 0.60 0.63 0.60
Residual Std.
Error 0.24 (df = 26441) 0.22 (df = 4705) 0.24 (df = 31164)
F Statistic 1,480.52*** (df = 27;
26441)
464.21*** (df = 17;
4705)
1,768.89*** (df = 27;
31164)
Note: Dummy values have been used for Month and Counties. Full table in Table A5 in Appendix. *p**p***p<0.01
5. Discussion, limitations and future research
By analyzing prices of different object types on the Swedish housing market for structural
breaks, we can conclude that breakpoints are found at different dates for different object types
in different counties. Building on the exploratory notion that a sharp increase in trading volumes
for object type ‘Fritidshus’ in entire Sweden occurred during the summer of 2020, the structural
break analysis finds evidence that a structural break in prices also occurred during this time.
These findings support that two hedonic regression models, split by the breakpoint date 2020-
08-01 and with housing price determinants as regressors, better explain the development in the
price for object type ‘Fritidshus’ compared to a single regression model for the full period.
Evidence for structural breaks during the summer of 2020 were also found for object type
‘Lägenhet’ in entire Sweden, ‘Villa’ in Stockholm county and ‘Radhus’ in entire Sweden, and
Stockholm county and Skåne county individually. As these structural breaks are consistently
dated short after the COVID-19 outbreak, the findings could indicate at a pandemic infused
structural break for these housing prices.
Extending the analysis from the determined structural breaks, for object types
‘Fritidshus’ and ‘Lägenheter’ throughout Sweden and for ‘Villa’ in Stockholm county, the
hedonic regression comparison yielded some interesting result to further dwell on. The
significant and important determinants living area and population density were both affected in
a similar way by the shift in the regression model from before to after the break for object types
Page 25
‘Lägenhet’ and ‘Villa’. Object type ‘Fritidshus’, on the other hand, showed a direct opposite
effect for these variables. While the living area increased in impact in relation to the price after
the breakpoint date for object types ‘Lägenhet’ and ‘Villa’, it decreased for object type
‘Fritidshus’ and while the impact of population density decreased for object types ‘Lägenhet’
and ‘Villa’ after the break, it increased for object type ‘Fritidshus’. As these object types differs
in character, the decreased impact of population density for object types ‘Lägenhet’ and ‘Villa’
could, in combination with the increased impact for living area, be an indicated pandemic effect
the people might increasingly value living area at the cost of location to a certain degree for
their permanent housing during the pandemic. This also suggests that the temporary object types
Fritidshus, on the other hand, is not valuable as much for its size, perhaps more for its
functionality as a second home while travelling options are limited.
One of the limitations of this research is the dependability of external data. In particular,
even though Booli has the largest collection of housing prices in Sweden, it is difficult to
evaluate to what extent their housing data is representable for all property sales in Sweden. An
option is to bring together various sources of housing data to increase representability.
Furthermore, a significant time of this research was used to enhance the fit of the regression
model’s fit. Since time was limited, future research should invest more resources in increasing
the level fitness in order to achieve even more profound results. Also, only a few individual
counties were investigated due to time constraints and as an effect all potentially interesting
breakpoints might not have been found. Future research could also put effort in studying
indicated effects more in detail for a single object type in a single specified county, perhaps
adding variables of specific interest available for the chosen area. Future research could also
focus on applying other methods of specifying the locational attributes of the housing object,
as we did not get enough quality in the data for variables like distance to transportations hubs,
central business districts, schools or hospitals. This could give further insights to what variables
affects the housing prices and hence what determinant variables changes at a breakpoint.
Regarding the methods used, the Chow test and Quandt-Andrews Sup-F statistic test was used
to test for a single structural break in the time series analysis. The structural break models
assume parameters to shift directly at a specific breakpoint, and as this direct shift might not
always be the case there are other models that allows for different parameters shifts. As time
was limited this research focused solely on structural break, and future research is therefore
encouraged to evaluate other model like for example time-varying parameter models, Markov-
switch or threshold autoregression models. Finally, it should be noted that the hedonic model
Page 26
utilized to compare the coefficients shifted impact before and after the structural break only
allows for conclusions regarding correlation and not causality. Hence all indicated results
discussed in the paper is derived from observed correlational effects of the pandemic related
breakpoint and cannot with certainty be causally linked to be an effect of the COVID-19
pandemic itself.
