Empirical evidence on the Euler equation for

consumption and output in the US

Guido Ascari Leandro M. Magnusson

University of Oxford University of Western Australia

Sophocles Mavroeidis
University of Oxford

February 15, 2016

Abstract

We apply recently developed econometric methods that are robust to weak in-
struments and exploit information in possible time-variation in the dynamics of the
economy, to study the empirical evidence on the Euler equation for consumption and
output using aggregate US post-war data. We consider several extension of the baseline
Euler equation model proposed in the literature that include habits, hand-to-mouth
consumers and credit constraints. We find that (i) the baseline Euler equation does
not fit the dynamics of consumption; (ii) all of the proposed extensions of the baseline
model fit the data and they are empirically indistinguishable; (iii) identification of the
intertemporal elasticity of substitution (IES) is very weak; (iv) high values of the IES
are consistent with the data only when a very high degree of habits/hand to mouth or
credit constraints is present; (v) there is no evidence of structural instability in any of
the models.
Keywords: Consumption, Euler Equation, New Keynesian IS curve.
JEL classification: E12, C12

1
1 Introduction

The Euler equation model for consumption and output is a very important ingredient
in most modern macroeconomic models. It features in most dynamic stochastic general
equilibrium (DSGE) models used for policy analysis. Yet, the empirical evidence on the
model seems rather mixed. Numerous studies have pointed out that the baseline Euler
equation model does not fit the consumption data. The main reason is that in the data
consumption moves too much with output and appears completely unresponsive to the
real interest rate. To address these shortcomings, various extension to the baseline model
have been proposed in the literature. One strand of the literature has introduced habits
in consumption, which generate history dependence and can account for unrepsonsiveness
to the real interest rate (e.g., Furher, 2000; Amato and Laubach, 2004). Another strand
of the literature, pioneered by Campbell and Mankiw (1989),1 introduced hand-to-mouth
consumption, which can explain the strong comovement between consumption and output.
Other papers have emphasized the role of credit constraints which can also account for a
large sensitivity of consumption to output (e.g., Iacoviello and Neri, 2010).
With very few exceptions, most papers in the literature estimated the Euler equation
model using econometric methods that are not robust to the problem of weak instruments,
which is known to be important in this application, see Stock, Wright and Yogo (2002), .
Yogo (2004) and Fuhrer and Rudebusch (2004) are the only papers we are aware of that
have addressed this issue from the finance and macro perspectives, respectively. Of those,
only Yogo (2004) used methods of inference that are fully robust to weak instruments,2 and
his results are based on the baseline model with homoskedastic and serially uncorrelated
errors. In our analysis, we relax those assumptions, and we also estimate several extensions
of the baseline model. In addition, we use the recently developed econometric methods in
1
See also, e.g., Mankiw (2000), Galı́ et al. (2004, 2007), Bilbiie (2008).
2
Fuhrer and Rudebusch (2004) use methods that have better finite sample properties than the standard
2-step GMM estimator of Hansen (1982) commonly employed in the literature on Euler equation models,
but these estimators are not fully robust to weak instruments.

2
Magnusson and Mavroeidis (2014) which use subsample information in order to test the
stability of the model’s parameters and can potentially improve the identification of the
stable parameters of the model.
Our findings can be summarized as follows. First, we corroborate the evidence in the
literature that the basic Euler equation model does not fit the dynamics of the consumption
data. In the case of GDP, the model fits with an intertemporal elasticity of substitution
(IES) equal to zero, suggesting that output is completely unresponsive to the real interest
rate. Second, all of the proposed extensions in the literature that we study fit the data,
and they are empirically indistinguishable. Third, identification of the IES parameter is
very weak. Forth, high values of the IES are consistent with the data only when a very
high degree of habits, hand-to-mouth consumers, or credit constraints is present. Finally,
there is no evidence of structural instability in any of the models.
The above results show that there is a clear need to move away from the basic Euler
equation model for consumption, but there is very little scope for discriminating amongst
the various extensions proposed in the literature on the basis of aggregate macro data.
Moreover, once we introduce additional features to the basic model, the identification
of the IES becomes very weak. The intuition for that is that identification of the IES
comes from the rational unconstrained consumers, and if the parameters are such that this
fraction becomes small, there will be little relevant variation in the data to identify the
IES. On the positive side, there is no evidence to suggest that the Euler equation model
is structurally unstable, but, unfortunately, structural change or time-variation in other
aspects of the economy (e.g., a time-varying trend inflation), do not seem to help improve
the identification of the IES, unlike the case of the New Keynesian Phillips curve studied
in Magnusson and Mavroeidis (2014).

3
2 Models

We consider the baseline Euler equation model for consumption and a number of exten-
sions of that model, listed in Table 1. Detailed description of each model is given in the
subsections that follow.

