Drivers of household credit demand before and during the crisis

Abstract

This paper analyses the fundamental determinants of the household borrowing decisions in Hungary and Poland, using the Euro Survey data conducted by the Austrian central bank. The previous researches, based on the Euro Survey, are supplemented and enhanced in several respects. It bears utmost importance that instead of using simple ad hoc methods to manage the problem of missing data we got more reliable results by applying multiple imputation technic. We show that former relations of different factors change in the crisis, so different periods should be examined individually as well. According to the results, credit demand was determined in both countries by the trust in the institutional system, the financial awareness/education and the economic expectations. However, the impact of the various socio-demographic variables differs in the two countries. E.g. in Hungary households in poorer regions tend to have higher credit demand before the crisis and lower demand in the crisis. The paper also presents that in Hungary the negative experiences related to foreign currency loans have a major effect on the borrowing decisions. Finally, it is demonstrated that self-selection has a material role in household lending as well, i.e. the customers who assume that they are not creditworthy, do not even try to apply for loans. JEL: C35, C58, D14, G20, G21

Full text

Drivers of household credit demand before and during the crisis: Micro-level evidence
from Hungary and Poland
This version: 15 September 2017
Ádám Banaia
1
a Corresponding author. Magyar Nemzeti Bank (Central Bank of Hungary)
Budapest, Szabadsag Tér 8-9, 1054, Hungary; banaia@mnb.hu;
Nikolett Vágób
2
b Magyar Nemzeti Bank (Central Bank of Hungary)
Budapest, Szabadsag Tér 8-9, 1054, Hungary; vagon@mnb.hu
Abstract
This paper analyses the determinants of households’ borrowing decisions in Hungary and
Poland, using the Euro Survey. Previous studies based on the Euro Survey are supplemented
and enhanced in several respects. Instead of using simple ad-hoc methods to manage missing
data, we applied more reliable multiple imputation techniques. We demonstrate that
households’ behaviour changed significantly in the crisis, and thus different periods should be
examined individually as well. The main drivers of credit demand are: trust in the institutional
system, financial awareness/literacy and economic expectations. In addition, negative
experiences related to foreign currency loans strongly influenced households’ behaviour, but
only in Hungary. Finally, we demonstrate that self-selection plays an important role in
household borrowing, i.e. customers who assume that they are not creditworthy do not apply
for loans.
JEL classification: C35; C58; C83; D14; G21
Keywords: credit demand; financial behaviour of households; crisis; Central and Eastern
European Countries; multiple imputation; microdata
1
Ádám Banai is Head of the Applied Research and Stress Testing Department at the Central Bank of Hungary.
2
Nikolett Vágó is Senior economist at the Applied Research and Stress Testing Department at the Central Bank
of Hungary.

2/46
1. Introduction
Since the beginning of the crisis lending to the household and the corporate sector has been
continuously shrinking in the Hungarian banking system. The focus was mainly on lending to
the corporate sector, as its impact on the real economy is direct and greater than that of
household lending. Moreover, the general view is that the large-scale contraction of the
household loan stock is the result of an inevitable adjustment. However, eight years after the
onset of the crisis, positive lending dynamics in the household segment would be increasingly
important, as this is essential both for consumption and real estate investment over a longer
horizon. By contrast, even after 2008 outstanding household debt in Poland has been
continuously increasing, albeit not at the same pace as before the crisis.
Using data from the Euro Survey conducted by the central bank of Austria (hereinafter: OeNB),
we attempt to identify the determinants of households’ borrowing decisions in Hungary and
Poland and determine the demand-side factors behind these different trends.
Euro Survey data have been used as the basis for several earlier papers on households’ FX
credit demand in the CEE region (e.g. Beckmann et al., 2011; Fidrmuc et al., 2013; Beckmann
et al., 2015). These papers significantly promoted a better understanding of households’
financial decisions, though it is worth going beyond them in some aspects. Most of them make
pooled estimations for the whole region (about ten countries), assuming that the behaviour of
households in these countries is homogeneous. In our paper, by comparing two countries
(which are very similar from several perspectives), we examine this assumption and show that
there are some interesting differences between the two countries. Earlier papers using Euro
Survey data did not place much emphasis on the problem of missingness which – according to
the literature – can cause significant bias (Jones, 1996; Schafer and Graham, 2002; Allison,
2009). Due to this possible bias, several household surveys, such as the Survey on Consumer
Finances (SCF) managed by the Federal Reserve Board (Kennickel, 1991 and 1998), the

3/46
Spanish Survey on Household Finances (HFCS) managed by the Bank of Espana (Barceló,
2006) and the Household Finance and Consumption Survey managed by Eurostat (HFCN,
2013), use advanced methods to correct for missing answers. Thus, from a methodological
point of view, one great added value of our paper is the testing of different methods on Euro
Survey data. We show the effect of different methodologies on the estimated results. Our paper
can function as a reference for making estimations on survey data with different levels of
missingness.
The main reference for our paper is Fidrmuc et al. (2013), as we are both interested in the
drivers of household borrowing, and credit demand is measured in the same way. One
important difference is that we analyse the drivers of credit demand irrespective of the currency
denomination, while they focus on FX loans. In addition to (i) managing missing data
differently, and (ii) examining Hungary and Poland separately, we add two more aspects: (iii)
the distinction between the crisis and non-crisis period underlines that the effect of different
factors on credit demand changed during the crisis; and (iv) we show that households’
exchange rate expectations are not a good measure of risk awareness, although this is usually
assumed in household survey-based research studies.
Our paper proceeds as follows. In the next chapter, we provide an overview of the existing
literature. Chapter 3 describes the dataset in detail and presents a factor analysis, which helps
to understand better the meaning of the variables. Methodology is elaborated in Chapter 4.
Chapter 5 contains all of the relevant results of our estimations for different periods. Finally,
the conclusions are presented in Chapter 6.

4/46
2. Literature review
Drivers of lending have been widely examined, especially since the onset of the crisis. One
major area of the research has focused on the factors explaining the trends in lending to the
entire non-financial sector. A large part of the studies mostly dealt with the supply side. Several
economists concluded that the high non-performing loan (NPL) portfolio may hinder lending
and constrain activity (e.g. Everaert et al., 2015; Temesváry and Banai, 2017), while a better
solvency situation was accompanied by stronger lending during the crisis (Popov and Udell,
2012; Frey and Kerl, 2011). It was also demonstrated that significant liquidity risk restrains
bank lending (e.g. Cornett et al., 2011).
Analysis of the demand side is more problematic in many respects, as less information is
available on the corporate sector, and particularly on the household sector, than on banks.
Models aimed at identifying demand-side and supply-side effects in the CESEE region
primarily focused on the corporate sector (e.g. Sóvágó, 2011; Brown et al., 2012). Everaert et
al. (2015) attempted to separate the demand-side and supply-side effects of household lending
using bank-level data and macro indicators, but they emphasised the uncertainty of their
findings. Although their estimation based on the entire CESEE region suggests the dominance
of supply effects, in the case of certain countries (e.g. Poland and Romania) they found that the
demand effects were stronger. Hence, it is crucially important to examine the motivations of
the households appearing on the demand side.
Relatively little information is available on the factors influencing household demand for
credit, and thus even in the literature the relevant studies are mainly based on questionnaires.
Beer et al. (2010) used a financial wealth survey of Austrian households conducted in 2004.
Their main question was what determines whether somebody plans to borrow in foreign
currency. Their results reflect that typically risk-taker, affluent and married customers show
preference to CHF housing loans. In addition, it should be noted that the borrowers of housing

5/46
loans are usually customers with a better income position. That is, those who took out a CHF
housing loan are in a better position than the population as a whole. Pellényi and Bilek (2009)
specifically examined the motivations of Hungarian foreign currency loan-holders based on a
custom household survey conducted in 2008. Their findings differed considerably from those
of Beer et al. (2010). Their results do not prove that foreign currency loan-holders have higher
qualifications or higher willingness to take risks. They demonstrated that it was very common
among the Hungarian households to borrow in foreign currency, mainly due to the high interest
rate difference and underestimation of the exchange rate risks.
Since 2007, the OeNB has conducted a survey (Euro Survey) in ten countries of the Central
and Eastern European region regarding households’ financial knowledge, attitudes and
expectations. This survey has served as a basis for several studies related to FX loan demand
(e.g. Beckmann et al., 2011; Fidrmuc et al., 2013; Beckmann et al.,2015). Fidrmuc et al. (2013)
supplemented the literature on household FX lending with a new aspect, when – in addition to
macro factors – they analysed households’ motivations. Most of the socio-demographic
variables used in the estimations were not significant. Only age had significant explanatory
power. Expectations about the stability of the local currency had a significant effect on the
willingness to borrow in FX. In addition, lack of trust in local financial institutions also had a
significant explanatory power. Those who planned to borrow in foreign currency regarded the
exchange rate more stable and income in foreign currency was more typical of them and they
deemed the introduction of the euro more likely in the medium term.
Literature on households’ credit demand in the CEE region is rather scarce. Almost every
influential paper examines the CEE region as a group of very homogenous countries. Changes
in behaviour over the last decade were also not examined. This paper would like to fill this gap
by comparing two countries and different periods. In addition, by using multiple imputation

6/46
techniques, this paper has significant added value for the literature both from an economics and
methodological point of view.
3. Data
3.1. Data description
The research was based on the “Euro Survey” conducted by the OeNB. The questionnaire has
been completed biannually
3
in 10 countries in the CEE region since autumn 2007. Starting
from 2012, the survey only included the relevant questions in the spring wave. The OeNB made
available for us the data from 12 questionnaires in the period between autumn 2007 and autumn
2014, both in the case of Hungary and Poland. The questionnaire, which is completed in both
countries by roughly 1,000 respondents in every period, is representative in terms of sex, age
and region of residence. The questionnaires completed in Poland are special in the sense that
before 2012 they cover the regions of the 10 largest cities and not the entire country (after
which the survey is representative for the country). Different respondents participated in the
different periods, and thus the database cannot be treated as a panel.
The questionnaire consists of five blocks (Fidrmuc et al., 2013):
4
In the first block, participants
are asked about their opinion on the economic prospects. In addition, this block also contains
questions on the respondents’ trust in specific institutions (domestic and foreign banks,
government, EU, police). The second group of questions focuses on savings and its
denomination, as well as on the cash holdings of the households. The third block covers plans
to borrow and existing loans. Within that, it differentiates between contracts of various
denominations. It should be noted that in terms of planned borrowing the questionnaire makes
no differentiation between loan categories. That is, the planned borrowing may include both
consumer and housing loans. The questions in block four focus on the use of foreign currency
3
Spring wave surveys are conducted in April and May, while fall wave surveys in October and November.
4
The definition of the variables used for our research can be found in Table A.1 of the Appendix.

