The ABC of Simulation Estimation with Auxiliary Statistics

Jean-Jacques Forneron∗ Serena Ng†

November 2015

Abstract
The frequentist method of simulated minimum distance (SMD) is widely used in economics
to estimate complex models with an intractable likelihood. In other disciplines, a Bayesian
approach known as Approximate Bayesian Computation (ABC) is far more popular. This
paper connects these two seemingly related approaches to likelihood-free estimation by means of
a Reverse Sampler that uses both optimization and importance weighting to target the posterior
distribution. Its hybrid features enable us to analyze an ABC estimate from the perspective of
SMD. We show that an ideal ABC estimate can be obtained as a weighted average of a sequence
of SMD modes, each being the minimizer of the deviations between the data and the model.
This contrasts with the SMD, which is the mode of the average deviations. Using stochastic
expansions, we provide a general characterization of frequentist estimators and those based on
Bayesian computations including Laplace-type estimators. Their differences are illustrated using
analytical examples and a simulation study of the dynamic panel model.

JEL Classification: C22, C23.

Keywords: Indirect Inference, Simulated Method of Moments, Efficient Method of Moments,

Laplace Type Estimator.

∗
Department of Economics, Columbia University, 420 W. 118 St., New York, NY 10025. Email:
[email protected]
†
Department of Economics, Columbia University, 420 W. 118 St. Room 1117, New York, NY 10025. Email
Serena.Ng at Columbia.edu. Financial support is provided by the National Science Foundation, SES-0962431.
We thank Richard Davis for discussions that initiated this research, Neil Shephard, Christopher Drovandi, two
anonymous referees, and the editors for many helpful suggestions. Comments from seminar participants at Columbia,
Harvard/MIT, UPenn, and Wisconsin are greatly appreciated. All errors are our own.
1 Introduction

As knowledge accumulates, scientists and social scientists incorporate more and more features into
their models to have a better representation of the data. The increased model complexity comes at
a cost; the conventional approach of estimating a model by writing down its likelihood function is
often not possible. Different disciplines have developed different ways of handling models with an
intractable likelihood. An approach popular amongst evolutionary biologists, geneticists, ecologists,
psychologists and statisticians is Approximate Bayesian Computation (ABC). This work is largely
unknown to economists who mostly estimate complex models using frequentist methods that we
generically refer to as the method of Simulated Minimum Distance (SMD), and which include such
estimators as Simulated Method of Moments, Indirect Inference, or Efficient Methods of Moments.1
The ABC and SMD share the same goal of estimating parameters θ using auxiliary statistics
ψb that are informative about the data. An SMD estimator minimizes the L2 distance between
ψb and an average of the auxiliary statistics simulated under θ, and this distance can be made as
close to zero as machine precision permits. An ABC estimator evaluates the distance between ψb
and the auxiliary statistics simulated for each θ drawn from a proposal distribution. The posterior
mean is then a weighted average of the draws that satisfy a distance threshold of δ > 0. There are
many ABC algorithms, each differing according to the choice of the distance metric, the weights,
and sampling scheme. But the algorithms can only approximate the desired posterior distribution
because δ cannot be zero, or even too close to zero, in practice.
While both SMD and ABC use simulations to match ψ(θ) to ψb (hence likelihood-free), the rela-
tion between them is not well understood beyond the fact that they are asymptotically equivalent
under some high level conditions. To make progress, we focus on the MCMC-ABC algorithm due to
Marjoram et al. (2003). The algorithm applies uniform weights to those θ satisfying kψb − ψ(θ)k ≤ δ
and zero otherwise. Our main insight is that this δ can be made very close to zero if we combine
optimization with Bayesian computations. In particular, the desired ABC posterior distribution
can be targeted using a ‘Reverse Sampler’ (or RS for short) that applies importance weights to a
sequence of SMD solutions. Hence, seen from the perspective of the RS, the ideal MCMC-ABC
estimate with δ = 0 is a weighted average of SMD modes. This offers a useful contrast with the
SMD estimate, which is the mode of the average deviations between the model and the data. We
then use stochastic expansions to study sources of variations in the two estimators in the case
of exact identification. The differences are illustrated using simple analytical examples as well as
simulations of the dynamic panel model.
Optimization of models with a non-smooth objective function is challenging, even when the
1
Indirect Inference is due to Gourieroux et al. (1993), the Simulated Method of moments is due to Duffie and
Singleton (1993), and the Efficient Method of Moments is due to Gallant and Tauchen (1996).

1
model is not complex. The Quasi-Bayes (LT) approach due to Chernozhukov and Hong (2003)
uses Bayesian computations to approximate the mode of a likelihood-free objective function. Its
validity rests on the Laplace (asymptotic normal) approximation of the posterior distribution with
the goal of valid asymptotic frequentist inference. The simulation analog of the LT (which we call
SLT) further uses simulations to approximate the intractable relation between the model and the
data. We show that both the LT and SLT can also be represented as a weighted average of modes
with appropriately defined importance weights.
A central theme of our analysis is that the mean computed from many likelihood-free poste-
rior distributions can be seen as a weighted average of solutions to frequentist objective functions.
Optimization permits us to turn the focus from computational to analytical aspects of the poste-
rior mean, and thus provides a bridge between the seemingly related approaches. Although our
optimization-based samplers are not intended to compete with the many ABC algorithms that are
available, they can be useful in situations when numerical optimization of the auxiliary model is
fast. This aspect is studied in our companion paper Forneron and Ng (2015) in which implemen-
tation of the RS in the overidentified case is also considered. The RS is independently proposed in
Meeds and Welling (2015) with suggestions for efficient and parallel implementations. Our focus
on the analytical properties complements their analysis.
The paper proceeds as follows. After laying out the preliminaries in Section 2, Section 3 presents
the general idea behind ABC and introduces an optimization view of the ideal MCMC-ABC. Section
4 considers Quasi-Bayes estimators and interprets them from an optimization perspective. Section
5 uses stochastic expansions to study the properties of the estimators. Section 6 uses analytical
examples and simulations to illustrate their differences. Throughout, we focus the discussion on
features that distinguish the SMD from ABC which are lesser known to economists.2

2 Preliminaries

As a matter of notation, we use L(·) to denote the likelihood, p(·) to denote posterior densities, q(·)
for proposal densities, and π(·) to denote prior densities. A ‘hat’ denotes estimators that correspond
to the mode and a ‘bar’ is used for estimators that correspond to the posterior mean. We use (s, S)
and (b, B) to denote the (specific, total number of) draws in frequentist and Bayesian type analyses
respectively. A superscript s denotes a specific draw and a subscript S denotes the average over
∂
S draws. For a function f (θ), we use fθ (θ0 ) to denote ∂θ f (θ) evaluated at θ0 , fθθj (θ0 ) to denote
2
The class of SMD estimators considered are well known in the macro and finance literature and with apologies,
many references are omitted. We also do not consider discrete choice models; though the idea is conceptually similar,
the implementation requires different analytical tools. Smith (2008) provides a concise overview of these methods.
The finite sample properties of the estimators are studied in Michaelides and Ng (2000). Readers are referred to the
original paper concerning the assumptions used.

2
 ∂ ∂2
∂θj fθ (θ) evaluated at θ0 and fθ,θj ,θk (θ0 ) to denote ∂θj θk fθ (θ) evaluated at θ0 .
Throughout, we assume that the data y = (y1 , . . . , yT )0 are covariance stationary and can
be represented by a parametric model with probability measure Pθ where θ ∈ Θ ⊂ RK . There
is a unique model parameterized by θ0 . Unless otherwise stated, we write E[·] for expectations
taken under Pθ0 instead of EPθ0 [·]. If the likelihood L(θ) = L(θ|y) is tractable, maximizing the
log-likelihood `(θ) = log L(θ) with respect to θ gives

θbM L = argmaxθ `(θ).

Bayesian estimation combines the likelihood with a prior π(θ) to yield the posterior density

L(θ) · π(θ)
p(θ|y) = R . (1)
Θ L(θ)π(θ)dθ

For any prior π(θ), it is known that θbM L solves argmaxθ `(θ) = limλ→∞ R θ exp(λ`(θ))π(θ) . That
Θ exp(λ`(θ))π(θ)dθ
is, the maximum likelihood estimator is a limit of the Bayes estimator using λ → ∞ replications
of the data y3 . The parameter λ is the cooling temperature in simulated annealing, a stochastic
optimizer due to Kirkpatrick et al. (1983) for handling problems with multiple modes.
In the case of conjugate problems, the posterior distribution has a parametric form which makes
it easy to compute the posterior mean and other quantities of interest. For non-conjugate problems,
the method of Monte-Carlo Markov Chain (MCMC) allows sampling from a Markov Chain whose
ergodic distribution is the target posterior distribution p(θ|y), and without the need to compute the
normalizing constant. We use the Metropolis-Hastings (MH) algorithm in subsequent discussion.
In classical Bayesian estimation with proposal density q(·), the acceptance ratio is
 L(θb+1 )π(θb+1 )q(θb |θb+1 ) 
ρBC (θb , θb+1 ) = min ,1 .
L(θb )π(θb )q(θb+1 |θb )

When the posterior mode θbBC = argmaxθ p(θ|y) is difficult to obtain, the posterior mean
B Z
1 X b
θBC = θ ≈ θp(θ|y)dθ
B Θ
b=1

is often the reported estimate, where θb are draws from the Markov Chain upon convergence. Under
quadratic loss, the posterior mean minimizes the posterior risk Q(a) = Θ |θ − a|2 p(θ|y)dθ.
R

2.1 Minimum Distance Estimators

The method of generalized method of moments (GMM) is a likelihood-free frequentist estimator

developed in Hansen (1982); Hansen and Singleton (1982). It allows, for example, the estimation
3
See Robert and Casella (2004, Corollary 5.11), Jacquier et al. (2007).

3
of K parameters in a dynamic model without explicitly solving the full model. It is based on a
vector of L ≥ K moment conditions gt (θ) whose expected value is zero at θ = θ0 , ie. E[gt (θ0 )] = 0.
Let g(θ) = T1 Tt=1 gt (θ) be the sample analog of E[gt (θ)]. The estimator is
P

T
θbGM M = argminθ J(θ), J(θ) = · g(θ)0 W g(θ) (2)
2
where W is a L × L positive-definite weighting matrix. Most estimators can be put in the GMM
framework with suitable choice of gt . For example, when gt is the score of the likelihood, the
maximum likelihood estimator is obtained.
√ d
Let ψb ≡ ψ(y(θ
b 0 )) be L auxiliary statistics with the property that T (ψb − ψ(θ0 ))−→N (0, Σ).
It is assume that the mapping ψ(θ) = limT →∞ E[ψ(θ)] b is continuously differentiable in θ and
locally injective at θ0 . Gourieroux et al. (1993) refers to ψ(θ) as the binding function while Jiang
and Turnbull (2004) uses the term bridge function. The minimum distance estimator is a GMM
estimator which specifies
g(θ) = ψb − ψ(θ),

with efficient weighting matrix W = Σ−1 . Classical MD estimation assumes that the binding
function ψ(θ) has a closed form expression so that in the exactly identified case, one can solve for
θ by inverting g(θ).

