n = 2; 3; 4; 5:

TAG:  credit card rate 

Published Time: -

Filetype: pdf

Filesize: 80513

1 Consumer Credit Conditions in the U.K.
Emilio Fernandez-Corugedo* and John Muellbauer** This version 8 May 2005 (SP)
* Centre for Central Banking Studies, Bank of England.
E-mail: emilio.fernandez.corugedo@bankofengland.co.uk

** Nuffield College.
E-mail: john.muellbauer@nuf.ox.ac.uk




This paper represents the views and analysis of the authors and should not be thought to represent
those of the Bank of England or Monetary Policy Committee members. We are grateful to
Sebastian Barnes, Pru Cox, Robert Hamilton, Philip Thomas and Jan Vlieghe for discussions
about our data, to Charles Goodhart and seminar participants at the Bank of England, the Federal
Reserve Board, the Money, Macro and Finance Conference, and at Oxford for comments, and to
Ken Wallis for encouragement. We are also grateful to Janine Aron, Ian Bond, Jens Larsen,
Simon Whitaker and two anonymous referees for extensive comments, to Rachel Pigram for
excellent research assistance, to Mary Gregory and Sarah Voitchovsky for extracting data from
the New Earnings Survey, and to Bruno Muellbauer for technical assistance. Mark Boleat
advised on the historical record. Research by John Muellbauer was supported partly by the
ESRC under grant no. R000237500.





Copies of working papers may be obtained from Publications Group, Bank of England,
Threadneedle Street, London, EC2R 8AH; telephone 020 7601 4030, fax 020 7601 3298, e-mail
mapublications@bankofengland.co.uk

Working papers are also available at www.bankofengland.co.uk/wp/index.html



The Bank of England's working paper series is externally refereed.
© Bank of England 2005 ISSN 1368-5562 2 3 Contents Abstract 5 Summary 7 1 Introduction 9 2 Financial liberalisation in the U.K. 11 3 Analysing information on credit conditions from the Survey of Mortgage Lenders 12 3.1 Distributions of loan-to-income and loan-to value ratios and limits set by lenders 12 3.2 The empirical methodology 15 3.3 Economic factors impinging on PLIR 16 3.4 Economic factors impinging on PLVR 18 3.5 Functional forms for PLIR and PLVR equations 19 4 Economic factors impinging on aggregate unsecured and mortgage debt 20 5 Empirical results: the basic model 22 5.1 The equation for unsecured debt 24 5.2 The equation for mortgage debt 27 5.3 The equations for PLIR 30 5.4 The equations for PLVR 31 5.5 The estimated RISK 34 5.6 The estimated CCI 34 6 Empirical results: interaction effects with the Credit Conditions Index and RISK 36 7 Conclusion 39 Appendix 42 Figures 46 Tables of results 49 References 62 4 5 Abstract


It is widely perceived that credit supply conditions faced by U.K. consumers, particularly in the
mortgage market, have been liberalised since the late 1970s, with implications for the housing
market and consumer spending. This paper examines quarterly micro-data from the Survey of
Mortgage Lenders (SML) to learn about changes in credit conditions from loan-to-value ratios
(LVRs) and loan-to-income ratios (LIRs) of first-time buyers (classified by region and age). It
combines data on the proportions of high LVR and high LIR loans with aggregate information on
U.K. consumer credit and mortgage debt to give ten quarterly series for 1975-2001. These are
modeled in a ten-equation system. A comprehensive set of economic and demographic
influences on the demand and supply of credit, applying relevant sign restrictions, are controlled
for. A single time-varying index of credit conditions captures the common variation in the ten
credit indicators which cannot be explained by the economic and demographic controls. The
broad coverage of credit market indicators and thorough investigation of economic forces driving
the credit market should make the resulting Credit Conditions Index more robust than previous
estimates. The index increases in the 1980s, peaking towards the end of the decade. It retraces
part of this rise in the early 1990s, before increasing again to levels, for one of the two measures,
exceeding the previous peak. The index will be useful in modelling consumption and the housing
market, and in interpreting current monetary conditions. An important by-product of the paper is
the model for consumer credit and mortgage debt developed here.



JEL classification: C32, E44, E51, G21. 6 7 Summary
It is widely perceived that credit supply conditions for U.K. consumers have been liberalised
since the late 1970s, with implications for the housing market and consumer spending. For
example, the evidence is that consumption and the housing market (with changes in credit
availability likely to have contributed) were important factors in the economic boom of the late
1980s and the subsequent recession of the early 1990s.

The need for a credit-conditions index (CCI), which measures credit availability, other than
through the level of interest rates, has been recognised in the consumption literature. Proxies
such as unsecured credit to income ratios and interest rate spreads have been used in empirical
work. However, such proxies are unsatisfactory because they depend on the economic
environment. This paper constructs a CCI, which, as far as possible, is free of this endogeneity
criticism because it controls for the effects of the economic environment. The usage of the term
‘Credit Conditions Index’ rather than ‘Financial Liberalisation Index’ is due to the former’s
connotations of outcome, rather than process. Nevertheless, the process of financial deregulation
and other changes in financial architecture have had an important bearing on credit availability.

The paper constructs a CCI for households between 1976 and 2001. The index is derived as a
common factor in ten credit indicators. Priors derived from the history of institutional changes in
the credit markets are used to guide the estimation of the CCI. Two of the ten credit indicators
are aggregate unsecured debt and mortgages (secured debt). The remaining eight consist of the
fractions of high loan-to-income and high loan-to-value mortgages for U.K. first-time house
buyers split by age and regions. The paper argues that mortgage defaults largely arise from the
coincidence of having a poor debt/equity position and experiencing cash flow problems.
Mortgage lenders thus limit initial loan-to-value and loan-to-income ratios to control default risk.
The paper uses these arguments to derive econometric models for the fractions of first-time
buyers with respectively high loan-to-value ratios and high loan-to-income ratios. It also derives
specifications for aggregate unsecured and mortgage debt in the context of previous literature.
The attention to expectations and risk factors distinguish these models from previous research.

To ensure that, as far as possible, the CCI is not subject to the criticism that it is endogenous, this
paper tests for and includes an exhaustive set of economic controls. Working with such general
specifications is made feasible by careful consideration of sign priors. As far as is feasible, the
CCI should measure credit availability, i.e. the position of the supply function for credit faced by
a typical household, once all these economic and demographic influences have been removed.
The econometric results produce two credit condition indices. In one model, the CCI has only a
level effect; in the other, it interacts with other variables. Both indices increase in the 1980s,
peaking towards the end of the decade. They fall partway back in the early 1990s, before
increasing again to a level just below the previous peak for one of the indices, and to an all time
high for the index which includes interaction effects. All equations include a common risk factor,
RISK, that depends upon a measure of inflation volatility, the change in the unemployment rate
and a downside risk measure rate of housing return, all in the previous two years, and the
mortgage possessions rate in the previous three years. New models for unsecured debt and for
mortgage debt are important by-products of this research. 8 9 1. Introduction It is widely perceived that credit conditions facing U.K. consumers, particularly in the mortgage
market, have been liberalised since the 1970s. The implications for the housing market and
consumer spending have been important: the evidence is that consumption and the housing
market were the principal agencies in the economic boom of the late 1980s and the subsequent
recession of the early 1990s. Changes in credit availability during that period are likely to have
contributed to the boom and subsequent retrenchment in consumption. The U.K. experience of
financial liberalisation and its macro-economic aftermath can be seen in the wider international
context discussed e.g. by Caprio et al (2001) and Goodhart et al (2004). The need for a credit-conditions index (CCI), which measures credit availability, i.e. the position
of the supply function for credit faced by a typical household, other than via the level of interest
rates, has been widely recognised in the consumption literature. Indeed, proxies such as
unsecured credit to income ratios and interest rate spreads have been used in empirical work (see
Bayoumi, 1993a,b and Sarno and Taylor, 1998 as examples of the former; and Scott, 1996, for
the latter). However, such proxies are unsatisfactory because they are too dependent on interest
rates, asset prices, incomes, expectations, risk perceptions and other aspects of the economic
environment. The key aim of the present paper is to construct a CCI, which, as far as possible, is
free of this endogeneity criticism because it controls for the effects of the economic environment. A first attempt in this direction for the U.K. was made by Muellbauer and Murphy (1993), in their
‘flib’ index, based on the analysis of average loan-to-value ratios (LVRs) in the U.K. for first-time
buyers from the 5 percent sample of Building Society mortgages. 1 For many of these buyers, the LVRs are at ceilings set by mortgage lenders. Effectively the method involved regressing the log
average LVR on the log of the mortgage interest rate, the log house price to income ratio and the
real mortgage interest rate, for 1969-1980 data, and using the post-1980 residuals as a measure of
the easing of credit conditions due to financial liberalisation. Caporale and Williams (2001) and
Fernandez-Corugedo and Price (2002) have extended the method to more recent data and have
applied it to modelling consumption. However, the method is somewhat fragile since it relies on a
single indicator, and plausible changes in the specification of the relationship can cause notable
changes in the implied ‘flib’ estimates. Muellbauer (1997) used annual regional data for first-time
buyers on average LVRs and on the proportion of LVRs over 0.9 to bring additional information to
bear, using dummies to trace out the post-1980 easing of mortgage credit. As he acknowledges,
see also Muellbauer (2002), there is a sample selection problem in relying on a sample, which,
before 1992, contained only building societies. Thus, when the banks aggressively entered the
mortgage market, loan terms offered by building societies became much less representative of the
market as a whole. From 1975 to 2001, there are over a million observations on mortgages for first-time buyers in
the Survey of Mortgage Lenders (SML), and in its predecessor before 1992, the 5 percent Sample 1 LVRs have been used in studies of mortgage default as an indicator of lending quality (Breedon and Joyce (1992), Brookes, Dicks and Pradhan (1994)). Lending quality and ‘ease of credit’ tend to be negatively related, though better
screening of individuals’ credit histories and other characteristics could improve quality and access to credit. 10 of Building Society Mortgages. We extract quarterly data on loan and household characteristics
from this source, excluding sitting tenants and others buying at a price discount. The present
paper uses quarterly data on distributions of LVRs and also loan-to-income ratios (LIRs), with a
regional and age split, and combines these with aggregate data on mortgage and non-mortgage
household debt to generate time series data for 1975Q1 to 2001Q4 on ten credit indicators. These
are modelled in a ten-equation system with extensive economic controls, and a common factor,
the Credit Conditions Index, is extracted. The common factor restriction is also used to model the
effects of uncertainty for the debt market environment, using a wide range of uncertainty proxies
combined into a single index. Since the aggregate data are not subject to sample selection
problems, we can use them to identify sample selection factors for the micro data.