Page 27
REFERENCES
Ahoniemi, K., & Putkuri, H. (2020). Nordic countries are vulnerable to housing market risks
aggravated by the coronavirus pandemic. http://urn.fi/URN:NBN:fi:bof-202005222147
Aktay, A., Bavadekar, S., Cossoul, G., Davis, J., Desfontaines, D., Fabrikant, A., Gabrilovich,
E., Gadepalli, K., Gipson, B., Guevara, M., & others. (2020). Google COVID-19
community mobility reports: Anonymization process description (version 1.0). ArXiv
Preprint ArXiv:2004.04145.
Andrews, D. W. K. (1993). Tests for Parameter Instability and Structural Change With
Unknown Change Point. Econometrica, 61(4), 821. https://doi.org/10.2307/2951764
Andrews, D. W. K., & Ploberger, W. (1994). Optimal Tests when a Nuisance Parameter is
Present Only Under the Alternative. Econometrica, 62(6), 1383.
https://doi.org/10.2307/2951753
Barot, B. (2001). An econometric demand--supply model for Swedish private housing.
European Journal of Housing Policy, 1(3), 417–444.
Bartik, A., Cullen, Z., Glaeser, E. L., Luca, M., & Stanton, C. (2020). What Jobs are Being
Done at Home During the COVID-19 Crisis? Evidence from Firm-Level Surveys. SSRN
Electronic Journal. https://doi.org/10.2139/ssrn.3634983
Booli.se. (n.d.). Om oss. Retrieved May 29, 2020, from https://www.booli.se/p/om-booli/
Chau, K. W., & Chin, T. L. (2003). A Critical Review of Literature on the Hedonic Price
Model. International Journal for Housing Science and Its Applications, 2(27), 145–165.
Chen, J.-H., Ong, C. F., Zheng, L., & Hsu, S.-C. (2017). Forecasting spatial dynamics of the
housing market using support vector machine. International Journal of Strategic
Property Management, 21(3), 273–283.
Chow, G. C. (1960). Tests of Equality Between Sets of Coefficients in Two Linear
Regressions. Econometrica, 28(3), 591. https://doi.org/10.2307/1910133
Claussen, C. A. (2013). Are Swedish houses overpriced? International Journal of Housing
Markets and Analysis, 6(2), 180–196. https://doi.org/10.1108/IJHMA-12-2011-0056
Englund, P., & Ioannides, Y. M. (1997). House Price Dynamics: An International Empirical
Perspective. Journal of Housing Economics, 6(2), 119–136.
https://doi.org/10.1006/jhec.1997.0210
Finansinspektionen. (2020). Banks may grant all mortgagors amortisation exemption.
https://www.fi.se/en/published/press-releases/2020/banks-may-grant-all-mortgagors-
Page 28
amortisation-exemption/
Galati, G., Teppa, F., & Alessie, R. J. M. (2012). Macro and Micro Drivers of House Price
Dynamics: An Application to Dutch Data. SSRN Electronic Journal, 288.
https://doi.org/10.2139/ssrn.1950606
García, J. A. B., & Hernández, J. E. R. (2004). User Cost Changes, Unemployment and
Home-ownership: Evidence from Spain. Urban Studies, 41(3), 563–578.
Garrod, G. D., & Willis, K. G. (1992). Valuing goods’ characteristics: An application of the
hedonic price method to environmental attributes. Journal of Environmental
Management, 34(1), 59–76. https://doi.org/https://doi.org/10.1016/S0301-
4797(05)80110-0
Girouard, N., et al. (2005). Recent House Price Developments: The Role of Fundamentals.
OECD Economics Department Working Papers, 475.
https://doi.org/https://doi.org/10.1787/864035447847
Google. (2021). Community Mobility Reports Help. https://support.google.com/covid19-
mobility/answer/9824897?hl=en&ref_topic=9822927#zippy=%2Chow-the-headline-
number-date-relates-to-the-report-date
Helsingen, L. M., Refsum, E., Gjøstein, D. K., Løberg, M., Bretthauer, M., Kalager, M., &
Emilsson, L. (2020). The COVID-19 pandemic in Norway and Sweden--threats, trust,
and impact on daily life: a comparative survey. BMC Public Health, 20(1), 1–10.