7/46
and are not included in the questionnaire in each period. Block five includes questions on
banks’ lending behaviour. These questions were also included in the questionnaire only in some
of the periods; however, in order to obtain a full picture of the borrowers’ behaviour, it was
important to take them into consideration. Separate estimations were made for these periods,
as shown later.
The dependent variable describing credit demand was the answer to the question: “Do you plan
to take a loan in the next 1 year?” In Hungary, 8-9 per cent of the respondents answered yes to
this question before the crisis, after which this ratio decreased significantly. In Poland, it was
slightly higher, and the decline was also of a lesser degree than in Hungary. After 2009, it fell
from the pre-crisis level of 12-14 per cent to 9-12 per cent. For more details on each period
analysed, please see the descriptive statistics in Table F.1 – Table F.5 of the supplementary
materials.
3.2. Factors
It is often difficult to interpret the actual content of certain variables, which somewhat
complicates the use of the survey. Trust in specific institutions or giving preference to certain
banking products may be attributable to the level of education or the income position. For that
very reason, in order to ease the interpretation of the future results of the estimates, the variables
were also examined by a short factor analysis.
5
Based on the factors made for the various survey waves, it is clear both for Hungary and Poland
that the holding of current accounts and saving deposits is correlated with respondents’
education and income position. That is, respondents with higher income and higher education
are more likely to have a current account and saving deposits. In Hungary, households’ saving
ability seems to be connected to the aforementioned factors. Based on the Polish results, a
5
For a description of the applied methodology see Appendix B. The results of the factor analysis can be found in Table F.6 – Table F.11 of
the supplementary materials.

8/46
family is more likely to report savings capacity if it expects an improvement both in the
country’s economic situation and the family’s financial conditions. Based on the Hungarian
results this nexus does not appear to be obvious.
Based on the factors, it can also be established that in Hungary there is a strong correlation
between the opinion on economic prospects and exchange rate expectations. Respondents who
anticipate strengthening of the domestic currency also have positive macroeconomic
expectations. These respondents usually also gave a positive answer to the following question:
“Currently, the domestic currency is a very stable and trustworthy currency?” This may imply
that exchange rate expectations are determined by general impressions of the macroeconomy,
rather than by precise views about the FX market. Therefore, papers on FX lending based on
household surveys may misinterpret households’ exchange rate expectations: probably it is
more a general view on the economy and not a forecast of the exchange rate, which means that
it can be hardly used as the measure of risk awareness. The results also suggest that the question
about the stability of the domestic currency measures the trust in the currency, rather than
inflation or exchange rate expectations, and so in this case stability probably means that a
collapse of the exchange rate is unlikely. In addition, in several waves the household’s own
financial position also has a strong relation to the economic expectations. In Poland, there is
no clear indication that exchange rate expectations have a similar systematic relationship with
the economic outlook.
From the survey waves in 2010 H2 and 2011 H1, it appears that both in Hungary and in Poland
people who know somebody whose foreign currency loan instalment has increased are more
likely to (i) believe that foreign currency borrowing became riskier, (ii) perceived the
tightening of the lending conditions, and (iii) have risk-averse behaviour. In Hungary, the nexus
between risk-aversion and the experience regarding FX loans also proved to be notable in the
period between 2011 H2 and 2013 H2. It suggests that in Hungary – in contrast to Poland –

9/46
negative experiences regarding the FX loans had a longer lasting effect on risk-taking. It can
be explained – at least partly – by the fact that in Hungary households’ instalment did not only
rise because of the depreciating exchange rate, but also because of the higher interest rates
which were the result of banks’ practice of (legally) changing the interest rate on outstanding
FX loans. This may have further worsened households’ experience in relation to FX loans.
Finally, trust in various institutions also has a strong relation. Typically, those respondents who
trust the government, the police and the EU also trust the domestic and the foreign banks. For
Polish respondents these five factors in particular returned a very strongly separated factor. In
Hungary, this trust often correlates to trust in the financial system and confidence in the stability
of bank deposits. In addition, it should be noted that starting from 2009 (since the time when
these questions were available) the trust factor proved to be the strongest in all periods. In
Hungary especially it can be also seen that trust in the government moves together with the
economic prospects.
4. Methodology
One important feature of the database containing the responses to the Euro Survey – used for
the analysis and described above – is that many data points are missing. Moreover, the degree
of missingness is far from being negligible (Table A.2 of the Appendix, Table F.1 – Table F.5
of the supplementary materials). In the case of the variables included in the final estimation 0.1
to 22.1 per cent and 0.1 to 40.1 per cent of the data points are missing in the Hungarian and in
the Polish survey, respectively. If the regression estimation is performed on an incomplete
database in the simplest way, i.e. after deleting the cases that contain missing data (“listwise
deletion”), about 27 to 37 per cent and 32 to 65 per cent (depending on the examined time
horizon) of the sample narrowed to the Hungarian and Polish data, respectively, would be lost.
An analysis based on an incomplete database can lead to accurate estimates and correct
conclusions, if the problem arising from the missing data is properly handled; accordingly, this

10/46
chapter first briefly presents the advantages and disadvantages of the various approaches and
then describes the methodologies used for the analyses.
4.1. Handling missing data
There are several methods for handling the problem of missing data. It is a general conviction
(see e.g. Allison, 2009; Graham, 2009; Schafer and Graham, 2002) that a methodology
performs well if (i) it minimises the degree of bias, (ii) it relies on the widest possible
information base, and (iii) it provides a fair view of the degree of the estimation uncertainty,
i.e. it returns accurate standard error, confidence intervals and significance levels. While the
methods designated by Allison (2009) as conventional
6
breach some of the criteria mentioned
above, the (now widely accepted) advanced methods, such as the model-based Maximum
Likelihood (ML) and the Multiple Imputation (MI) perform well on the basis of the above
criteria. Under the same assumptions both MI and ML methods return consistent,
asymptotically efficient and asymptotically normally distributed estimations (Allison, 2009;
Allison, 2012). The MI and ML approaches are more efficient than the conventional methods
(Schafer and Graham, 2002; Graham, 2009), and therefore application of the latter ones is not
advisable even if an almost unbiased result could be expected of them subject to the fulfilment
of certain conditions.
Both methods assume that the missing data mechanism is random (MAR: missing at random),
i.e. the probability of a data point to be missing depends on the observations of the other
variables in the database, but does not depend on other unobserved information. That is, in the
case of MAR data mechanism the probability that a respondent answers the question depends
6
The conventional, simpler methods include, among others, the following: the deletion of the cases containing missing data in full (listwise
deletion) or by pair (pairwise deletion), creating a new category denoting the missing values (dummy-variable adjustment), inverse
probability weighting, the mean substitution and the regression mean imputation. For a detailed description of the conventional missing
data approaches see e.g. Schafer and Graham (2002), Graham (2009).

11/46
only on his own characteristics, included in the database (Schafer and Graham, 2002).
Although the missing data mechanism cannot be tested in a straightforward manner, the MI
and ML approaches often return unbiased result even if the MAR assumption does not hold
perfectly (Schafer and Graham, 2002), and it often hardly has any effect on the parameter
estimates and the standard errors (Collins et al., 2001). Furthermore, in the case of MI the
credibility of the MAR assumption is strengthened (Buuren et al., 1999), if a broad enough
range of independent variables is included in the imputation models.
The two most commonly used ML approaches are the Full Information Maximum Likelihood
(FIML) and the Expectation Maximisation (EM) algorithm (see Appendix C.). According to
Graham (2009) the best estimation for the average, the standard deviation, and the correlation
and covariance matrices is returned by the maximum likelihood estimate performed by the EM
algorithm, and he also emphasises that in case of missing data the covariance matrix obtained
by the EM estimate can be used very well for exploratory factor analysis. Therefore, we
handled the missing data problem during the factor analysis using this method. However, due
to the absence of appropriate standard errors it is not suitable for reliable hypothesis testing
(Graham, 2009), and hence for logit estimations.
The MI methodology is related to the study of Rubin (1987). The main idea is that – in contrast
to the simpler methods, which substitute the missing observation with a single value and after
imputation of the database treat the formerly missing observations as known – the missing data
points are substituted in multiple (m) ways and m databases are created as a result. Thereafter,
the model estimation (in our case a logistic regression) is performed on each of the created, m
complete databases, and finally the results are combined according to the so-called Rubin rule
(Rubin, 1987). In this way, with the multiple imputations the MI provides a more reliable
picture on the uncertainties of the imputations, as it considers not only the average estimation
uncertainties of the individual imputations, but – since it is not possible to fill in the missing

12/46
information in a straightforward manner – also the uncertainties between the imputations
(Rubin, 1996; Schafer and Graham, 2002; Stuart et al., 2009). During imputation, almost all
the information in the database is considered, since the purpose of the multiple imputation
(Little and Rubin, 1987; Rubin, 1987; Rubin, 1996) is to preserve the characteristics of the joint
distribution of the observed and unobserved data points and the associations among the
variables.
There are several methods within the MI approach. In order to select the most appropriate
method for our analysis, the number of variables containing missing observations, the pattern
of missingness (Lee and Carlin, 2010), and the type of the variables to be imputed must be
taken into consideration. Since there are several variables with incomplete data points in all
periods, we selected a multivariate MI approach. The pattern of missingness in the
questionnaire is arbitrary
7
, meaning that which variables are missing for a certain observation
is random (no specific pattern). Accordingly, use of the Joint Multivariate Normal (JMVN)
method based on the assumption that the joint distribution of the variables is multivariate
normal or the iterative Multiple Imputation by Chained Equations (MICE) method built on
univariable conditional distributions can be appropriate for the imputation of the database.
According to experience, the MICE approach works better for binary and category variables
than the JMVN method, as more reliable results can be obtained from the (here: logit) model
which is estimated in the analysis step of MICE (for more details see Appendix D.).
The MICE method
8
works similarly as the Gibb sampling algorithm, (see the description by
e.g. Casella and George, 1992) belonging to the family of the Markov chain-based Monte Carlo
(MCMC) approaches. MICE primarily differs from it in the sense that it does not assume the
7
The different types of missing data patterns are presented in detail for example by Schafer and Graham (2002).
8
The methodology is described by e.g. Stuart et al. (2009), Abayomi et al. (2008), Buuren et al. (1999), StataCorp (2013).

13/46
existence of a genuine multivariate distribution for all variables participating in the imputation
that correspond to the conditional specifications (StataCorp, 2013).
Instead of drawing simultaneously from a multivariate distribution, the MICE algorithm creates
the univariate distributions of all incomplete variables conditional on the other variables. For
all of the incomplete variables an assumption is needed with regard to the conditional
distribution of the missing data points: binary logistic and ordinal logistic regression is used
for binary and ordinal variables, respectively. The algorithm generates the imputed values that
substitute the missing observations by randomly drawing from the conditional distributions.
For further information on the methodology, see Appendix C.
4.2. Logit model
Based on the review of the methodological literature, we decided to use the MICE approach.
The broadest possible range of data was involved in the imputation step (for more details see
Appendix D.). We chose 25 and 50 as the number of imputations for Hungary and Poland,
respectively, except for the two last periods in the case of Poland, where – due to the larger
level of missingness – 75 and 65 imputations were run (Table A.2 of the Appendix).
In the analysis step, a logit model was fitted to the binary variable of the following question:
“Do you plan to take out a loan within the next year?” That is, similarly to Fidrmuc et al.
(2013), credit demand was separated from supply effects by measuring it with the intention to
borrow at the household level. The logit model can be expressed by the following equation:
          ,
where  contain the independent variables, while  is the K element vector of the coefficients,
and  is the distribution function belonging to the standard logistical distribution:   

Among the independent variables there is a question regarding households’ banking relations,
and there are questions related to the institutional and financial system, focusing on trust, as
well as expectations with regard to economic performance and the household’s financial

14/46
conditions. The variables regarding a household’s own expectations can better capture demand-
side factors than macroeconomic variables with historical data, as households’ opinion on
future economic prospects influence their borrowing decision to a greater extent than
aggregated macro-level data based on past events. In addition, all estimation equations include
a wide range of socio-demographic variables (namely education, income, region of residence,
employment status and size of the household) and time dummy variables which incorporate
the effect of the overall macroeconomic environment.
As there are several different variables for similar characteristics, we run numerous estimations
for both country and period. All of the estimations contain a similar set of socio-demographic
controls (namely education, income, region of residence, employment status and size of the
household), time dummy and different variables focusing on financial literacy or connection
with the financial system, trust in the financial system and institutional system, economic
expectations. In the next chapter, we show the best models for each period and each country.
We also present the best specifications for Hungary estimated on Polish data and vice versa.
5. Results
Estimates were made in respect of five different periods both for Hungary and Poland. The
obtained results are compared by periods. Model one was made for the full period of 2007 H2
– 2014 H2, model two for the pre-crisis period (2007 H2 – 2008 H2)
9
, model three (2009 H1
– 2010 H1) and model four (2009 H1 – 2014 H2) as post-crisis specifications, while model
five for 2010 H1 – 2011 H1. The differentiation of five periods is justified by two factors. On
the one hand, the crisis may have also resulted in an important change in borrowers’ behaviour,
and thus it is worth examining the period right before the crisis and after the crisis separately.
This is also supported by the fact that – according to the survey – the intention to borrow
9
This period was considered as the non-crisis period, since CEE countries were only hit by the crisis in the fourth quarter of 2008.

