Inattention and Inertia in Household Finance: Evidence from the

Inattention and Inertia in Household Finance:
Evidence from the Danish Mortgage Market
S. Andersen, J.Y. Campbell, K.M. Nielsen, and T. Ramadorai
Copenhagen, Harvard, HKUST, Oxford
May 20, 2015
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
1 / 27
Introduction
Inertia in Household Finance
Households respond slowly to changed circumstances.
I
I
Participation, saving, and asset allocation in retirement savings plans
(Agnew, Balduzzi, and Sunden 2003, Choi, Laibson, Madrian, and
Metrick 2002, 2004, Madrian and Shea 2001).
Portfolio rebalancing in risky asset markets (Bilias, Georgarakos, and
Haliassos 2010, Brunnermeier and Nagel 2008, Calvet, Campbell, and
Sodini 2009).
An important example: Mortgage re…nancing.
I
I
Inertia (“woodheads”) in prepayment models and MBS pricing
(Stanton 1995, Deng, Quigley, and Van Order 2000, Gabaix,
Krishnamurthy, and Vigneron 2007).
Cross-subsidies from sluggish to prompt re…nancers (Miles 2004,
Campbell 2006, Gabaix and Laibson 2006).
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
2 / 27
Introduction
Mortgage Re…nancing Inertia: Questions
Do prompt re…nancers look di¤erent from sluggish re…nancers?
I
I
US HMDA tracks borrowers at origination, so we don’t observe
non-re…nancers.
American Housing Survey and other survey data are very noisy
(Schwartz 2006).
Does the opposite of inertia (too-hasty re…nancing) also exist?
I
I
Optimal re…nancing solves a di¢ cult real options problem (Agarwal,
Driscoll, and Laibson 2013).
Errors of “commission” and “omission” when only re…nancers are
observed (Agarwal, Rosen, and Yao 2012).
Can household constraints explain sluggish re…nancing?
I
In the US, re…nancing requires positive home equity and su¢ ciently
high credit score: inevitably imperfectly measured (Archer, Ling, and
McGill 1996, Campbell 2006, Schwartz 2006, Keys, Pope, and Pope
2014).
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
3 / 27
Introduction
Mortgage Data from Denmark
We use high-quality administrative data from Denmark to surmount
many of these obstacles.
Denmark has predominantly FRMs, like the US, but with important
special features:
I
I
I
I
Funding with covered bonds, …xed-rate maturity-matched bonds with
integer coupons.
Re…nancing does not require positive home equity or a credit check
provided there is no cash-out.
Re…nancing involves buying back the underlying mortgage bond, either
at market value or face value.
When buying back at face value, the re…nancing incentive is the bond’s
coupon rate less the current mortgage yield.
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
4 / 27
Data
Administrative Data from Denmark
All mortgages from 5 largest mortgage banks (out of 7) with a 94%
market share.
Demographic information from Civil Registration System.
Income and wealth from the Customs and Tax Administration.
Education from the Ministry of Education.
Medical treatments from the National Board of Health.
Start with 2.7 million households.
I
I
I
I
Match education and income: 2.5 million.
953,000 households have mortgages in 2009 and 703,000 have a single
mortgage.
282,000 households have a …xed-rate mortgage in 2009 and 272,000
have one in 2010.
60,000 households re…nance in 2009 and 23,000 re…nance in 2010.
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
5 / 27
Data
Summary Statistics (Table 1)
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
6 / 27
Data
4.6
4.4
4
4.2
Interest rate
4.8
5
30000
20000
10000
0
Number of Refinancing Households
Re…nancing by Coupon (Figure 4)
2010:1
2010:3
2011:1
Old Coupon <5%
Old Coupon 6%
Interest Rate
Andersen et al (2015)
Inattention and Inertia
2011:3
2012:1
Old Coupon 5%
Old Coupon 7+%
Mortgage Design 2015
7 / 27
Data
Re…nancers and Non-Re…nancers (Table 3)
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
8 / 27
A Mixture Model of Re…nancing Types
Re…nancing Types
pih,t (yi ,t = 1jνh , βh , σe ) = pih,t (νh + e βh Ih (zi ,t ) + ei ,t > 0).
Household i has type h, re…nancing is event yi ,t = 1.
Parameter νh governs base re…nancing rate, βh governs response to
incentive Ih (zi ,t ), zit contains mortgage characteristics.
Stochastic choice error ei ,t is logistic (as in standard logit model).
Woodheads: re…nance at …xed rate νW , ignore incentives so
IW (zi ,t ) = 0 and βW = 0.
Levelheads: respond rationally to incentives with some βL > 0, but
νL = 0.