The broad coverage of credit market indicators and the thorough and comprehensive investigation
of economic forces driving the credit market should make the resulting index more robust than
previous estimates. The economic variables we control or check for include factors likely to
influence perceived risk: unemployment rates, measures of inflation and interest rate volatility,
asymmetric rate of return measures for the housing market, and the recent history of mortgage
possessions rates. They also include yield gaps to reflect interest rate expectations, one year
ahead income growth to reflect income expectations, as well as survey based consumer
confidence measures, and more conventional demography, interest rate, wealth and income
effects. Our index is intended as a scalar measure of credit availability, i.e. of the position of the
supply function for credit faced by a typical household, once all these economic and demographic
influences have been removed. We use the term ‘Credit Conditions Index’ rather than ‘Financial Liberalisation Index’ because of
the latter’s connotations of process, rather than outcome. 2 Nevertheless, the process of financial deregulation and other changes in financial architecture, described in Section 2, have had an
important bearing on the credit availability outcome. We return in the conclusions to the question
of whether our index solves the ‘endogeneity’ problem. As well as its use in econometric
modelling of consumption, debt and the housing market, our estimated index has applicability in
interpreting the current state of credit markets. New models for unsecured and mortgage debt are
an important by-product of our research. We should make clear at the outset, that when the marginal impact of our index on debt levels
and the other debt indicators is measured, these are not the general equilibrium effects of a shock
to credit supply on these indicators. Our models are all conditional on interest rates, asset prices,
income and other variables, which themselves will be influenced directly or indirectly by such a
shock. General equilibrium calculations of this type can only be carried out inside a much larger
model that endogenises variables which depend on the credit conditions index. The layout of the paper is as follows. Section 2, examines the various aspects of liberalisation
and other changes in U.K. credit markets since the 1970s. Section 3 discusses the information
content of the Survey of Mortgage Lenders and its predecessor. It discusses reasons why lenders 2 Our index of credit supply conditions refers to shifts in the supply of credit function. We also refer to this index as measuring the non-price aspect of credit availability to remind the reader that the index is capturing shifts in the
supply of credit rather than movements along the supply curve. 11 use credit ceilings, such as limits on loan-to-income and loan-to-value ratios. The information we
extract consists of the proportions, by age and region, of first-time buyers with loan-to-income
ratios of 2.5 or more and the corresponding proportions with loan-to-value ratios of 0.9 or more.
The remainder of Section 3 explains the empirical methodology and various economic influences
on the proportions of high loan-to-income and high loan-to-value ratios. Section 4 discusses the
economic influences on unsecured and mortgage debt, in the context of the previous literature.
Section 5 presents empirical estimates of the ten-equation system assuming there are no
interaction effects between the Credit Conditions Index and the other economic variables.
Section 6 discusses possible interaction effects of this type and presents estimates of the
generalised model including interaction effects. Both versions of the Credit Conditions Index can
be seen in Figure 13. Section 7 concludes and discusses possible applications of the Credit
Conditions Index. A data appendix provides details of data construction and sources. 2. Credit market liberalisation in the U.K. In the 1970s, a period of negative after-tax real interest rates, the U.K. authorities attempted to
control credit with stringent liquidity ratios on banks, special deposits (the ‘corset’), regulations
on minimum deposits and maximum repayment periods on hire purchase credit, and directives
and persuasion aimed at building societies to limit lending and/or to keep nominal interest rates
low. There were several key events in the evolution of financial liberalisation under the Thatcher
Government, which came to power in 1979. Exchange controls were removed in 1979, opening
the banking sector to greater foreign competition and giving domestic institutions access to the
developing Eurodollar markets. This was an important step in integrating the U.K. into rapidly
expanding international capital markets. 3 The logical step after removing exchange controls was to abolish the minimum deposit requirement on banks, or ‘corset’ on bank lending. Banks could
enter the mortgage market from 1980. Increased competition in the mortgage market led to the relaxation of rules on building societies
(eg their access to wholesale financial markets) and the break-up in 1983 of the interest rate-
fixing agreements. The Building Societies Act (1986) formalised the relaxation of rules. One
consequence is likely to have been charging higher interest rates for more risky loans. 4 A second phase of new entry into the mortgage market from 1985 was heralded by the influx of centralised
mortgage lenders without high street branches. 5 The Basel I Accords on capital adequacy ratios for banks, agreed in 1988, gave mortgage loans a
preferred status, with a 50 percent risk weighting relative to other loans. This may have caused a
further easing to the mortgage-lending regime in the U.K. 3 It is noteworthy that measures of globalisation in financial markets, such as the ratio of gross capital flows between OECD economies, divided by OECD GDP, expanded rapidly in the 1980s. 4 We have microeconomic evidence from the Survey of Mortgage Lenders and its predecessor of a significant positive correlation in cross-sections of loan to value ratios for first time buyers from 1983 with the mortgage rate
charged. This correlation is missing before 1983. However, these correlations are relatively weak or entirely
missing for loan to income ratios. We control for borrower income in these analyses. 5 These included Allied Irish Bank, Credit Lyonnais and other foreign banks . 12 The first major de-mutualisation of a building society occurred in 1989 (Abbey National),
demonstrating the new fluidity of the mortgage market. This was followed by a spate of others
over the next decade. Initially banks and building societies served somewhat different markets,
but by 1990, differences in the average loan or the average income of borrowers between bank
and building society customers had become relatively insignificant. After 1990, following the start of the mortgage repossessions crisis, credit market liberalisation
was partly reversed. The Building Society Commission increased prudential advice in 1991.
Mortgage indemnity insurers moved the terms of insurance policies sharply against mortgage
lenders, not just in pricing, but also in risk sharing. One symptom of the tougher conditions of the
early 1990s was the loss of market share of centralised mortgage lenders (with higher default
rates than building societies). Improvements in the mortgage lenders’ credit scoring methods, data and arrears management and
probably in pricing for risk occurred after the early 1990s. The late 1990s saw another type of
new entry - the internet mortgage lenders – and significant innovation in new products, such as
fixed rate mortgages over longer terms and ‘flexible mortgages’. The latter permit borrowers to
repay loans more quickly, take payment holidays or extend loans flexibly, as long as loan-to-
value ratios remain within pre-set bounds, see Munro et al (2001). The spate of special offers to
new customers tended to improve mortgage terms for those willing to undergo the inconvenience
of re-mortgaging, one symptom of the strength of competitive pressures, see Samuels (2001). In 1998, a significant change in pricing behaviour by mortgage lenders occurred. Following the
lead of the largest lender, the Halifax, lenders gave borrowers exemption from mortgage
indemnity insurance (which insures lenders against mortgage default) if loan-to-value ratios were
below 0.9. This gave borrowers a considerable incentive to reduce mortgages to below this level. Figure 1 illustrates two aspects of this brief history. It shows the value share of banks in
mortgages outstanding (Abbey National continues to be treated as a building society after 1988).
The rapid rise from 1980 is notable, but after 1990, this share has little more meaning as a sign of
competitive pressure since the lending profiles of banks and building societies had become very
similar. Figure 1 also shows the value share of centralised mortgage lenders, demonstrating the
post 1985 rise (see the left axis for the scale), and renewed growth in recent years. Figure 2 shows consumption to income and house price to income ratios rising strongly in the
1980s. Part of the co-movement is almost certainly due to credit market liberalisation rather than
the causal effect of house prices on consumption. Figure 3 shows mortgage and unsecured debt to
income ratios more than doubling between 1980 and 1990 and rising to new heights recently. As mentioned in the Introduction, another type of evidence is available from surveys of mortgage
lenders, carried out since the end of 1968. We turn next to the use of data from this source. 3. Extracting information on credit conditions from the Survey of Mortgage Lenders 3.1 Distributions of loan-to-income and loan-to-value ratios and limits set by lenders 13 The key reason for mortgage lenders applying limits to loan-to-income ratios (LIRs) and loan-to-
value ratios (LVRs) is to avoid the risk of default, both in payment arrears, and, more seriously,
mortgage possession. Mortgage possession, where the lender seizes the housing collateral of
mortgages in default, can be seen as the intersection of two events: the ‘debt/equity ratio rising
above some threshold’ and ‘a trigger function (of debt service ratio, income shocks) exceeding
another threshold’. The first of these events makes it difficult or impossible for the borrower to
trade down to cheaper housing or out into the rental sector, given the difficulty of obtaining
substantial unsecured debt. Borrowers and lenders have a common interest in avoiding default,
even with a bad debt/equity ratio, since mortgage possession in the U.K. is unpleasant for
borrowers. The latter are liable for the lenders’ transactions costs, and subject to pursuit in the
courts for years and denial of access to future credit. However, if unexpected cash flow problems
arise - the trigger function exceeding some threshold, there will often be no alternative to default. The probability of default, i.e. the intersection of the two bad events, equals the probability of a
bad debt/equity ratio multiplied by the probability of a bad trigger, given a bad debt/equity ratio. 6 By limiting the initial LVR, lenders can reduce the probability of later bad debt-equity ratios, that
can arise through a fall in house prices, and/or through accumulated payment arrears. Limiting initial LIRs, reduces the probability of a later bad ‘trigger’. For example, with a LIR of
2.5, the initial debt service ratio is 2.5*r, where r is the tax-adjusted mortgage interest rate plus
pro-rated loan repayments, initially a small fraction of monthly mortgage payments, plus pro-
rated insurance costs. For example, with r at 10 percent, 25 percent of income is committed to
debt service. However, with r at 15 percent, 37.5 percent of income would be committed to debt
service, a percentage many households would find hard to tolerate, particularly if they had not
planned for it. Thus, in an environment of high nominal interest rates, lenders are likely to apply
tighter LIR criteria, while in a low interest rate environment the opposite will be the case. It would seem logical for lenders to offer borrowers a trade-off between LVR and LIR limits,
since it should be possible to raise the LIR when LVR is lowered, keeping overall risk constant.
We are not aware of formal, ex-ante offer schedules of this type being published by any of the
major mortgage lenders. However, lenders do operate some trade-offs in practice, in making
mortgage offers in specific circumstances, having evaluated all the information provided by a
mortgage applicant. For example, some lenders offer self-certification mortgages where, for
LVRs below 70 or 80 percent, the lender does not carry out income checks but relies on the
information given by the borrower. In any case, even if any given mortgage lender offered only a