2.1 The Baseline Euler Equation

P∞ i u(C
Households maximize a standard intertemporal utility function of the form Ut = i=0 β t+i ),

where β is the subjective discount factor. The instantaneous utility function exhibits
1−1/σ
Ct+i
CRRA, i.e., u(Ct+i ) = 1−1/σ , where C represents consumption and σ the inverse of the
degree of risk aversion. The first order condition for the household optimal consumption
reads  
1 uc (Ct+1 ) Pt
= Et β . (1)
1 + it uc (Ct ) Pt+1

Log-linearizing this expression yields the standard Euler equation:

ct = Et (ct+1 ) − σrt , (2)

where ct denotes the log-deviation of consumption from steady state and rt = it − Et πt+1
denotes the log-deviation of the ex-ante real interest rate from steady state. Equation (2)
is the baseline specification that we estimate by adding an error term to replace (2) with

ct = Et (ct+1 ) − σrt + εt .]

[Sophocles: we need to say more to motivate the rest and link with the literature, e.g.]
This basic model has been studied thoroughly in both the macroeconomics and finance
literatures (two review article would be essential here). The general message from these
studies is that (2) does not work well because consumption growth is both unresponsive
to the real interest rate and seems to be predictable by lags of other series, e.g., output

4
growth, e.g., Campbell and Mankiw (1989).

2.2 Habits

Many papers in the literature point to the need for adding habits in consumption to
the baseline model in order to match some features of the data, as the hump-shaped
inertial response of consumption to various shocks (e.g., Fuhrer, 2000). Moreover, habits
in consumption are now a standard feature in medium-scale DSGE New Keynesian models
(e.g., Smets and Wouters, 2003, 2007, Christiano et al., 2005)?. Dennis (2009) analyzes
the different approaches to modelling consumption habits in the literature. We consider
here the two main ways of modelling habits: ‘internal’ and ‘external’. Following Fuhrer
(2000), we assume multiplicative habits, so that the utility function takes the following
form:  1−1/σ
ct
Ht
u(ct ) = (3)
1 − 1/σ

and that the stock of habits Ht evolves as:

 γ
1−D
Ht = cD C
t−1 t−1 , (4)

where Ct denotes aggregate consumption, γ ∈ [0, 1] is the parameter that governs the
degree of habits,3 and D = {0, 1} indicates whether habits are internal or external. If
D = 0 then the consumer is concerned with the level of her current consumption relative
to the aggregate consumption in the previous period, so we are in the case of external habits
where the consumer wants to “keep up with the Joneses”. If D = 1 then the consumer
is concerned with the level of her current consumption relative to her consumption in the
previous period, so we are in the case of internal habits where the consumer is concerned
with keeping up with her own standard of living.
3
γ = 0 corresponds to the baseline model, and γ > 1 is inadmissible because in that case steady state
utility would be decreasing in consumption, see Fuhrer (2000, p. 370).

5
 In the case of external habits, the log-linearized Euler equation can be written as

1 σ
∆ct = (Et (ct+1 ) − ct−1 ) − rt , (5)
1 + γ (1 − σ) 1 + γ (1 − σ)

see Dennis (2009) for the derivation.4 In the case of internal habits, the log-linearized Euler
equation can be written as:

γ (1 − σ) σ (1 − γβ)
Et ∆ct+1 = (∆ct + βEt ∆ct+2 ) + rt . (6)
1 + γβ (γ − σ (1 + γ)) 1 + γβ (γ − σ (1 + γ))

Following Fuhrer and Rudebusch (2004), we also estimate a flexible expanded Euler
Equation, that nests the previous equations, to allow for two lags in consumption:

ct = φEt (ct+1 ) − σrt + α1 ct−1 + α2 ct−2 , (7)

with φ ∈ [0, 1].

2.3 Hand-to-mouth consumers

Assuming hand-to-mouth consumers (HTMC) is another recent modification introduced

in the literature to model consumption behavior (Galı́ et al., 2004 , 2007, Bilbiie, 2008,
Ascari et al., 2014) . We will assume that the economy is populated by HTMC that
simply consume their income. Aggregate consumption in levels is thus given by: Ct =
 
1 − λ̃ Cs,t +λ̃CH,t where Cs,t , CH,t is consumption of optimizing consumers and of HTMC,
respectively, and λ̃ is the fraction of HTMC. Log-linearizing this expression yields: ct =
λ̃CH
(1 − λ) cs,t +λcH,t where λ = C is the fraction of total consumption consumed by HTMC.
In the case of no habits, the simple Euler equation above (2) will apply to optimizing
4
Note that Dennis (2009) imposes the restriction that σ ∈ (0, 1), i.e., risk aversion coefficient strictly
greater than 1. This is necessary for the joint concavity of the utility function (3). Moreover, as (5) and (6)
show, this is also a sufficient condition for the intertemporal elasticity of substitution to have the correct
sign with respect to the ex-ante real interest rate.