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
9 / 27
A Mixture Model of Re…nancing Types
A Mixture Model
Household i has mixing weight δhi on type h, where 0 < δhi < 1 and
h
∑h δi = 1.
We model
h
h
δhi = e ξ i / ∑ e ξ i ,
where ξ hi can be a function of household characteristics.
We can capture dynamic e¤ects using issuing quarter and current
quarter dummies (interactions of these dummies have almost no
explanatory power).
I
I
I
Pure time e¤ects (e.g. from media coverage of re…nancing
opportunities).
Age e¤ects (burn-in and burn-out).
Currently working on modeling the persistence of type assignments.
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
10 / 27
A Mixture Model of Re…nancing Types
0
.1
Refinancing Probability
.2
.3
.4
.5
.6
A Basic Mixture Model (Figure 1)
-3
-2
-1
0
1
2
3
4
Incentives
Observed Refinancing Probability
(i) Woodheads
(ii) Levelhead
(iii) Model-Predicted Refinancing Probability
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
11 / 27
A Mixture Model of Re…nancing Types
The Re…nancing Incentive
I (zit ) = Citold
Yitnew
O (zit ).
Interest saving is old bond coupon less new mortgage bond yield.
Use Agarwal, Driscoll, and Laibson (2013) approximate closed-form
solution for threshold:
s
q
σκ it
2(ρ + λit ).
O (zit )
mit (1 τ )
σ interest rate volatility, τ mortgage interest tax deduction, ρ
discount rate, κ it …xed plus variable re…nancing cost, mi ,t size of
mortgage, λit base rate of principal reduction, which includes
termination probability.
I
We estimate termination probability: median 8.4%, mean 11.0%,
standard deviation 8.7% (ADL suggest 10%).
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
12 / 27
Re…nancing Incentives and Household Behavior
Summary of the Evidence
Danish mortgage rates have fallen substantially since their peak in
2008.
About 23% of household-quarters have positive re…nancing incentives.
Almost 90% of these do not re…nance (errors of omission).
About 2% of the households with negative incentives do re…nance,
but about half of these cash out or extend maturity so only 1%
appear to be mistakes (errors of commission).
Most demographic characteristics shift re…nancing up or down and
therefore move these errors in opposite directions.
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
13 / 27
Re…nancing Incentives and Household Behavior
0
0
.1
.2
Refinancing Probability
Number of Observations (10,000s)
5
10
15
20
.3
25
Incentives and Re…nancing (Figure 6)
-3
-2
-1
0
Incentives
1
2
3
4
Number of Observations (10,000s)
Observed Refinancing Probability
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
14 / 27
Re…nancing Incentives and Household Behavior
Errors of Omission and Commission (Table 5 Panel A)
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
15 / 27
Re…nancing Incentives and Household Behavior
Who Makes These Errors?
Most household demographic characteristics have o¤setting e¤ects on
the two types of errors (Table 5 Panel B).
Characteristics that are associated with increased re…nancing in Table
3 increase errors of commission and reduce errors of omission.
This suggests that a pure inattention model will not …t the data
(since pure inattention would increase both types of error).
Errors of omission are costly (Table 6): 1.9% of the outstanding
mortgage balance for the average error-prone household, and about
0.25% of all outstanding mortgages (using 0.25 cuto¤, across both
years).
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
16 / 27
Estimating the Mixture Model
Mixture Model Results (1)
Baseline model with no history dependence or demographic e¤ects
delivers sensible estimates (Figure 1):
I
I
88% of household-quarters are woodheads who re…nance with
probability 0.8%.
12% are levelheads who re…nance with probability 10% when the
incentive is -0.88%, 25% when the incentive is -0.43%, 50% when the
incentive is zero, 75% when the incentive is 0.43%, and 90% when the
incentive is 0.88%.
History dependence and demographics greatly increase model’s
explanatory power from initial pseudo R 2 = 8.5%.
Issuing quarter e¤ects are intuitive (Figure 8):
I
I
Woodhead re…nancing probability increases initially, then remains ‡at
on average (as in the PSA model used in the US).
Levelhead probability declines in mortgage age, except for mortgages
with few lifetime chances to be re…nanced at attractive rates.
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
17 / 27
Estimating the Mixture Model
0
Refinancing Probability
.1
.2
.3
Mixture Model Fit (Figure 7)
-3
-2
-1
0
1
2
3
4
Incentives
Observed Refinancing Probability
Model-Predicted Refinancing Probability
Fraction of Levelheads
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
18 / 27
Estimating the Mixture Model
Mixture Model Results (2)
Full mixture model has pseudo R 2 = 15.7%.