Jumadi, N. (2010). The Relationship Between Demographic Factors and Housing
Affordability. Malaysian Journal of Real Estate, 5(1), 49–58.
Kiel, K. A., & Zabel, J. E. (2008). Location, location, location: The 3L Approach to house
price determination. Journal of Housing Economics, 17(2), 175–190.
https://doi.org/10.1016/j.jhe.2007.12.002
Koren, M., & Pető, R. (2020). Business disruptions from social distancing. Plos One, 15(9),
e0239113.
Lancaster, K. J. (1966). A New Approach to Consumer Theory. Journal of Political
Economy, 74.
Li, M. M., & Brown, H. J. (1980). Micro-Neighborhood Externalities and Hedonic Housing
Prices. Land Economics, 56(2), 125–141. https://doi.org/10.2307/3145857
Linneman, P. (1980). Some empirical results on the nature of the hedonic price function for
the urban housing market. Journal of Urban Economics, 8(1), 47–68.
https://doi.org/https://doi.org/10.1016/0094-1190(80)90055-8
Liu, S., & Su, Y. (2020). The impact of the Covid-19 pandemic on the demand for density:
Page 29
Evidence from the US housing market. Available at SSRN 3661052.
Luitel, H. S., & Mahar, G. J. (2015). A Short Note on the Application of Chow Test of
Structural Break in US GDP. International Business Research, 8(10).
https://doi.org/10.5539/ibr.v8n10p112
Nuredini, B. (2020). Impact of the COVID 19 pandemic on the global real estate market.
Geo-SEE Institute, 15, 54–65.
Oikarinen, E. (2009). Household borrowing and metropolitan housing price dynamics -
Empirical evidence from Helsinki. Journal of Housing Economics, 18(2), 126–139.
https://doi.org/10.1016/j.jhe.2009.04.001
Olanrewaju, A., Tat, L., Tan, S. Y., Zafarullah, N., Abdul-Aziz, A.-R., & NAOTO, M.
(2016). Analysis of Economic Determinants of affordable housing prices (Issue
November) [Universiti Tunku Abdul Rahman, Malaysia].
https://doi.org/10.14455/ISEC.res.2016.122
Olofsson, P., & Andersson, M. (2012). Probability, Statistics, and Stochastic Processes
(Second). Wiley.
Quandt, R. E. (1960). Tests of the Hypothesis That a Linear Regression System Obeys Two
Separate Regimes. Journal of the American Statistical Association, 55(290), 324–330.
https://doi.org/10.1080/01621459.1960.10482067
Rosen, S. (1974). Hedonic Prices and Implicit Markets: Product Differentiation in Pure
Competition. Journal of Political Economy, 82(1), 34–55.
Statistics Sweden. (n.d.). DeSO – Demografiska statistikområden. Retrieved May 30, 2021,
from https://www.scb.se/hitta-statistik/regional-statistik-och-kartor/regionala-
indelningar/deso---demografiska-statistikomraden/
Swedish Central Bank. (2020). Financial stability report (Issue 2).
https://www.riksbank.se/en-gb/financial-stability/financial-stability-
report/2020/financial-stability-report-20202/
Turk, R. (2015). Housing Price and Household Debt Interactions in Sweden. IMF Working
Papers, 15(276), 1. https://doi.org/10.5089/9781513586205.001
Wang, B. (2021). How Does COVID-19 Affect House Prices? A Cross-City Analysis.
Journal of Risk and Financial Management, 14(2), 47.
Wang, X. R., Hui, E. C. M., & Sun, J. X. (2017). Population migration, urbanization and
housing prices: Evidence from the cities in China. Habitat International, 66, 49–56.
https://doi.org/10.1016/j.habitatint.2017.05.010
Wong, G. (2008). Has SARS infected the property market? Evidence from Hong Kong.
Page 30
Journal of Urban Economics, 63(1), 74–95.
Zeileis, A., Kleiber, C., Walter, K., & Hornik, K. (2003). Testing and dating of structural
changes in practice. Computational Statistics and Data Analysis, 44(1–2), 109–123.
https://doi.org/10.1016/S0167-9473(03)00030-6
Zeileis, A., Leisch, F., Hornik, K., & Kleiber, C. (2002). Strucchange: An R package for
testing for structural change in linear regression models. Journal of Statistical Software,
7, 1–38. https://doi.org/10.18637/jss.v007.i02
Zhao, Y. (2020). US Housing Market during COVID-19: Aggregate and Distributional
Evidence. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.3677651
Page 31
APPENDIX A
Table A1. Description of data.