specific LVR, LIR pair, in the market as a whole, borrowers will find variations in these offers.
Borrowers therefore face a schedule of LVR, LIR possibilities - the envelope of best offers. 7 6 The U.K. differs in important ways from many U.S. states where borrowers’ liabilities end as soon as they return the keys to the mortgaged property to the lender. Thus, in the U.S., the debt/equity ratio should have a more
dominant role in defaults, largely decided on by borrowers. In the U.S., the ‘rational default model’, (Vandell, 1995),
which applies option pricing theory to find the bad debt/equity threshold, beyond which the household defaults, in
absence of transactions costs and credit restrictions, is widely applied. It is unlikely to apply in the U.K., where
most defaults were instigated by lenders, not borrowers. Unpublished research on arrears and possessions for a large
mortgage lender in the U.K. (Cameron, Hendry and Muellbauer) provides evidence consistent with these points. 7 A best offer is a pair of LVR, LIR offers where no alternative offer is higher in both LVR and LIR at the same interest rate. In practice, there can also be variations, such as first-time-buyer discounts, in interest rates and other
mortgage charges such as conditions associated with fixed rate offers. 14 For given interest rates, house prices etc, financial liberalisation as occurred in the 1980s, is likely
to have raised both types of limits on loans as reflected in this schedule of best offers. Note that,
while borrowers and lenders have significant common interests in wanting to avoid loan defaults
or coming under severe financial pressure from debt service costs, there seems no reason why
financial liberalisation should cause borrowers to want to raise these limits. 8 The rise in LIR and LVR limits, which occurred, and which cannot be explained by conventional demand side
variables, is therefore likely to have been a shift on the credit supply-side. Muellbauer (1997) has discussed the structure of decision-making behind the observable LIRs
and LVRs. A credit-unconstrained household chooses a mortgage loan M d and a house of value V d by maximizing utility subject to its budget constraint and the housing quality-house price trade-off generating LVR d = M d /V d and LIR d = M d /Y (Y is the household’s current income). The household is faced with the schedule of best LVR, LIR offers described above. There are two
possibilities illustrated in Figures 4a and 4b. Figure 4a illustrates the case of a household whose voluntary LVR, LIR choice is below the best
offer schedule available in the market. The fact that there are such households means that the
market share of mortgage lenders offering lower LVR, LIR limits than available from other
lenders is not zero. The household indifference curves illustrated are the contours of a utility hill
whose peak is at H. This coincides with the chosen equilibrium point E. ADD LIT REFS FROM CONFERENCE Households are making temporal and inter-temporal consumption, housing and risk trade-offs.
Note that the value of the house purchased is V=(LIR/LVR)Y. For a given income Y, V can
increase if LIR increases for a given LVR, or if LVR falls for a given LIR. Thus, in Figure 4, a
move to the right and/or down represents more housing. Such a move, however, implies less
non-housing consumption, at least in the current period and near future, though if the user cost of
housing is expected to be negative, future consumption can be expected to be higher. In
particular, with a given LIR, the sacrifice in current consumption to increase the down-payment
inherent in increasing V and reducing LVR, soon becomes untenable - unless there is access to
unsecured credit. Such credit is generally more expensive. 9 Figure 4b illustrates the case of the household constrained in the mortgage market. Point H is not
attainable and the household chooses the point E on the LVR, LIR offer envelope where it can
reach its highest indifference curve. As illustrated, the shape of the utility contours is independent 8 Whilst we make the point that financial liberalisation should not lead to an increase in the ceilings chosen by borrowers, habit formation may suggest the contrary. With credit market liberalisation, the consumption of housing
goods increased in the economy as agents were able to borrow to purchase a house. As the consumption of housing
goods increased, individuals falling behind would have come under reference group pressure to ‘keep up with the
Joneses’ and so increase their housing consumption and debt levels. However, note that the story begins with credit
market liberalisation. It certainly seems plausible that the diffusion process by which it affected behaviour could
have been via consumer habits as well as their and the lenders’ information sets. These channels cannot be
empirically distinguished in our estimates of the long-run impact of the credit conditions index on behaviour. 9 The different risk characteristics to the borrower in the event of payment difficulties and lumpy transactions costs, can make holding both kinds of debt rational, even if the household were able to borrow sufficiently in the mortgage
market. 15 of the position of the LVR, LIR offer envelope. This would be the case if mortgage borrowers
also held some more expensive non-mortgage debt with different liquidity and risk
characteristics, see footnote 9. Kinks in the indifference curves and corresponding behavioural
switches could otherwise arise and could also result from limits on non-mortgage credit. 3.2 The empirical methodology We use 26 years of quarterly data on mortgage credit conditions from the Survey of Mortgage
Lenders (SML) and its predecessor. 10 Specifically, we examine distributions of LVR and LIR for first-time buyers (FTBs), concentrating particularly on vulnerable tails for LVR > 0.9 and LIR >
2.5. We examine data by age (under 27 and 27 plus) and by region (North/South) giving 8 series
on the proportion of FTBs in these respective vulnerable tails. By including also aggregate data
on mortgage debt and unsecured consumer credit (see Figure 3) we have 10 series. We then use
economic and demographic variables to control for variations in credit demand. We combine
relevant economic risk indicators into a single RISK index, which enters all the equations. The
common unobserved supply shift component, our Credit Conditions Index (CCI), also enters all
equations. The set-up of the 10-equation system is detailed in Section 5. Briefly, the ten variables to be explained are aggregate unsecured and secured debt, four
measures by region and age of the proportions of high LVRs, and four measures of the
proportions of high LIRs. The explanatory variables are economic and demographic variables
(with allowance for lagged adjustment), and two unobserved factors - CCI and RISK. CCI is
constructed with a spline function driven by year dummies. The RISK equation defines the
common risk factor as a combination of detailed risk indicators. Identification is possible
because CCI and RISK enter each of the ten equations. The CCI index might subsequently be used to model consumption, house prices, housing equity
withdrawal, housing turnover and subsequent loan defaults in separate equations. Alternatively,
such equations could be added to the system. Usable electronic records for the LIR and LVR distributions begin in 1975 but average LIRs and
LVRs for first-time buyers (excluding discounted ‘right to buy’ sales to social housing tenants)
are available back to 1969 and are illustrated in Figure 5. The LIR graph shows an early peak in
1972 during the first of the post-war house price booms, a strong rise between 1980 and 1990,
and a weaker upward drift in more recent years. Despite some definitional issues, to be discussed,
the data suggest a considerable positive correlation between average house price/income ratios
and LIRs. 11 The graph of the average LVR ratio suggests the opposite correlation with average 10 The survey and its predecessor consisted of a 5 percent sample until 2002. The survey includes information on income, size of loan, value of house being purchased, previous tenure, the age of the main borrower, whether the
price was discounted, type and duration of mortgage, and the interest rate charged. From 1983, single borrowers and
multiple borrowers, such as couples, are distinguished, but not so in earlier years. 11 They also contain at least a hint that the temporary early 1970s "Competition and Credit Control” policy shift by the Bank of England was associated with some easing of credit market conditions. WASN’T COMPETITION AND 16 house price/income ratios, and a strong rise between 1980 and 1984, with levels thereafter
remaining higher than before, but otherwise no immediately obvious pattern emerging. Figures 6 and 7 show PLIR, the percentage of FTBs with LIRs of 2.5 or more by age and region,
rising from under 10 percent in 1980 to over 60 percent in 1989 in the case of the South and to
over 35 percent in North. The correlation with average house price to income ratios is apparent
as in Figure 5: both the 1989-90 peak and the early 1990s decline in the South are consistent with
the pronounced Southern boom/bust in house prices in this period. The systematic tendency of
PLIR to be higher in the South than the North is consistent with higher house price/income ratios
in the South. Differences by age are less pronounced than by region: in both regions, older FTBs
tend to have slightly lower percentages of high LIR mortgages. 12 Figures 8 and 9 show PLVR, the percentage of FTBs with LVRs of 0.9 or more. The graphs
show a very clear difference between borrowers aged under 27 and those aged over 27: in both
regions, systematically higher percentages of younger borrowers have LVRs of 0.9 or more. One
should expect such differences since younger borrowers tend to have lower cash resources and so
are less able to provide substantial deposits. For these younger borrowers, PLVR rose from
averages of 25-30 percent in 1975-80, to 60-80 percent and 50-70 percent for North and South,
respectively, for 1984-2000. The decline since 1998 is notable. 3.3 Economic factors impinging on PLIR The structure of decision-making underlying the observed LIR and LVR distributions was
discussed in Section 3.1. In many respects mortgage lenders and households have the same
interest in avoiding default.REF? The direction of effects of interest rate and risk factors on the
proportion of high LIR loans will therefore be the same. The directions of most of the economic
forces operating on the proportion of high LIR loans are easy to understand, see Muellbauer
(1997) for more microeconomic detail. The key economic variables and the signs of their
expected effects on PLIR, are now listed: 13 Nominal interest rate: to avoid uncomfortably high debt service ratios, see Section 3.1. (-) Real interest rate: a high real rate raises the probability of mortgage arrears and lower house
prices and so default risk. (-)
CREDIT CONTROL A LEGISLATIVE FRAMEWORK? CAN WE CITE A REFERENCE SHOWING BOE
POLICY SHIFTED? IF NOT, BETTER TO DROP. 12 We have also examined plots of the PLIR type using thresholds of 3 and 3.5. These have broadly similar shapes but clearly respond somewhat differently to house prices and risk factors. Given that we control for these, we expect
the estimated CCI to be fairly insensitive to the precise choice of threshold. The choice of of 2.5 was influenced by
the fact that, in the important 1980-5 period, the proportions above the higher thresholds are substantially lower,
which is likely to increase sampling variability. 13 Note that the signs in brackets and in bold indicate the impact of the chosen variable on the dependent variable. 17 Interest rate expectations: the nominal yield gap between gilts at durations of one or more years
minus short rates, should reflect the market view of the direction of movement of short rates. (-) Interest spread between unsecured and mortgage debt: a higher spread increases the price
advantage of mortgage debt and so raises PLIR. (+) House price/income ratio: a high ratio puts pressure on borrowers to get the highest possible loan
(and so helps explain higher PLIR in South). (+) Consumer confidence: greater confidence in economic prospects should increase the willingness
of lenders to lend and borrowers to borrow. (+) Perceived risk: we examine five indicators, which enter a risk factor common to all ten equations.
The first two are inflation and interest rate volatility. Greater historical volatility is likely to be
interpreted, by lenders and borrowers, as a sign of greater riskiness and should discourage
borrowing on high multiples. The third risk indicator is the four-quarter change in the