15/46
decreased significantly after the onset of the crisis. On the other hand, some variables are only
available for period five, but due to their relevant information content it was important to
involve them in the analysis.
5.1. Full period (2007 H2 – 2014 H2)
In the model covering the entire period, in the case of both Hungary and Poland similar factors
determine whether or not the given respondent plans to borrow (Table 1). In Hungary,
households which have bank relation in the form of a transaction account or saving deposits
are more likely to borrow. This alone does not mean that these households have significant
savings, i.e. the wealth is not the key driver. The existence of banking relations may imply
higher financial awareness and literacy or even stronger trust in the banks. The first is supported
by our factor analysis. Moreover, according to the literature (e.g. Hastings et al., 2013;
Beckmann and Stix, 2015), a causal relationship can be found between financial literacy and
banking relations (personal experience with either financial products or the banking system).
The situation is similar in Poland as well. In Poland, respondents with transaction accounts are
more likely to contemplate borrowing.
Expectations bear utmost importance in both countries. This is reflected in Hungary by the fact
that respondents who expected the exchange rate to strengthen were more likely to plan to
borrow. Naturally, this is also attributable to the fact that in the case of borrowing in foreign
currency the strengthening of the exchange rate is favourable. On the other hand, as this survey
suggests, most of the Hungarian borrowers became indebted in foreign currency at the bank’s
recommendation, and thus this fact alone may also imply the role of economic outlook. This is
also supported by the co-movements seen in the factors: the exchange rate expectations and
future economic outlooks were typically included in the same factor. Moreover, in the Polish
model the opinion on economic prospects appears explicitly as an independent variable which
has a significant effect on borrowing plans.

16/46
Table 1. Regression results (2007 H2 – 2014 H2)
2007 H2 – 2014 H2
Hungarian
specifications
Polish specifications
Hungarian
specifications on
Polish data
Polish specifications
on Hungarian data
H1
H2
P1
P2
P1'
P2'
H1'
H2'
HUF or PLN stable
0.019***
0.019***
0.016**
0.013**
0.019**
0.019***
0.022***
0.021***
(0.0071)
(0.0070)
(0.0070)
(0.0075)
(0.0071)
(0.0071)
(0.0056)
(0.0057)
Transaction account
0.015***
0.013***
0.058***
0.057***
0.051***
0.052***
0.016***
0.016***
(0.0048)
(0.0050)
(0.0092)
(0.0093)
(0.0073)
(0.0073)
(0.0059)
(0.0059)
HUF or PLN stay or
appreciate
0.021***
0.021***
0.008
0.007
(0.0048)
(0.0048)
(0.0080)
(0.0080)
EUR stable
0.018***
-0.004
(0.0068)
(0.0047)
FX cash
0.032***
0.036***
0.036***
0.036***
0.034***
0.034***
(0.0083)
(0.0109)
(0.0109)
(0.0109)
(0.0086)
(0.0086)
Saving deposit
0.017***
0.013
(0.0054)
(0.0091)
Economic outlook
0.018**
0.001
(0.0075)
(0.0047)
Employment status
Student
0.011
0.015*
-0.032***
-0.033***
-0.031***
-0.03***
0.012
0.012
(0.0078)
(0.0081)
(0.011)
(0.011)
(0.0110)
(0.0110)
(0.0079)
(0.0079)
Unemployed
0.03***
0.032***
0.05***
0.055***
0.055***
0.056***
0.029***
0.029***
(0.0067)
(0.0068)
(0.0142)
(0.0142)
(0.0141)
(0.0141)
(0.0066)
(0.0066)
Working
0.04***
0.042***
0.04***
0.04***
0.04***
0.041***
0.04***
0.04***
(0.0046)
(0.0045)
(0.0093)
(0.0093)
(0.0092)
(0.0092)
(0.0046)
(0.0047)
Self-employed
0.019*
0.021**
0.033*
0.033**
0.034**
0.036**
0.019*
0.019*
(0.0100)
(0.0101)
(0.0167)
(0.0168)
(0.0168)
(0.0169)
(0.0101)
(0.0101)
Region (HU)
Central Transdanubia
-0.005
0.005
-0.004
-0.004
(0.0072)
(0.0073)
(0.0073)
(0.0073)
West Transdanubia
-0.025***
-0.025***
-0.025***
-0.025***
(0.0062)
(0.0063)
(0.0062)
(0.0062)
South Transdanubia
-0.016**
-0.016**
-0.016**
-0.016**
(0.0073)
(0.0074)
(0.0072)
(0.0072)
North Hungary
-0.003
-0.004
-0.002
-0.002
(0.0074)
(0.0074)
(0.0074)
(0.0074)
North Great Plain
-0.005
-0.008
-0.005
-0.005
(0.0069)
(0.0068)
(0.0068)
(0.0068)
South Great Plain
-0.010
-0.012*
-0.010
-0.010
(0.0068)
(0.0068)
(0.0067)
(0.0067)
Region (POL)
Lodz
0.038***
0.040***
0.038***
0.041***
(0.0130)
(0.0132)
(0.0131)
(0.0131)
Trojmiasto –
Pomerania
-0.010
-0.010
-0.010
-0.009
(0.0118)
(0.0118)
(0.0118)
(0.0118)
Szczecin - West
Pomerania
0.029*
0.029*
0.029*
0.034**
(0.0152)
(0.0153)
(0.0153)
(0.0155)
Silesian Agglomeration
– Silesia
0.017
0.015
0.014
0.016
(0.0101)
(0.0100)
(0.0100)
(0.0099)
Cracow - Lesser Poland
0.052***
0.051***
0.050***
0.052***
(0.0142)
(0.0141)
(0.0141)
(0.0140)
Poznan - Greater
Poland
0.004
0.004
0.003
0.004
(0.0132)
(0.0132)
(0.0132)
(0.0131)

17/46
Wroclaw - Lower
Silesia
0.018
0.018
0.017
0.020
(0.0129)
(0.0129)
(0.0129)
(0.0128)
Bydgoszcz - Kuyavia-
Pomerania
0.021
0.019
0.019
0.022
(0.0162)
(0.0160)
(0.0161)
(0.0163)
Lublin
-0.027**
-0.028**
-0.029**
-0.028**
(0.0124)
(0.0123)
(0.0123)
(0.0121)
Household size
Household size (3 or
more)
0.017***
0.017***
0.002
0.003
0.003
0.003
0.017***
0.017***
(0.0045)
(0.0045)
(0.0069)
(0.0069)
(0.0069)
(0.0069)
(0.0045)
(0.0045)
Education
Medium education
-0.007
-0.006
-0.012
-0.011
-0.011
-0.011
-0.006
-0.006
(0.0057)
(0.0056)
(0.0106)
(0.0106)
(0.0106)
(0.0106)
(0.0056)
(0.0056)
High education
0.003
0.004
-0.006
-0.006
-0.006
-0.005
0.005
0.005
(0.0072)
(0.0072)
(0.0122)
(0.0122)
(0.0122)
(0.0122)
(0.0072)
(0.0073)
Income level
Medium income
-0.001
-0.002
-0.006
-0.007
-0.007
-0.007
-0.001
-0.001
(0.0058)
(0.0058)
(0.0080)
(0.0080)
(0.0080)
(0.0080)
(0.0057)
(0.0058)
High income
0.004
0.004
-0.001
-0.002
-0.002
-0.002
0.005
0.005
(0.0060)
(0.0060)
(0.0097)
(0.0097)
(0.0096)
(0.0097)
(0.0060)
(0.0060)
Note: Average partial effects are listed in the first row, standard errors are reported in the row below in parentheses
and the corresponding significance levels are in the adjacent column. * significant at 10 percent, ** significant at
5 percent, *** significant at 1 percent. Time dummies were included in all specifications. Base categories: retired
(employment status), Central Hungary (region HU), Warsaw (region POL), less than three (household size), low
(education), low (income).
Our findings are also in line with the study of Fidrmuc et al. (2013). In Hungary, similarly to
the expectations, trust in the institutional system and in the economy also determines loan
applications. Based on the previously presented factor analysis, it is quite likely that the
question related to the stability of forint (“the local currency is stable and trustworthy”)
measures trust in the domestic currency, rather than inflation or exchange rate expectations.
Otherwise, in this case stability does not mean that the currency cannot depreciate, but suggests
that a collapse of the exchange rate is unlikely. For this very reason, the positive coefficient
confirms that trust in the institutions and in the economy increases credit demand.
Previous studies based on the Euro Survey typically focused on the entire Central and Eastern
European region and thus did not fully utilise the existing socio-demographic variables, e.g.
the information on the residence of the respondent. Our results based on the full period show
significant divergence between different regions, but there is no clear pattern. We cannot see
any evidence on the connection between economic performance of certain regions and the
borrowing intentions of their citizens. The employment classification showed a strong effect.

18/46
Loan demand was larger in all categories than that of the retired people and students in
Hungary, with the employed having the strongest intentions. Even unemployed people
exhibited higher loan demand than retired persons, students or the self-employed, which may
imply that the intention to borrow was also influenced by liquidity constraints. By contrast, in
Poland students’ credit demand was relatively low and the borrowing intention among
unemployed people was even stronger than among employed ones.
5.2. Pre-crisis period (2007 H2 – 2008 H2)
The crisis changed not only the situation of the banking system, but also the behaviour of the
demand side. For example, the intention to borrow fell dramatically from 2009. Fidrmuc et al.
(2013) controlled for the crisis period with a dummy variable. However, we think that the crisis
not only switched the level of demand, but at the same time relations between variables may
have also changed (several papers have shown the effect of the crisis on different participants,
e.g. Cull and Martínez Pería, 2013; De Haas and van Lelyveld, 2014). For this very reason, we
also examined the three biannual periods immediately preceding the crisis, as well as the three
half-years after the crisis. In the case of Hungary, the results of the pre-crisis period resemble
the estimation made on the full sample in many respects (Table 2). Those with a transaction
account are more likely to borrow than those without a transaction account. Those who expect
no exchange rate depreciation and assess the economic outlook positively are also more likely
to plan to borrow. Finally, the trust in the domestic currency also boosts the intention to borrow.
Similar statements can be made about Poland as well, since the transaction account and positive
economic expectations had significant positive explanatory power in the model.