6
 ct −λcH,t
consumers, so that: cs,t = Et cs,t+1 − σrt . Using cs,t = 1−λ yields:

ct = Et (ct+1 − λ∆cH,t+1 ) − (1 − λ) σrt . (8)

Since HTMC simply consume their income, it follows that cH,t = lH,t + ωH,t , where
ωH,t is the real wage and lH,t is hours worked of HTMC. Substituting for Et ∆cH,t+1 =
Et (∆lH,t+1 + ∆ωH,t+1 ) we get

ct = Et (ct+1 ) − λEt (∆lH,t+1 + ∆ωH,t+1 ) − (1 − λ) σrt . (9)

This specification is very similar to Campbell and Mankiw’s (1989), where ∆lH,t+1 +
∆ωH,t+1 represents the expected future change in income of HTMC.
We can proceed in the same way in the case the utility function features habits in
ct −λcH,t
consumption, by using cs,t = 1−λ to substitute for cs,t in the optimizing consumers
Euler equation to get an expression that depends only on aggregate consumption and the
consumption of HTMC. Combining optimizing consumers with external habits and HTMC
therefore yields the following expression:
 
1 (1 − λ) σ 1
∆ct = Et (∆ct+1 ) − rt + λ ∆cH,t − Et (∆cH,t+1 ) , (10)
γ (1 − σ) γ (1 − σ) γ (1 − σ)

while assuming internal habits for optimizing consumers gives:

γ (1 − σ) σ (1 − γβ) (1 − λ)
Et ∆ct+1 = (∆ct + βEt (∆ct+2 )) + rt + (11)
1 + γβ (γ − σ (1 + γ)) 1 + γβ (γ − σ (1 + γ))
 
γ (1 − σ)
+ λ Et (∆cH,t+1 ) − (∆cH,t + β∆cH,t+2 ) .
1 + γβ (γ − σ (1 + γ))

The above expressions feature cH,t , i.e., the consumption of HTMC. Following the
literature (e.g., Galı́ et al., 2004, Bilbiie and Straub, 2012, 2013), we can make assump-
tions which will enable us to substitute for cH,t and thus obtain equations containing only

7
observable aggregate variables. We follow Campbell and Mankiw (1989, p. 188) in assum-
ing that a constant fraction of total labor income goes to HTMC. HTMC has only labor
income and they just consume their labor income each period by assumption. Hence, since
consumption and income are in log-deviations, it follows that we can replace ∆cH,t with
the change in aggregate labor income. This is what we use in the empirical work below.

2.4 Wealth

Finally we will also estimate model specifications that includes wealth measures. There
are two theoretical models that suggest wealth measures could enter the Euler equation:
overlapping generations (OLG) models and borrowing constraints models. A standard
OLG framework a la Blanchard-Yaari will deliver a dynamic equation for consumption
dynamics which is similar to the standard Euler equation but with an extra term which is
aggregate wealth (e.g., Castelnuovo and Nistico, 2010). In this framework, the coefficient
on the wealth measures can be interpreted as a measure of people nonparticipating in the
financial markets.5 In a model with borrowing constraints, instead, the Euler Equation
for constrained agents will be the same as the one for unconstrained optimizing agents
augmented by the multiplier attached to the borrowing constraint (e.g., Iacoviello and
Neri, 2010). The multiplier measures the shadow value of relaxing the constraint, i.e., the
shadow value of one unit of wealth. Thus, it seems sensible to assume that the multiplier is
a function of wealth, or rather of that part of wealth used as a collateral in the borrowing
constraint (e.g., housing wealth). Proceeding as above for the HTMC, the coefficient on the
measure of wealth could then be interpreted as the fraction of people subject to borrowing
constraints.
We will therefore also estimate equations (2), (5 ) and (6) augmented with a wealth
5
To be more precise, in an OLG model the parameter can be interpreted as the percentage of agents
trading in the financial market that are replaced each period by newcomers holding zero-wealth. It can also
be interpreted as the effective average planning horizon of households when they trade in financial assets
(see Castelnuovo and Nistico, 2010).

8
variable as follows
µ
ct = Et (ct+1 ) − σrt − ∆wt+1 , (12)
1−µ
1 σ µ
∆ct = Et (ct+1 − ct−1 ) − rt − ∆wt+1 (13)
1 + γ (1 − σ) 1 + γ (1 − σ) 1−µ

and

γ (1 − σ)
ct = Et (ct+1 ) − (∆ct + βEt (∆ct+2 ))
1 + γβ (γ (1 − σ) − σ)
σ (1 − γβ) µ
− rt − ∆wt+1 , (14)
1 + γβ (γ (1 − σ) − σ) 1−µ

where µ ∈ [0, 1).