unemployment rate. The fourth risk indicator is an asymmetric indicator 14 of returns on housing and the fifth is the rate of mortgage possessions (these are two or three year moving averages). (-) Deviation of real income from trend: one aspect of favourable economic conditions. (+) Expected income growth: using actual income growth over the next 4 quarters as a proxy. (+) Dummy for cut in ISMI (income support for mortgage interest): in 1995 such income support was
sharply reduced, increasing the risk of cash flow problems in the event of unemployment. (-) Dummy for ‘pricing’ mortgage indemnity premia: in 1998, Halifax removed mortgage indemnity
premia for LVR < 0.9 and the market followed. This gave borrowers an incentive to bring LVRs
below 0.9, e.g. by increasing unsecured borrowing to raise cash deposits. The effect on PLIR is
less clear: a negative effect operates via those mortgages, which would previously have been in
the group where LVR is greater than 0.9; but a positive effect on PLIR also operates since
mortgages with LVR below 0.9 became cheaper. Moreover, with PLVR lower, risk is reduced in
the house price risk dimension, giving scope for relaxation in the income risk dimension, so
raising PLIR. (?) Share of couples: since lenders apply lower LIR ceilings to joint incomes, a rise in the share of
couples among first-time buyers, would lower the proportion of high LIR loans. (-) Sample selection due to the increasing market share of banks: before 1992, when the mortgage
lenders’ survey refers to building societies only, an increase in the share of banks (and later, in
centralised mortgage lenders) makes the building society sample less representative of the whole
market. In other words, while our dependent variable should be the log odds-ratio of PLIR for all
mortgage lenders, before 1992 it was based only on building societies (including Abbey 14 This indicator is defined as the rate of return when this is negative and zero when the rate of return is positive. 18 National). We therefore need to add a correction factor to the equation. If the proportion of high
LIR loans for banks was above/below that for building societies, the correction factor would be
negative/positive, and it also needs to be weighted by the share of banks. From 1983, data are
available on average loans advanced by banks and building societies. The data show average
bank loans per borrower to be around 30 percent higher than building society loans in 1983-5, but
declining to be close to the level of building societies by 1990. For 1980-82, we assume bank
advances to be 40 percent higher. We define our sample selection bias correction factor to be a
coefficient to be estimated times (average bank advance/average building society advance –1)*
(annual change in the share of banks in total mortgages outstanding), and zero after 1990, when
banks and building societies have very similar lending profiles. The change in the share of banks
is a proxy for the share in the volume of new advances to FTBs by banks, which is not available. Thus, when banks first entered the market, the average mortgage advance by banks was higher
than that of building societies. The banks initially catered to the upper end of the market,
particularly to existing bank customers with stable jobs, some liquidity, and known credit and
income records and hence with relatively low credit risk. This drove the building societies down-
market. One might have expected both PLIRs and PLVRs to be lower for higher income groups
and so expect a positive coefficient for our sample selection correction factor. Indeed, our cross-
section evidence is that throughout the sample, there is a negative correlation between the LIR
and the borrower’s income. However, as noted, for its own customers, a bank’s information
asymmetry will be lower than for a random customer and this would suggest high LIRs to be
offered to such customers. This could result in a negative coefficient for our sample selection
correction. Since we do not know which tendency dominates, we do not have a sign prior for this
case. By the late 1980s, the average loans from banks and building societies had become fairly
similar, and the sample selection effect fades away. (?) Sample selection due to increasing market share of centralised mortgage lenders: these entered
the market from 1985, obtaining access to customers through financial advisers, estate agents and
others because they did not have an established presence on the U.K.’s high streets. Their
subsequent mortgage possessions rates were around three times as high as those of the building
societies (Ford (1994)) suggesting a riskier lending profile, and hence a higher proportion of high
LIR and high LVR loans. The annual change in the share of loans outstanding accounted for by
centralised mortgage lenders is a proxy for their proportion of new advances and hence a sensible
sample selection proxy. We expect negative effects on PLIR and PLVR. 15 (-) 3.4 Economic factors impinging on PLVR The economic factors affecting PLVR should work similarly to those affecting PLIR, with the
following exceptions: 15 Note that the information from aggregate mortgage debt data is the key for identifying such sample selection effects. For example, while the centralised mortgage lenders are gaining market share, this is reflected in rising
aggregate mortgage debt/income ratios - even though the building society data on PLIR and PLVR apparently
suggests no rise or even a contraction of credit conditions. 19 House price to income ratio: this should act on PLVR in the opposite direction to the effect on
PLIR. There are two mechanisms: a high house price indicates a greater probability of a fall in
house prices, other things being equal. Second, in areas with high house price/income ratios,
households are more likely to be pushed to high LIR levels. To control overall levels of risk,
lenders should be more cautious about offering very high LVRs to borrowers with high LIRs (see
Section 3.1). (-) Rate of change of house prices: valuations by surveyors for mortgage lenders are likely to be
conservative, tending to lag behind the market when prices are rising strongly. IS THERE
EVIDENCE FOR THIS? Loan offers are based on these valuations but LVRs reported for
completed transactions are based on prices actually paid, which will tend to exceed mortgage
valuations in rising markets. The time lag in the mortgage approvals process can induce similar
effect in rising markets, where the incidence of ‘gazumping’ - the seller demanding a higher price
than initially agreed - is higher. Note that this effect should be absent from the PLIR equation. (-) ‘Pricing’ mortgage indemnity premia: the relaxation from 1998 would have had a particularly
strong effect for mortgages whose LVRs would otherwise have been only just above 0.9, where
the cost saving would substantially exceed the higher marginal cost of unsecured borrowing;
PVLRs should fall. (-) Average age: though we divide borrowers into under 27, and 27 and over, age groups, variations
do occur of average ages within these groups, especially in the open-ended 27 and over group,
which drifted up in the 1990s, and especially since 1998. Since the accumulation of financial
assets available for housing deposits increases with age, we expect a negative effect on PLVR. 16 Figure 15 shows the age deviations for borrowers over 27 years of age. (-) Sample selection with respect to the rising share of banks: the discussion of sample selection in
Section 3.3 suggests a positive effect on PLVR for building society mortgages in the early 1980s
from the rising market share of banks. It is likely that existing bank customers, particularly with
above average incomes, would have had above average liquidity, so that the incidence of high
LVRs would be lower, so pushing the observed PLVR for building societies above the market
average. One could argue for an offsetting effect arising from lower information asymmetry for
banks, as discussed in Section 3.3. However, the information banks have is likely to be about the
security and prospects for the incomes of their customers and is more relevant for PLIR than for
PLVR. This makes a positive sample selection effect more likely. (+?) Sample selection due to increasing market share of centralised mortgage lenders: a negative
sample selection effect for centralised mortgage borrowers is expected for reasons discussed in
Section 3.3. (-) 3.5 Functional forms for PLIR and PLVR equations 16 One would expect a negative effect on PLIR also. However, given the stylised fact from Figures 4 and 5, of only a small age difference in PLIR between the under 27 and 27+ groups, it seems likely that this effect will be weak. 20 Section 3.1 set out the decision structure behind the observed LIRs and LVRs. We do not have
precise ideas about the functional forms for LIRs and LVRs corresponding to household choices
not constrained by limits imposed by lenders, see Figure 4a, and how these differ from choices
made subject at lenders’ limits, see Figure 4b, and about the stochastic structure of disturbances
at the level of individual households and mortgage lenders. Identification of these structural
relationships is hopeless. Moreover, the observed distributions of LIRs and LVRs are for
completed transactions of first-time buyers. Some potential FTBs may not have been able to
obtain finance at all, or have been unsuccessful in housing search, or have encountered sellers
unable to transact within the relevant period. We know that the number of transactions by first-
time buyers has fluctuated considerably over the last 26 years (Holmans (1996, 2001)) and it is
possible that the shape of the LIR and LVR distributions may have been affected by the
transactions volume. However, we use a rich set of controls to model reduced form equations for
PLIR and PLVR. Suppose reduced forms 17 for observed LIRs and LVRs at the individual level are given by it log LIR ( ) t it f x ? = + (4) it log LVR ( ) t it g x ? = + (5) where ? i and ? i are household specific error terms with zero means. Then t PLIR ( ( ) log 2.5) t it P f x ? = + ? (6) and t PLVR ( ( ) log 0.9) t it P g x ? = + ? (7) If the distributions of ? it and ? it could be approximated by logistics, so that, for example P( ? it > z t ) = 1/(1 + exp( ?z t )) (8) Then exp( ) (1 ) / t t t z P P ? = ? and log( /1 ) t t t P P z ? ? = ? (9) This suggests we should use this “log-odds” ratios as the dependent variables. As we shall see,
we allow for a possible mis-specification of equation (6), by introducing a cubic in z t as well. 4. Economic factors impinging on aggregate unsecured and mortgage debt to income ratios 17 We call these ‘reduced forms’ because of the mix between decisions by lenders and by borrowers outlined in Section 3.1. In contrast, the aggregate debt equations are closer to structural demand equations, conditional on
variations in the position of the credit supply function, measured through CCI. 21 One would expect the economic variables affecting debt to income ratios to work as follows: Demography: proportions of working age individuals in the key house buying age-groups, see
Figure 10 for the ratio of 20-34 year olds to 20-69 year olds. (+) Income: higher income should allow households to service debt more easily. 18 (+) Expected income growth: if individuals are consumption smoothers and expect higher income
growth in the future, they will increase their consumption of housing and non-housing goods for a
given level of income and will therefore be more likely to get into debt. (+) The change in the unemployment rate: this may proxy income expectations and uncertainty. 19 (-) Liquid financial wealth: at the individual level, greater liquid wealth reduces the need to borrow.
At the level of the economy, a higher level of household liquid assets/income suggests a greater
ability for the financial system to recycle assets into debt, though with financial deregulation,
household deposits would no longer constrain lending to households. In other words, controlling
for the Credit Conditions Index, greater liquid wealth should reduce indebtedness. (-) Illiquid financial wealth, e.g. tied up in pensions, provides long-term asset backing for debt, and
so should have a positive effect on the demand for debt. (+) Housing wealth: the greater gross housing wealth, the greater the available collateral for
mortgage debt. However, for unsecured debt, the sign is ambiguous. Against the collateral
argument is the argument that, given substitution between the two kinds of debt, having less
collateral induces a switch towards unsecured borrowing. (?/+) Consumer confidence: the consumer confidence measure used corresponds to the GfK survey and
measures consumers’ confidence about their finances and the state of the economy. An increase
in confidence should be consonant with households taking on more debt. (+) Change in consumer credit controls: consumer credit controls, which regulated down-payments
and repayment periods for ‘hire purchase’ borrowing to buy durable goods, were an important