2.2 SMD Estimators

Simulation estimation is useful when the asymptotic binding function ψ(θ0 ) is not analytically
tractable but can be easily evaluated on simulated data. The first use of this approach in economics
appears to be due to Smith (1993). The simulated analog of MD, which we will call SMD, minimizes
the weighted difference between the auxiliary statistics evaluated at the observed and simulated
data:

θbSM D = argminθ JS (θ) = argminθ g 0S (θ)W g S (θ).

where
S
1 X bs s s
g S (θ) = ψb − ψ (y (ε , θ)),
S
s=1

ys ≡ ys (εs , θ)
are data simulated under θ with errors ε drawn from an assumed distribution Fε ,
and ψ (θ) ≡ ψ (ys (εs , θ)) are the auxiliary statistics computed using ys . Of course, g S (θ) is also
bs b s

the average of over S deviations between ψb and ψbs (ys (εs , θ)). To simplify notation, we will write
ys and ψbs (θ) when the context is clear. As in MD estimation, the auxiliary statistics ψ(θ) should
‘smoothly embed’ the properties of the data in the terminology of Gallant and Tauchen (1996).
But SMD estimators replace the asymptotic binding function ψ(θ0 ) = limT →∞ E[ψ(θ
b 0 )] by a finite

4
sample analog using Monte-Carlo simulations. While the SMD is motivated with the estimation
of complex models in mind, Gourieroux et al. (1999) shows that simulation estimation has a bias
reduction effect like the bootstrap. Hence in the econometrics literature, SMD estimators are used
even when the likelihood is tractable, as in Gourieroux et al. (2010).
The steps for implementing the SMD are as follows:

0 For s = 1, . . . , S, draw εs = (εs1 , . . . , εsT )0 from Fε . These are innovations to the structural
model that will be held fixed during iterations.

1 Given θ, repeat for s = 1, . . . S:

a Use (εs , θ) and the model to simulate data ys = (y1s , . . . , yTs )0 .

b Compute the auxiliary statistics ψbs (θ) using simulated data ys .
1 PS bs (θ). Minimize JS (θ) = g (θ)0 W g (θ).
2 Compute: g S (θ) = ψ(y)
b − S s=1 ψ S S

The SMD is the θ that makes JS (θ) smaller than the tolerance specified for the numerical optimizer.
In the exactly identified case, the tolerance can be made as small as machine precision permits.
When ψb is a vector of unconditional moments, the SMM estimator of Duffie and Singleton (1993) is
obtained. When ψb are parameters of an auxiliary model, we have the ‘indirect inference’ estimator
of Gourieroux et al. (1993). These are Wald-test based SMD estimators in the terminology of Smith
(2008). When ψb is the score function associated with the likelihood of the auxiliary model, we have
the EMM estimator of Gallant and Tauchen (1996), which can also be thought of as an LM-test
based SMD. If ψb is the likelihood of the auxiliary model, JS (θ) can be interpreted as a likelihood
ratio and we have a LR-test based SMD. Gourieroux and Monfort (1996) provides a framework that
unifies these three approaches to SMD estimation. Nickl and Potscher (2010) shows that an SMD
based on non-parametrically estimated auxiliary statistics can have asymptotic variance equal to
the Cramer-Rao bound if the tuning parameters are optimally chosen.
The Wald, LM, and LR based SMD estimators minimize a weighted L2 distance between the
data and the model as summarized by auxiliary statistics. Creel and Kristensen (2013) considers a
class of estimators that minimize the Kullback-Leibler distance between the model and the data.
Within this class, their MIL estimator maximizes an ‘indirect likelihood’, defined as the likelihood
of the auxiliary statistics. Their BIL estimator uses Bayesian computations to approximate the
mode of the indirect likelihood. In practice, the indirect likelihood is unknown. A SBIL estimator
then uses Bayesian computations in conjunction with simulations and non-parametric estimation.
The latter step estimates the indirect likelihood by kernel smoothing of the simulated data. Gao
and Hong (2014) shows that using local linear regressions instead of kernel estimation can reduce
the variance and the bias. These SBIL estimators actually correspond to two implementations of

5
ABC considered in Beaumont et al. (2002). The SBIL provides a link between ABC and the SMD
to the extent that the SBIL can be seen as a Kullback-Leibler distance-based SMD estimator. In
the sequel, we take the more conventional L2 definition of SMD as given above.

3 Approximate Bayesian Computations

The ABC literature often credits Donald Rubin to be the first to consider the possibility of esti-
mating the posterior distribution when the likelihood is intractable. Diggle and Gratton (1984)
proposes to approximate the likelihood by simulating the model at each point on a parameter grid
and appears to be the first implementation of simulation estimation for models with intractable like-
lihoods. Subsequent developments adapted the idea to conduct posterior inference, giving the prior
an explicit role. The first ABC algorithm was implemented by Tavare et al. (1997) and Pritchard
et al. (1996) to study population genetics. Their Accept/Reject algorithm is as follows: (i) draw θb
from the prior distribution π(θ), (ii) simulate data using the model under θb (iii) accept θb if the
auxiliary statistics computed using the simulated data are close to ψ.
b As in the SMD literature, the
auxiliary statistics can be parameters of a regression or unconditional sample moments. Heggland
and Frigessi (2004), Drovandi et al. (2011, 2015) use simulated auxiliary statistics.
Since simulating from a non-informative prior distribution is inefficient, subsequent work sug-
gests to replace the rejection sampler by one that takes into account the features of the posterior
distribution. The general idea is to set as a target the intractable posterior density

p∗ABC (θ|ψ)
b ∝ π(θ)L(ψ|θ)
b

and approximate it using Monte-Carlo methods. Some algorithms are motivated from the per-
spective of non-parametric density estimation, while others aim to improve properties of the
Markov chain.4 The main idea is, however, using data augmentation to consider the joint den-
b ∝ L(ψ|x,
sity pABC (θ, x|ψ) b θ)L(x|θ)π(θ), putting more weight on the draws with x close to ψ. b
When x = ψ, b L(ψ|b ψ,
b θ) is a constant, pABC (θ, ψ| b ∝ L(ψ|θ)π(θ)
b ψ) b and the target posterior is
recovered. If ψb are sufficient statistics, one recovers the posterior distribution associated with the
intractable likelihood L(θ|y), not just an approximation.
To better understand the ABC idea and its implementation, we will write yb instead of yb (εb , θb )
and ψbb instead of ψbb (yb (εb , θb )) to simplify notation. Let Kδ (ψbb , ψ|θ)
b ≥ 0 be a kernel function that
weighs deviations between ψb and ψbb over a window of width δ. Suppose we keep only the draws
that satisfy ψbb = ψb and hence δ = 0. Note that K0 (ψbb , ψ|θ)
b = 1 if ψb = ψbb for any choice of the
4
Recent surveys on ABC can be found in Marin et al. (2012), Blum et al. (2013) among others. See Drovandi et
al. (2015, 2011) for differences amongst ABC estimators.

6
kernel function. Once the likelihood of interest
Z
L(ψ|θ)
b = L(x|θ)K0 (x, ψ|θ)dx
b

is available, moments and quantiles can be computed. In particular, for any measurable function
ϕ whose expectation exists, we have:
h i R ϕ(θ)π(θ)L(ψ|θ)dθ
b
R R
ϕ(θ)π(θ)L(x|θ)K0 (x, ψ|θ)dxdθ
b
b Θ
E ϕ(θ)|ψb = ψb = R = ΘR R
Θ π(θ)L(ψ|θ)dθ Θ π(θ)L(x|θ)K0 (x, ψ|θ)dxdθ
b b

Since ψbb |θb ∼ L(.|θb ), the expectation can be approximated by averaging over draws from L(·|θbb ).
More generally, draws can be taken from an importance density q(·). In particular,
PB b )K (ψ b π(θbb )
bb , ψ|θ)
b=1 ϕ(θ 0
h i
b ϕ(θ)|ψb = ψbb q(θ )
E = .
PB bb b π(θb )
b=1 K0 (ψ , ψ|θ) q(θb )

The importance weights are then

b
b π(θ ) .
w0b ∝ K0 (ψbb , ψ|θ)
q(θb )
h i h i
b ϕ(θ)|ψb → E ϕ(θ)|ψb as B → ∞.
By a law of large number, E
There is, however, a caveat. When ψb has continuous support, ψbb = ψb is an event of measure
zero. Replacing K0 with Kδ where δ is close to zero yields the approximation:
R R
h
b
i ϕ(θ)π(θ)L(x|θ)Kδ (x, ψ|θ)dxdθ
b
E ϕ(θ)|ψ = ψ
b b ≈ ΘR R .
π(θ)L(x|θ)Kδ (x, ψ|θ)dxdθ
Θ
b

Since Kδ (·) is a kernel function, consistency of the non-parametric estimator for the conditional
expectation of ϕ(θ) follows from, for example, Pagan and Ullah (1999). This is the approach
considered in Beaumont et al. (2002), Creel and Kristensen (2013) and Gao and Hong (2014). The
b ψbb ) = I b bb
case of a rectangular kernel Kδ (ψ, corresponds to the ABC algorithm proposed in
kψ−ψ k≤δ
Marjoram et al. (2003). This is the first ABC algorithm that exploits MCMC sampling. Hence we
refer to it as MCMC-ABC. Our analysis to follow is based on this algorithm. Accordingly, we now
explore it in more detail.

Algorithm MCMC-ABC Let q(·) be the proposal distribution. For b = 1, . . . , B with θ0 given,

1 Generate θb+1 ∼ q(θb+1 |θb ).

2 Draw εb+1 from Fε and simulate data yb+1 . Compute ψbb+1 .

7
 3 Accept θb+1 with probability ρABC (θb , θb+1 ) and set it equal to θb with probability 1 −
ρABC (θb , θb+1 ) where
 π(θb+1 )q(θb |θb+1 ) 
ρABC (θb , θb+1 ) = min Ikψ−
b ψbb+1 k≤δ ,1 . (3)
π(θb )q(θb+1 |θb )

As with all ABC algorithms, the success of the MCMC-ABC lies in augmenting the posterior with
simulated data ψbb , ie. p∗ (θb , ψbb |ψ)
ABC
b b , ψbb )L(ψbb |θb )π(θb ). The joint posterior distribution
b ∝ L(ψ|θ
that the MCMC-ABC would like to target is
 
p0ABC θb , ψbb |ψb ∝ π(θb )L(ψbb |θb )Ikψbb −ψk=0
b

since integrating out εb would yield p∗ABC (θ|ψ).

b But it would not be possible to generate draws
such that kψbb − ψk
b equals zero exactly. Hence as a compromise, the MCMC-ABC algorithm allows
δ > 0 and targets
 
pδABC θb , ψbb |ψb ∝ π(θb )L(ψbb |θb )Ikψbb −ψk≤δ
b .

The adequacy of pδABC as an approximation of p0ABC is a function of the tuning parameter δ.