Feature Description Example Origin
DACT Day active before sold 36 Booli.se
MONTH Time dummy (1 = January and 12 = December) 1 Booli.se
OBJT Type of property Lägenhet Booli.se
LIVA Property size in square meters 102 Booli.se
SPRI Property sale price (SEK) 2 600 000 Booli.se
COUN Name of län. Transformed to dummy variable Stockholms län SCB
DDIM DeSo area mean disposable income 249.5 SCB
DPSE % in DeSo area with higher education 0.38 SCB
DPRD % of dwellings in DeSo area that are rental 0 SCB
DPHC % of households in DeSo area that has children 0.30 SCB
DPEM % of people in DeSo area gainfully employed 0.778 SCB
DM2P Square meter per person in DeSo area 77.27 SCB
IDLQ % development of disposable income in Sweden 0.061 SCB
GDPL % development of GDP in Sweden 3 SCB
UNEM % unemployed in Sweden by 0.075 SCB
DESB Area dummy for centrality (A, B and C exists) 1 SCB
DESC Area dummy for centrality (A, B and C exists) 0 SCB
HAFS Households’ average financial situation 6.56 NIER
EI12 Households’ expectation on inflation in 1 year 2.2 NIER
ER12 Households’ expectation on interest rates in 1 year 2.21 NIER
Page 32
Table A2. Translation table
Figure A1. Sup-𝐹 statistic for object type ‘Fritidshus’ in Sweden.
Figure A2. Sup-𝐹 statistic for object type ‘Lägenhet’ in Sweden.
Swedish English
Lägenhet Apartment
Fritidshus Holiday home
Villa Villa
Radhus Town house
Kedjehus Chain house
Län County
Kommun Municipality
Page 33
Figure A3. Sup-𝐹 statistic for object type ‘Villa’ in Stockholm.
Table A3. Full Hedonic model comparison for object type ‘Fritidshus’ Sweden by breakpoint
date 2020-08-01.
Dependent variable:
Log of sale price ‘Fritidshus’ Before After Pooled
DACT -0.0003274*** -0.0007591*** -0.0003895*** (0.0000603) (0.0002024) (0.0000577)
LIVA 0.01277*** 0.01262*** 0.01275*** (0.000337) (0.0008201) (0.0003121)
DPEM 0.11 1.16*** 0.26* (0.16) (0.41) (0.15)
DPSE 0.40*** 0.47** 0.41*** (0.08) (0.20) (0.08)
DDIM 0.003397*** 0.002946*** 0.003339*** (0.0001648) (0.0004271) (0.0001538)
DPRD 0.31*** 0.48*** 0.33*** (0.05) (0.12) (0.05)
DPHC -1.01*** -1.00*** -1.02*** (0.09) (0.24) (0.09)
DM2P 0.00000001206 0.00000002744* 0.00000001513** (0.00000000790) (0.00000001613) (0.000000007087)
Page 34
UNEM 2.79** -4.34 7.16*** (1.19) (6.87) (0.91)
IDLQ -0.35 8.97 -0.08 (0.30) (26.02) (0.29)
GDPL 0.004 -0.01 0.01** (0.005) (0.04) (0.004)
HAFS 0.001 0.06* -0.002 (0.004) (0.03) (0.004)