policy instrument in the 1950s to 1970s. A tightening in controls should have a negative impact
on unsecured debt in the 1970s, in addition to its role as a proxy for CCI at this time. (-) Stock of debt in the previous period/income: the higher is debt, the more debt has to be repaid
each period under typical debt contracts. Another aspect is equilibrium correction: the tendency 18 See Ludvigson (1999) for a theoretical model of unsecured debt as a function of consumers’ income. In that model an increase in the debt to income ceiling enables consumers to get into more debt, a fact Ludvigson finds consistent
with the data. Also see Japelli (1990) for the finding that a major reason households in the Survey of Consumer
Finances are denied credit is because they had insufficient income. 19 Carroll and Dunn (1997) argue that unemployment expectations are a good predictor of consumption. In the U.K., changes in the unemployment rate are persistent, so that they could have an expectational interpretation. The
uncertainty role of changes in the unemployment rate is also allowed for through the common risk factor. 22 of the steady state level of debt not to be exceeded. For the rate of change of debt the lagged
stock effect is therefore negative. (-) Nominal interest rate: with a higher nominal interest rate the debt service ratio is likely to reduce
the amount of debt that individuals are likely to undertake and to which lenders will agree. (-) Real interest rate: higher real interest rates should lower debt through two channels. In the first,
higher interest rates make debt less affordable. The second channel is through saving as higher
real interest rates raise the relative price of current consumption. (-) Interest rate expectations: the yield gap between gilts at durations of one or more years minus
short rates, should reflect the market view of the direction of movement of short rates. (-) Spreads between the credit card interest rate and mortgage rate: a negative effect on unsecured
debt; a positive effect on mortgage debt. (-/+) Perceived risk: we use the five indicators discussed in Section 3.3, entering a risk factor common
to all equations. (-) Cut in ISMI (income support for mortgage interest): the increase in the risk of future cash flow
problems should reduce demand for mortgages at the margin, but increase demand for unsecured
loans, by reducing their relative disadvantage. (+/-) Ratio of credit cards outstanding to adult population: a positive effect on unsecured credit, see
Figure 10. The more credit cards are available to consumers, the higher the probability that these
will be used. (+) Mortgage indemnity premium pricing dummy: a positive effect on unsecured credit from 1998 as
lenders abolished the premium for LVRs under 0.9 giving an incentive for those whose LVRs
would have exceeded 0.9, particularly those who would have been only just over the 0.9 level, to
use unsecured debt for marginal finance. Our prior is that this effect dominates an effect in the
opposite direction from the overall lowering of mortgage costs relative to unsecured loans from
the abolition of the premium for most mortgages. The effect on mortgage debt should also be
positive: our prior is that the lowering of costs for most mortgages (including, of course, to
previous owner-occupiers) outweighed the incentive to switch marginal funding to unsecured
loans for those whose LVRs would have been over 0.9. (+) 5. Empirical results: the base-line model Sections 3.2 to 3.4 have outlined the economic variables expected to impact on loan-to-income
ratios (LIRs) and loan-to-value ratios (LVRs) for first-time buyers, as represented by the eight
series on PLIR and PLVR, and on aggregate unsecured and mortgage debt, shown in Figure 3 as
ratios to income. As explained in Section 3.2, the effect of the altered credit supply environment,
linked to the institutional changes discussed in Section 2, is introduced in each equation through
the Credit Conditions Index, CCI, common to all ten equations. This is represented by a linear 23 spline function, which apart from the year 1980, consists of connected straight-line segments,
which can change slope at the beginning of each year. 20 The CCI also depends on the change in consumer credit controls, phased out in 1983. In addition, in the unsecured debt equation, we
incorporate the ratio of the number of credit cards outstanding to the number of adults, to capture
changes in credit supply not reflected in CCI, the latter being more tuned to the mortgage market.
The data appendix details the data used. Even without interaction effects (see Section 6), estimating this 10-equation system was not
trivial: we imposed our extensive prior expectations on sign patterns of coefficients outlined in
Sections 3 and 4, as well as on the broad outline of the CCI index, given the institutional
evolution described in Section 2. The sign priors are imposed sequentially by setting to zero
parameters which violate these priors, restricting first the parameter which most strongly violates
the prior. Moreover, to the extent practicable, we allowed the lag structures (e.g. of responses to
interest rates and house price changes) to be determined empirically. For example, in the general
specifications, we include quarterly interest rates with lags up to 4, and check for 4-quarter
changes or 4-quarter moving averages at lags of 4 and 8. We also check whether current year
effects can be simplified into 4-quarter changes or moving averages. It is well-known that the
order followed by reductions from general to simpler models can affect the final model chosen.
We therefore check for alternative reduction paths by varying the order in which terms are
simplified and priors imposed, and go back and re-check the restrictions. 21 The priors were helpful in reducing flexible general specifications to a parsimonious one. The priors, together
with restrictions, such as zero coefficients on the pre-1980 time dummies in CCI, and
normalisations about to be discussed, are important for achieving identification. We now explain our framework and some identification issues. Our equations are: ?y it = ? i ( ? i CCI t + µ i RISK t + ?? ij x jt – y it-1 ) + ? it for i=1,10 (10) CCI t = ? ? s Dum st + ? 1 ? 4 CC t + ? 2 liqr t-1 (11) RISK t = ? v j z jt (12) The y variables in (10) are the two log debt measures and the eight log odds-ratios of LVRs and
LIRs exceeding given threshold values, by region and age. We model each as an equilibrium
correction model (ECM), with the dependent variable in quarterly change form, ? y. Here ? i is the speed of adjustment. For those x jt, which can be given a long-run interpretation, ? j is the long-run coefficient. Note that the x’</i>s include variables in ? form, and so are not in the long-run solution. We also impose some homogeneity across equations, for example, setting slope coefficients to be
the same across regions and age for the four PLVR and the four PLIR equations, respectively. 20 For the year 1980, we check the quarterly timing and find that CCI rises only in 1980Q4. 21 For example, in an early version of the unsecured debt equation, we found a positive effect from our not very well measured interest rate and on base rate. This could be interpreted as an aspect of increased credit availability, since
this will tend to be correlated with higher interest rates and since we want the CCI to capture credit availability, we
set this coefficient to zero. In our final specification, however, as other elements of the model reached a
parsimonious, interpretable structure we were able to find a specification with a (small) negative base rate effect. 24 The definition of the credit conditions index CCI given in (11) incorporates split trends - the
Dums, the four-quarter change in consumer credit controls, ? 4 CC and liqr, defined as the liquidity ratio of building societies up to 1980Q3 minus its 1980Q4 value, and zero thereafter. Section 5.6
below gives more details on the construction and estimation of CCI. The definition of the RISK
factor incorporates measures of inflation volatility, interest rate volatility, the four-quarter rate of
change of unemployment, a measure of negative returns in housing, and the rate of mortgage
possessions. The final specification of the risk indicator, (12), is as follows: RISK t = (1/(1 + ? ))( v 1 (aainfma t + ? aainfma t-4 ) + v 2 ( ? 4 ur t + ? ? 4 ur t-4 ) + (13) v 3 (nrorma t + ? nrorma t-4 ) + v 4 (possesma t-2 + ? 1 possesma t-6 + ? 12 possesma t-10 )/(1 + ? 1 + ? 12 ) Here ma is the four-quarter moving average throughout. Inflation volatility is proxied by aainf,
the absolute value of the difference between the current four-quarter inflation rate and its value
one year previously. To be precise, 4 4 4 aainf ( ) t t t abs lpc lpc ? = ? ? ? where lpc is the log of consumer expenditure deflator, and abs( ) indicates the absolute value. ? 4 ur is the four-quarter change in the unemployment rate. Downside risk in housing returns is proxied by nror, equal to
ror, the rate of return in housing when this is negative and equal to zero when the rate of return is
positive. We define ror t = ? 4 lhp t-1 +0.02 – abmr t /100, lagging the house price appreciation term to avoid possible endogeneity bias. Another aspect of housing risk is posses, the possessions rate,
see Figure 11. This is the quarterly interpolation of a series published every six months. We lag
it by two quarters to reflect the typical information lag. This lag is also supported by the
empirical evidence. More general forms of the function were investigated, including an interest
rate volatility measure, and checking for longer lags in the first three terms. However, these were
found to be insignificant. To define interest rate volatility, we define 4 ( ), t t albr abs lbr = ? where lbr is the log of the base rate, and then take the four-quarter moving average. It is clear that the ? i and therefore the ? ij are identified in (10). However, a scalar multiple of the coefficients in (11) cannot be separated from a similar multiple of ? i , so that either one of the coefficients in (11) or one of the ? i has to be normalised at some value. Exactly the same applies to µ i and the coefficients of (12) or (13). We choose to set ? i =10 and µ i =10, for the PLVR equations. Estimation is performed in Hall and Cummins (1997) TSP 4.5 using system maximum
likelihood, estimating the 10x10 co-variance matrix of equation disturbances. 5.1 The equation for unsecured debt. Previous research on unsecured household debt in the U.K. is relatively sparse. 22 The most recent study is that of Chrystal and Mizen (2001), who examine a simultaneous system including 22 At the level of theory, most of the research on unsecured debt has been undertaken with respect to the effects that unsecured debt can have on consumption/saving decisions. See Antzoulatos (1994), Scott (1996), Ludvigson (1999),
Carroll (2001), Fernandez-Corugedo (2002) for the theoretical effects of the relaxation of liquidity constraints on
consumption decisions. The these papers show that relaxing credit ceilings allows agents to increase their
consumption for a given level of cash-on-hand. Maki (2000) provides an excellent summary of some of the U.S.
empirical literature on consumer credit and the household debt service burden. Recent papers include Murphy
(1999), who demonstrates that the debt burden of households is helpful in forecasting future consumption growth,
and in particular durable consumption growth. Gross and Souleles (2001) examine credit demand and supply using
US credit card data. They find that an increase in the credit limit leads to an immediate and significant rise in debt. 25 consumption, household M4 and unsecured credit using a VECM approach. They identify 3
cointegrating vectors. That for unsecured lending has the form (Table C): log ud/pc = 1.41 log real income – 0.68 (credit card interest rate – bank rate) +
0.65 log real net worth – 2.89 ? log pc (14) where ud is unsecured debt and pc is the consumer expenditure deflator. In the context of the
VAR, the Johansen approach finds empirically that the ECM terms for consumption and broad
money also enter the model for ? log ud/pc with significant coefficients, see their Table C. In a model, which conditions on the growth rate of current consumption and in current M4, the former
enters negatively with a coefficient of –1.6 (t=3.5) and the latter positively with a coefficient of
3.9 (t=6.9), while the change in the unemployment rate has a positive coefficient, 0.039 (t=4.5)
and consumer confidence has a significantly positive effect. The ECM for log real ud enters with
a coefficient of –0.11 (t=5.2), see their Table E. Such an equation cannot be given a conventional
demand for credit interpretation: unsecured credit can only be understood as part of the system. Our approach is more conventional: unsecured credit is interpreted in terms of a function of