19/46
Table 2. Regression results (2007 H2 – 2008 H2)
2007 H2 – 2008 H2
Hungarian
specifications
Polish specifications
Hungarian
specifications on
Polish data
Polish specifications
on Hungarian data
H1
H2
P1
P2
P1'
P2'
H1'
H2'
HUF or PLN stable
0.035**
0.035**
0.015
0.014
(0.0171)
(0.0171)
(0.0144)
(0.0146)
Transaction account
0.023**
0.025**
0.060***
0.059***
0.059***
0.024**
(0.0113)
(0.0112)
(0.0143)
(0.0144)
(0.0144)
(0.0114)
HUF or PLN stay or
appreciate
0.033**
0.034**
-0.013
-0.012
(0.0139)
(0.0139)
(0.0175)
(0.0146)
Prefer cash
0.017*
0.014
(0.0105)
(0.0142)
Economic outlook
0.026*
0.033**
(0.0143)
(0.013)
Employment status
Student
0.004
0.005
-0.078***
-0.029*
-0.069***
-0.069***
0.003
0.004
(0.0142)
(0.0146)
(0.0227)
(0.0158)
(0.023)
(0.023)
(0.0134)
(0.0146)
Unemployed
0.025*
0.025*
0.018
0.022
0.024
0.024
0.026*
0.023
(0.0141)
(0.014)
(0.028)
(0.028)
(0.0282)
(0.0281)
(0.014)
(0.0141)
Working
0.067***
0.068***
0.02
0.019
0.02
0.02
0.071***
0.063***
(0.0105)
(0.0105)
(0.0215)
(0.0212)
(0.0212)
(0.0212)
(0.0108)
(0.0106)
Self-employed
0.068*
0.063*
-0.035
-0.039
-0.041
-0.039
0.076*
0.069*
(0.0357)
(0.0345)
(0.0588)
(0.0560)
(0.0548)
(0.0553)
(0.0388)
(0.0366)
Note: Average partial effects are listed in the first row, standard errors are reported in the row below in parentheses
and the corresponding significance levels are in the adjacent column. * significant at 10 percent, ** significant at
5 percent, *** significant at 1 percent. Time and regional dummies were included in all specifications. Controls
on education, household size and income level were also included. The base category is retired in the case of
employment status.
However, the role of the place of residence differs substantially from that seen in the estimation
made for the entire period. In Hungary, the South and North Great Plain, as well as the Central
Transdanubia regions – all of which are significantly poorer than the Central Hungary region
– showed stronger credit demand before the crisis. This confirms the allegation that in the
period immediately preceding the crisis the banking system financed riskier borrowers (e.g.
Balás et al., 2015 or Banai et al., 2010). Moreover, this also implies that the shift to the riskier
segments was justified by strengthening demand. In Poland, in contrast to the full sample, no
significant difference can be seen between regions. Only in Wroclaw was the intention to
borrow stronger than in the Warsaw region. In the case of Poland before the crisis, only
students’ demand differed significantly (negatively) from that of pensioners. In Hungary, it

20/46
was just the opposite: the employed, self-employed and unemployed persons all exerted a
significant (positive) effect on credit demand.
5.3. Post-crisis periods (2009 H1 – 2010 H1 and 2009 H1 – 2014 H2)
The Euro Survey was launched in 2007 and so we have much longer dataset for the post crisis
period than we have for the pre-crisis era. For this reason, we decided to investigate two periods
separately: the first period (from 2009 H1 to 2010 H1) contains information on the abrupt effect
of the crisis; while the second period (from 2009 H1 to 2014 H1) may highlight the potential
regime shift after the crisis.
10
There are significant differences between the two countries in the post-crisis period. In
Hungary, most of the factors examined in this study had a much stronger effect just after the
crisis than over the longer term in the post-crisis world. Our results show that, in the models
run on Hungarian data, significance levels are lower and coefficients are smaller for almost
every variable when estimating for the longer period between 2009 H1 and 2014 H1. In the
case of Poland, this is not obvious since some variables became more important for the longer
period, while some others had strong explanatory power only for the shorter one.
After the onset of the crisis, trust in the financial system dropped significantly in almost every
country. For this reason, we conducted a separate analysis on the effect of trust in the banking
system (Table A.3 of the Appendix). The Polish results confirm that trust in the banking system
is an important driver of household credit demand. This was true in Poland both over the shorter
horizon and over the longer horizon as well after the onset of the crisis. In Hungary, on the
contrary, trust was only an important driver just after the onset of the crisis. It is also important
that in the Hungarian case only the trust in foreign banks had a significant effect on the intention
to borrow.
10
Polish specifications on Hungarian data and vice versa can be found in Table F.12 and Table F.13 of the supplementary materials.

21/46
Some other factors were also examined for both countries. In the case of Hungary, the
economic outlook (expected development of the forint exchange rate) and expectations of the
respondents' own financial situation appeared distinctly (Table 3), in addition to banking
relations (transaction account, saving deposit). Respondents who expected their financial
position to improve are more likely to want to borrow than others. The significant effect of the
exchange rate in Hungary again supports our view on the possible explanation of this variable.
As we stated in the factor analysis, it captures a general view on the economy. Since there was
no demand for FX loans from 2009 new borrowers were not affected by exchange rate
movement. The role of exchange rate expectation cannot be interpreted as risk awareness.

Table 3. Regression results for Hungary (2009 H1 – 2010 H1 and 2009 H1 – 2014 H2)
Hungary
2009 H1 – 2010 H1 specifications
2009 H1 –2014 H2 specifications
Specification:
H1 short
H1 long
H2 short
H2 long
H1 short
H1 long
H2 short
H2 long
H3 short
H3 long
HUF stable
0.002
0.012*
0.003**
0.012*
(0.0136)
(0.0072)
(0.0148)
(0.0071)
Transaction account
0.022**
0.009*
0.020**
0.008
0.023**
0.009*
(0.0099)
(0.0053)
(0.0100)
(0.0055)
(0.0096)
(0.0054)
HUF stay or appreciate
0.033***
0.017***
0.032***
0.016***
0.037***
0.018***
0.038***
0.017***
0.030***
0.016***
(0.0098)
(0.0048)
(0.0097)
(0.0048)
(0.0095)
(0.0047)
(0.0102)
(0.0047)
(0.0098)
(0.0048)
FX cash
0.038**
0.031***
0.044***
0.035***
0.042**
0.036***
0.041****
0.028***
(0.0158)
(0.0089)
(0.0164)
(0.0092)
(0.0168)
(0.0093)
(0.0164)
(0.0088)
Future financial situation
0.038***
0.022***
0.035***
0.020***
0.040***
0.021***
(0.0129)
(0.0060)
(0.0126)
(0.0059)
(0.0131)
(0.0060)
Risk aversion
-0.009**
-0.012*
(0.0141)
(0.0065)
Saving deposit
0.008**
0.011*
(0.0106)
(0.0057)
Trust in foreign banks
Neutrality in relation to foreign banks
-0.026*
-0.001
-0.024*
0.000
-0.026**
-0.001
-0.029**
-0.001
(0.0131)
(0.0062)
(0.0129)
(0.0061)
(0.0132)
(0.0061)
(0.0143)
(0.0062)
Distrust in foreign banks
-0.037***
-0.011*
-0.034**
-0.009
-0.038***
-0.011*
-0.047***
-0.013**
(0.0136)
(0.0062)
(0.0137)
(0.0062)
(0.0139)
(0.0063)
(0.0149)
(0.0063)
Employment status
Student
0.012
0.013
0.008
0.011
0.012
0.013
0.015
0.012
0.010
0.013
(0.0178)
(0.0090)
(0.0173)
(0.0087)
(0.0185)
(0.0090)
(0.0193)
(0.0087)
(0.0174)
(0.0089)
Unemployed
0.048***
0.031***
0.047***
0.031***
0.050***
0.032***
0.056***
0.031***
0.050***
0.032***
(0.0155)
(0.0076)
(0.0156)
(0.0076)
(0.0160)
(0.0076)
(0.0174)
(0.0075)
(0.0156)
(0.0076)
Working
0.034***
0.031***
0.033***
0.030***
0.034***
0.031***
0.041***
0.032***
0.037***
0.032***
(0.0118)
(0.0051)
(0.0120)
(0.0052)
(0.0118)
(0.0051)
(0.0126)
(0.0051)
(0.0117)
(0.0050)
Self-employed
-0.005
0.007
-0.007
0.005
-0.006
0.006
-0.001
0.007
-0.003
0.006
(0.0164)
(0.0093)
(0.0157)
(0.0088)
(0.0158)
(0.0090)
(0.0172)
(0.0091)
(0.0163)
(0.0088)

Note: “Short” refers to the data between 2009 H1 and 2010 H1, while “Long” refers to the data between 2009 H1
and 2014 H2. Average partial effects are listed in the first row, standard errors are reported in the row below in
parentheses and the corresponding significance levels are in the adjacent column. * significant at 10 percent, **
significant at 5 percent, *** significant at 1 percent. Time and regional dummies were included in all
specifications. Controls on education, household size and income level were also included. The base category is
retired in case of employment status.
In Poland, immediately after the onset of the crisis the credit demand was determined by trust
in banks or trust in the EU, by existing banking relations and by the fact whether the respondent
prefers cash (Table 4). The signs also developed in accordance with the intuitions: the trust in
foreign banks and in the EU, and existing banking relations all have a positive effect on credit
demand. Examining the full post-crisis period, households’ expectations about their future
financial situation and the economic outlook also appeared to be significant factors.
The role of the place of residence changed during the crisis in both countries. While in Hungary
before the crisis credit demand increased in the regions with poorer economic performance,
during the crisis it was clearly the most developed Central Hungary region where the demand
was the strongest. In Poland before the crisis only one region, i.e. the relatively well-performing
Wroclaw, had a significant positive effect on credit demand, while after the crisis two less
developed regions, i.e. Lodz and Krakow, showed stronger intent to borrow than the others.
Prior to the crisis, in Hungary employed persons were more likely to plan to borrow compared
to the unemployed, while after the crisis the intention to borrow was already higher among the
unemployed compared to the employed. In Poland, before the crisis the employed and
unemployed people’s credit demand did not differ significantly from that of pensioners, but
during the crisis – similarly to the Hungarian results – the unemployed had relatively stronger
credit demand compared to the employed. Based on the results, the tightening of liquidity
constraint probably raised credit demand for the unemployed. Implicitly, this may also imply
that the respondents anticipated quick resolution of their poor labour market position.