To understand why this algorithm works, we follow the argument in Sisson and Fan (2011).
If the initial draw θ1 satisfies kψb − ψb1 k ≤ δ, then all subsequent b > 1 draws are such that
Ikψbb −ψk≤δ
b = 1 by construction. Furthermore, since we draw θb+1 and then independently simulate
data ψbb+1 , the proposal distribution becomes q(θb+1 , ψbb+1 |θb ) = q(θb+1 |θb )L(ψbb+1 |θb+1 ). The two
observations together imply that

π(θb+1 )q(θb |θb+1 ) Ikψ− bb+1 k≤δ π(θ b+1 )q(θ b |θ b+1 ) L(ψ
bb+1 |θb+1 ) L(ψbb |θb )
Ikψ−
b ψ
=
b ψbb+1 k≤δ
π(θb )q(θb+1 |θb ) Ikψ−
b ψbb k≤δ
b b+1
π(θ )q(θ |θ ) b
L(ψbb |θb ) L(ψbb+1 |θb+1 )
Ikψ−
b ψbb+1 k≤δ π(θ b+1 )L(ψbb+1 |θb+1 ) q(θb |θb+1 )L(ψbb |θb )
=
Ikψ−
b ψbb k≤δ π(θb )L(ψbb |θb ) q(θb+1 |θb )L(ψbb+1 |θb+1 )
 
pδABC θb+1 , ψbb+1 |ψb q(θb , ψbb |θb+1 )
=   .
δ b bb
pABC θ , ψ |ψ b q(θb+1 , ψbb+1 |θb )

The last equality shows that the acceptance ratio is in fact the ratio of two ABC posteriors times
the ratio of the proposal distribution. Hence the MCMC-ABC effectively targets the joint posterior
distribution pδABC .

3.1 The Reverse Sampler

1 PS bs (θ)k no larger
Thus far, we have seen that the SMD estimator is the θ that makes kψb − S s=1 ψ
than the tolerance of the numerical optimizer. We have also seen that the feasible MCMC-ABC

8
accepts draws θb satisfying kψb − ψbb (θb )k ≤ δ with δ > 0. To view the MCMC-ABC from a different
perspective, suppose that setting δ = 0 was possible. Then each accepted draw θb would satisfy:

ψbb (θb ) = ψ.
b

For fixed εb and assuming that the mapping ψbb : θ → ψbb (θ) is continuously differentiable and
one-to-one, the above statement is equivalent to:
 0  
θb = argminθ ψbb (θ) − ψb ψbb (θ) − ψb .

Hence each accepted θb is the solution to a SMD problem with S = 1. Next, suppose that instead
of drawing θb from a proposal distribution, we draw εb and solve for θb as above. Since the mapping
ψbb is invertible by assumption, a change of variable yields the relation between the distribution of
ψbb and θb . In particular, the joint density, say h(θb , εb ), is related to the joint density L(ψbb (θb ), εb )
via the determinant of the Jacobian |ψbθb (θb )| as follows:

h(θb , εb |ψ)
b = |ψbb (θb )|L(ψbb (θb ), εb |ψ).
θ
b

Multiplying the quantity on the right-hand-side by wb (θb ) = π(θb )|ψbθb (θb )|−1 yields π(θb )L(ψ,
b εb |θb )
since ψbb (θb ) = ψb and the mapping from θb to ψ b (θb ) is one-to-one. This suggests that if we solve
the SMD problem B times each with S = 1, re-weighting each of the B solutions by wb (θb ) would
give the target the joint posterior p∗ (θ|ψ)
b after integrating out εb .
ABC

Algorithm RS

1 For b = 1, . . . , B and a given θ,

i Draw εb from Fε and simulate data yb using θ. Compute ψbb (θ) from yb .

ii Let θb = argminθ J1b (θ), J1b (θ) = (ψb − ψbb (θ))0 W (ψb − ψbb (θ)).

iii Compute the Jacobian ψbθb (θb ) and its determinant |ψbθb (θb )|. Let wb (θb ) = π(θb )|ψbθb (θb )|−1 .
PB b (θ b )θ b wb (θb )
2 Compute the posterior mean θRS = b=1 w where wb (θb ) = PB c c
.
c0=1 w (θ )

The RS has the optimization aspect of SMD as well as the sampling aspect of the MCMC-ABC. We
call the RS the reverse sampler for two reasons. First, typical Bayesian estimation starts with an
evaluation of the prior probabilities. The RS terminates with evaluation of the prior. Furthermore,
we use the SMD estimates to reverse engineer the posterior distribution.
Consistency of each RS solution (ie. θb ) is built on the fact that the SMD is consistent even
with S = 1. The RS estimate is thus an average of a sequence of SMD modes. In contrast, the SMD

9
is the mode of an objective function defined from a weighted average of the simulated auxiliary
statistics. Optimization effectively allows δ to be as close to zero as machine precision permits.
This puts the joint posterior distribution as close to the infeasible target as possible, but has the
consequence of shifting the distribution from (yb , ψbb ) to (yb , θb ). Hence a change of variable is
required. The importance weight depends on Jacobian matrix, making the RS an optimization
based importance sampler.
∂ψ b (θ)
Lemma 1 Suppose that ψ : θ → ψbb (θ) is one-to-one and ∂θ is full column rank. The poste-
rior distribution produced by the reverse sampler converges to the infeasible posterior distribution
p∗ (θ|ψ)
ABC
b as B → ∞.

By convergence, we mean that for any measurable function ϕ(θ) such that the expectation exists,
a law of large number implies that B b b b a.s. b b 1
P
b=1 w (θ )ϕ(θ )−→Ep∗ (θ|ψ)
b (ϕ(θ)). In general, w (θ ) 6= B .
Results for moments computed from the RS draws can be interpreted as draws from p∗ABC , the
posterior distribution had the likelihood p(ψ|θ)
b been available.
We mainly use the RS as a conceptual framework to understand the differences between the
MCMC-ABC and SMD in what follows. While it is not intended to compete with existing likelihood-
free estimators, it can nonetheless be useful in situations when numerical optimization of the
auxiliary model is easy. Properties of the RS are further analyzed in Forneron and Ng (2015).
Meeds and Welling (2015) independently proposes an algorithm similar to the RS, and shows how
it can be implemented efficiently by embarrassingly parallel methods.

4 Quasi-Bayes Estimators

The GMM objective function J(θ) defined in (2) is not a proper density. Noting that exp(−J(θ)) is
the kernel of the Gaussian density, Jiang and Turnbull (2004) defines an indirect likelihood (distinct
from the one defined in Creel and Kristensen (2013)) as

b ≡ √1 |Σ|−1 exp(−J(θ)).
LIN D (θ|ψ)
2π
Associated with the indirect likelihood is the indirect score, indirect Hessian, and a generalized
information matrix equality, just like a conventional likelihood. Though the indirect likelihood is
not a proper density, its maximizer has properties analogous to the maximum likelihood estimator
provided E[gt (θ0 )] = 0.
In Chernozhukov and Hong (2003), the authors observe that extremum estimators can be dif-
ficult to compute if the objective function is highly non-convex, especially when the dimension
of the parameter space is large. These difficulties can be alleviated by using Bayesian computa-
tional tools, but this is not possible when the objective function is not a likelihood. Chernozhukov

10
and Hong (2003) take an exponential of −J(θ), as in Jiang and Turnbull (2004), but then com-
bines exp(−J(θ)) with a prior density π(θ) to produce a quasi-posterior density. Chernozhukov
and Hong initially termed their estimator ‘Quasi-Bayes’ because exp(−J(θ)) is not a standard
likelihood. They settled on the term ‘Laplace-type estimator’ (LT), so-called because Laplace sug-
gested to approximate a smooth pdf with a well defined peak by a normal density, see Tierney and
Kadane (1986). If π(θ) is strictly positive and continuous over a compact parameter space Θ, the
‘quasi-posterior’ LT distribution
exp(−J(θ))π(θ)
pLT (θ|y) = R ∝ exp(−J(θ))π(θ) (4)
Θ exp(−J(θ)π(θ))dθ
is proper. The LT posterior mean is thus well-defined even when the prior may not be proper. As
discussed in Chernozhukov and Hong (2003), one can think of the LT under a flat prior as using
simulated annealing to maximize exp(−J(θ)) and setting the cooling parameter τ to 1. Frequentist
inference is asymptotically valid because as the sample size increases, the prior is dominated by the
pseudo likelihood which, by the Laplace approximation, is asymptotically normal.5
In practice, the LT posterior distribution is targeted using MCMC methods. Upon replacing
the likelihood L(θ) by exp(−J(θ)), the MH acceptance probability is
 exp(−J(ϑ))π(ϑ)q(θb |ϑ) 
ρLT (θb , ϑ) = min ,1 .
exp(−J(θb ))π(θb )q(ϑ|θb )
The quasi-posterior mean is θLT = B1 B b b
P
b=1 θ where each θ is a draw from pLT (θ|y). Chernozhukov
and Hong suggest to exploit the fact that the quasi-posterior mean is much easier to compute than
the mode and that, under regularity conditions, the two are first order equivalent. In practice,
the weighting matrix can be based on some preliminary estimate of θ, or estimated simultaneously
with θ. In exactly identified models, it is well known that the MD estimates do not depend on the
choice of W . This continues to be the case for the LT posterior mode θbLT . However, the posterior
mean is affected by the choice of the weighting matrix even in the just-identified case.6
The LT estimator is built on the validity of the asymptotic normal approximation in the second
order expansion of the objective function. Nekipelov and Kormilitsina (2015) shows that in small
samples, this approximation can be poor so that the LT posterior mean may differ significantly
from the extremum estimate that it is meant to approximate. To see the problem in a different
√
b 0) −
light, we again take an optimization view. Specifically, the asymptotic distribution T (ψ(θ
d
ψ(θ0 ))−→N (0, Σ(θ0 )) ≡ A∞ (θ0 ) suggests to use
Ab (θ0 )
ψbb (θ) ≈ ψ(θ) + ∞
√
T
5
R
For loss function d(·), the LT estimator is θbLT (ϑ) = argminθ Θ d(θ − ϑ)pLT (θ|y)dθ. If d(·) is quadratic, the
posterior mean minimizes quasi-posterior risk.
6
Kormiltsina and Nekipelov (2014) suggests to scale the objective function to improve coverage of the confidence
intervals.

11
where Ab∞ (θ0 ) ∼ N (0, Σ(θ)).
b Given a draw of Ab∞ , there will exist a θb such that (ψbb (θ) −
b 0 W (ψbb (θ) − ψ)
ψ) b is minimized. In the exactly identified case, this discrepancy can be driven
to zero up to machine precision. Hence we can define

θb = argminθ kψbb (θ) − ψk.

b

Arguments analogous to the RS suggests the following will produce draws of θ from pLT (θ|y).

1 For b = 1, . . .B:
Ab∞ (θ)
i Draw Ab∞ (θ0 ) and define ψbb (θ) = ψ(θ) + √
T
.

ii Solve for θb such that ψbb (θb ) = ψb (up to machine precision).

iii Compute wb (θb ) = |ψbθb (θb )|−1 π(θb ).

wb (θb )
wb (θb )θb , where wb =
P
2 θLT = PB c c
.
c=1 w (θ )

Seen from an optimization perspective, the LT is a weighted average of MD modes with the de-
terminant of the Jacobian matrix as importance weight, similar to the RS. It differs from the RS
in that the Jacobian here is computed from the asymptotic binding function ψ(θ), and the draws
are based on the asymptotic normality of ψ.
b As such, simulation of the structural model is not
required.