EI12 0.03* -0.29 0.06*** (0.01) (0.22) (0.01)
ER12 0.17** 0.78 -0.005 (0.08) (2.78) (0.07)
MONTH02 -0.004 0.05 0.04 (0.04) (0.09) (0.04)
MONTH03 -0.01 0.07** (0.04) (0.04)
MONTH04 0.01 0.03 (0.04) (0.04)
MONTH05 -0.0001 0.001 (0.04) (0.03)
MONTH06 -0.06* -0.07** (0.04) (0.03)
MONTH07 -0.003 0.02 (0.04) (0.03)
MONTH08 -0.02 0.02 (0.04) (0.03)
MONTH09 0.03 0.08** (0.04) (0.03)
MONTH10 -0.0001 0.08** (0.04) (0.04)
MONTH11 0.001 0.06* (0.05) (0.04)
MONTH12 0.07 0.15*** (0.06) (0.04)
``COUN_Dalarnas län`` 0.14*** 0.25** 0.15*** (0.04) (0.10) (0.04)
``COUN_Gävleborgs län`` 0.03 0.29*** 0.07* (0.04) (0.11) (0.04)
``COUN_Gotlands län`` 0.60*** 0.81*** 0.63*** (0.05) (0.13) (0.05)
Page 35
``COUN_Hallands län`` 0.45*** 0.53*** 0.47*** (0.04) (0.11) (0.04)
``COUN_Jämtlands län`` 0.19*** 0.46*** 0.24*** (0.05) (0.12) (0.04)
``COUN_Jönköpings län`` -0.11** 0.06 -0.09** (0.05) (0.13) (0.04)
``COUN_Kalmar län`` 0.14*** 0.29*** 0.16*** (0.04) (0.11) (0.04)
``COUN_Kronobergs län`` -0.21*** -0.20 -0.22*** (0.06) (0.15) (0.06)
``COUN_Norrbottens län`` -0.03 -0.01 -0.03 (0.05) (0.12) (0.04)
``COUN_Örebro län`` -0.17*** -0.03 -0.15*** (0.04) (0.12) (0.04)
``COUN_Östergötlands län`` -0.02 0.21* 0.01 (0.04) (0.11) (0.04)
``COUN_Skåne län`` -0.22*** 0.07 -0.18*** (0.04) (0.10) (0.03)
``COUN_Södermanlands län`` 0.20*** 0.50*** 0.24*** (0.04) (0.10) (0.04)
``COUN_Stockholms län`` 0.41*** 0.63*** 0.44*** (0.04) (0.10) (0.03)
``COUN_Uppsala län`` 0.14*** 0.39*** 0.17*** (0.04) (0.11) (0.04)
``COUN_Värmlands län`` 0.06 0.16 0.07* (0.04) (0.11) (0.04)
``COUN_Västerbottens län`` 0.18*** 0.35*** 0.21*** (0.05) (0.12) (0.04)
``COUN_Västernorrlands
län`` 0.04 0.21 0.06
(0.05) (0.13) (0.05)
``COUN_Västmanlands län`` -0.12*** 0.16 -0.08** (0.04) (0.12) (0.04)
``COUN_Västra Götalands
län`` 0.16*** 0.37*** 0.19***
(0.04) (0.10) (0.03)
DESB 0.17*** 0.12** 0.16*** (0.02) (0.06) (0.02)
DESC 0.02 -0.11** 0.004 (0.02) (0.05) (0.02)
Constant 11.39*** 10.56 11.23***
Page 36
(0.22) (6.46) (0.21)
Observations 9,332 1,589 10,921
R2 0.35 0.37 0.35
Adjusted R2 0.34 0.35 0.35
Residual Std. Error 0.48 (df = 9284) 0.48 (df = 1551) 0.48 (df = 10873)
F Statistic 104.14*** (df = 47;
9284)
24.39*** (df = 37;
1551)
125.74*** (df = 47;
10873)
*p**p***p<0.01
Table A4. Full Hedonic model comparison for object type ‘Lägenhet’ in Sweden by breakpoint
date 2020-08-12.