Most demographic variables move levelhead proportion and woodhead
re…nancing probability, or equivalently inattention and inertia, in the
same direction.
I
Inertia and inattention as …tted from demographics have a
cross-sectional correlation of 0.67; we can reject perfect correlation.
Age reduces attention while education and income increase it among
younger, less educated, and poorer households.
Financial wealth and housing wealth have opposite e¤ects
I
Highest attention for households with large houses relative to their
…nancial wealth.
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
19 / 27
Estimating the Mixture Model
84
70
65
60
56
52
48
44
39
34
26
-.02
-.01
Change in Probability
0
.01
.02
.03
.04
E¤ects of Age (Figure 9A)
Rank of Age (Years)
Levelhead
Andersen et al (2015)
Woodhead Refinancing
Inattention and Inertia
Mortgage Design 2015
20 / 27
Estimating the Mixture Model
20
16
15
12
10
7
-.02
-.01
Change in Probability
0
.01
.02
.03
.04
E¤ects of Education (Figure 10A)
Rank of Education (Years)
Levelhead
Andersen et al (2015)
Woodhead Refinancing
Inattention and Inertia
Mortgage Design 2015
21 / 27
Estimating the Mixture Model
61
0
1.
.9
34
.7
84
.6
95
.6
27
.5
59
.4
76
.3
93
.3
20
.2
38
0.
13
6
-.02
-.01
Change in Probability
0
.01
.02
.03
.04
E¤ects of Income (Figure 11A)
Rank of Income (Million DKR)
Levelhead
Andersen et al (2015)
Woodhead Refinancing
Inattention and Inertia
Mortgage Design 2015
22 / 27
Estimating the Mixture Model
67
2.
0
.5
68
.3
02
.1
72
.0
87
.0
3
-.0
32
-.1
24
-.2
41
-.4
23
-1
.3 -.02
05
-.01
Change in Probability
0
.01
.02
.03
.04
E¤ects of Financial Wealth (Figure 12A)
Rank of Financial Wealth (Million DKR)
Levelhead
Andersen et al (2015)
Woodhead Refinancing
Inattention and Inertia
Mortgage Design 2015
23 / 27
Estimating the Mixture Model
60
0
5.
71
0
2.
12
7
2.
78
9
1.
53
4
1.
33
0
1.
17
6
1.
01
2
1.
.8
38
.6
54
0.
39
0
-.02
-.01
Change in Probability
0
.01
.02
.03
.04
E¤ects of Housing Wealth (Figure 13A)
Rank of Housing Wealth (Million DKR)
Levelhead
Andersen et al (2015)
Woodhead Refinancing
Inattention and Inertia
Mortgage Design 2015
24 / 27
Conclusion
Re…nancing in Household Finance
We propose a mixture model of household types to capture
heterogeneity in propensity to re…nance.
I
I
Distinguish inattention (low levelhead probability) and inertia (low
woodhead re…nancing probability).
Household characteristics generally move inertia and inattention in the
same direction.
Demographic e¤ects are intuitive.
I
I
Inertia and inattention increase with age, decrease with education and
income.
Financial wealth (proxy for cost of time?) and housing wealth have
opposite e¤ects.
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
25 / 27
Conclusion
Next Steps
Enriching the set of household types, looking for active behavioral
patterns.
For example, “roundheads” re…nance when interest saving or coupon
reduction reaches a round number.
I
I
I
We …nd some evidence for a “new bond available with 2% lower
coupon” e¤ect.
But the improvement in the overall model …t is modest, because few
households reach this point.
Demographic patterns discussed above are robust to this change in
model speci…cation.
Also working on a better model of type persistence. Ultimate goal is
a richer dynamic characterization of multiple household types.
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
26 / 27
Conclusion
Some Thoughts on Mortgage Policy
The Danish mortgage system is impressively well designed.
But it still places the burden of the re…nancing decision on
households.
I
I
Many people, particularly poorer and less educated people, get this
wrong.
Errors of omission can be expensive for these people.
Errors of omission increase the value of mortgage bonds, lowering
yields in equilibrium.
I
I
Thus, sophisticated borrowers gain at the expense of the less
sophisticated.
A troublesome phenomenon in an age of inequality.
This cross-subsidy makes it harder for individual mortgage lenders to
introduce new products (Gabaix and Laibson 2006).
I
I
An automatically re…nancing “ratchet” bond would help the
unsophisticated but hurt the sophisticated, who would otherwise be the
natural early adopters.
In this situation there is a case for public policy to force the issue.
Andersen et al (2015)
Inattention and Inertia
Mortgage Design 2015
27 / 27

Download Report