Table 4. Regression results for Poland (2009 H1 – 2010 H1 and 2009 H1 – 2014 H2)
Poland
2009 H1 – 2010 H1 specifications
2009 H1 – 2014 H2 specifications
P1 short
P1 long
P2 short
P2 long
P3 short
P3 long
P1 short
P1 long
P2 short
P2 long
PLN stable
-0.033**
0.016**
-0.039**
0.011
-0.014
0.018**
(0.0153)
(0.0078)
(0.0159)
(0.0079)
(0.0153)
(0.0077)
Transaction account
0.100***
0.051***
0.099***
0.052***
0.097***
0.051***
0.099***
0.050***
0.099***
0.050***
(0.0137)
(0.0086)
(0.0137)
(0.0086)
(0.0138)
(0.0086)
(0.0138)
(0.0087)
(0.0138)
(0.0087)
EUR stable
0.041**
0.024***
0.036**
0.028***
(0.0165)
(0.0080)
(0.0157)
(0.0077)
Prefer cash
0.029*
-0.002
(0.0160)
(0.0079)
Trust in EU
-0.073***
-0.032***
-0.065***
-0.028***
(0.0161)
(0.0078)
(0.016)
(0.0078)
Trust in foreign banks
Neutrality in relation to foreign banks
-0.019
-0.016*
(0.0188)
(0.0092)
Distrust in foreign banks
-0.052***
-0.038***
(0.0192)
(0.0092)
PLN stay or appreciate
-0.014*
0.015*
-0.009
0.016*
(0.0189)
(0.0088)
(0.0183)
(0.0087)
FX cash
0.028*
0.047***
0.030***
0.047***
(0.0275)
(0.0136)
(0.0275)
(0.0136)
Economic outlook
0.008
0.020**
(0.0159)
(0.0079)
Employment status
Student
-0.027
-0.015
-0.028
-0.015
-0.029
-0.015
-0.028
-0.016
-0.027
-0.015
(0.0261)
(0.0129)
(0.0261)
(0.0129)
(0.0258)
(0.0129)
(0.0261)
(0.0129)
(0.0260)
(0.0129)
Unemployed
0.100***
0.074***
0.099***
0.075***
0.104***
0.076***
0.103***
0.073***
0.103***
0.072***
(0.0364)
(0.0166)
(0.0363)
(0.0167)
(0.0365)
(0.0167)
(0.0364)
(0.0166)
(0.0363)
(0.0165)
Working
0.040*
0.051***
0.039*
0.050***
0.040*
0.051***
0.039*
0.050***
0.041*
0.051***
(0.0210)
(0.0098)
(0.0212)
(0.0098)
(0.0211)
(0.0098)
(0.0212)
(0.0098)
(0.0211)
(0.0098)

25/46
Self-employed
0.033
0.049***
0.034
0.048***
0.036
0.048***
0.042
0.046***
0.044
0.047***
(0.0351)
(0.0172)
(0.0354)
(0.0172)
(0.0356)
(0.0171)
(0.0372)
(0.0172)
(0.0372)
(0.0172)
Poland
2009 H1 – 2014 H2 specifications
P3 short
P3 long
P4 short
P4 long
P5 short
P5 long
P6 short
P6 long
Transaction account
0.100***
0.050***
0.097***
0.049***
0.096***
0.048***
0.099***
0.054***
(0.0137)
(0.0087)
(0.0139)
(0.0088)
(0.0140)
(0.0088)
(0.0136)
(0.0085)
EUR stable
0.045***
0.031***
(0.0154)
(0.0076)
PLN stay or appreciate
-0.015
0.019*
-0.016
0.017**
-0.017
0.017**
(0.0178)
(0.0085)
(0.0176)
(0.0085)
(0.0177)
(0.0085)
FX cash
0.025
0.047***
0.027
0.048***
0.027
0.048***
(0.0271)
(0.0136)
(0.0272)
(0.0136)
(0.0273)
(0.0136)
Trust in EU
-0.068***
-0.034***
(0.0154)
(0.0076)
Trust in foreign banks
Neutrality in relation to foreign banks
-0.026
-0.018*
(0.0191)
(0.0092)
Distrust in foreign banks
-0.062***
-0.041***
(0.0189)
(0.0090)
Future financial situation
0.011
0.014*
(0.0157)
(0.0077)
Employment status
Student
-0.028
-0.016
-0.030
-0.016
-0.030
-0.016
-0.028
-0.014
(0.0261)
(0.0129)
(0.0262)
(0.0129)
(0.0258)
(0.0129)
(0.0261)
(0.0131)
Unemployed
0.102***
0.073***
0.097***
0.073***
0.104***
0.073***
0.102***
0.074***
(0.0365)
(0.0166)
(0.0364)
(0.0166)
(0.0367)
(0.0166)
(0.0363)
(0.0166)
Working
0.039*
0.050***
0.037*
0.051***
0.039*
0.051***
0.040*
0.050***
(0.0213)
(0.0098)
(0.0213)
(0.0098)
(0.0212)
(0.0098)
(0.0211)
(0.0099)
Self-employed
0.041
0.045***
0.032
0.045***
0.034
0.045***
0.043
0.049***
(0.0368)
(0.0171)
(0.0355)
(0.0171)
(0.0358)
(0.0171)
(0.0371)
(0.0173)

Note: “Short” refers to the data between 2009 H1 and 2010 H1, while “Long” refers to the data between 2009 H1
and 2014 H2. Average partial effects are listed in the first row, standard errors are reported in the row below in
parentheses and the corresponding significance levels are in the adjacent column. * significant at 10 percent, **
significant at 5 percent, *** significant at 1 percent. Time and regional dummies were included in all
specifications. Controls on education, household size and income level were also included.
5.4. Period between 2010 H1 and 2011 H1
In 2010–2011 in Hungary banking relations, economic expectations and trust in the forint still
had significant explanatory power (Table 5). In this period, the opinion on banks’ lending
policy, and the experiences related to foreign currency loans could be also considered. The
latter clearly has a negative effect on households’ credit demand. This may be attributable to
the fact that in Hungary foreign currency borrowers’ increased burdens due to exchange rate
movements and increased interest rates was already a widely discussed topic in 2010. In
addition, it is also an important result that those clients who deem the banks’ standards to have
tightened also exhibit weaker credit demand. Households which regard banks’ credit conditions
as strict are less inclined to borrow. This justifies the phenomenon, which has been widely
examined especially in relation to enterprises (e.g. MNB, 2015 or ECB, 2014), according to
which self-selection has a strong effect on lending. Households do not even try to apply for
credit because they feel that they would not get it anyway.
The estimation result for Poland differs to the greatest degree from that of Hungary in this
period. The most important explanatory factors are still trust in the institutional system,
expectations regarding the economic performance and own financial position and banking
relations. Moreover, household’s risk appetite also matters: households which can be
characterised by risk-averse behaviour are more likely to plan to borrow. In the Polish case,
the problems related to foreign currency loans substantially lagged behind those seen in
Hungary (as explained in Chapter 3.2.), and thus it does not influence their borrowing
decisions. In addition, the views on lending conditions also do not have a significant effect on
intention to borrow. The role of the regions also differs in the two countries: in Hungary credit

27/46
demand in most of the regions is clearly lower than in the Central Hungary region, by contrast
in the case of Poland credit demand was the strongest in two regions with relatively poor
economic performance, while other regions’ borrowing plans did not differ significantly from
Warsaw’s.
To summarise, we can see that there are numerous similarities in the factors that drive credit
demand in the two countries, but there are also differences in several areas. Banking relations
(which may imply financial awareness), trust in the institutional system, risk appetite and
macroeconomic expectations have significant explanatory power. In addition, in Hungary two
other important factors also appeared after the crisis: the negative experiences with foreign
currency loans and self-selection. The role of the place of residence also represents a difference
between the two countries. The region of the respondent often had a significant effect on credit
demand, and thus taking that into consideration carries extra information. The effect of
exchange rate expectations also represents a difference. In Hungary, when respondents
expected a stable or strengthening exchange rate it had a positive effect on credit demand
throughout all periods. Based on the factors, it seemed that exchange rate expectations exhibit
strong correlation with general economic expectations, and thus this is not surprising. Our
models supported this finding since the outlook for the exchange rate significantly affected
loan demand not only before the crisis (when it was reasonable due to potentially decreasing
instalments on FX loans), but from 2009 as well. On the other hand, in Poland the positive
effect of the exchange rate expectations did not appear in the pre-crisis period, it became
significant in 2010-2011. The differences mentioned above confirm the importance of creating
models separately by countries, as well as by periods.
Finally, we analysed the robustness of our results from many aspects: (i) we checked whether
the drivers of credit demand are different in the case of a sample containing only those
households that do not have a loan; (ii) some of our estimations were rerun using two additional

28/46
methods; and (iii) we changed the number of imputations in case of the benchmark multiple
imputation method. Our results proved to be robust according to these robustness tests (detailed
analysis can be found in Appendix E).
Table 5. Regression results (2010 H1 – 2011 H1)
2010 H1 – 2011 H1
Hungarian specifications
Polish specifications
H1
H2
H3
H4
P1
P2
P3
HUF or PLN stable
0.023*
0.021*
0.025*
0.024*
(0.0127)
(0.0127)
(0.0130)
(0.0130)
Transaction account
0.018*
0.018*
0.019**
0.019**
0.054***
0.055***
0.053***
(0.0094)
(0.0094)
(0.0091)
(0.0091)
(0.0168)
(0.0165)
(0.0165)
EUR stable
0.051***
(0.0149)
Distrust in EU
-0.031**
-0.033**
(0.0146)
(0.0146)
Trust in foreign banks
Neutrality in relation to foreign banks
-0.003
(0.0175)
Distrust in foreign banks
-0.054***
(0.0173)
HUF or PLN stay or appreciate
0.031***
0.031***
0.031***
0.030***
0.054***
0.056***
0.054***
(0.0072)
(0.0072)
(0.0072)
(0.0072)
(0.0162)
(0.0161)
(0.0163)
FX cash
0.039**
0.038**
(0.0152)
(0.0151)
Future financial situation
0.032*
0.038**
0.034**
(0.0166)
(0.0166)
(0.0154)
Risk aversion
-0.057**
-0.044**
(0.0224)
(0.0213)
Saving ability
-0.031**
-0.027*
(0.0159)
(0.0160)
Saving deposit
0.02*
0.019**
(0.0101)
(0.0103)
Tighter loan standards
-0.031**
-0.028**
(0.0144)
(0.0141)
FX loan experience
-0.026*
-0.022*
(0.0137)
(0.0130)
2010 H1 – 2011 H1
Hungarian specifications on Polish data
Polish specifications on Hungarian
data
P1'
P2'
P3'
P4'
H1'
H2'
H3'
HUF or PLN stable
0.009
0.008
0.008
0.008
(0.0142)
(0.0142)
(0.0142)
(0.0141)
Transaction account
0.054***
0.055***
0.053***
0.054***
0.020**
0.020**
0.020**
(0.0166)
(0.0165)
(0.0166)
(0.0166)
(0.0089)
(0.0089)
(0.0088)
EUR stable
-0.013
(0.0082)

29/46
Distrust in EU
-0.007
-0.009
(0.0082)
(0.0084)
Trust in foreign banks
Neutrality in relation to foreign banks
0.003
(0.0104)
Distrust in foreign banks
-0.011
(0.0103)
HUF or PLN stay or appreciate
0.062***
0.061***
0.062***
0.061***
0.034***
0.035***
0.033***
(0.0160)
(0.0162)
(0.0160)
(0.0162)
(0.0074)
(0.0074)
(0.0074)
FX cash
0.002
0.001
(0.0218)
(0.0219)
Future financial situation
0.004
0.002
0.004
(0.0084)
(0.0085)
(0.0082)
Risk aversion
-0.021*
-0.021*
(0.0121)
(0.0118)
Saving ability
0.009
0.008
(0.0108)
(0.0107)
Saving deposit
-0.008
-0.008
(0.0179)
(0.0179)
Tighter loan standards
-0.023
-0.024
(0.0216)
(0.0216)
FX loan experience
-0.013
-0.013
(0.0161)
(0.0161)
Note: Average partial effects are listed in the first row, standard errors are reported in the row below in parentheses
and the corresponding significance levels are in the adjacent column. * significant at 10 percent, ** significant at
5 percent, *** significant at 1 percent. Time and regional dummies were included in all specifications. Controls
on education, household size, income level and employment were also included.
6. Conclusion
This paper identifies drivers of households’ credit demand in Hungary and Poland. Several
logit models were estimated, where dependent variable was credit demand measured by the
binary variable of households’ borrowing intention in the next year. We used the Euro Survey
data created by the OeNB, which was available for the period between 2007 H2 and 2014 H2.
This paper relies most on the work of Fidrmuc et al. (2013), though the main focus is not
households’ FX borrowing, but merely on borrowing in general, and it goes beyond that paper
in some respects.
One key challenge is the high proportion of missing observations, which reduces the reliability
of the results if the problem is not handled properly. We used multiple imputation, an advanced
method that is more reliable and efficient than conventional methods such as the listwise