Dependent variable:
Log of sale price ‘Lägenhet’ Before After Pooled
DACT 0.0000695*** 0.00002781 0.00005345*** (0.00001081) (0.00001778) (0.00000928)
LIVA 0.009412*** 0.01021*** 0.00954*** (0.00003477) (0.00007851) (0.00003194)
DPEM 0.53*** 0.69*** 0.56*** (0.01) (0.02) (0.01)
DPSE 1.58*** 1.64*** 1.59*** (0.004) (0.01) (0.004)
DDIM 0.0007735*** 0.0007996*** 0.0007803*** (0.000007703) (0.00001861) (0.000007153)
DPRD 0.34*** 0.37*** 0.34*** (0.002) (0.01) (0.002)
DPHC -0.37*** -0.32*** -0.36*** (0.01) (0.01) (0.01)
DM2P 0.000000335300*** 0.00000022310*** 0.0000003113*** (0.000000007411) (0.00000001083) (0.000000005997)
UNEM 1.61*** -2.31*** 4.06*** (0.11) (0.60) (0.07)
IDLQ -0.42*** -2.77 -0.27*** (0.02) (2.40) (0.02)
GDPL 0.003*** 0.02*** 0.01*** (0.0004) (0.004) (0.0004)
Page 37
HAFS 0.002*** 0.01** 0.0003 (0.0003) (0.003) (0.0003)
EI12 0.02*** -0.003 0.05*** (0.001) (0.02) (0.001)
ER12 0.13*** -0.20 0.02*** (0.01) (0.25) (0.01)
MONTH02 -0.01*** 0.01 -0.01*** (0.002) (0.01) (0.002)
MONTH03 0.01*** 0.02*** (0.002) (0.002)
MONTH04 -0.003 0.003 (0.002) (0.002)
MONTH05 -0.002 -0.01*** (0.002) (0.002)
MONTH06 -0.02*** -0.04*** (0.002) (0.002)
MONTH07 0.01*** -0.01** (0.003) (0.002)
MONTH08 0.04*** 0.04*** (0.002) (0.002)
MONTH09 0.04*** 0.05*** (0.003) (0.002)
MONTH10 0.04*** 0.06*** (0.003) (0.002)
MONTH11 0.03*** 0.04*** (0.003) (0.002)
MONTH12 0.02*** 0.04*** (0.003) (0.003)
``COUN_Dalarnas län`` 0.20*** 0.20*** 0.20*** (0.01) (0.02) (0.01)
``COUN_Gävleborgs län`` 0.14*** 0.13*** 0.14*** (0.01) (0.02) (0.01)
``COUN_Gotlands län`` 0.51*** 0.54*** 0.51*** (0.01) (0.02) (0.01)
``COUN_Hallands län`` 0.52*** 0.55*** 0.53*** (0.01) (0.02) (0.01)
``COUN_Jämtlands län`` 0.14*** 0.24*** 0.16*** (0.01) (0.02) (0.01)
``COUN_Jönköpings län`` 0.33*** 0.30*** 0.33*** (0.01) (0.02) (0.01)
Page 38
``COUN_Kalmar län`` 0.13*** 0.11*** 0.13*** (0.01) (0.02) (0.01)
``COUN_Kronobergs län`` 0.18*** 0.11*** 0.17*** (0.01) (0.02) (0.01)
``COUN_Norrbottens län`` 0.09*** 0.05** 0.08*** (0.01) (0.02) (0.01)
``COUN_Örebro län`` 0.16*** 0.15*** 0.16*** (0.01) (0.02) (0.01)
``COUN_Östergötlands
län`` 0.27*** 0.26*** 0.27***
(0.01) (0.02) (0.01)
``COUN_Skåne län`` 0.26*** 0.31*** 0.27*** (0.01) (0.02) (0.01)
``COUN_Södermanlands
län`` 0.27*** 0.25*** 0.26***
(0.01) (0.02) (0.01)
``COUN_Stockholms län`` 0.77*** 0.70*** 0.76*** (0.01) (0.02) (0.01)
``COUN_Uppsala län`` 0.43*** 0.38*** 0.42*** (0.01) (0.02) (0.01)
``COUN_Värmlands län`` 0.06*** 0.08*** 0.06*** (0.01) (0.02) (0.01)
``COUN_Västerbottens
län`` 0.13*** 0.17*** 0.14***
(0.01) (0.02) (0.01)
``COUN_Västernorrlands
län`` 0.003 -0.08*** -0.01
(0.01) (0.02) (0.01)
``COUN_Västmanlands
län`` 0.18*** 0.14*** 0.17***
(0.01) (0.02) (0.01)
``COUN_Västra Götalands
län`` 0.53*** 0.50*** 0.53***
(0.01) (0.02) (0.01)
DESB -0.04*** -0.12*** -0.06*** (0.01) (0.02) (0.01)
DESC -0.13*** -0.23*** -0.15*** (0.01) (0.01) (0.01)
Constant 11.58*** 12.82*** 11.58*** (0.02) (0.60) (0.02)
Observations 214,283 42,296 256,579