4.1 The SLT

When ψ(θ) is not analytically tractable, a natural modification is to approximate it by simulations

as in the SMD. This is the approach taken in Lise et al. (2015). We refer to this estimator as the
Simulated Laplace-type estimator, or SLT. The steps are as follows:

0 Draw structural innovations εs = (εs1 , . . . , εsT )0 from Fε . These are held fixed across iterations.

1 For b = 1, . . . , B, draw ϑ from q(ϑ|θb ).

i. For s = 1, . . . S: use (ϑ, εs ) and the model to simulate data ys = (y1s , . . . , yTs )0 . Compute
ψbs (ϑ) using ys .
PS
ii. Form JS (ϑ) = g S (ϑ)0 W g S (ϑ), where g S (ϑ) = ψ(y)
b − 1
S
bs (ϑ).
s=1 ψ

iii. Set θb+1 = ϑ with probability ρSLT (θb , ϑ), else reset ϑ to θb with probability 1 − ρSLT
where the acceptance probability is:
 exp(−J (ϑ))π(ϑ)q(θb |ϑ) 
S
ρSLT (θb , ϑ) = min ,1 .
exp(−JS (θb ))π(θb )q(ϑ|θb )

12
 b PB b.
2 Compute θSLT = b=1 θ

The SLT algorithm has two loops, one using S simulations for each b to approximate the asymptotic
binding function, and one using B draws to approximate the ‘quasi-posterior’ SLT distribution
exp(−JS (θ))π(θ)
pSLT (θ|y, ε1 , . . . , εS ) = R ∝ exp(−JS (θ))π(θ) (5)
Θ exp(−JS (θ))π(θ)dθ

The above SLT algorithm has features of SMD, ABC, and LT. Like SMD and ABC, SLT
also requires simulations of the full model. As a referee pointed out, the SLT resembles the
ABC algorithm when used with a Gaussian kernel. But exp(−JS (θ)) is not a proper density
and pSLT (θ|y, ε1 , . . . , εS ) is not a conventional likelihood-based posterior distribution. While the
SLT targets the pseudo likelihood, ABC algorithms target the proper but intractable likelihood.
Furthermore, the asymptotic distribution of ψb is known from a frequentist perspective. In ABC
estimation, lack of knowledge of the likelihood of ψb motivates the Bayesian computation.
The optimization implementation of SLT presents a clear contrast with the ABC.

1 Given εs = (εs1 , . . . , εsT )0 for s = 1, . . . S, repeat for b = 1, . . . B:

1 PS Ab∞ (θ)
i Draw ψbb (θ) = S s=1 ψ
bs (θ) + √
T
.

ii Solve for θb such that ψbb (θb ) = ψb (up to machine precision).

iii Compute wb (θb ) = |ψbθb (θb )|−1 π(θb ).

wb (θb )
wb (θb )θb , where wb =
P
2. θSLT = PB c c
.
c=1 w (θ )

While the SLT is a weighted average of SMD modes, the draws of ψbb (θ) are taken from the (fre-
quentist) asymptotic distribution of ψb instead of solving the model at each b. Gao and Hong (2014)
use a similar idea to make draws of what we refer to as g(θ) in their extension of the BIL estimator
of Creel and Kristensen (2013) to non-separable models.
The SMD, RS, ABC, and SLT all require specification and simulation of the full model. At a
practical level, the innovations ε1 , . . . εs used in SMD and SLT are only drawn from Fε once and
held fixed across iterations. Equivalently, the seed of the random number generator is fixed so that
the only difference in successive iterations is due to change in the parameters to be estimated. In
contrast, ABC draws new innovations from Fε each time a θb+1 is proposed. We need to simulate
B sets of innovations of length T , not counting those used in draws that are rejected, and B is
generally much bigger than S. The SLT takes B draws from an asymptotic distribution of ψ.
b Hence
even though some aspects of the algorithms considered seem similar, there are subtle differences.

13
5 Properties of the Estimators

This section studies the finite sample properties of the various estimators. Our goal is to compare
the SMD with the RS, and by implication, the infeasible MCMC-ABC. To do so in a tractable way,
1
we only consider the expansion up to order T.
As a point of reference, we first note that under
assumptions in Rilstone et al. (1996); Bao and Ullah (2007), θbM L admits a second order expansion

AM L (θ0 ) CM L (θ0 ) 1
θbM L = θ0 + √ + + op ( ).
T T T

where AM L (θ0 ) is a mean-zero asymptotically normal random vector and CM L (θ0 ) depends on the
curvature of the likelihood. These terms are defined as

AM L (θ0 ) = E[`θθ (θ0 )]−1 ZS (θ0 ) (6a)

 K 
1X
CM L (θ0 ) = E[−`θθ (θ0 )]−1 ZH (θ0 )ZS (θ0 ) − (−`θθθj (θ0 ))ZS (θ0 )ZS,j (θ0 ) (6b)
2
j=1

where the normalized score √1 `θ (θ0 ) and centered hessian√1 (`θθ (θ0 ) − E[`θθ (θ0 )]) converge in
T T
distribution to the normal vectors ZS and ZH respectively. The order T1 bias is large when Fisher
information is low.
Classical Bayesian estimators are likelihood based. Hence the posterior mode θbBC exhibits a
bias similar to that of θbM L . However, the prior π(θ) can be thought of as a constraint, or penalty
since the posterior mode which maximizes log p(θ|y) = log L(θ|y) + log π(θ). Furthermore, Kass
et al. (1990) shows that the posterior mean deviates from the posterior mode by a term that
depends on the second derivatives of the log-likelihood. Accordingly, there are three sources of bias
in the posterior mean θBC : a likelihood component, a prior component, and a component from
approximating the mode by the mean. Hence
 
AM L (θ0 ) 1 πθ (θ0 ) P M 1
θBC = θ0 +
b √ + CBC (θ0 ) + CBC (θ0 ) + CBC (θ0 ) + op ( ).
T T π(θ 0 ) T

Note that the prior component is under the control of the researcher.
In what follows, we will show that posterior means based on auxiliary statistics ψb generically
have the above representation, but the composition of the terms differ.

5.1 Properties of θbSM D

Minimum distance estimators depend on auxiliary statistics ψ.

b Its properties has been analyzed
in Newey and Smith (2004a, Section 4.2) within an empirical-likelihood framework. To facilitate
subsequent analysis, we follow Gourieroux and Monfort (1996, Ch.4.4) and directly expand around

14
 √
ψ,
b under the assumption that it admits a second-order expansion. In particular, since ψb is T
consistent for ψ(θ0 ), ψb has expansion
A(θ0 ) C(θ0 ) 1
ψb = ψ(θ0 ) + √ + + op ( ). (7)
T T T

It is then straightforward to show that the minimum distance estimator θbM D has expansion
h i−1
AM D (θ0 ) = ψθ (θ0 ) A(θ0 ) (8a)
i−1  K 
h 1X
CM D (θ0 ) = ψθ (θ0 ) C(θ0 ) − ψθ,θj (θ0 )AM D (θ0 )AM D,j (θ0 ) . (8b)
2
j=1

The bias in θbM D depends on the curvature of the binding function and the bias in the auxiliary
b C(θ0 ). Then following Gourieroux et al. (1999), we can analyze the SMD as follows. In
statistic ψ,
view of (7), we have, for each s:
As (θ) Cs (θ) 1
ψbs (θ) = ψ(θ) + √ + + op ( ).
T T T

satisfies ψb = S1 Ss=1 ψbs (θbSM D ) and has expansion θbSM D = θ0 + ASM√DT(θ0 ) +

P
The estimator θbSM D
CSM D (θ0 )
T + op ( T1 ). Plugging in the Edgeworth expansions give:

A(θ0 ) C(θ0 ) S 
As (θbSM D ) Cs (θbSM D )

1 1X 1
ψ(θ0 ) + √ + + Op ( ) = ψ(θbSM D ) + √ + + op ( ) .
T T T S T T T
s=1

Expanding ψ(θbSM D ) and As (θbSM D ) around θ0 and equating terms in the expansion of θbSM D ,
 −1  S 
1X s
ASM D (θ0 ) = ψθ (θ0 ) A(θ0 ) − A (θ0 ) (9a)
S
s=1
 −1  S S 
1X s 1X s
C(θ0 ) − C (θ0 ) − Aθ (θ0 ) ASM D (θ0 )


CSM D (θ0 ) = ψθ (θ0 ) (9b)
S S
s=1 s=1
 −1 XK
1
− ψθ (θ0 ) ψθ,θj (θ0 )ASM D (θ0 )ASM D,j (θ0 ).
2
j=1
PB
The first order term can be written as ASM D = AM D + B1 [ψθ (θ0 )]−1 b (θ
0 ), the last term has
b=1 A
 
variance of order 1/B which accounts for simulation noise. Note also that E S1 Ss=1 Cs (θ0 ) =
P

E[C(θ0 )]. Hence, unlike the MD, E[CSM D (θ0 )] does not depend on the bias C(θ0 ) in the auxiliary
statistic. In the special case when ψb is a consistent estimator of θ0 , ψθ (θ0 ) is the identity map and
the term involving ψθθj (θ0 ) drops out. Consequently, the SMD has no bias of order T1 when S → ∞
and ψ(θ) = θ. In general, the bias of θbSM D depends on the curvature of the binding function as
 K
−1 X  
S→∞ 1
E[CSM D (θ0 )] → − ψθ (θ0 ) ψθ,θj (θ0 )E AM D (θ0 )AM D,j (θ0 ) . (10)
2
j=1

15
This is an improvement over θbM D because as seen from (8b),
 −1  K
−1 X  
1
E[CM D (θ0 )] = ψθ (θ0 ) C(θ0 ) − ψθ (θ0 ) ψθ,θj (θ0 )E AM D (θ0 )AM D,j (θ0 ) . (11)
2
j=1

The bias in θbM D has an additional term in C(θ0 ).

5.2 Properties of θRS

The ABC literature has focused on the convergence properties of the algorithms and less on the
theoretical properties of the estimates. Dean et al. (2011) establishes consistency of the ABC in
the case of hidden Markov models. The analysis considers a scheme so that maximum likelihood
estimation based on the ABC algorithm is equivalent to exact inference under the perturbed hidden
Markov scheme. The authors find that the asymptotic bias depends on the ABC tolerance δ. Calvet
and Czellar (2015) provides an upper bound for the mean-squared error of their ABC filter and
studies how the choice of the bandwidth affects properties of the filter. Under high level conditions
and adopting the empirical likelihood framework of Newey and Smith (2004b), Creel and Kristensen
(2013) shows that the infeasible BIL is second order equivalent to the MIL after bias adjustments,
while MIL is in turn first order equivalent to the continuously updated GMM. The feasible SBIL
(which is also an ABC estimator) has additional errors compared to the BIL due to simulation noise
and kernel smoothing, but these errors vanish as S → ∞ for an appropriately chosen bandwidth.
Gao and Hong (2014) shows that local-regressions have better variance properties compared to
kernel estimations of the indirect likelihood. Both studies find that a large number of simulations
are needed to control for the stochastic approximation error.
The results of Creel and Kristensen (2013) and Gao and Hong (2014) shed light on the rela-
tionship between SMD and ABC from the perspective of non-parametric estimation. The difficulty
in analyzing ABC algorithms comes in the fact that simulation introduces additional randomness
which interacts with smoothing bias induced by non-parametric estimation of the density. These
effect are difficult to make precise. We present an optimization/importance sampling perspective
by appealing to an implication of Proposition 1.(i), which is that θRS is the weighted average of a
sequence of SMD modes. Analysis of the weights wb (θb ) requires an the expansion of ψbb (θb ) around
θ
ψθ (θ0 ). From such a analysis, shown in the Appendix, we find that
B
X ARS (θ0 ) CRS (θ0 ) 1
θRS = wb (θb )θb = θ0 + √ + + op ( )
T T T
b=1

16
where
B  −1  B 
1 X b 1 X b
ARS (θ0 ) = ARS (θ0 ) = ψθ (θ0 ) A(θ0 ) − A (θ0 ) (12a)
B B
b=1 b=1
B B
" #
1 X b πθ (θ0 ) 1 X b
CRS (θ0 ) = CRS (θ0 ) + (ARS (θ0 ) − ARS (θ0 ))AbRS (θ0 ) + CRS
M
(θ0 ).(12b)
B π(θ0 ) B
b=1 b=1

Proposition 1 Let ψ(θ)

b be the auxiliary statistic that admits the expansion as in (7) and suppose
that the prior π(θ) is positive and continuously differentiable around θ0 . Then E[ARS (θ0 )] = 0 but
E[CRS (θ0 )] 6= 0 for an arbitrary choice of prior.