30/46
deletion and the creation of a new missing category which were used in earlier studies. Our
results confirm that the outcome of the two methods differs partially, i.e. proper treatment of
the missing variables is indeed important. This can be a particularly important result for
researchers using household surveys. In addition, the earlier picture is further refined by the
fact that Hungary and Poland were examined separately. Finally, separate estimates were made
for different periods, which helped understand the impacts of the crisis better.
Our results show that both in Hungary and in Poland borrowing decisions are primarily
determined by three factors: existing banking relations, which have a strong relation to
financial awareness and financial literacy; macroeconomic expectations, which also have a
connection with the personal financial situation; and trust in the institutional system. Regarding
the latter, trust in the EU is particularly important. Credit demand is stronger among those who
trust the EU.
The effects of the place of residence on borrowing plans appeared to be strong. In Hungary,
the results of the estimations regarding the regions’ role suggests that in the pre-crisis period
the credit market was indeed flooded by less and less creditworthy borrowers. During the crisis,
in Hungary the labour market situation also appeared among the factors that considerably
influenced credit demand. Unemployed people’s intention to borrow was stronger immediately
following the onset of the crisis than that of employed persons, which may have been
attributable to the strong liquidity constraints. Households’ economic expectations, banking
relations and their trust in the institutional system had a much stronger effect on their credit
demand just after the crisis than the whole post-crisis period, particularly in Hungary. This
implies that households’ credit demand – particularly in a crisis period – can be influenced
through affecting their trust in the financial system and their beliefs about the country’s
economic prospect.

31/46
The largest difference between the Polish and Hungarian results appeared in the later phase of
the crisis (2010 H2 – 2011 H2). There were two factors that proved to be significant for
Hungary and not for Poland. First, negative experience with the foreign currency loans reduced
credit demand. This may be attributable to the fact that in Hungary the problems related to
foreign currency loans were already a widely discussed topic in 2010. This was because in
Hungary CHF lending was very widespread and much bigger problems stemmed from it. In
order to mitigate the probability of the occurrence of such a phenomenon that has long-lasting
negative impacts, in the future banks should lend in a more responsible manner, while
households’ confidence in the financial system could be increased by strengthening and
expanding the institution of financial consumer protection. Second, results confirm that in
Hungary self-selection may also play a strong role in household lending. Those households
that regard banks’ credit conditions as strict are less inclined to borrow.
Acknowledgements:
We would like to thank the Austrian central bank for providing us with Euro Survey. We thank
Peter Backe, Elisabeth Beckmann, Thomas Scheiber and Helmut Stix for their help and useful
comments. We are also grateful for the participants of the 2016 SSEM Conference in Porto,
the 2016 Seminar in the Austrian National Bank and the 2017 ECEE Conference Tallinn. We
also express our gratitude to the colleagues at the Financial System Analysis Directorate of the
Magyar Nemzeti Bank (the central bank of Hungary) for their useful comments.

References
Allison, P.D. 2001. Missing data - Quantitative applications in the social sciences, Thousand
Oaks, CA: Sage, Vol. 136.
Allison, P.D. 2009. Missing data. In: Millsap, R. E., Maydeu-Olivares, A. (Eds.), The SAGE
Handbook of Quantitative Methods in Psychology, pp. 72-89. Thousand Oaks, CA: Sage
Publications Inc.
Allison, P.D. 2012. Handling Missing Data by Maximum Likelihood, keynote presentation at
the SAS Global Forum, April 23, 2012, Orlando, Florida.
Balás, T., Banai, A., Hosszú, Zs. 2015. Modelling probability of default and optimal PTI level
by using a household survey. Acta Oeconomica, 2015, vol. 65, issue 2, 183-209.
Banai, Á., Király, J. and Nagy, M. 2010. The demise of the halcyon days in Hungary: foreign
and local banks – before and after the crisis. BIS Papers No. 54, December 2010.
Barceló, C. 2006. Imputation of the 2002 Wave of the Spanish Survey on Household Finances
(EFF). Documentos Ocasionales, No 0603. Banco de España.
Bodner, T.E. 2008. What Improves with Increased Missing Data Imputations? Structural
Equation Modeling: A Multidisciplinary Journal, 15(4):651–75.
Beckmann, E., Roitner, A., Stix, H. 2015. A Local or a Foreign Currency Loan? Evidence on
the Role of Loan Characteristics, Preferences of Households and the Effect of Foreign Banks.
Focus on European Economic Integration. Oesterreichische Nationalbank (Austrian Central
Bank). Issue 1, pages 24-48.
Beckmann, E., Scheiber, T., Stix, H. 2011. How the Crisis Affected Foreign Currency
Borrowing in CESEE - Microeconomic Evidence and Policy Implications, Focus on European
Economic Integration, 2011 Q1. Available at: https://www.oenb.at/dam/jcr:f966cd52-b7f4-
44eb-9f9e-d9937d4a781a/feei_2011_q1_studies_02_tcm16-224902.pdf

33/46
Beckmann, E. and Stix, H. 2015. Foreign currency borrowing and knowledge about exchange
rate risk. Journal of Economic Behavior & Organization. Elsevier, vol. 112(C), pages 1-16.
Beer, C., Ongena, S., Peter, M., 2010. Borrowing in foreign currency: Austrian households as
carry traders. Journal of Banking and Finance 34, 2198–2211.
Brown, M., Ongena, S., Popov, A., Yesin, P. 2012. Who Needs Credit and Who Gets Credit in
Eastern Europe? ECB Working Paper No. 1421.
Casella, G. and George, E.I. 1992. Explaining the Gibbs Sampler. The American Statistician,
Vol. 46, No. 3., pp.167-174.
Collins, L.M., Schafer, J.L. and Kam, C.K. 2001. A comparison of inclusive and restrictive
strategies in modern missing-data procedures. Psychol Methods, 2001;6(4):330–351.
Cornett, M.M., McNutt, J.J., Strahan, P.E., Tehranian, H. 2011. Liquidity risk management and
credit supply in the financial crisis. Journal of Financial Economics 101, pp.297-312.
Cull, R. and Martínez Pería, M.S. 2013. Bank ownership and lending patterns during the 2008-
2009 financial crisis: Evidence from Latin America and Eastern Europe. Journal of Banking &
Finance 37(12), pp. 4861-4878.
De Haas, R. and van Lelyveld, I. 2014. Multinational banks and the global financial crisis:
Weathering the perfect storm. Journal of Money, Credit and Banking 46(1), pp.333-364.
Dempster, A.P., Laird, N.M., and Rubin, D.B. 1977. Maximum likelihood estimation from
incomplete data via the EM algorithm (with discussion). Journal of the Royal Statistical
Society, Series B, 39, pp.1-38.
Everaert, G., Che, N., Geng, N., Gruss, B., Impavido, G., Lu, Y., Saborowski, C.,
Vandenbussche, J. and Zeng, L. 2015. Does Supply or Demand Drive the Credit Cycle?
Evidence from Central, Eastern, and Southeastern Europe. IMF Working Paper WP/15/15.

34/46
Fidrmuc, J., Hake, M., Stix, H. 2013. Households’ foreign currency borrowing in Central and
Eastern Europe. Journal of Banking & Finance, Volume 37, Issue 6, June 2013, Pages 1880-
1897.
Frey, R. and Kerl, C. 2011. Multinational banks in the crisis: Foreign affiliate lending as a
mirror of funding pressure and competition on the internal capital market. Journal of Banking
& Finance 50, pp.52-68.
Graham J.W., Olchowski, A.E. and Gilreath, T.D. 2007. How many imputations are really
needed? Some practical clarifications of multiple imputation theory. Prevention science: the
official journal of the Society for Prevention Research, 2007;8(3):206–213.
Graham, J.W. 2009. Missing data analysis: making it work in the real world. Annual Review
of Psychology 60, 549–576.
Hastings, J., Madrian, B., Skimmyhorn, W., 2013. Financial Literacy, Financial Education and
Economic Outcomes. Annual Review of Economics 5, 347-373.
Household Finance and Consumption Network. 2013. The Eurosystem Household Finance and
Consumption Survey - Methodological Report for the First Wave. ECB Statistics paper series
No. 1.
Jones, M P 1996. Indicator and stratification methods for missing explanatory variables in
multiple linear regression. Journal of the American Statistical Association, 91: 222–230.
Kaiser, H.F. 1958. The varimax criterion for analytic rotation in factor analysis. Psychometrika
23: 187–200.
Kennickell, A. 1991. Imputation of the 1989 Survey of Consumer Finances: Stochastic
Relaxation and Multiple Imputation. Available at: https://www.irs.gov/pub/irs-soi/1991-
1992preprintar30.pdf
Kennickell, A. 1998. Multiple Imputation in the Survey of Consumer Finances. Available at:
https://www.federalreserve.gov/econresdata/scf/files/impute98.pdf

35/46
Kropko, J., Goodrich, B., Gelman, A. and Hill, J. 2014. Multiple Imputation for Continuous
and Categorical Data: Comparing Joint Multivariate Normal and Conditional Approaches.
Political Analysis (Autumn 2014) 22 (4): 497-519.
Lee, K.J. and Carlin, J.B. 2010. Multiple Imputation for Missing Data: Fully Conditional
Specification Versus Multivariate Normal Imputation. American Journal of Epidemiology,
171(5): 624-632.
Little, R.J.A. and Rubin, D.B. 1987. Statistical analysis with missing data. New York: John
Wiley & Sons.
Magyar Nemzeti Bank 2015. Senior loan officer survey on bank lending practices. November
2015, Magyar Nemzeti Bank.
Pellényi, G., Bilek, P., 2009. Foreign Currency Borrowing: The Case of Hungary. Working
Paper D.5.4, FINESS.
Popov, A. and Udell, G.F. 2012. Cross-border banking, credit access and the financial crisis.
Journal of International Economics 87 (2012), pp.147-161.
Rao, C.R. 1955. Estimation and tests of significance in factor analysis. Psychometrika, Vol.
20, pp.93-111.
Rubin, D.B. 1987. Multiple imputation for nonresponse in surveys. New York: John Wiley &
Sons.
Rubin, D.B. 1996. Multiple imputation after 18+ years, Journal of the American Statistical
Association, 91, pp.473-489.
Schafer, J.L. 1997. Analysis of incomplete multivariate data. London: Chapman & Hall.
Schafer, J.L. and Graham, J.W. 2002. Missing data: our view of the state of the art.
Psychological Methods, 2002 7(2):147-177.
Sóvágó, S. 2011. Identifying supply and demand in the Hungarian corporate loan market. MNB
Occasional Papers, No. 94.