income, lagged assets, interest rates etc, which, unlike current consumption and end-of-quarter
money holdings, can plausibly be regarded as given to the individual household. Though at the
micro level, asymmetric information is endemic, so that lenders use rules to limit their risk
exposure, information about macro aggregates should be broadly symmetric between lenders and
borrowers. Data on unsecured credit held by households will therefore incorporate lending rules,
as well as what credit demand could have been in the absence of lending ceilings, both, in turn,
reflecting the aggregate data on income, lagged assets etc. As already noted, the distinctive
feature of our approach is in the treatment of credit conditions through the CCI measure. The difficulties of modelling unsecured debt are considerable. The first is that this is a far from
homogeneous category. It includes hire purchase debt – often secured on the value of a car or
other expensive durable purchase. The duration of such debt can be as long as four years, and the
interest rate can sometimes be discounted as part of a purchase package. It also includes personal
bank loans, with not dissimilar durations. Student loans, however, tend to have longer durations.
These are the main types of closed-end loans. The other important ingredients are in the nature of
‘revolving’ credit, with short durations. Maki (2000) points out that, in the U.S., revolving credit
has grown from around 1 percent of personal disposable income in 1970 to around 8 percent in
recent years, now accounting for around 40 percent of total unsecured consumer credit. It seems
likely that much of this growth is accounted for by the growth of credit card debt, the rest being
largely bank overdrafts up to pre-arranged ceilings. The second difficulty, that of measuring the
relevant interest rate, stems from the first. Not only do interest rates differ by type of loan and by
lender, but much of credit card debt, where bills are fully paid off monthly, is interest free. The
third difficulty, measuring debt-service ratios, which add interest costs to repayment rates, is
related. The Bank of England calculates estimates of debt service ratios, excluding repayments of
principal, but only for the last few years, see Financial Stability Review, June 2002, p.82. As
Maki (2000) makes clear, the U.S. estimates are, in part, based on crude assumptions such as that
the minimum monthly payment on credit card debt is 2.5 percent of the outstanding debt, and on
approximate data on the durations of closed-end loans. 26 We model unsecured debt as an ECM, with the dependent variable, the log-change of unsecured
debt, ? log UD. We estimate the equation in the form (10), where ? u is the speed of adjustment. A key component of the ECM, is the log ratio of ud(-1)/income. But the general form of the ECM
includes other levels effects: log per capita real income, several alternative measures of the log
nominal interest rate 23 and of the real interest rate, interest rate spreads, the rate of return in housing, the log of the ratio of the number of credit cards outstanding to the adult population, log
ratios of liquid, illiquid financial and housing assets to income, the proportion of population in
the 20-35 age group, consumer confidence, and various risk indicators discussed above entering
the common RISK factor. Dynamic terms included the change in the unemployment rate, next
year’s income growth, and demographic change. Three dummies are included (see below). Table 1a shows the base line model estimated for 1976Q1 to 2001Q4. The parameter estimates
are consistent with almost all the sign priors stated in Section 4, though some effects are
insignificant. We begin our discussion with interest rate effects, where we find two to be very
significant. The first is the lagged spread between Barclays Bank credit card rate and the
mortgage rate, with, as expected, a negative coefficient. The second is the rate of return in
housing, ror t-1 defined above. The two effects can be re-parameterised into a negative spread effect, a conventional negative real credit card interest rate effect and the real rate of change of
house prices. 24 Given this parsimonious specification, we checked for other interest rate effects, including other forms of the spread using the personal loan rate from another bank and bank base
rate, log nominal rates and real rates. All proved insignificant. For the record, we show the least
insignificant, the effect of the real bank base rate. The real rate of change of house prices can,
perhaps, be thought of as a general confidence effect, empirically more relevant than the
insignificant confidence index. 25 Interest rate expectations, proxied by the spreads between the one-year gilt yield and the base rate, and the five-year gilt yield minus base rate, proved
insignificant in this equation, and indeed in all other equations. Other features of this equation are the long-run income elasticity of around 1.5, close to Chrystal
and Mizen’s (2001), despite including both the significant CCI effect and a significant effect of
the log-ratio of the number of credit cards outstanding to the adult population. Real per capita
income is shown in Figure 11. Expected income growth, proxied by the actual growth rate of
income four quarters ahead, has a significantly positive effect on unsecured debt, consistent with
intertemporal consumption smoothing. The asset effects suggest a negative coefficient for liquid assets, consistent with our prior. The
illiquid financial asset effect is positive, and the housing assets effect negative, but insignificant.
Liquid asset and illiquid financial asset to income ratios are shown in Figure 12. 23 The repayment part of the debt service ratio is implicitly proxied by a constant proportion of the lagged stock itself, already included in the equation. 24 Note that since the effects are approximately –0.6(credit card rate-mortgage rate) -0.3(mortgage rate –inflation rate) +0.3(real rate of change in house prices), they can be written as –0.3(credit card rate –mortgage rate) –0.3(real
credit card rate) +0.3(real rate of change of house prices). 25 We entered this as a linear combination of current and lagged values in a common factor in all ten equations and found it to be insignificant throughout. 27 No seasonal dummies are significant, not surprising given unsecured debt is seasonally adjusted.
A dummy consisting of 1 in 1995Q1 and zero elsewhere proxies the announcement of coming
windfalls from building society de-mutualisations, and a temporary rise in unsecured debt. A
more permanent change occurred in early 1998, when mortgage lenders decided to exempt
mortgages with LVRs below 0.9 from mortgage indemnity insurance premia. This created an
incentive, at the margin, to increase the unsecured component of borrowing. According to our
estimates, this raised the long-run stock of unsecured credit by 7.3 percent. In 1995, income
support for those with mortgages who become unemployed (ISMI) was sharply reduced. As we
shall see, this caused the mortgage stock to be lower, in the long-term than it would have been
otherwise, and by making mortgage borrowing relatively less advantageous, raised the unsecured
stock by around 5 percent in the long-run. The speed of adjustment is around 0.3, suggesting relatively rapid adjustment, or short average
loan duration for unsecured borrowing. Our composite indicator, RISK is very significant. 26 Given our definition of the RISK component (13), this suggests that the climate of low inflation,
the stable monetary policy framework of recent years and a long period of positive returns in
housing, and low possessions rates, has encouraged borrowing in recent years. The standard error of the equation is around one third of the Chrystal-Mizen unsecured debt
equation, even after a degree of freedom adjustment, reflecting our richer specification, as well as
the use of the CCI. The long-run direct effect of the latter can be computed by taking its close to
peak value of 0.203 in 2001Q4, multiplying by its long-run coefficient of 2.08 to give 0.42. The
implication is that about 0.42 of the rise in the log of unsecured debt from 1980 to 2001 can be
attributed to the rise in the CCI. This corresponds to a 52 percent rise. Between 1980Q3 and
2001Q4, our estimated risk factor declined by 0.063. With a long-run coefficient of 1.91, this
explains around 0.12 of the rise in the log of unsecured debt over the period, or 12.5 percent. The
log-ratio of unsecured debt to income rose by 1.6 over the period, so the remainder of the rise is
explained largely by the growth of real income per capita and by the rise in the number of credit
cards per capita, see Figures 10 and 11. However, it must be pointed out that the above long-run
effects are not general equilibrium effects, since a rise in CCI will itself influence interest rates,
asset prices and other variables. They are partial effects, conditional on the interest rates, asset
prices and other variables on which we argue unsecured debt depends. Tests for serial independence up to the 4th order and homoscedasticity of the residuals are all
satisfactory. A check on parameter stability is provided by the last two columns of Table 1a,
showing the estimates over the 1976Q1 to 1992Q4 sample. The standard error is very similar for
the shorter sample and the great majority of the parameters are under one standard deviation from
the full-sample estimates. 27 5.2 The equation for mortgage debt 26 Note that interest rate volatility per se made no significant contribution to the RISK factor. 27 However, we have set the parameter of the mortgage possessions rate in the RISK factor to its full sample value since the mortgage possessions rate has insufficient variation to be properly identified over this shorter sample. 28 There is a large volume of previous research on the determination of the U.K. stock of building
society mortgages, but this peters out by the mid-1980s, as the entry of the banks into the
mortgage market made this both harder to model and less relevant. Anderson and Hendry (1984),
Meen (1985) and Wilcox (1985) are the last significant studies. Meen reviews previous work
comprehensively. As he makes plain, all the previous studies took into account that before 1980
mortgage demand had been in an almost continuous excess demand state. Indeed, Meen estimates
a measure of excess demand and so the severity of mortgage rationing, MRAT, to be the
percentage deviation between the flow demand for mortgages and the actual mortgage flow,
which is positive for 1963Q1 to 1980Q3, before turning negative. Meen’s model does not give a
long-run solution for mortgage demand, though there is a solution for the long-run stock of
building society mortgages conditional on the stock of deposits in building societies. Meen’s
equation for the rate of growth of building society mortgages, estimated for 1963 to 1977, has a
standard error of around 0.0026. Wilcox (1985) follows a different approach. He takes the average LVR for first-time buyers
(FTBs) as an indicator of mortgage rationing, following Kent (1980). The long-run solution for
his equation, estimated for quarterly data for 1969-1983 on building society mortgages, can be
parameterised as: log (bsd/pdi) = constant +0.4*log real pdi -0.32*log abmr + log housing stock + 1.4 log LVR where bsd is the stock of building society mortgages, pdi is personal disposable income, abmr is
the tax-adjusted mortgage interest rate, the housing stock is defined as housing wealth scaled by
the mix-adjusted house price index and LVR is the average loan to value ratio for first-time
buyers for building societies. The equation, estimated in equilibrium correction form, also
contains dynamic terms in log house prices, the interest rate and LVR and has a coefficient of
–0.062 on log bsd(-1), measuring the speed of adjustment. The equation standard error is 0.0029. Section 4 outlined the relevant variables in our equation for the total stock of mortgage debt. The
equation has the same general form as equation (10), with the dependent variable ? log SD, where SD is the stock of mortgages held by the personal sector. After extensive reduction from a more
general dynamic specification, we arrive at the model shown in Table 1b. This has a speed of
adjustment of 0.061, almost equal to that of Wilcox (1985), and less than one quarter of that
corresponding to unsecured debt. A key component is log real per capita income where, after testing, we impose the same long-run
coefficient as in the unsecured debt equation 28 . The long-run income elasticity of around 1.59 is close to Wilcox’s 1.4. The long-run coefficient on the log of the adjusted building society
mortgage interest rate is -0.39, marginally higher than the estimate of Wilcox. For example, a
rise in the nominal mortgage rate from 6 to 7 percent, other things being equal, would reduce the