We also consider an extension of the Fuhrer and Rudebusch model (7) that includes
wealth to capture credit constraints:

ct = φEt (ct+1 ) − σrtt+1 − µL ∆wt+1 + α1 ct−1 + α2 ct−2 . (15)

3 Data

We use quarterly aggregate time series data for the US over the period 1956q1 to 2014q1
collected from variety of sources. For the nominal interest rate it we use the quarterly
average of the effective Federal Funds rate. For ct we use either real personal consumption
expenditures or real GDP in the nonfarm business sector. Inflation πt is measured by the
deflator that corresponds to each of the series used for ct . Real wealth wt is measured by the
household net worth level obtained obtained from the Federal Reserve Board. For aggregate
labor income we use the labor share time real GDP in the nonfarm business sector. Detailed
description of the data, sources and transformation is given in the supplementary appendix.

9
4 Econometric methodology

All of the specification in Table 1 fit into the generalized instrumental variables framework
of Hansen and Singleton (1982), where unobserved expectations terms are replaced by their
realizations, and orthogonality conditions are obtained by assuming that the residuals are
uncorrelated with any variables that are predetermined at time t. This follows from rational
expectations and the assumption that εt is an innovation process. The models can then be
estimated by the Generalized Method of Moments (GMM) using predetermined variables
as instruments.
It is well-known that estimation of Euler equation models for consumption or output
may be suffer from the problem of weak instruments, see Stock Wright and Yogo (2002).
Thus, our econometric analysis relies solely on methods of inference that are robust to weak
instruments. Specifically, we construct confidence intervals, or more generally confidence
sets, for the structural parameters that contain their true values with at the least as high
probability (asymptotically) as their confidence level.6 One such method is the S set of
Stock and Wright (2000), which is constructed as follows. We specify a grid of points
within the parameter space. For each of those points, we test whether the identifying
restrictions of the model hold using a Wald test (Stock and Wright call this the S test).
The 95%, say, confidence set that is known as the S set consists of all the points in the
grid that have not been rejected by a 5% level S test.
There are other weak-identification robust methods that have been proposed in the
literature, see Stock et al. (2002). Some (e.g. Moreira’s 2003 Conditional Likelihood
Ratio test) are more powerful than the S method in the case of the homoskedastic linear
instrumental variables regression model, but this power property has not been extended
to their GMM counterparts (e.g., Kleibergen’s 2005 version of the CLR test). In addition,
these methods are more involved computationally, and not as straightforward to explain,
so we follow Mavroeidis, Plagborg-Moller and Stock (2014) in reporting results only based
6
That is, a 90% confidence interval includes the true value at least 90% of the time in repeated samples.

10
on the S sets.7
In addition to the S sets, we also apply a new method developed by Magnusson and
Mavroeidis (2014), the qLL-S set. The qLL-S set combines the average information on
the moment conditions over the sample, which is what the S set uses, with information
on the validity of the moment conditions over subsamples. It can be thought of as using
subsample information as additional instruments. This subsample information is relevant
in two cases: (i) when the parameters of the model are unstable, or (ii) when the parameters
of the model are constant but there is time variation in other parts of the distribution of
the data. This can arise, for example, if there are changes in monetary policy, such as
time-variation in the inflation target, or changes in the policy rule, that do not affect the
preference parameters that determine the Euler equation for consumption. In case (i), the
qLL-S test can be interpreted as a joint parameter stability and identification test, so a
nonrejection is an indication of parameter stability. In case (ii), the qLL-S test can have
more power than the corresponding S test if the information that comes from structural
change elsewhere in the economy is sufficiently strong. Hence, in either case, the qLL-S
sets will usefully complement the results of the S sets.

5 Empirical Results

We report 90% and 95% confidence set on the structural parameters based on the S test of
Stock and Wright (2000), and the qLL-S test of Magnusson and Mavroeidis (2014) over the
full sample period 1956-2014. Results over the post 1984 sample are very similar and can
be found in the supplementary Appendix. The results in Figures 1 through 4 use aggregate
real non-durable consumption expenditure for ct , while the results in Figures 5 through 8
use aggregate real GDP in the nonfarm business sector.
7
Unreported resutls show that the confidence sets based on Kleibergen’s (2005) extention of Moreira’s
(2003) conditional likelihood ratio test to GMM are qualitatively very similar to the S sets reported.