36/46
StataCorp 2013. STATA Multiple Imputation Reference Manual, Release 13. College Station,
TX: Stata Press.
Stuart, E.A., Azur, M., Frangakis, C. and Leaf, P. 2009. Multiple Imputation With Large Data
Sets: A Case Study of the Childrens Mental. American Journal of Epidemiology, Vol. 169,
Issue 9, pp.1133-1139.
Temesváry, J. and Banai, Á. 2017. The Drivers of Foreign Bank Lending in Central and Eastern
Europe: The Roles of Parent, Subsidiary and Host Market Traits. Journal of International
Money and Finance. Accepted Manuscript. Available at:
http://www.sciencedirect.com/science/article/pii/S0261560617301699
Truxillo, C. 2005. Maximum likelihood parameter estimation with incomplete data. SAS
Institute Inc. 2005, Proceedings of the Thirtieth Annual SAS® Users Group International
Conference, Paper 111-30, Cary, NC: SAS Institute Inc.
Van Buuren, S., Boshuizen, H.C. and Knook D.L. 1999. Multiple Imputation of Missing Blood
Pressure Covariates in Survival Analysis. Statistics in Medicine, 18(6): 681-694.

37/46
Appendix A. Tables
Table A.1 Definition of variables
Name
Definition
Loan plan
1=plan to take out a loan within the next year; 0=otherwise.
Economic outlook
1=in the next 5 years the economic situation in my country will improve; 0=do not agree.
HUF/PLN stable
1=the local currency is stable and trustworthy; 0=do not agree.
Bank deposit safety
1=depositing money at banks is very safe in their country; 0=do not agree.
Prefer cash
1=prefer to hold cash rather than a savings account; 0=do not agree.
EUR stable
1=the euro is a very stable and trustworthy currency; 0=do not agree.
FX cash
1=have FX cash holdings; 0=otherwise.
Transaction account
1=have a transaction account; 0=otherwise.
Savings deposit
1=have a savings deposit; 0=otherwise.
Debit card
1= have a debit card; 0=otherwise.
Exchange rate expectation
1=the HUF or PLN exchange rate versus EUR will strengthen or remain unchanged in the next 5 years; 0=do
not agree.
Higher FX deposit interest
rate
1= interest rates on foreign currency savings deposits are higher than interest rates on local currency
saving deposits; 0=do not agree.
PPP of FX savings deposits
1= savings deposits in foreign currency are better to safeguard the value of my money than savings
deposits in local currency; 0=do not agree.
Future financial situation
1= over the next 12 months I expect the financial situation of my household to improve; 0=do not agree.
Good financial situation
1= currently the financial situation of my household is good; 0=do not agree.
High inflation expectation
1= over the next years, prices will rise strongly in my country; 0=do not agree.
Inflation expectation
prices compared to the previous 12 months will: 1= increase faster; 2=increase at the same rate;
3=increase at a slower rate; 4=remain unchanged; 5=decrease.
Stable financial system
1= banks and the financial system are stable in my country; 0=do not agree.
FX deposits are usual
1= in my country it is very common to hold foreign currency deposits; 0=do not agree.
Foreign bank deposit is safer
1= savings deposits at foreign banks are much safer than those at domestic banks; 0=do not agree.
Distrust in government
1= trust in the government; 2=neutral to the government; 3=do not trust in the government.
Distrust in police
1= trust in the police; 2=neutral to the police; 3=do not trust in the police.
Distrust in domestic banks
1= trust in the domestic banks; 2=neutral to the domestic banks; 3=do not trust in the domestic banks.
Distrust in foreign banks
1= trust in the foreign banks; 2=neutral to the foreign banks; 3=do not trust in the foreign banks.
Distrust in EU (ordinal)
1= trust in the EU; 2=neutral to the EU; 3=do not trust in the EU.
Distrust in EU
1=do not trust in the EU or neutral to it; 0=otherwise.
Saving ability
1= my household is able to save money; 0=otherwise.
Risk aversion
1= in financial matters, I prefer save investments over risky investments; 0=do not agree.
Tighter loan standards
1= over the last 2 years, banks have become very strict in granting loans; 0=do not agree.
EUR loan riskiness
1= over the last 2 years taking out a loan in EUR has become riskier because of possible exchange rate
depreciations; 0=do not agree.
EUR loan preference
1= taking everything into account, loans in EUR are more attractive than local currency loans; 0=do not
agree.
FX loan experience
1=I know someone who ran into difficulties with a foreign currency loan because repayments became
much higher than expected; 0=do not agree.
Education
1=low education; 2=medium-level education; 3=high-level education.
Income
1=low income; 2=medium income; 3 =high income.
Employment status
1=retired; 2=student; 3=unemployed; 4=working; 5=self-employed.
Household size (3 or more)
1=my household is comprised of more than two persons; 0=otherwise.
Head of household
1=the respondent is the breadwinner; 0=otherwise.
Region (HU)
1=Central Hungary; 2=Central Transdanubia; 3=West Transdanubia; 4=South Transdanubia; 5=North
Hungary; 6=North Great Plain; 7=South Great Plain.
Region (POL)
1=Warsaw - Masovia; 2=Lodz; 3=Trojmiasto - Pomerania; 4=Szczecin - West Pomerania; 5=Silesian
Agglomeration - Silesia; 6=Krakow - Lesser Poland; 7=Poznan - Greater Poland; 8=Wroclaw - Lower Silesia;
9=Bydgoszcz - Kuyavia-Pomerania; 10=Lublin.

Table A.2 Missingness and number of imputations applied for each period
Hungary
Number of
observations
(Logit)
Number of
observations
(MI logit)
Missingness
in the sample
(%)
Missingness of the variables (%)
Number of
imputations
Minimum
Median
Maximum
2007 H2 – 2014 H2
8140
12117
33
0
1
21
25
2007 H2 – 2008 H2
2218
3050
27
0
2
31
25
2009 H1 – 2010 H1
1986
3032
34
0
5
28
25
2010 H1 – 2011 H1
1960
3034
35
0
7
30
25
2009 H1 – 2014 H2
5690
9067
37
0
4
22
25
Poland
Number of
observations
(Logit)
Number of
observations
(MI logit)
Missingness
in the sample
(%)
Missingness of the variables (%)
Number of
imputations
Minimum
Median
Maximum
2007 H2 – 2014 H2
7185
12380
42
0
7
32
50
2007 H2 – 2008 H2
2098
3105
32
0
10
43
50
2009 H1 – 2010 H1
1694
3113
46
0
16
45
50
2010 H1 – 2011 H1
1100
3134
65
0
13
39
75
2009 H1 – 2014 H2
4267
9275
54
0
10
33
65
Table A.3 Regression results: domestic vs foreign banks
2009 H1 - 2010 H1
2009 H1 - 2014 H2
Hungary (H1)
Poland (P3)
Hungary (H1)
Poland (P5)
Foreign
Domestic
Foreign
Domestic
Foreign
Domestic
Foreign
Domestic
HUF or PLN stable
0.012*
0.011
(0.0072)
(0.0072)
Transaction account
0.022**
0.023**
0.097***
0.095***
0.009*
0.009
0.048***
0.049***
(0.0099)
(0.0096)
(0.0138)
(0.0153)
(0.0054)
(0.0054)
(0.0088)
(0.0087)
EUR stable
0.036**
0.038**
(0.0157)
(0.0157)
Trust in banks
Neutrality in relation
to banks
-0.026*
-0.002
-0.019
0.000
-0.001
-0.002
-0.018*
-0.008
(0.0131)
(0.0111)
(0.0188)
(0.0170)
(0.0061)
(0.0056)
(0.0092)
(0.0085)
Distrust in banks
-0.037***
-0.002
-0.052***
-0.059***
-0.011*
-0.009
-0.041***
-0.045***
(0.0136)
(0.0120)
(0.0192)
(0.0172)
(0.0063)
(0.0059)
(0.0090)
(0.0086)
HUF or PLN stay or
appreciate
0.033***
0.031***
0.018***
0.018***
0.017**
0.015*
(0.0098)
(0.0102)
(0.0047)
(0.0048)
(0.0085)
(0.0084)
FX cash
0.035***
0.036***
0.048***
0.048***
(0.0092)
(0.0092)
(0.0136)
(0.0138)
Future financial
situation
0.038***
0.043***
(0.0129)
(0.0134)
Note: Average partial effects are listed in the first row, standard errors are reported in the row below in parentheses
and the corresponding significance levels are in the adjacent column. * significant at 10 percent, ** significant at
5 percent, *** significant at 1 percent. Time dummies and controls on regions, employment status, education,
income level and household size were included in all specifications.

39/46
Table A.4 Robustness tests on the estimation methodology
LOGIT
MI MVN LOGIT
MI MICE LOGIT
MI MICE LOGIT (5 imp)
Hungary
2007 H2 – 2014 H2
Observations: 8140
Observations: 12117
Observations: 12117
Observations: 12117
Imputations: 0
Imputations: 50
Imputations: 50
Imputations: 5
APE
Std. Err.
APE
Std. Err.
APE
Std. Err.
APE
Std. Err.
HUF stable
0.026
***
0.0087
0.015
**
0.0061
0.019
***
0.0071
0.018
***
0.0071
Transaction account
0.012
*
0.0063
0.015
**
0.0062
0.015
***
0.0048
0.015
***
0.0048
Exchange rate expectation
0.019
***
0.0054
0.022
***
0.0049
0.021
***
0.0048
0.021
***
0.0050
FX cash
0.035
***
0.0097
0.028
***
0.0063
0.032
***
0.0084
0.032
***
0.0083
Hungary
2010 H1 – 2011 H1
Observations: 1960
Observations: 3034
Observations: 3034
Observations: 3034
Imputations: 0
Imputations: 75
Imputations: 75
Imputations: 5
APE
Std. Err.
APE
Std. Err.
APE
Std. Err.
APE
Std. Err.
HUF stable
0.042
**
0.0181
0.024
**
0.0098
0.019
**
0.0091
0.025
*
0.0130
Transaction account
0.031
***
0.0108
0.023
*
0.0128
0.025
*
0.0130
0.019
**
0.0091
Exchange rate expectation
0.030
***
0.0091
0.030
***
0.0085
0.031
***
0.0074
0.030
***
0.0073
FX cash
0.042
**
0.0194
0.035
***
0.0104
0.039
***
0.0153
0.040
***
0.0152
Tighter loan standards
-0.035
**
0.0179
-0.024
**
0.0104
-0.029
**
0.0144
-0.028
**
0.0141
2007 H2 – 2014 H2
Poland
Observations: 7185
Observations: 12380
Observations: 12380
Observations: 12380
Imputations: 0
Imputations: 50
Imputations: 50
Imputations: 5
APE
Std. Err.
APE
Std. Err.
APE
Std. Err.
APE
Std. Err.
Transaction account
0.049
***
0.0113
0.055
***
0.0089
0.058
***
0.0093
0.060
***
0.0100
EUR stable
0.021
**
0.0084
0.018
**
0.0072
0.018
***
0.0069
0.017
**
0.0070
FX cash
0.035
***
0.0131
0.030
***
0.0091
0.036
***
0.0110
0.035
***
0.0108
PLN stable
0.009
0.0081
0.015
**
0.0069
0.016
**
0.0070
0.014
**
0.0068
2010 H1 – 2011 H1
Poland
Observations: 1100
Observations: 3134
Observations: 3134
Observations: 3134
Imputations: 0
Imputations: 75
Imputations: 75
Imputations: 5
APE
Std. Err.
APE
Std. Err.
APE
Std. Err.
APE
Std. Err.
Transaction account
-0.003
0.0367
0.052
**
0.0209
0.051
***
0.0168
0.053
***
0.0174
EUR stable
0.091
***
0.0225
0.050
***
0.0162
0.051
****
0.0149
0.057
***
0.0146
Exchange rate expectation
0.035
0.0240
0.045
**
0.0181
0.054
***
0.0162
0.056
***
0.0173
Risk aversion
-0.033
0.0327
-0.050
***
0.0177
-0.057
**
0.0224
-0.048
**
0.0235
Saving ability
-0.043
*
0.0244
-0.027
0.0172
-0.031
**
0.0159
-0.036
**
0.0156
Future financial situation
0.035
0.0261
0.031
*
0.0169
0.032
*
0.0166
0.031
*
0.0175
Neutrality in relation to
foreign banks
0.022
0.0265
-0.001
0.0162
-0.003
0.0175
-0.002
0.0188
Distrust in foreign banks
-0.022
0.0269
-0.054
***
0.0189
-0.054
***
0.0173
-0.054
***
0.0183
Note: Time dummies and controls on regions, employment status, education, income level and household size
were included in all specifications. APE means average partial effects. * significant at 10 percent, ** significant
at 5 percent, *** significant at 1 percent.