1
The SMD and RS are first order equivalent, but θRS has an order T bias. The bias, given by
M (θ ) term (defined in Appendix A) can be traced directly
CRS (θ0 ), has three components. The CRS 0

to the weights, or to the interaction of the weights with the prior, and is a function of ARS (θ0 ).
Some but not all the terms vanish as B → ∞. The second term will be zero if a uniform prior is
chosen since πθ = 0. A similar result is obtained in Creel and Kristensen (2013). The first term is

B  −1 XB  K 
1 X b 1 1X
CRS (θ0 ) = ψθ (θ0 ) C(θ0 ) − Cb (θ0 ) − ψθθj (θ0 )AbRS (θ0 )AbRS,j (θ0 ) − Abθ (θ0 )AbRS (θ0 ) .
B B 2 j=1
b=1 b=1

C(θ0 ) − B1 Cb (θ0 ) is exactly the same as in CSM D (θ0 ). The middle term involves
PB
The term b=1
ψθθj (θ0 ) and is zero if ψ(θ) = θ. But because the summation is over θb instead of ψbs ,
B
1 X b
→ E[Abθ (θ0 )AbRS (θ0 )] 6= 0.
Aθ (θ0 )AbRS (θ0 ) B→∞
B
b=1

As a consequence E[CRS (θ0 )] 6= 0 even when ψ(θ) = θ. In contrast, E[CSM D (θ0 )] = 0 when
ψ(θ) = θ as seen from (10). The reason is that the comparable term in CSM D (θ0 ) is
 S 
1X s
Aθ (θ0 ) ASM D (θ0 ) S→∞
→ E[Asθ (θ0 )]ASM D (θ0 ) = 0.
S
s=1

The difference boils down to the fact that the SMD is the mode of the average over simulated
auxiliary statistics, while the RS is a weighted average over the modes. As will be seen below, this
difference is also present in the LT and SLT and comes from averaging over θb . The result is based
on fixing δ at zero and holds for any B. Proposition 1 implies that the ideal MCMC-ABC with
δ = 0 also has a non-negligible second-order bias.
1
In theory, the order T bias can be removed if π(θ) can be found to put the right hand side of
C RS (θ0 ) defined in (12b) to zero. Then θRS will be second order equivalent to SMD when ψ(θ) = θ
and may have a smaller bias than SMD when ψ(θ) 6= θ since SMD has a non-removable second order

17
bias in that case. That the choice of prior will have bias implications for likelihood-free estimation
echoes the findings in the parametric likelihood setting. Arellano and Bonhomme (2009) shows in
the context of non-linear panel data models that the first-order bias in Bayesian estimators can be
eliminated with a particular prior on the individual effects. Bester and Hansen (2006) also shows
1
that in the estimation of parametric likelihood models, the order T bias in the posterior mode
and mean can be removed using objective Bayesian priors. They suggest to replace the population
quantities in a differential equation with sample estimates. Finding the bias-reducing prior for the
RS involves solving the differential equation:

b πθ (θ0 )
0 = E[CRS (θ0 )] + E[(AbRS (θ0 ) − ARS (θ0 ))AbRS (θ0 )] + E[CRS
M
(θ0 )]
π(θ0 )
M (θ , π(θ )) that is not present in Bester and
which has the additional dependence on π in CRS 0 0

Hansen (2006). A closed-form solution is available only for simple examples as we will see Section
4.1 below. For realistic problems, how to find and implement the bias-reducing prior is not a trivial
problem. A natural starting point is the plug-in procedure of Bester and Hansen (2006) but little is
known about its finite sample properties even in the likelihood setting for which it was developed.
Finally, this section has studied the RS, which is the best that the MCMC-ABC can achieve
in terms of δ. This enables us to make a comparison with the SMD holding the same L2 distance
between ψb and ψ(θ) at zero by machine precision. However, the MCMC-ABC algorithm with δ > 0
will not produce draws with the same distribution as the RS. To see the problem, suppose that the
RS draws are obtained by stopping the optimizer before kψb − ψ(θb )k reaches the tolerance guided
by machine precision. This is analogous to equating ψ(θb ) to the pseudo estimate ψb + δ. Inverting
the binding function will yield an estimate of θ that depends on the random δ in an intractable way.
The RS estimate will thus have an additional bias from δ 6= 0. By implication, the MCMC-ABC
with δ > 0 will be second order equivalent to the SMD only after a bias adjustment even when
ψ(θ) = θ.

5.3 The Properties of LT and SLT

The mode of exp(−J(θ))π(θ) will inherit the properties of a MD estimator. However, the quasi-
posterior mean has two additional sources of bias, one arising from the prior, and from approxi-
mating the mode by the mean. The optimization view of θLT facilitates an understanding of these
effects. As shown in Appendix B, each draw θLT b has expansion terms
h i−1  
AbLT (θ0 ) = ψθ (θ0 ) A(θ0 ) − Ab∞ (θ0 )
 
i−1 K
h
C(θ0 ) − 1
ψθ,θj (θ0 )(AbLT (θ0 )AbLT,j (θ0 ) − Ab∞,θ (θ0 )AbLT (θ0 ) .
X
b
CLT (θ0 ) = ψθ (θ0 )
2
j=1

18
Even though the LT has the same objective function as MD, simulation noise enters both AbLT (θ0 )
and C b (θ0 ). Compared to the extremum estimate θbM D , we see that ALT = 1 B Ab (θ0 ) 6=
P
LT B b=1 LT
AM D (θ0 ) and CLT (θ0 ) 6= CM D (θ0 ). Although CLT (θ0 ) has the same terms as CRS (θ0 ), they are
different because the LT uses the asymptotic binding function, and hence AbLT (θ0 ) 6= AbRS (θ0 ).
b
A similar stochastic expansion of each θSLT gives:
S
!
h i−1 1X s
AbSLT (θ0 ) = ψθ (θ0 ) A(θ0 ) − A (θ0 ) − Ab∞ (θ0 )
S
s=1
 
i−1 S K
h
C(θ0 ) − 1 1
Cs (θ0 ) −
X X
b
CSLT (θ0 ) = ψθ (θ0 ) ψθ,θj (θ0 )AbSLT AbSLT,j ]
S 2
s=1 j=1
S
!
h i−1 1 X
− ψθ (θ0 ) (Aθ (θ0 ) + A∞,θ (θ0 ))ASLT (θ0 )
s b b
S
s=1

Following the same argument as in the RS, an optimally chosen prior can reduce bias, at least in
theory, but finding this prior will not be a trivial task. Overall, the SLT has features of the RS (bias
does not depend on C(θ0 )) and the LT (dependence on Ab∞ ) but is different from both. Because
PS
the SLT uses simulations to approximate the binding function ψ(θ), E[C(θ0 ) − S1 s=1 C
s (θ
0 )] = 0.
The improvement over the LT is analogous to the improvement of SMD over MD. However, the
AbSLT (θ0 ) is affected by estimation of the binding function (the term with superscript s) and of
the quasi-posterior density (the terms with superscript b). This results in simulation noise with
variance of order 1/S plus another of order 1/B. Note also that the SLT bias has an additional
term
B S B
!
1 X 1X s S→∞ 1
(Aθ (θ0 ) + Ab∞,θ (θ0 ))AbSLT (θ0 ) Ab∞,θ (θ0 )AbLT (θ0 ).
X
→
B S B
b=1 s=1 b=1

The main difference with the RS is that Ab is replaced with Ab∞ . For S = ∞ this term matches
that of the LT.

5.4 Overview

We started this section by noting that the Bayesian posterior mean has two components in its bias,
one arising from the prior which acts like a penalty on the objective function, and one arising from
the approximation of the mean for the mode. We are now in a position to use the results in the
foregoing subsections to show that for d=(MD, SMD, RS, LT) and SLT and D = (RS,LT,SLT)
these estimators can be represented as

1d∈D πθ (θ0 )
 
Ad (θ0 ) Cd (θ0 ) 1
θbd = θ0 + √ + + CdP (θ0 )) + CdM (θ0 ) + op ( ) (13)
T T T π(θ0 ) T

19
  
where with Abd (θ0 ) = [ψθ (θ0 )]−1 A(θ0 ) − Abd (θ0 ) ,
B
 1 X b 
Ad (θ0 ) = [ψθ (θ0 )]−1 A(θ0 ) − Ad (θ0 )
B
b=1
K
 1X 
Cd (θ0 ) = [ψθ (θ0 )]−1 C(θ0 ) − Cd (θ0 ) − ψθ,θj (θ0 )Abd (θ0 )Abd,j (θ0 ) − Abd,θ Abd (θ0 )
2
j=1
B
1 X
CdP (θ0 ) = (Abd (θ0 ) − Ad (θ0 ))Abd (θ0 ),
B
b=1

The term CdP (θ0 ) is a bias directly due to the prior. The term CdM (θ0 ), defined in the Appendix,
depends on Ad (θ0 ), the curvature of the binding function, and their interaction with the prior.
Hence at a general level, the estimators can be distinguished by whether or not Bayesian compu-
tation tools are used, as the indicator function is null only for the two frequentist estimators (MD
and SMD). More fundamentally, the estimators differ because of Ad (θ0 ) and Cd (θ0 ), which in turn
depend on Abd (θ0 ) and Cd (θ0 ). We compactly summarize the differences as follows:

d Abd (θ0 ) Cd (θ0 ) var(Ad (θ0 )) E[C(θ0 ) − Cd (θ0 )]

MD 0 0 0 E[C(θ0 )]
LT Ab∞ (θ0 ) 0 1
B var[ Ab∞ (θ0 )] E[C(θ0 )]
Ab (θ0 ) Cb (θ0 ) Ab (θ0 )]
1 PB 1
RS B b=1 B var[ 0
As (θ0 ) Cs (θ0 ) As (θ0 )]
1 PS 1 PS 1
SMD S s=1 S s=1 S var[ 0
ASM D (θ0 ) + AbLT (θ0 ) Cs (θ0 ) var[ASMD (θ0 )] + var[ALT (θ0 )]
1 PS
SLT S s=1 0

The MD is the only estimator that is optimization based and does not involve simulations.
Hence it does not depend on b or s and has no simulation noise. The SMD does not depend on
b because the optimization problem is solved only once. The LT simulates from the asymptotic
binding function. Hence its errors are associated with parameters of the asymptotic distribution.
The MD and LT have a bias due to asymptotic approximation of the binding function. In such
cases, Cabrera and Fernholz (1999) suggests to adjust an initial estimate θe such that if the new
estimate θb were the true value of θ, the mean of the original estimator equals the observed value
θ.
e Their target estimator is the θ such that EP [θ] b = θ.
e While the bootstrap directly estimates
θ

the bias, a target estimator corrects for the bias implicitly. Cabrera and Hu (2001) show that the
bootstrap estimator corresponds to the first step of a target estimator. The latter improves upon
the bootstrap estimator by providing more iterations.
An auxiliary statistic based target estimator is the θ that solves EPθ [ψ(y(θ))]
b = ψ(y(θ
b 0 )). It

replaces the asymptotic binding function limT →∞ E[ψ(y(θ0 ))] by EPθ [ψ(y(θ))] and approximates
b b

20
the expectation under Pθ by stochastic expansions. The SMD and SLT can be seen as target
estimators that approximate the expectation by simulations. Thus, they improve upon the MD
estimator even when the binding function is tractable and is especially appealing when it is not.
However, the improvement in the SLT is partially offset by having to approximate the mode by the
mean.