11
Mnemonic Formula

Baseline (2) ct = Et (ct+1 ) − σrt + εt ,

1 σ
Ext. habits (5) ∆ct = 1+γ(1−σ) (Et (ct+1 ) − ct−1 ) − 1+γ(1−σ) rt + εt
γ(1−σ) σ(1−γβ)
Int. habits (6) Et ∆ct+1 = 1+γβ(γ−σ(1+γ)) (∆ct + βEt ∆ct+2 ) + 1+γβ(γ−σ(1+γ)) rt + εt

HTMC (9) ct = Et (ct+1 ) − λEt (∆lst+1 + ∆yt+1 ) − (1 − λ) σrt + εt

1 (1−λ)σ
HTMC + EH (10) ∆ct = γ(1−σ) Et (∆ct+1 ) − γ(1−σ) rt
 
1
+λ ∆cH,t − γ(1−σ) Et (∆cH,t+1 ) + εt
γ(1−σ) σ(1−γβ)(1−λ)
HTMC + IH (11) Et ∆ct+1 = 1+γβ(γ−σ(1+γ)) (∆ct + βEt (∆ct+2 )) + 1+γβ(γ−σ(1+γ)) rt
 
γ(1−σ)
+λ Et (∆cH,t+1 ) − 1+γβ(γ−σ(1+γ)) (∆cH,t + β∆cH,t+2 )
µ
Wealth (12) ct = Et (ct+1 ) − σrt − 1−µ ∆wt+1
1 σ µ
Wealth + EH (13) ∆ct = 1+γ(1−σ) Et (ct+1 − ct−1 ) − 1+γ(1−σ) rt − 1−µ ∆wt+1
γ(1−σ)
Wealth + IH (14) ct = Et (ct+1 ) − 1+γβ(γ(1−σ)−σ) (∆ct + βEt (∆ct+2 ))
σ(1−γβ) µ
− 1+γβ(γ(1−σ)−σ) rt − 1−µ ∆wt+1

FR (7) ct = φEt (ct+1 ) − σrt + α1 ct−1 + α2 ct−2 + εt

FR + Wealth (15) ct = φEt (ct+1 ) − σrtt+1 − µL ∆wt+1 + α1 ct−1 + α2 ct−2

Table 1: List of models estimated in the paper

12
5.1 Results based on consumption data

Consider first the results reported in Figures 1 and 2. The former reports S sets, while
the latter reports the corresponding qLL-S sets. Panel (a) in both figures reports the
respective one-dimensional 95% confidence set (interval) for σ in the baseline model (2) if
nonempty. Panels (b) and (c) report 2-dimensional 95% (light) and 90% (dark) confidence
sets for (σ, γ) in the models with external (5) and internal (6) habits, respectively. Panel
(d) reports 2-dimensional 95% (light) and 90% (dark) confidence sets for (σ, λ) in the
model with HTMC (9). Panels (e) and (f) report 3-dimensional 95% confidence sets for
(σ, γ, λ) in the models with HTMC combined with external (10) and internal (11) habits,
respectively. Panel (g) reports 2-dimensional 95% (light) and 90% (dark) confidence sets
for (σ, µ) in the model with wealth (12), while panels (h) and (i) report the corresponding
3-dimensional confidence sets for (σ, γ, µ) in the model with wealth and habits. Note that
the λ = 0 and γ = 0 contours in panel (e) correspond to the 95% confidence sets in panels
(b) and (d), respectively, and similarly for the other 3D plots.
The S sets use average information from the full sample, while the qLL-S sets combine
the information used by the S sets with information from subsamples. This is akin to
using more instruments, so qLL-S sets have the potential to be more informative than the
corresponding S sets if the information in the subsamples is sufficiently strong, but the
converse is also possible. By comparing Figures 1 and 2, we notice that the qLL-S sets
are bigger than their S counterparts, which suggests that the subsample information is
not particularly useful in this application. However, because the qLL-S test has power
also against instability of the parameters of the model, the fact that the qLL-S sets are
non-empty can also be interpreted as evidence that the parameters of the model are stable.
We now discuss the results for each of the models in turn, starting from the baseline
model. The 95% S set in Figure 1(a) is empty, meaning that there is no value of σ ≥ 0
for which the identifying restrictions of the model are statistically acceptable at the 5%
level of significance. This is consistent with previous findings in the literature, e.g., Yogo

13
 No Habits External Habits Internal Habits
(a) (b) (c)
baseline

(d) (e) (f)

hand-to-mouth

(g) (h) (i)

wealth

Figure 1: 90% and 95% S confidence sets (Stock and Wright, 2000) for σ, γ λ and µ in
the Euler Equations listed in Table 1. Instruments: constant, three lags of ∆ct , and three
lags of rt−1 , where ct is consumption expenditure. Newey and West (1987) HAC. Period:
1956q1-2014q1. Labor income is used for cH,t , and household net worth level is used for
wealth wt .