mortgage stock by 6.2 percent in the long-run. The spread between the credit card rate and the
mortgage interest rate has a positive coefficient indicating that a fall in the mortgage rate relative 28 We also examined an alternative specification with an income elasticity of one but including the lagged ratio of the number of owner occupied houses to the number of adults between 20 and 70. The fit is a little worse, but the CCI
estimated from the whole model is very close to that reported below. 29 to the credit card rate encourages a switch from unsecured to mortgage borrowing, other things
being equal, though the effect is smaller than the parallel effect in the unsecured debt equation. 29 The real rate of interest has a significant, but small effect: a 1 percentage point rise reduces the
mortgage stock by 0.7 percent in the long-run. The implications are important: a decline in
inflation, with the real rate constant, reducing nominal rates and the ‘front-end loaded’ current
debt service ratio, will increase mortgage debt. A decline in inflation, with no change in the
nominal rate, and so a rise in the real rate, leaves the mortgage debt to income ratio only
marginally lower, even though future debt service ratios are higher with lower inflation. This
finding suggests that households suffer from an element of inflation illusion, or at least an
excessive concern about their short-term ability to finance debt. The Miles Reviews of the
mortgage market (2003, 2004) come to similar conclusions, arguing that both consumers and
financial advisers are poorly informed about longer-run costs and risks. An alternative
interpretation is that low inflation is associated with lower risks of interest rate rises and of
income surprises, and that this is encouraging households to carry heavier debt burdens.
However, we have already controlled for at least part of this effect though our volatility measure
of inflation included in RISK, to which it makes an important contribution. The long-run solution also includes the proportion of adults in the 20-34 year old age group,
plotted in Figure 10, and wealth effects. The log ratio of liquid assets to income has a significant
negative effect on log mortgage debt, since the greater the liquidity of households, the lower the
need for debt. Illiquid financial wealth has a positive but insignificant effect, but log housing
wealth/income has a long-run coefficient of 0.29. The RISK factor is significant, indicating that
mortgage borrowing, like unsecured borrowing, is encouraged by a low volatility environment. Between 1980Q3 and 2001Q4, the log ratio of mortgage debt to disposable household non-
property income rose by 1.02 (equivalent to a 177 percent rise in the ratio). The long-run partial
effect of the CCI on log mortgage debt in this period is 0.95 (the product of the rise in CCI of
0.203 and its long-run coefficient of 4.68). Given that positive effects on the long-run level of log
mortgage debt/income are also coming from real per capita income (given the income elasticity
of 1.59), from the fall in the RISK factor, contributing a long-run effect of 0.13, from the fall in
the nominal mortgage rate, contributing 0.30 in the long run, and from the rise in the housing
wealth to income ratio (to which, however, adjustment will have been far from complete by
2001Q4) one might wonder what are the offsetting factors. Over the period 1980Q3 to 2001Q4,
the rise in the log liquid assets/income is the biggest, ceteris paribus, implying a decline of 0.74.
The fall in the fraction of the population aged 20-34, and the rise in the real interest rate
contribute declines of 0.12 and 0.05. In the long run, we estimate that the reduction in income
support for mortgage interest (the ISMI dummy) in 1995 reduced mortgage debt by around 15
percent, other things being equal. This was, however, largely cancelled by the elimination of
mortgage indemnity insurance premia in 1998 for mortgages with LVR below 0.9, which we
estimate increased mortgage debt by around 15 percent in the long run. It seems that, as
expected, this reduction in mortgage costs for the bulk of mortgages, strongly outweighed the
switching effect to unsecured debt discussed in Section 4. 29 We expect the effect to be smaller in absolute size since, for a given switch of funds from one source of borrowing to the other, the percentage response for larger mortgage debt should be smaller. 30 The dynamics include the demographic change variable reflecting changes in the proportion of
the population in the key mortgage borrowing age group, and the four-quarter change in the
unemployment rate, which plays a direct role as well as its indirect role through RISK. Expected
income growth proved insignificant, suggesting that mortgage debt is less relevant for inter-
temporal consumption smoothing at a one-year horizon than is unsecured debt. However, the
hypothesis that the coefficient is the same as in the unsecured debt equation cannot be rejected,
and since the long-run income elasticity was found to be the same in the two equations, this
seems a reasonable restriction to impose. The equation includes seasonals (the mortgage debt data
are not seasonally adjusted), implying that the second and third quarters experience greater
mortgage growth. A dummy taking the form of 0.25 in 1988Q2 followed by 1 in 1988Q3, and
zero thereafter, measures the effect of Chancellor Nigel Lawson’s announcement in March 1988
that on August 1 st multiple mortgage interest tax relief would be abolished. The equation standard error is 0.00260, without correcting for degrees of freedom. Assuming
that 21 degrees of freedom are lost, from the 18 parameters in the function, plus another 3
apportioned to the contribution of that equation in the estimation of the CCI, gives a standard
error of 0.00288. 30 This is the same as for Wilcox’s model, but higher than Meen’s equation, but since these models covered only the more homogeneous building society component of
mortgages, this can be considered to be satisfactory. Tests for serial independence up to the 4th order and homoscedasticity of the residuals are all
satisfactory. A check on parameter stability is provided by the last two columns of Table 1b,
which shows the estimates over the 1976:1 to 1992:4 sample. The asymptotic standard error is
the same for the shorter sample, but correcting for degrees of freedom, it is higher. As for the
unsecured debt equation, the majority of the parameters are under one standard deviation from
the full-sample estimates, and all are under two standard deviations away. 5.3 The equations for PLIR We know of no previous work modelling the proportion of high loan-to-income mortgages to
first-time buyers. We have data on PLIR classified by age (under/over 27) and region
(North/South). The equation has the form 3 0 1 ( ( ) ( ( ) ) ) t t t t y f z f z f y ? ? ? ? = + ? ? (15) where y is the log odds-ratio of PLIR, f(z t ) is a linear function of the various drivers of y, and f 0 is the average across age and regions of the maximum observed value of y. When f(z t ) = f 0, the long-run value of y is f 0 , and near this value the non-linearity is unimportant. When ? is positive, as f(z t ) falls further and further below this value, the cubic term becomes more and more negative, so that y falls below the value otherwise predicted by f(z t ) and y t-1 . Without this non- linearity, the model finds it slightly more difficult to capture the low values of PLIR reached in 30 This suggests that the reported t-ratios in Table 1b should be scaled down by a factor 0.895 to incorporate the degree of freedom correction. 31 1980, before credit conditions eased and at a time of high interest rates and recession. Otherwise,
the type of curvature, implied by the log odds-ratio, seems to capture well the behaviour of PLIR The speed of adjustment ? is estimated at 0.44. The estimated long-run effects of CCI and the RISK factor are both highly significant, though lower than in the PLVR equations, discussed in
the next sub-section. The long-run coefficient on the log income/house price ratio is –1.8, though
the negative income effect is offset by the coefficient of 3.0 on log real income minus trend. The
implication is that high real house prices tend to force up LIRs, as argued in Section 3.3. The log
of the mortgage rate has a strong negative effect with a long-run coefficient of –0.89, suggesting
that debt-service considerations are relevant for LIR rules followed by lenders and for borrowers
themselves. But increases in mortgage rates over the previous two years also have a strong
negative effect on PLIR. It is possible that the effect being captured is not just on interest rate
expectations, since changes in interest rates tend to be positively auto-correlated, but perhaps also
on economic conditions in the labour and housing markets. This may be why the change in the
unemployment rate, relevant in the mortgage debt equation, proves to be insignificant here,
except via the RISK factor. The ISMI dummy has a negative coefficient, though insignificant and set to zero here. The
dummy for the 1998 elimination of mortgage indemnity premia for LVRs below 0.9 would be
expected to have a larger effect in the PLVR equation. Indeed, its coefficient in the PLIR
equation is negative, though not precisely estimated. The proportion of couples in the sample has
the expected negative effect, reflecting the lower LIRs offered by lenders on joint incomes. An important consideration when modelling mortgage data is the sample selection issue. We
argued in Section 3.3 that the variable representing the sample selection effect of the entry of
banks is difficult to sign a priori, while the variable representing the sample selection effect of the
entry of centralised mortgage lenders should have a negative coefficient. Both effects prove
negative, though only the latter is significant. Figures 6 and 7 confirm that PLIR for building
societies rose at the time. However, our results suggest that it is likely that PLIR for building
societies and banks together rose by even more. The dynamics also include the change in the log odds-ratio of PLIR lagged one quarter, with a
negative coefficient, suggesting smoothing in the dependent variable. Other dynamic terms are
the rate of change and the rate of acceleration of regional house prices, partly compensating for
the fact that the log ratio of income/house price enters at a one-quarter lag. Between 1980Q3 and 2001Q4, the log odds ratio of PLIR rose, on average, by around 2.2 31 . Of this rise, the long-run partial contribution of CCI is 0.57 and of the decline in RISK is 0.27. The
decline in the nominal mortgage interest rate contributes 0.67, and most of the rest is explained
by the rise in the average house price/income ratio (a long-run effect of 0.27) and by the higher
value of real income minus trend in 2001. 31 Based on a 4-quarter moving average of data for 1980Q3 and 2001Q4, and taking the simple average of the data over the two regions and age groups. 32 There are some symptoms of heteroscedastic residuals. The latter can be traced to large residuals
in 1980, when the proportion of high LIR loans fell to the lowest level in the sample. Our
reported t-ratios are robust, incorporating a heteroscedasticity correction. A check on parameter
stability is provided by the last two columns of Table 1b, which show the estimates over the
1976:1 to 1992:4 sample. The standard errors tend to be higher for the shorter sample and the
great majority of the estimated parameters are under two standard deviations from the full-sample
estimates. 5.4 The equations for PLVR The only previous work modelling the proportion of high LVR loans, of which we know, is
Muellbauer (1997). 32 . Muellbauer analyses annual data, on 11 U.K. regions, for PLVR for 1971- 1995, extrapolating missing PLVR data for 1971-3 from average regional LVR data and a simple
econometric model. His model, incorporating a CCI, confirms most of the priors set out in
section 3.4. More specifically, t t-1 it t (9.0) (4.7) log (PLVR/(1-PLVR)) = 0.33* log (PLVR/(1-PLVR)) - 0.083*log (hp /pc ) t (6.1) (6.1) (12.7) (4.9) (2.8) + 0.33* log 0.16* log 1.88* 0.48* log 0.08* (Gallup/100) it it it it ry hp abmr pc ? ? ? ? + ? + ? t-1 t-1 (6.5) (2.3) - 0.15*arorse ** + 0.12*rorsem ** (16) Here, hp refers to the house price in the ith region, pc is the consumer expenditure deflator for the
U.K., ry is real regional non-property income, abmr is the tax-adjusted mortgage interest rate
using the ith region tax adjustment, Gallup is the Gallup Poll measure of consumer confidence in
December of the previous year, and arorse** and rorsem** are risk factors proxying the level and
volatility of housing returns. Absolute values of t-ratios are shown in parentheses. Thus, the log real house price index and its nominal rate of change have negative coefficients; the