6 Two Examples

The preceding section can be summarized as follows. A posterior mean computed through auxiliary
statistics generically has a component due to the prior, and a component due to the approximation
of the mode by the mean. The binding function is better approximated by simulations than
asymptotic analysis. It is possible for simulation estimation to perform better than ψbM D even if
ψ(θ) were analytically and computationally tractable.
In this section, we first illustrate the above findings using a simple analytical example. We then
evaluate the properties of the estimators using the dynamic panel model with fixed effects.

6.1 An Analytical Example

We consider the simple DGP yi ∼ N (m, σ 2 ). The parameters of the model are θ = (m, σ 2 )0 . We
focus on σ 2 since the estimators have more interesting properties.
The MLE of θ is
T T
1X 2 1X
m
b = yt , σ
b = (yt − y)2 .
T T
t=1 t=1

While the posterior distribution is dominated by the likelihood in large samples, the effect of
the prior is not negligible in small samples. We therefore begin with a analysis of the effect of the
prior on the posterior mean and mode in Bayesian analysis. Details of the calculations are provided
in appendix D.1.
We consider the prior π(m, σ 2 ) = (σ 2 )−α Iσ2 >0 , α > 0 so that the log posterior distribution is
 T 
−T 1 X
log p(θ|y) = log p(θ|m,
b σb )∝ 2 2 2
log(2πσ ) − α log σ − 2 (yt − m) Iσ2 >0 .
2
2 2σ
t=1

Tσb2 Tσb2
The posterior mode and mean of σ 2 are σmode
2 = T +2α
2
and σmean = T +2α−5 . respectively. Using
(T −1) 2
the fact that σ2]
E[b = T σ , we can evaluate 2
σmode , 2
σmean and their expected values for different
α. Two features are of note. For a given prior (here indexed by α), the mean does not coincide with
the mode. Second, the statistic (be it mean or mode) varies with α. The Jeffrey’s prior corresponds
to α = 1, but the bias-reducing prior is α = 2. In the Appendix, we show that the bias reducing
1
prior for this model is π R (θ) ∝ σ4
.

21
 Table 1: Mean θBC vs. Mode θbBC

α θBC θbBC E[θBC ] E[θbBC ]

0 b2 T T−5
σ b2
σ σ 2 TT −1
−5 σ 2 T T−1
1 b2 T T−3
σ b2 T T+2
σ σ 2 TT −1
−3
−1
σ 2 TT +2
2 b2 T T−1
σ b2 T T+4
σ σ2 σ 2 TT −1
+4

3 b2 T T+1
σ b2 T T+6
σ σ 2 TT −1
+1
−1
σ 2 TT +6

Next, we consider estimators based on auxiliary statistics:

 
b 0= m
ψ(y) 2
b .
b σ

As these are sufficient statistics, we can also consider (exact) likelihood-based Bayesian inference.
bS2 ) = ( S1 Ss=1 m
b s , S1 Ss=1 σ
b2,s ). The LT quasi-likelihood using
P P
For SMD estimation, we let (m b S, σ
the variance of preliminary estimates of m and σ 2 as weights is:

b − m)2 (b σ 2 − σ 2 )2
  
2 T (m
exp(−J(m, σ )) = exp − + .
2 b2
σ σ4
2b

The LT posterior distribution is p(m, σ 2 |m,

b σb2 ) ∝ π(m, σ 2 ) exp(−J(m, σ 2 )). Integrating out m gives
p(σ 2 |m,
b σb2 ). We consider a flat prior π U (θ) ∝ Iσ2 ≥0 and the bias-reducing prior π R (θ) ∝ 1/σ 4 Iσ2 ≥0 .
The RS is the same as the SMD under a bias-reducing prior. Thus,

2 b2
σ
σ
bSM D = 1 PS PT s − es )2
ST s=1 t=1 (et
2,R b2
σ
σ
bRS = 1 PB PT b b 2
BT b=1t=1 (et − e )
σ 2
B b
[ T b b 2 2
P
2,U t=1 (et −e ) /T ]
X
σ
bRS = PB 1
.
b=1 b0 =1 PT (eb0 −eb0 )2 /T
t=1 t

For completeness, the parametric Bootstrap bias corrected estimator σ 2

bBootstrap σ 2 −EBootstrap (b
= 2b σ2)
is also considered:

2 T −1 1
σ
bBootstrap σ2 − σ
= 2b b2 b2 (1 + ).
=σ
T T
σ 2 ) computes the expected value of the estimator replacing the true value σ 2 with σ
EBootstrap (b b2 ,
1
σ 2 (1 +
the plug-in estimate. In this example the bias can be computed analytically since E(b T )) =

22
 Figure 1: ABC vs. RS Posterior Density

0.6
Posterior Density

0.4

0.2

0.0
0 2 4 6
Weights = Prior x Jacobian

True RS Re−Weighted

0.6
Posterior Density

0.4

0.2

0.0
0 2 4 6
Weights = Prior Only

True RS Re−Weighted

σ 2 (1 − T1 )(1 + T1 ) = σ 2 (1 − T12 ). While the bootstrap does not involve inverting the binding function,
this computational simplicity comes at the cost of adding a higher order bias term (in 1/T 2 ).
A main finding of this paper is that the reverse sampler can replicate draws from p∗ABC (θ0 ),
which in turn equals the Bayesian posterior distribution if ψb are sufficient statistics. The weight for
each SMD estimate is the prior times the Jacobian. To illustrate the importance of the Jacobian
transformation, the top panel of Figure 1 plots the Bayesian/ABC posterior distribution and the
one obtained from the reverse sampler. They are indistinguishable. The bottom panel shows an
incorrectly constructed reverse sampler that does not apply the Jacobian transformation. Notably,
the two distributions are not the same.
The properties of the estimators are summarized in Table 2. It should be reminded that
increasing S improves the approximation of the binding function in SMD estimation while increasing
B improves the approximation to the target distribution in Bayesian type estimation. For fixed T ,
only the Bayesian estimator with the bias reducing prior is unbiased. The SMD and RS (with bias
reducing prior) have the same bias and mean-squared error in agreement with the analysis in the
previous section. These two estimators have smaller errors than the RS estimator with a uniform
prior. The SLT posterior mean differs from that of the SMD by κSLT that is not mean-zero. This
term, which is a function of the Mills-ratio, arises as a consequence of the fact that the σ 2 in SLT

23
 Table 2: Properties of the Estimators

Estimator Prior E[θ]

b Bias Variance
2
θbM L - σ 2 T T−1 − σT 2σ 4 TT−1
2

−1 2σ 2 −1
θBC 1 σ 2 TT −5 T −5 2σ 4 (TT−5)2
R 1
θBC 1/σ 4 σ2 0 2σ 4 T −1
U −1 2σ 2 −1
θRS 1 σ 2 TT −5 T −5 2σ 4 (TT−5)2
R B(T −1) 2σ 2
θRS 1
σ4
σ 2 B(T −1)−2 B(T −1)−2 2σ 4 Tκ−1
1

S(T −1) 2σ 2
θbSM D - σ 2 S(T −1)−2 S(T −1)−2 2σ 4 Tκ−1
1

U σ2
θLT 1 σ 2 T T−1 (1 + κLT ) σ 2 T T−1 κLT − T 2σ 4 TT−1
2 (1 + κLT )
2

S(T −1) σ2 2 T −1
U
θbSLT 1 σ 2 S(T −1)−2 + κSLT S(T −1)−2 +σ T E[κSLT ] 2σ 4 TκLT
−1 + ∆SLT
−σ 2
θbBootstrap - σ 2 (1 − 1
T2
) T2
2σ 4 TT−1
2 (1 +
1 2
T)

φ(x)
Notes to Table 2: Let M (x) = 1−Φ(x)
be the Mills ratio.
(S(T −1))2 (T −1+S(T −1)−2)
i κ1 (S, T ) = (S(T −1)−2)2 (S(T −1)−4)
> 1, κ1 tends to one as B, S tend to infinity.
ii κLT = c−1
LT M (−cLT ), c2LT = T
2
, κLT → 0 as T → ∞.
iii κSLT = κLT · S · T · Invχ2S(T −1) , ∆,SLT = 2σ 4 var(κSLT ) + 4σ 4 TT−1 2
2 cov(κSLT , S · T InvχS(T −1)) ).

are drawn from the normal distribution and then truncated to ensure positivity.

6.2 The Dynamic Panel Model with Fixed Effects

The dynamic panel model yit = αi + ρyit−1 + σeit is known to suffer from severe bias when T is
small because the unobserved heterogeneity αi is imprecisely estimated. Various analytical bias
corrections have been suggested to improve the precision of the least squares dummy variable
(LSDV) estimator β.b Instrumental variable estimators have also been considered. Hsiao (2003)
provides a detailed account on the treatment of this incidental parameter problem. Gourieroux
et al. (2010) suggests to exploit the bias reduction properties of the indirect inference estimator
using the dynamic panel model as auxiliary equation. That is, ψ(θ) = θ. The authors reported
estimates of βb that are sharply more accurate in simulation experiments that hold σ 2 fixed. The
results continue to be impressive when an exogenous regressor and a linear trend is added to the
model. We reconsider their exercise but also estimate σ 2 .
With θ = (ρ, β, σ 2 )0 , we simulate data from the model:

yit = αi + ρyit−1 + βxit + σεit .

24
Let A = IT − 1T 10T /T A = A ⊗ IT , y = A vec(y), y −1 = A vec(y−1 ), x = A vec(x), where y−1 are
the lagged y, we use the following moment conditions:
 
y −1 (y − ρy −1 − βx)
 
2
g(ρ, β, σ ) =  x(y − ρy −1 − βx) .
 
 
2 2
(y − ρy −1 − βx) − σ (1 − 1/T )

with g(b
ρ, β,
bσ b2 ) = 0. The MD estimator is thus also the LSDV. The quantity g S (θ) for SMD and
g b (θ) for ABC are defined analogously. For this model, Bayesian inference is possible since the
likelihood in de-meaned data
N
!
1 1 X
L(y, x|θ) = p N
exp − 2 (y i − ρy i,−1 − βxi )0 Ω−1 (y i − ρy i,−1 − βxi )
2π|σ 2 Ω| 2σ
i=2

where Ω = IT −1 − 1T −1 10T −1 /T . For LT, ABC-MCMC, SMD the weighting matrix is computed
P 0
as: W = ( N1T i,t git git − g 0 g)−1 . Recall that while the weighting matrix is irrelevant to finding
the mode in exactly identified models, W affects computation of the posterior mean. The prior
is π(θ) = Iσ2 ≥0,ρ∈[−1,1] . For SMD, the innovations εs used to construct ψbs are drawn from the
standard normal distribution once and held fixed.