14
 No Habits External Habits Internal Habits
(a) (b) (c)
baseline

(d) (e) (f)

hand-to-mouth

(g) (h) (i)

wealth

Figure 2: 90% and 95% qLL-S confidence sets (Magnusson and Mavroeidis, 2000) for σ, γ
λ and µ in the Euler Equations listed in Table 1. Instruments: constant, three lags of ∆ct ,
and three lags of rt−1 , where ct is consumption expenditure. Newey and West (1987) HAC.
Period: 1956q1-2014q1. Labor income is used for cH,t , and household net worth level is
used for wealth wt .

15
(2004, Table 3), who used the same method. All other confidence sets in Figures 1 and 2
are nonempty, indicating that all of the extensions of the baseline model that we consider
are consistent with the data, and there is no evidence of parameter instability.
Turning to the model with external habits in Figure 1(b) we see that γ is significantly
different from zero (in fact, greater than 0.5 at the 90% level), while σ is not significantly
different from zero, and the confidence set is very tight in the direction of σ – it excludes all
σ > 0.08. This shows that even with substantial external habits, the model requires a very
high coefficient of risk aversion σ −1 to fit the data. On that dimension, the fit of the model
is somewhat better with internal habits, Figure 1(c), since, with γ = 1, the confidence set
for σ contains much larger values.8 Note that this comes at the cost of more estimation
uncertainty, since the confidence sets are a lot bigger than with external habits.
Next, consider the results for the model with HTMC, Figure 1(d). The fraction of
HTMC λ is significantly greater than zero, but otherwise not very well identified. Specifi-
cally, the 95% confidence set includes λ = 1, at which point σ becomes completely uniden-
tified: if all consumers are HTMC, then aggregate consumption will not respond at all to
the real interest rate at any level of IES σ. The upshot from this is that the model is
now able to fit the data with reasonable levels of risk aversion, or sufficiently high IES σ,
provided the fraction of HTMC is sufficiently high.
We now move to the model that combines HTMC with external or internal habits,
Figures 1(e) and (f), respectively. The 3-dimensional confidence sets are both very large,
containing a significant part of the parameter space, with the one for the model with
internal habits being larger. We now find that the model with internal habits can fit the
data with significantly higher σ when we also allow for HTMC (λ > 0). The 3D pictures
also confirm that either λ > 0 or γ > 0 is necessary for the model to fit the data at the
95% level.
Figure 1(g) gives the 2-dimensional 95% (light) and 90% (dark) confidence sets for (µ, σ)
8
The one-dimensional, projection confidence set for σ is disjoint in this case, specifically, it includes
σ ≤ 0.7 and σ ≥ 1.7.

16
in the model with wealth (12). We notice that µ > 0 in those confidence sets, and that σ
can be high if the coefficient on wealth µ, which measures the intensity of the borrowing
constraint, is sufficiently high, though this seems to be caused by weaker identification of
σ. The result is very similar to the HTMC specification. Combining borrowing constraints
with habits, Figures 1(h) and (i), has very similar implications to the corresponding HTMC
specifications, Figures 1(e) and (f). The joint 3-dimensional 95% confidence sets for (µ, γ, σ)
exclude µ = γ = 0, but they are generally very large (covering close to half of the parameter
space), indicating that these parameters are fairly weakly identified.
The results for the Fuhrer and Rudebusch (2004) inspired specifications in Table 1 are
given separately in Figures 3 and 4. The former reports the S sets and the latter reports
the qLL-S sets. Panel (b) reports a 2-dimensional confidences set for (φ, σ) in (7) and
Panel (c) reports 3-dimensional 95% confidence sets for (φ, µ, σ) in (15), where, in both
cases, the coefficients on lagged consumption growth α1 and α2 have been concentrated out.
Figure 3(b) shows that the FR model (7) fits the data at the 95% level with a coefficient
of 0 < φ < 1, but σ is extremely low and not signficantly different from zero (σ < 0.1).
Adding wealth to the baseline FR model allows it to fit the data with a much higher σ,
albeit still significantly less that 1 at the 95% (Figure 3(c)). The results from the qLL-S sets
in 4 are similar to those from the S sets, with some minor differences that don’t change the
basic conclusions that the baseline FR Euler equation is very flat, and that the coefficient
on future consumption, measuring the degree of forward-lookingness, is poorly identified.

5.2 Results based on GDP

We now turn to the results that are based on the GDP as a measure of ct in the models.
Figures 5 and 6 give the S and qLL-S confidence sets, respectively, for the various models
in Table 1 except for the Fuhrer-Rudebusch specifications given in Figures 7 and 8. The
description of each graph is the same as for the corresponding to Figures 5 through 4 in the
previous subsection. All in all, the results based on the GDP measure are very similar to

17
 no lags 2 lags 2 lags and wealth
(a) (b) (c)
Fuhrer-Rudebusch

Figure 3: 90% and 95% S confidence sets (Stock and Wright, 2000) for σ, φ and µ in the
Fuhrer-Rudebusch equations listed in Table 1. Instruments: constant, three lags of ∆ct ,
and three lags of rt−1 , where ct is consumption expenditure. Newey and West (1987) HAC.
Period: 1956q1-2014q1. Labor income is used for cH,t , and household net worth level is
used for wealth wt .