growth rate of income has a positive coefficient; the nominal interest rate has a negative
coefficient, while inflation has a positive effect, suggesting an element of real as well as nominal
interest rate effects. The annual change in an index of consumer confidence has a positive
coefficient, while two risk indicators based on the rate of return in housing in the South East have
the expected signs, suggesting that greater volatility of returns dampens the proportion of high
LVRs, and that a recent history of negative returns also dampens high LVRs. As Muellbauer
(1997) acknowledges, a defect of the estimated model is its failure to deal with the sample
selection problem, (the estimated CCI turns down temporarily in 1983 and 1987). Also, the
model is based only on LVR data instead of the more comprehensive data of this paper. 32 Wilcox (1985) and Muellbauer and Murphy (1993) had modelled the average LVR for FTBs. Wilcox, however, does not control for market conditions, except through the ratio of building society deposits to building society
mortgages, a proxy for the societies’ liquidity. 33 Turning to the current study, the ECM for the log odds ratio of PLVR was specified in a similar
form to that for the PLIR, and incorporates a similar non-linearity through a cubic term, which
proved insignificant. Again, there are fixed effects by age and region. The speed of adjustment
at 0.52 is a little higher than that of the PLIR equations. The coefficients on CCI and on the RISK
factor are set to the value 10. As explained above, the CCI and RISK coefficients in one equation
must be set, to achieve identification, given that the parameters of the spline function, which
governs the shape of the CCI function, and the parameters of the RISK function are estimated. The coefficient on the log income/house price ratio is 0.67 and significant, consistent with the
posited negative real house price effect. The rate of growth of house prices also has the posited
negative effect. The log mortgage rate enters with a four-quarter moving average, with a
significant negative coefficient, and there is also a significant negative ‘shock’ effect of the
current change in the mortgage rate, perhaps also forecasting further rate rises. The real
mortgage rate has a negative but insignificant coefficient. The step dummy for 1998Q1, capturing the abandonment of mortgage indemnity insurance
payments for those with LVRs below 0.9, has a strong negative coefficient (and enters as a
moving average with the same weights in all ten equations). As expected, this pricing shift
created an incentive for borrowers over this threshold, especially those who would have been
only just above, to bring their LVRs below 0.9, and pushed down PLVR. The 1995 ISMI dummy
also has a significant negative coefficient, reflecting the increased risks faced by borrowers, with
the tightening of income support for the unemployed with mortgage commitments. The effect is
substantially larger than for the PLIR equations, for no obvious reasons. As Figures 8 and 9 illustrate, PLVRs for older borrowers tend to be substantially lower than for
younger borrowers, while the PLIR differences are much less pronounced. Clearly younger
borrowers have had less opportunity to save for a deposit. From the early 1990s, there has been a
substantial rise in the average age of FTBs. 33 Since the under 27 category is bounded by the age 27, the rise has been more noticeable in the over 27 age group. One would expect this to account
for some of the downward drift in PLVRs in the second half of the 1990s. Indeed, when we enter
the deviation in the average age of the over 27 group in each quarter from the average over the
whole sample, we find a significant negative coefficient. Finally, to turn to the sample selectivity proxies discussed in Section 5.3, the proxy for sample
selection due to the entry of banks is positive, consistent with the view that initially the banks
targeted existing customers with more significant deposits, so leaving building societies with
cash-poor borrowers and so higher PLVR. A negative coefficient for centralised mortgage
lenders is also consistent with prior expectations. Thus, when the down-market centralised
mortgage lenders gained market share in the second half of the 1980s, this pulled down the
PLVRs reported for building societies, and made them a negatively biased estimate of those for
the whole market. 33 Note that ‘first-time buyer’ is defined as someone whose previous tenure was not in owner-occupation. Some in this category may therefore have had a spell as owner-occupiers, returned to renting, before switching back
(Holmans (2001)). 34 Comparing 1980Q3 and 2001Q4, the average log odds ratio has risen by around 1.79 (using the
four-quarter moving average and averaging over age and region). The long-run partial
contribution of CCI over the period to this rise is 2.03, the fall in nominal mortgage rates
contributes 1.05, and RISK 0.63. The key offsetting effects are three: the negative effect (of 0.84
in the long-run) of the abolition of the mortgage indemnity premium in 1998 for LVRs below 0.9,
the negative effect (of 0.61 in the long-run) of the tightening of ISMI (income support for
mortgage interest) in 1995, and the negative effect (around 0.1 in the long-run) of the rise in
house price/income ratios. Tests for serial independence up to the 4th order and homoscedasticity of the residuals are all
satisfactory. A check on parameter stability is provided by the last two columns of Table 1d,
which shows the estimates over the 1976Q1 to 1992Q4 sample. The standard errors tend to be
higher for the shorter sample and, as for the other debt equations, the great majority of the
parameters are under two standard deviations from the full-sample estimates. 5.5 The estimated RISK We fix the coefficient of RISK (equation (13)) in the PLVR equation at 10 so that we expect the
v i ’s to be negative, except for the negative housing return risk proxy. The results are in Table 1e. The most significant of the three terms is v 1 : a fall in inflation volatility has sizeable positive effects on both types of debt and on the proportion of high LVRs and LIRs. The proxy for
negative housing return risk is also very significant, while the change in the unemployment rate
and the possessions rate are more marginal. The ? coefficient is estimated to be quite close to 0.5, while ? 1 is estimated to be around 1, a restriction we impose. This means that the possessions rate enters as the 12 quarter moving average, lagged two quarters. Figure 14 shows the estimated
RISK term where a rise denotes a reduction in risk. 5.6 The estimated CCI Table 1f shows estimates of the parameters of the CCI function. Note that the last quarter of
1980 marks the start of the rise in CCI, handled by a fourth quarter dummy. Otherwise,
piecewise linear splines, shifting in quarter one of each year, are used to model CCI. These are
constructed by defining step dummies for each year, stepping from 0 to 1 in the first quarter, and
remaining at 1 thereafter. The four-quarter moving average converts the step into a trend going
from 0 in the previous year’s fourth quarter to 1 in the current year’s fourth quarter, thenceforth
remaining at 1. We estimate the coefficients on each of these terms, subject to the restriction that
these effects are all zero before 1980Q4, and that in the 1980-89 and 1994-2001 periods, no
negative reversals take place: such reversals can be ruled out given the institutional background
set out in Section 2. These restrictions imply a zero coefficient in 1983 and in 1987. Zero
restrictions were also imposed on the coefficients in 1994, 1996, 1997, 1998 and in 2000, either
because the unrestricted coefficients would have been negative, though none significantly so, or
because the estimated coefficients were very close to zero. To account for possible variations in CCI before 1980, we include the four-quarter change of the
consumer credit controls dummy. This should reflect the stance of the authorities to expansion of 35 credit more generally. Its coefficient is negative and significant. The lagged liquidity ratio of
building societies up to 1980Q3 is another indicator of ease of mortgage credit before the credit
markets liberalised. Building societies flush with liquidity are more likely to lend, lowering their
liquidity in the following period. The liquidity ratio lagged one quarter has a significant positive
coefficient. We also investigated the mortgage-rationing indicator, MRAT, from Meen (1995)
for the period up to 1980 Q3, after which the CCI dummies begin to operate. The coefficients are
positive for the current value and lags up to two quarters, and significantly so at a lag of one
quarter. This suggests that greater rationing is associated with higher values of debt growth and
looser ceiling on LVRs and LIRs. In our view, this casts some doubt on the short run dynamics
of Meen’s indicator, though it does show a fall at the end of 1980 and remains low, indicating an
easing of rationing at this time. Figure 13 shows our estimated CCI and the real tax-adjusted
mortgage interest rate. 34 An important question concerns the downturn in CCI from 1990, reaching a trough in 1993,
before turning up again. One might ask whether this is a genuine credit supply shift, or reflects
the risk perceptions and negative outlook on both sides of the market during this period when
mortgage possessions ran at record levels (Figure 11). We control for risk perceptions as noted
above, and we have included or tested for an extensive set of economic factors including
consumer confidence and next year’s income, but one can never be entirely sure that these
controls are adequate. The description of the evolution of credit conditions given in Section 2
suggests that the biggest source of a downward supply shift was the change in mortgage
indemnity insurance contracts available to lenders. In the post-war period, there had never before
been a time of sustained falls in nominal house prices, and with hindsight, the insurers had under-
priced credit risk in the late 1980s. However, the result of these misperceptions was that, for
given economic conditions, mortgage borrowers had had greater access to credit in the 1987-90
period than in 1991-94. For modelling consumption, house prices, housing turnover and
subsequent mortgage defaults, the interpretation of a fall in CCI for the period 1991-3, when the
insurers sharply raised their premia, does seem to point in the right direction. One possible area for concern is the question of house price endogeneity, even though we use
house price indices defined for all types of buyers and not just first-time buyers. The only place
where this could be a problem is in the PLIR and PLVR equations, where contemporaneous
values of house price changes enter. Everywhere else, including in housing wealth terms and
measures of rates of return in housing, we use lagged house price data. If we respecify the PLIR
and PLVR equations, replacing terms in ? log hp t with its lags and current and lagged changes in mortgage interest rates, the results are very little changed. Indeed the graphs of CCI are visually
virtually indistinguishable. Co-integration issues seem relatively straightforward, even though formal analysis using the
Johansen method is impossible in this set-up, given that CCI is estimated from the data. Augmented Dickey-Fuller tests confirm the I(1) status of the following variables: the log
unsecured debt and mortgage debt to income ratios, the log changes in nominal unsecured and 34 This helps to explain, given the positive correlation in the early 1980s, why it is common to find weak or perversely signed real interest rate effects in mortgage and in consumption equations which omit a CCI effect. 36 secured debt, the four log odds ratios for PLIR and PLVR respectively, log real per capita
income, the real mortgage interest rate, the log nominal mortgage rate, the three log asset to
income ratios, the age deviation for first-time buyers aged over 27 (see Figure 15) and the share
of couples among first time buyers. The proportion of the population aged 20 to 35 is I(2), in this
26-year sample, but since its variance is very low, this is unlikely to cause problems. The rate of
return in housing is I(1) but not far from the I(0) borderline. The composite RISK indicator is
clearly I(1), as are most of its ingredients individually, and so, of course is CCI. 35 Co-integration between the variables specified in each of the ten equations looks unproblematic, and the
extensive use of sign priors based on economic theory and institutional knowledge has played an
important part in arriving at a well-determined specification. We have constructed ecm terms consisting of the I(1) terms in each equation and weighted these by their estimated coefficients.
Dickey-Fuller tests applied to each of these confirm that they are all I(0). 6.