Figure 2: Frequentist, Bayesian, and Approximate Bayesian Inference for ρ

12.5

10.0

7.5
density

5.0

2.5

0.0

0.3 0.4 0.5 0.6 0.7

Posterior for rho

variable RS SLT BC SMD

pBC (ρ|ψ)
b is the likelihood based Bayesian posterior distribution,
pSLT (ρ|ψ) is the Simulated Laplace type quasi-posterior distribution.
b
pRS (ρ|ψ)
b is the approximate posterior distribution based on the RS .
The frequentist distribution of θbSM D is estimated by N (θbSM D , var(
c θbSM D )).

25
 Table 3: Dynamic Panel ρ = 0.6, β = 1, σ 2 = 2

Mean over 1000 replications

MLE LT SLT SMD ABC RS Bootstrap

Mean 0.419 0.419 0.593 0.598 0.588 0.599 0.419

ρb : SD 0.037 0.037 0.036 0.035 0.036 0.035 0.074
Bias -0.181 -0.181 -0.007 -0.002 -0.012 -0.001 -0.181

Mean 0.940 0.940 0.997 1.000 0.995 1.000 0.940

βb : SD 0.070 0.071 0.073 0.073 0.073 0.073 0.139
Bias -0.060 -0.060 -0.003 0.000 -0.005 0.000 -0.060

Mean 1.869 1.878 1.973 1.989 2.055 2.099 1.869

b2 :
σ SD 0.133 0.146 0.144 0.144 0.150 0.152 0.267
Bias -0.131 -0.122 -0.027 -0.011 0.055 0.099 -0.131

S – – 500 500 1 1 –
B – 500 500 – 500 500 500

Note: MLE=MD. The ABC is MCMC-ABC with δABC = 0.025, keeping every 500-th draw.

Table 3 report results from 5000 replications for T = 6 time periods and N = 100 cross-section
b2 are significantly biased. The mean estimate
units, as in Gourieroux et al. (2010). Both ρb and σ
β
of the long run multiplier 1−ρ (not reported) is only 1.6 when the true value is 2.5. The LT is the
same as the MD except that it is computed using Bayesian tools. Hence its properties are similar
to the MD. The simulation estimators have much improved properties. The properties of θRS are
similar to those of the SMD. Figure 2 illustrates for one simulated dataset how the posteriors for
RS /SLT are shifted compared to the one based on the direct likelihood.

26
7 Conclusion

Different disciplines have developed different estimators to overcome the limitations posed by an
intractable likelihood. These estimators share many similarities: they rely on auxiliary statistics
and use simulations to approximate quantities that have no closed form expression. We suggests
an optimization framework that helps understand the estimators from the perspective of classical
minimum distance estimation. All estimators are first-order equivalent as S → ∞ and T → ∞ for
any choice of π(θ). Nonetheless, up to order 1/T , the estimators are distinguished by biases due
to the prior and approximation of the mode by the mean, the very two features that distinguish
Bayesian and frequentist estimation.
We have only considered regular problems when θ0 is in the interior of Θ and the objective
function is differentiable. When these conditions fail, the posterior is no longer asymptotically
normal around the MLE with variance equal to the inverse of the Fisher Information Matrix.
Understanding the properties of these estimators under non-standard conditions is the subject for
future research.

27
Appendix
The terms A(θ) and C(θ) in θbM D are derived for the just identified case as follows. Recall that ψb has a
second order expansion:
A(θ0 ) C(θ0 ) 1
ψb = ψ(θ0 ) + √ + + op ( ). (A.1)
T T T
A(θ0 ) C(θ0 )
Now θb = θ0 + √
T
+ T + op ( T1 ). Thus expanding ψ(θ)
b around θb = θ0 :
 
A(θ0 ) C(θ0 ) 1
ψ(θ) = ψ θ0 + √
b + + op ( )
T T T
  K
A(θ0 ) C(θ0 ) 1 1 X 1
= ψ(θ0 ) + ψθ (θ0 ) √ + + op ( ) + ψθ,θj (θ0 )A(θ0 )Aj (θ0 ) + op ( ).
T T T 2T j=1 T

Equating with ψ(θ0 ) + A√(θ ) + C(θ ) + op ( 1 ) and solving for A, C

0 0
we get:
T T T

h i−1
A(θ0 ) = ψθ (θ0 ) A(θ0 )
 
i−1 K
h
C(θ0 ) − 1 X
C(θ0 ) = ψθ (θ0 ) ψθ,θj (θ0 )A(θ0 )Aj (θ0 ) .
2 j=1

ψθ,θj (θ0 )Abd,j (θ0 ) + Abd,θ (θ0 )]),

PK
For estimator specific Abd and abd , define abd = trace([ψθ (θ0 )]−1 [ j=1

πθ (θ0 )πθ (θ0 )0

 
πθ (θ0 )
CdM (θ0 ) = 2 Ad (θ0 )ad (θ0 )θ0 − ad (θ0 )2 θ0 − Ad (θ0 )0 Ad (θ0 )θ0
π(θ0 ) π(θ0 )2
B
1 X b
− (ad (θ0 ) − ad (θ0 ))Abd (θ0 ). (A.2)
B
b=1

PB
Where ad = B1 b=1 abd , Ad is defined analogously. Note that a(θ0 ) → 0 as B → ∞ if ψ(θ) = θ and the first
two terms drop out.

A.1 Proof of Proposition 1, RS

To prove Proposition 1, we need an expansion for ψbb (θb ) and the weights using
Ab (θ0 ) C b (θ0 )
θb = θ0 + √
T
+ T + op ( T1 ). (A.3)

28
i. Expansion of ψbb (θ0 ) and ψbθb (θ0 ):

Ab (θb ) Cb (θb ) 1
ψbb (θb ) = ψ(θb ) + √ + + op ( )
T T T
Ab (θ0 + A√(θT0 ) + C T(θ0 ) + op ( T1 ))
b b
b
A (θ0 ) C b (θ0 ) 1
= ψ(θ0 + √ + + op ( )) + √
T T T T
C (θ0 +
b Ab (θ0 )
√
T
+ C b (θ0 )
T + 1op ( T1 ))
+ + op ( )
T T
Ab (θ0 ) ψθ (θ0 )Ab (θ0 ) Cb (θ0 ) Abθ (θ0 )Ab (θ0 )
= ψ(θ0 ) + √ + √ + +
T T T T
K
1 X ψθ,θj (θ0 )Ab (θ0 )Abj (θ0 ) 1
+ + op ( ).
2 j=1 T T

Since ψbb (θb ) equals ψb for all b,

h i−1
Ab (θ0 ) A(θ0 ) − Ab (θ0 )


= ψθ (θ0 ) (A.4)
 
i−1 K
C(θ0 ) − Cb (θ0 ) − 1
h
ψθ,θj (θ0 )Ab (θ0 )Abj (θ0 ) − Abθ (θ0 )Ab (θ0 ) ,
X
C b (θ0 ) = ψθ (θ0 ) (A.5)
2 j=1

it follows that
Ab (θ0 ) C b (θ0 )
 
1
ψbθb (θb ) bb
= ψθ θ 0 + √ + + op ( )
T T T
 
 A θ0 + + b
θ + Ab (θ0 )
√
C b (θ0 )
op ( T1 )
Ab (θ0 ) C b (θ0 ) T

1 T
= ψθ θ 0 + √ + + op ( ) + √
T T T T
 
Cbθ θ0 + A√(θT0 ) + C T(θ0 ) + op ( T1 )
b b

1
+ + op ( )
T T
XK
ψθ,θj (θ0 )Abj (θ0 ) Abθ (θ0 ) 1 X K X K
ψθ,θj ,θk (θ0 )Abj (θ0 )Abk (θ0 )
= ψθ (θ0 ) + √ + √ +
j=1
T T 2 j=1 T
k=1
K
X ψθ,θj (θ0 )Cjb (θ0 ) K
X Abθ,θj (θ0 )Abj (θ0 ) Cb (θ0 ) 1
+ + + + op ( ).
j=1
T j=1
T T T

To obtain the determinant of ψbθb (θb ), let ab (θ0 ) = trace(Ab (θ0 )), ab2 (θ0 ) = trace(Ab (θ0 )2 ), cb (θ0 ) = trace(C b (θ0 )),
where
 
h i−1 XK
Ab (θ0 ) = ψθ (θ0 )  ψθ,θj (θ0 )Abj (θ0 ) + Abθ (θ0 )
j=1
K XK K K
!
b b b
h i−1 1 X ψ (θ )A (θ )A (θ ) ψ (θ )C (θ )
Abθ,θj (θ0 )Abj (θ0 ) + Cb (θ0 ) .
θ,θ ,θ 0 j 0 k 0 X θ,θ 0 j 0 X
C b (θ0 ) = ψθ (θ0 )
j k j
+ +
2 j=1 T j=1
T j=1
k=1

Now for any matrix X with all eigenvalues smaller than 1 we have: log(IK + X) = X − 21 X 2 + o(X).
Furthermore, for any matrix M the determinant |M | = exp(trace(log M ))). Together, these imply that for

29
arbitrary X1 , X2 :
X2
  
X1 X2 1 X1 X2 1
I+√ + + op ( ) = exp trace √ + + 1 + op ( )
T T T T T T T


trace (X1 ) trace (X2 ) trace X12 1
=1+ √ + + + op ( ).
T T T T
Hence the required determinant is
Ab (θ0 ) C b (θ0 ) ab (θ0 ) ab2 (θ0 ) cb (θ0 )
 
1 1
ψbθb (θb ) = ψbθ (θ0 ) I + √ + + op ( ) = ψbθ (θ0 ) 1+ √ + + + op ( ) .
T T T T T T T

ii. Expansion of wb (θb ) = |ψbθ (θb )|−1 π(θb ):

−1
ab (θ0 ) ab2 (θ0 ) cb (θ0 ) Ab (θ0 ) C b (θ0 )

−1 −1 1 1
ψbθb (θb ) b
π(θ ) = ψbθ (θ0 ) 1+ √ + + + op ( ) π(θ0 + √ + + op ( ))
T T T T T T T
ab (θ0 ) ab2 (θ0 ) cb (θ0 )
 
−1 1
= ψbθ (θ0 ) 1− √ − − + op ( )
T T T T
 
b b K
X πθ,θj (θ0 )Ab (θ0 )Abj (θ0 )
A (θ 0 ) C (θ 0 ) 1 1
× π(θ0 ) + πθ (θ0 ) √ + πθ (θ0 ) + + op ( ) 
T T 2 j=1 T T

−1 ab (θ0 ) πθ (θ0 ) Ab (θ0 ) ab2 (θ0 ) cb (θ0 )