no lags 2 lags 2 lags and wealth

(a) (b) (c)
Fuhrer-Rudebusch

Figure 4: 90% and 95% qLL-S confidence sets (Magnusson and Mavroeidis, 2014) for σ, φ
and µ in the Fuhrer-Rudebusch equations listed in Table 1. Instruments: constant, three
lags of ∆ct , and three lags of rt−1 , where ct is consumption expenditure. Newey and West
(1987) HAC. Period: 1956q1-2014q1. Labor income is used for cH,t , and household net
worth level is used for wealth wt .

18
the ones based on consumption data, and so the same qualitative conclusions hold. A few
notable differences are the following. First, the baseline Euler equation model with GDP
seems to fit the data at the 95% level, since the S set is nonempty – compare Figures 1(a)
and 5(a). The Euler equation is still estimated to be very flat and σ is not significantly
different from zero. Second, the habit parameter γ appears to be completely unidentified
in the GDP specifications – compare, e.g., Figures 1(b) and 5(b). The same is true for
coefficient on wealth, which measures borrowing constraints – compare, e.g., Figures 1(g)
and 5(g). This is, of course, related to the finding that the baseline Euler equation (2)
fits the GDP data, so neither habits nor borrowing constraints are needed for this. Third,
the fraction of HTMC is significantly less 1 at the 5% level – compare, e.g., Figures 1(d)
and 5(d). Finally, the measure of forward-lookingness φ in the Fuhrer and Rudebusch
specification is completely unidentified – compare Figures 3(b) and 7(b). So, overall, the
GDP data is somewhat less informative about the Euler equation model and its extensions
than the aggregate consumption data.

5.3 Conclusions

In this paper, we revisited the empirical evidence on the log-linear Euler Equation model
of consumption using aggregate US data over the period 1956 to 2014. Because of the
well-known problems with the identification of this model, our inference was based on both
well-established and new econometric methods that are robust to weak instruments.
The main findings of this paper can be summarized as follows. First, we corroborate
previous findings that the basic Euler equation model is at odds with the data, in the
sense that consumption growth does not react to the real interest rate and appears to be
backward-looking. In the case of output growth, the model can only fit the data with a very
low elasticity of substitution, so the New Keynesian IS curve is flat. Second, extensions of
the model that allow for habits, hand-to-mouth consumers or borrowing constraints can all
fit the data with a much higher elasticity of substitution. Intuitively, this happens because

19
 No Habits External Habits Internal Habits
(a) (b) (c)
baseline

(d) (e) (f)

hand-to-mouth

(g) (h) (i)

wealth

Figure 5: 90% and 95% GMM-S confidence sets for σ, γ λ and µ in the Output Gap Euler
Equations. Instruments: constant, three lags of ∆ct , and three lags of (rt−1 − πt ), where ct
is GDP non-farmer business sector. Newey and West (1987) HAC. Period: 1956q1-2014q1.
Labor income, measured by the GDP labor share, is a proxy for cH,t , and household net
worth level is used for wealth wt .

20
 No Habits External Habits Internal Habits
(a) (b) (c)
baseline

(d) (e) (f)

hand-to-mouth

(g) (h) (i)

wealth

Figure 6: 90% and 95% qLL-S confidence sets for σ, γ, λ and µ in the Output Gap Euler
Equations. Instruments: constant, three lags of ∆ct , and three lags of (rt−1 − πt ), where ct
is GDP non-farmer business sector. Newey and West (1987) HAC. Period: 1956q1-2014q1.
Labor income, measured by the GDP labor share, is a proxy for cH,t , and household net
worth level is used for wealth wt .

21
 no lags 2 lags 2 lags and wealth
(a) (b) (c)
Fuhrer-Rudebusch

Figure 7: 90% and 95% GMM-S confidence sets for σ, φ, and µ in the Fuhrer-Rudebusch
Equations. Instruments: constant, three lags of ∆ct , and three lags of (rt−1 − πt ), where ct
is GDP non-farmer business sector. Newey and West (1987) HAC. Period: 1956q1-2014q1.
Household net worth level is used for wealth wt

no lags 2 lags 2 lags and wealth

(a) (b) (c)
Fuhrer-Rudebusch

Figure 8: 90% and 95% qLL-S confidence sets for σ, φ and µ in the Fuhrer-Rudebusch
Equations. Instruments: constant, three lags of ∆ct , and three lags of (rt−1 − πt ), where ct
is GDP non-farmer business sector. Newey and West (1987) HAC. Period: 1956q1-2014q1.
Household net worth level is used for wealth wt .