40/46
Appendix B. Methodology of Factor Analysis
The FIML (also known as Direct Maximum Likelihood) and EM approaches, which are
described in detail by Allison (2009), assume multivariate normal distribution and MAR data
mechanism, and require a large sample for the estimation. We used the latter one for the factor
analysis, as the parameter estimates obtained by the two ML approaches are identical (Allison,
2009), and the FIML approach works better only in case of analysis that use standard errors
(e.g. hypothesis testing), but for the factor analysis there is not any advantage to use it.
The EM is a two-step, iterative method (Dempster et al., 1977), which in the first (expectation)
phase defines the expected value of the loglikelihood considering only the complete
observations, and then in the second (maximisation) phase maximises the likelihood function.
In the first step, the EM algorithm substitutes the missing observations by values estimated by
simple regression imputation built on the observed data. The regression equations, which are
built on all the incomplete variables, include all variables other than the dependent variable of
the given regression as independent variables. In the second step of the approach the parameter
estimations are performed based on the database imputed in the first step. Thereafter, using the
estimated covariance matrix, a new regression equation is created for each incomplete variable,
which results in new, more reliable estimations (with regard to the missing data points), which
are used by the algorithm during each iteration for the overwriting of the initially missing
observations. The two phases of the EM approach are repeated until the parameter estimates
generated by the consecutive iteration converge, i.e. the elements of the covariance matrix no
longer change substantially (Allison, 2001; Graham, 2009).
According to the baseline model of factor analysis the standardised variables can be expressed
as the sum of the linear combination of the common factors and the unique factors:
   ,

41/46
where  is the observations’ n*p matrix,  is the common factors’ n*q matrix,  is the factor
weights’ p*q matrix and  is the unique factors’ n*p matrix, where n denotes the number of the
observations, p the number of the variables and q the number of the factors.
It is assumed that (i) the shared factors are independent, (ii) the shared factors and the error
term are uncorrelated, and (iii) the error terms are independent of each other. The factor
analysis is based on the correlation matrix of the variables, from the definition of which the
baseline equation of the factor analysis can be derived, taking the aforementioned assumptions
into consideration:
   ,
where  is the variables’ p*p correlation matrix,  is the shared factors’ q*q correlation matrix
(presumed identity matrix), and  is the error terms’ p*p variance matrix.
Due to the special attributes of the database presented before, we did not apply the conventional
(pairwise deletion) approach here either, but rather estimated the variance-covariance matrix
of the variables with the aforementioned maximum likelihood-based EM algorithm and based
on that (similarly to Truxillo, 2005) performed the maximum likelihood factor analysis
(MLFA). Although the ML factor analysis assumes a multivariate normal distribution, the
approach is equivalent to Rao’s (1955) canonical factor analysis, and it can also be applied
when the normality assumption does not hold perfectly (StataCorp, 2013). In order to ease the
interpretation of the results, the factor matrix is rotated orthogonally, after which the common
factors remain uncorrelated. The varimax rotation (Kaiser, 1958) approach used by us
transforms the factor matrix containing factor loadings into a simpler structure by maximising
the variance explained by the factors.
Appendix C. Handling missing data problem in the case of binary variables
In the case of binary variables, although the assumption of the EM approach (used for the factor
analysis) does not hold perfectly with regard to the normal distribution, Allison (2009) – citing

42/46
a number of simulations and practical experiences (e.g. Schafer, 1997) – deems the application
of the EM approach appropriate in this case as well. In the example of Allison (2009), the
independent variables include both continuous and dummy variables and, although the standard
errors obtained by the EM approach are much higher than the real value, the parameter
estimates used in the factor analysis are accurate.
In case of missing data, multiple imputation methods are appropriate for the estimation of a
binary choice model. When choosing from these, it is worth bearing in mind that the dependent
variable of the model to be examined in the analysis step is binary, and the other variables of
the database are also not continuous. Allison (2009) points out that – although there are several
examples for the imputation of the category variables by the JMVN method – it is not
necessarily good if we intend to estimate a logistic regression model on a binary dependent
variable in the analysis step. Kropko et al. (2014) provide a very wide comparison of the JMVN
and MICE approaches. The MICE approach never performed worse than the JMVN approach
when multiple imputation was performed by the two approaches using a real database and the
assumption of multivariate normal distribution did not hold. In the case of continuous variables,
the two methods returned almost identical results, while in the case of the ordinal variables
MICE performed slightly better, and in the case of (non-ordinal) category variables it
performed much better than the MVN approach. In the case of a binary variable, which is
relevant for us, the two approaches performed very similarly, even though the JMVN method’s
assumption regarding the multivariate normal distribution was not fulfilled for the database.
Since within the framework of the JMVN method first an EM type maximum likelihood
estimation is performed to define the initial values of the data augmentation based the MCMC
(Markov Chain Monte Carlo) method (StataCorp, 2013), the result achieved by Kropko et al.
(2014) is important in the factor analysis. Presumably, during the EM-based maximum
likelihood estimate we perform mainly for binary variables, which returns the covariance

43/46
matrix necessary for the factor analysis, it does not cause a problem that the assumption of
multivariate normal distribution does not hold perfectly. The authors argue that if we want to
perform a regression analysis following the imputation step, use of the MICE method is
recommended, even though the JMVN approach returns a faster and similar result, because it
may occur that under the MICE method different conclusions are drawn based on the
significance tests built on the somewhat more accurately estimated coefficients and standard
errors than in the case of the JMVN method.
Appendix D. Imputation process
Apart from the variables included in the final regression – considering the correlation between
the variables and the ratio of missing observations – efforts were made to include in the
imputation the highest possible number of variables available in the database. Basically, it is
recommended to use all available information, as based on the studies by Buuren et al. (1999),
Allison (2009) and Collins et al. (2001) it can be stated that (i) the accuracy of the imputation
and thereby the efficiency of the parameter estimates may be substantially improved by the
independent variables included in the imputation (apart from the variables of the final
regression), and (ii) it helps reduce the uncertainty and minimise the degree of the bias (the
missingness can be related to other attributes of the household), and finally (iii) a sufficiently
broad range of independent variables strengthens the authenticity of the MAR assumption.
The order of the variables during the imputation depends on the ratio of missing observations.
The algorithm first imputes the variable with the smallest ratio of missingness, including all
complete variables in the model as independent variables. This is followed by the imputation
of the variable that has the second fewest missing observations, in the course of which the
imputation model also considers the first imputed variable as independent variable in addition
to the complete variables. The process is continued in this way until finally the variable with
the highest number of missing observations is also imputed, and by the end of the first iteration

44/46
a fully imputed database is created. During the first iteration different numbers of observations
were used for each incomplete variable. Due to this, upon creating each imputed database there
is a so-called burn-in period (let us indicate the number of iterations by b), at the end of which
the initially incomplete database is supplemented with the result of iteration b and thereby the
first imputed database is created. The purpose of the burn-in period is to achieve the stability
of the imputation process, i.e. to remove the effect of the order in which the imputation of the
incomplete variables is performed. The second imputation starts after iteration b, which will
also comprise of b number of iterations. Thus, in total the algorithm performs b*m iterations
(where m denotes the number of imputations), and in each iteration n regressions are estimated
(where n denotes the number of the incomplete variables participating in the imputation).
Initially, the literature deemed very few, 2-10 imputations to be sufficient (see e.g. Rubin, 1987;
Rubin, 1996) if a relatively small volume of information is lost. Graham et al. (2007) concluded
that far more imputations may be required than was previously deemed sufficient: with a
missingness of 50 per cent they thought at least 40 imputations necessary. Bodner (2008)
recommends that the number of the imputations to be performed be determined depending on
the percentage of the missingness. In this spirit, we decided to choose a different number of
imputations for the two countries, because the survey of Polish households is much more
incomplete than the survey of Hungarian households. As increasing the number of imputations
has no disadvantage other than the increase in the time needed for calculations, we determined
the number of imputations in a conservative manner.
Appendix E. Robustness checks
We presented a detailed analysis of our results in Chapter 5. Examination of different
specifications is of special importance, as it confirmed that our results are robust for different
sets of explanatory variables. In this section, we focus on the performance of the method we

45/46
used. Moreover, we made estimations on a restricted sample in which only households which
do not have a loan are included.
First, some of our estimations were also made by two additional methods, namely: a simple
logit regression using listwise deletion; and another multiple imputation method, the JMVN
method (Table A.3 of the Appendix). The first period covering the whole sample and the period
between 2010 H1 and 2011 H1 were examined this way. We took special interest in the latter,
since the missingness was the highest in this subsample. Accordingly, how to handle the
missing data problem is a particulary important question. We made two findings based on the
estimations: 1. Different multiple imputations give similar results. Significances and signs are
close to each other in the two estimations. Also, in terms of the size of the coefficients and
standard errors, the two MI methods are quite similar to each other. 2. There is a bigger
difference in case of the listwise deletion method, though the deviation is not drastic. While
the decision on the variables’ significances is almost the same, the listwise deletion method
results in rather different coefficient and standard error values compared to the MI methods.
Missingness is much higher in the Polish case, so applying the proper method should have even
more relevance. Our results support this idea. Especially for the estimations for the period
between 2010 H1 and 2011 H1 there is a huge difference between the applied methods. This is
true even for the two multiple imputation techniques and underlines the need for proper
examination of the database and a precise decision as to the applied method, as we
demonstrated in Chapter 4.
Next, we examined the effect of the number of imputations on the results (last column in Table
A.3 of the Appendix). We can see only minor differences between the benchmark MICE
estimations and the new estimations based on a database in which variables are imputed only
five times. Our conclusions regarding the significance of the analysed motives of credit demand
remain the same. This suggests that in case of large surveys where multiple imputation can be

46/46
very time-consuming – such as the SCF and HFCS – the typically applied number of 5
imputations seems to be sufficient.
Finally, we re-estimated several models using a sample that consists of only households which
do not have a loan. It is worth pointing out that in this case we lose approximately one-third of
the observations and the representativeness of the sample is questionable. Nonetheless, we
found that our results are quite robust to this change. There are only two motives that – contrary
to the results of the main models – do not influence households’ credit demand if they do not
have a loan, and both only appear in the Hungarian case: (i) households’ opinion about the
stability of the forint; and (ii) distrust in foreign banks. This suggests that these two drivers of
credit demand mainly influence the borrowing decisions of households which had a foreign
currency loan, as they were primarily interested in exchange rate fluctuations and they had
relationship with typically foreign banks since in Hungary most FX loans were provided by
foreign banks. It can be also stated that households’ personal experience – not just information
received from others – regarding FX lending is particularly important.