R2 0.75 0.74 0.75
Page 39
Adjusted R2 0.75 0.74 0.75
Residual Std. Error 0.22 (df = 214235) 0.22 (df = 42258) 0.22 (df = 256531)
F Statistic 13,932.63*** (df = 47;
214235)
3,243.31*** (df = 37;
42258)
16,340.89*** (df = 47;
256531)
*p**p***p<0.01
Table A5. Full hedonic model comparison for object type ‘Villa’ in Stockholm county by
breakpoint. 2020-07-31
Dependent variable:
Log of sale price Villa Stockholm county Before After Pooled
DACT -0.0004131*** -0.0002868*** -0.0004113*** (0.00004663) (0.00008211) (0.00004117)
LIVA 0.003075*** 0.003141*** 0.003091*** (0.00004761) (0.000106) (0.00004379)
DPEM -0.31*** -0.02 -0.26*** (0.06) (0.12) (0.05)
DPSE 1.64*** 1.53*** 1.62*** (0.02) (0.04) (0.02)
DDIM 0.0002896*** 0.0003993*** 0.0003026*** (0.00002234) (0.00005581) (0.00002076)
DPRD 0.15*** 0.20*** 0.16*** (0.01) (0.03) (0.01)
DPHC -0.17*** -0.08* -0.16*** (0.02) (0.05) (0.02)
DM2P 0.0000005051*** 0.000000323*** 0.0000004826*** (0.00000002223) (0.000000057460) (0.00000002071)
UNEM 0.85** -4.33** 5.10*** (0.36) (1.78) (0.24)
IDLQ -0.37*** 3.27 -0.14* (0.08) (6.91) (0.07)
GDPL 0.0002 0.05*** 0.01*** (0.001) (0.01) (0.001)
HAFS 0.002* -0.01 -0.001 (0.001) (0.01) (0.001)
EI12 0.01*** -0.05 0.06*** (0.004) (0.05) (0.004)
ER12 0.16*** 0.46 -0.02
Page 40
(0.02) (0.72) (0.02)
MONTH02 -0.001 0.03 -0.003 (0.01) (0.02) (0.01)
MONTH03 0.01 0.02*** (0.01) (0.01)
MONTH04 0.02* 0.02*** (0.01) (0.01)
MONTH05 0.02** -0.003 (0.01) (0.01)
MONTH06 0.02*** -0.02** (0.01) (0.01)
MONTH07 0.04*** 0.02** (0.01) (0.01)
MONTH08 0.05*** 0.05*** (0.01) (0.01)
MONTH09 0.04*** 0.06*** (0.01) (0.01)
MONTH10 0.04*** 0.07*** (0.01) (0.01)
MONTH11 0.03*** 0.06*** (0.01) (0.01)
MONTH12 0.05*** 0.07*** (0.01) (0.01)
DESB 0.001 -0.01 0.001 (0.01) (0.02) (0.01)
DESC 0.14*** 0.15*** 0.14*** (0.01) (0.01) (0.01)
Constant 13.95*** 13.78*** 13.90*** (0.07) (1.72) (0.06)
Observations 26,469 4,723 31,192
R2 0.60 0.63 0.61
Adjusted R2 0.60 0.63 0.60
Residual Std.
Error 0.24 (df = 26441) 0.22 (df = 4705) 0.24 (df = 31164)
F Statistic 1,480.52*** (df = 27;
26441)
464.21*** (df = 17;
4705)
1,768.89*** (df = 27;
31164)
*p**p***p<0.01
Page 41
Figure A3. Model diagnostic for object type ‘Fritidshus’ in Sweden, before structural break.
Figure A4. Model diagnostic for object type ‘Fritidshus’ in Sweden, after structural break.
Page 42
Figure A5. Model diagnostic for object type ‘Fritidshus’ in Sweden, pooled model.
Figure A6. Model diagnostic for object type ‘Lägenhet’ in Sweden, before structural break.
Page 43
Figure A7. Model diagnostic for object type ’Lägenhet’ in Sweden, after structural break.
Figure A8. Model diagnostic for object type ‘Lägenhet’ in Sweden, pooled model.
Page 44
Figure A9. Model diagnostics for object type ‘Villa’ in Stockholm, before structural break.
Figure A10. Model diagnostics for object type ‘Villa’ in Stockholm, after structural break.
Page 45
Figure A11. Model diagnostics for object type ‘Villa’ in Stockholm, pooled model.