22
if the share of consumption that is due to non-optimizing behavior becomes sufficiently
high, there is little relevant variation left over in the data to identify the behavior of the
optimizing or unconstrained consumers. Third, there appears to be too little information
in the data to distinguish between those alternative extensions of the basic model. Fourth,
the data is not very informative about the degree of forward-lookingness in the Fuhrer and
Rudebusch (2004) specification. Finally, there is no evidence of parameter instability in
any of those models.

References

Amato, J., and T. Laubach (2004): “Implications of Habit Formation for Optimal
Monetary Policy,” Journal of Monetary Economics, 51(2), 305–325.

Ascari, G., A. Colciago, and L. Rossi (2014): “Limited Asset Market Participation
and Optimal Monetary Policy,” Oxford University.

Bilbiie, F. O. (2008): “Limited Asset Market Participation, Monetary Policy and (In-
verted) Aggregate Demand Logic,” Journal of Economic Theory, 140, 162–196.

Bilbiie, F. O., and R. Straub (2012): “Changes in the output Euler equation and asset
markets participation,” Journal of Economic Dynamics and Control, 36, 1659–1672.

(2013): “Asset Market Participation, Monetary Policy Rules, and the Great
Inflation,” The Review of Economics and Statistics, 95, 377–392.

Campbell, J. Y., and G. N. Mankiw (1989): “Consumption, Income, and Interest

Rates: Reinterpreting the Time Series Evidence,” in NBER Macroeconomics Annual,
ed. by O. Blanchard, and S. Fisher, pp. 185–216. MIT Press, Cambridge, MA.

Castelnuovo, E., and S. Nisticò (2010): “Stock Market Conditions and Monetary
Policy in a DSGE Model for the U.S.,” Journal of Economic Dynamics and Control,
34(9), 1700–1731.

23
Dennis, R. (2009): “Consumption Habits in a New Keynesian Business Cycle Model,”
Journal of Money, Credit and Banking, 41(5), 1015–1030.

Fuhrer, J., and G. Rudebusch (2004): “Estimating the Euler Equation for Output,”
Journal of Monetary Economics, 51(6), 1133–1153.

Furher, J. C. (2000): “Habit Formation in Consumption and its Implications for

Monetary-Policy Models,” American Economic Review, 90(3), 367–390.

Galı́, J., J. López-Salido, and J. Vallés (2004): “Rule-of-Thumb Consumers and the
Design of Interest Rate Rules,” Journal of Money, Credit and Banking, 36, 739–764.

(2007): “Understanding the Effects of Government Spending on Consumption,”

Journal of the European Economic Association, 5, 227–270.

Hansen, L. P. (1982): “Large Sample properties of Generalized Method of Moments

estimators,” Econometrica, 50, 1029–1054.

Iacoviello, M., and S. Neri (2010): “Housing Market Spillovers: Evidence from an
Estimated DSGE Model,” American Economic Journal: Macroeconomics, 2, 125–
163.

Kleibergen, F. (2005): “Testing parameters in GMM without assuming that they are
identified,” Econometrica, 73(4), 1103–1123.

Magnusson, L. M., and S. Mavroeidis (2014): “Identification using stability restric-

tions,” Econometrica, 82(5), 1799–1851.

Mankiw, N. G. (2000): “The Savers-Spenders Theory of Fiscal Policy,” American Eco-

nomic Review, 90(2), 120–125,.

Mavroeidis, S., M. Plagborg-Møller, and J. H. Stock (2014): “Empirical evidence

on inflation expectations in the New Keynesian Phillips curve,” Journal of Economic
Literature, 52(1), 124–188.

24
Moreira, M. J. (2003): “A conditional likelihood ratio test for structural models,” Econo-
metrica, 71(4), 1027–1048.

Newey, W. K., and K. D. West (1987): “A simple, positive semidefinite, heteroskedas-

ticity and autocorrelation consistent covariance matrix,” Econometrica, 55(3), 703–
708.

Smets, F., and R. Wouters (2003): “An Estimated Dynamic Stochastic General Equi-
librium Model of the Euro Area,” Journal of the European Economic Association, 1,
1123–1175.

Smets, F., and R. Wouters (2007): “Shocks and Frictions in US Business Cycles: A
Bayesian DSGE Approach,” American Economic Review, 97(3), 586–606.

Stock, J. H., and J. Wright (2000): “GMM with Weak Identification,” Econometrica,
68(5), 1055–1096.

Stock, J. H., J. H. Wright, and M. Yogo (2002): “A survey of weak instruments

and weak identification in generalized method of moments,” Journal of Business &
Economic Statistics, 20, 518–529.

Yogo, M. (2004): “Estimating the elasticity of intertemporal substitution when instru-

ments are weak,” Review of Economics and Statistics, 86(3), 797–810.

25