= ψbθ (θ0 ) π(θ0 ) 1 − √ + √ − −
T π(θ0 ) T T T
!
πθ (θ0 ) ab (θ0 )Ab (θ0 ) πθ (θ0 ) C b (θ0 ) 1 Ab (θ0 )πθ,θ0 (θ0 )Ab0 (θ0 ) 1
− + + + op ( ) .
π(θ0 ) T π(θ0 ) T 2 T T
1
PB 1 b
Now A(θ0 ) = B b=1 Ab (θ0 ). Similarly define C(θ0 ) = B C (θ0 ). Also, denote the term in 1/T by:
πθ (θ0 ) b πθ (θ0 ) b 1
eb (θ0 ) = −ab2 (θ0 ) − cb (θ0 ) − a (θ0 )Ab (θ0 ) + C (θ0 ) + Ab (θ0 )πθ,θ0 (θ0 )Ab0 (θ0 ).
π(θ0 ) π(θ0 ) 2
The normalized weight for draw b is:
−1 b b b
ψbθb (θb ) π(θb ) 1
 1 − a √(θT0 ) + ππ(θ
θ (θ0 ) A (θ0 )
0)
√
T
+ e (θ T
0)
+ op ( T1 ) 
b b
w (θ ) = P −1 = PB c c c
B
ψbθc (θc ) π(θc )
B 1 + B1 c=1 − a √(θT0 ) + ππ(θ θ (θ0 ) A (θ0 )
0)
√
T
+ e (θT
0)
+ op ( T1 )
c=1
ab (θ0 ) b b

1
1 − √
T
+ ππ(θ
θ (θ0 ) A (θ0 )
0)
√
T
+ e (θ
T
0)
+ op ( T1 ) 
=
B 1 − a(θ ) π (θ ) A(θ )
√0 + θ 0 √ 0 +
e(θ0 ) 1
T π(θ0 ) T T + op ( T )

1 ab (θ0 ) πθ (θ0 ) Ab (θ0 ) eb (θ0 )   1

a(θ0 ) πθ (θ0 ) A(θ0 ) e(θ0 ) 1 
= 1 − √ + √ + ) × 1+ √ −
+ op ( √ − + op ( )
B T π(θ 0 ) T T T T π(θ0 ) T T T
1 ab (θ0 ) − a(θ0 ) πθ (θ0 ) Ab (θ0 ) − A(θ0 )b b b
e (θ0 ) − e(θ0 ) a (θ0 )a(θ0 ) πθ (θ0 ) A (θ0 )a(θ0 )
= 1− √ + √ + − −
B T π(θ0 ) T T T π(θ0 ) T
πθ (θ0 ) A(θ0 )ab (θ0 ) πθ (θ0 )πθ (θ0 )0 Ab (θ0 )0 A(θ0 )
 
1 
− − + op ( ) .
π(θ0 ) T π(θ0 )2 T T
PB
The posterior mean is θRS = b=1 wb (θb )θb . Using θb defined in (A.3), A and C defined in (A.4) and (A.5):
B B B
1 X Ab (θ0 ) 1 X C b (θ0 ) πθ (θ0 ) 1 X (Ab (θ0 ) − A(θ0 ))Ab (θ0 ) 1
θRS = θ0 + √ + + + C M (θ0 ) + op ( ).
B T B T π(θ0 ) B T T
b=1 b=1 b=1

30
B.1 Proof of Results for LT
From
Ab (θ0 ) C b (θ0 ) 1
θ b = θ0 + √ + + op ( ),
T T T

we have, given that ψbb is drawn from the asymptotic distribution of ψb

Ab∞ (θb )
ψbb (θb ) = ψ(θb ) + √
T
Ab∞ (θ0 + A√(θT0 ) + C T(θ0 ) + op ( T1 ))
b b
b
A (θ0 ) C b (θ0 )
 
1
= ψ θ0 + √ + + op ( ) + √
T T T T
A ψθ (θ0 )A (θ0 ) A b b K
b
∞ (θ0 )
b
∞,θ (θ0 )A (θ0 ) 1 X ψθ,θj (θ0 )Ab (θ0 )Abj (θ0 ) 1
= ψ(θ0 ) + √ + √ + + op ( )
T T T 2 j=1
T T

which is equal to ψb for all b. Hence

h i−1
Ab (θ0 ) A(θ0 ) − Ab∞ (θ0 )


= ψθ (θ0 ) (B.1)
 
i−1 K
C(θ0 ) − 1
h
ψθ,θj (θ0 )Ab (θ0 )Abj (θ0 ) − Ab∞,θ (θ0 )Ab (θ0 ) .
X
C b (θ0 ) = ψθ (θ0 ) (B.2)
2 j=1

Note that the bias term C b depends on the bias term C. For the weights, we need to consider
 
Ab∞,θ θ0 + A√(θT0 ) + C T(θ0 ) + op ( T1 )
b b
b b
 
A (θ0 ) C (θ0 ) 1
ψbθb (θb ) = ψθ θ0 + √ + + op ( ) + √
T T T T
X ψθ,θj (θ0 )Aj (θ0 ) A∞,θ (θ0 ) X ψθ,θj (θ0 )Cj (θ0 ) X
K b b k b K
Ab∞,θ,θj Abj (θ0 )
= ψθ (θ0 ) + √ + √ + +
j=1
T T j=1
T j=1
T
K
1 X ψθ,θj ,θk (θ0 )Abj (θ0 )Abk (θ0 ) 1
+ + op ( ).
2 T T
j,k=1

Let
 
h i−1 K
Ab∞,θ (θ0 ) +
X
Ab (θ0 ) = ψθ (θ0 ) ψθ,θj (θ0 )Abj (θ0 )
j=1
 
i−1 X K K K XK
h 1
Ab∞,θ,θj (θ0 )Abj (θ0 ) +
X X
C b (θ0 ) = ψθ (θ0 )  ψθ,θj (θ0 )Cjb (θ0 ) + ψθ,θj ,θk (θ0 )Abj (θ0 )Abk (θ0 )
j=1 j=1
2 j=1 k=1
b b
a (θ0 ) = trace(A (θ0 )), ab2 (θ0 ) b
= trace(A (θ0 ) ),2 b
c (θ0 ) = trace(C (θ0 )). b

The determinant is
−1
Ab (θ0 ) C b (θ0 ) ab (θ0 ) ab2 (θ0 ) cb (θ0 )

−1 −1 1 −1 −1 1
ψbθb (θ0 ) = ψθ (θ0 ) I+ √ + + op ( ) = ψθ (θ0 ) 1+ √ + + + op ( )
T T T T T T T
ab (θ0 ) ab2 (θ0 ) cb (θ0 )
 
−1 1
= ψθ (θ0 ) 1− √ − − + op ( ) .
T T T T

31
The prior is

Ab (θ0 ) C b (θ0 )
 
b 1
π(θ ) = π θ0 + √ + + op ( )
T T T
Ab (θ0 ) C b (θ0 ) 1 Ab (θ0 )πθ,θ0 Ab0 (θ0 ) 1
= π(θ0 ) + πθ (θ0 ) √ + πθ (θ0 ) + + op ( ).
T T 2 T T
π
Let: eb (θ0 ) = −cb (θ0 ) − ab2 (θ0 ) + ππ(θ
θ (θ0 ) 0
0)
C b (θ0 ) + Ab (θ0 ) θ,θ b0
π (θ0 )A (θ0 ). After some simplification, the product
is
−1 −1  ab (θ0 ) πθ (θ0 ) Ab (θ0 ) eb (θ0 ) 1 
ψbθb (θ0 ) π(θb ) = ψθ (θ0 ) π(θ0 ) 1 − √ + √ + + op ( ) .
T π(θ0 ) T T T
Hence, the normalized weight for draw b is
−1
ab (θ0 ) πθ (θ0 ) Ab (θ0 ) b
ψbθb (θ0 ) π(θb ) 1 1 − √T + π(θ0 )
√
T
+ e (θT
0)
+ op ( T1 )
b b
w (θ ) = P −1 =
B B 1 − a(θ )
√0 +
πθ (θ0 ) A(θ0 )
√ + e(θT0 ) + op ( T1 )
c=1 ψbθc (θ0 ) π(θc ) T π(θ0 ) T

ab (θ0 ) πθ (θ0 ) Ab (θ0 ) eb (θ0 )

  
1 1 a(θ0 ) πθ (θ0 ) A(θ0 ) e(θ0 ) 1
= 1− √ + √ + + op ( ) 1+ √ − √ − + op ( )
B T π(θ0 ) T T T T π(θ0 ) T T T
π θ (θ 0 ) b πθ (θ0 )
1 ab (θ0 ) − a(θ0 ) πθ (θ0 ) Ab (θ0 ) − A(θ0 ) eb (θ0 ) − e(θ0 ) ab (θ0 )a(θ0 ) π(θ0 ) A (θ0 ) π(θ0 ) A(θ0 )
= 1− √ + √ + − −
B T π(θ0 ) T T T T
πθ (θ0 ) ab (θ0 )A(θ0 ) πθ (θ0 ) a(θ0 )Ab (θ0 ) 1 
+ + + op ( ) .
π(θ0 ) T π(θ0 ) T T
PB  b b

Hence the posterior mean is θLT = b=1 wb (θb )θb and θb = θ0 + A√(θT0 ) + C T(θ0 ) + op ( T1 ) . After simplifi-
cation, we have

B πθ (θ0 ) 2 B
A(θ0 ) C(θ0 ) 1 X (ab (θ0 ) − a(θ0 ))Ab (θ0 ) [ π(θ0 ) A(θ0 )] θ0 1 πθ (θ0 ) X (Ab (θ0 ) − A(θ0 ))Ab (θ0 )
θLT = θ0 + √ + − − +
T T B T T B π(θ0 ) T
b=1 b=1
a(θ0 )2 θ0 πθ (θ0 ) a(θ0 )A(θ0 )θ0 1
− +2 + op ( )
T π(θ0 ) T T
B
A(θ0 ) C(θ0 ) πθ (θ0 ) 1 X (Ab (θ0 ) − A(θ0 ))Ab (θ0 ) 1
= θ0 + √ + + + C M (θ0 ) + op ( ),
T T π(θ 0 ) B T T
b=1

where all terms are based on Ab (θ0 ) defined in (B.1) and C b (θ0 ) in (B.2).

C.1 Results for SLT:

From
S
1 X bs Ab∞ (θ)
ψbb (θ) = ψ (θ) + √
S s=1 T
A (θ) Cs (θ)
s
1
ψbs (θ) = ψ(θ) + √ + + op ( )
T T T
b
A (θ 0 ) C b (θ0 ) 1
θ b = θ0 + √ + + op ( ),
T T T

32
we have

Ab∞ (θ0 + A√(θT0 ) + C T(θ0 ) + op ( T1 ))

b b
S
Ab (θ0 ) C b (θ0 )
 
1 X bs 1
ψbs (θb ) = ψ θ0 + √ + + op ( ) + √
S s=1 T T T T

1 X As (θ0 )
S
Ab∞ (θ0 ) Ab (θ0 ) 1 X Asθ (θ0 )Ab (θ0 )
S
Ab∞,θ (θ0 )Ab (θ0 )
= ψ(θ0 ) + √ + √ + ψθ (θ0 ) √ + +
S s=1 T T T S s=1 T T

1
S
X Cs (θ0 ) 1
K
X Ab (θ0 )Abj (θ0 ) C b (θ0 ) 1
+ + ψθ,θj (θ0 ) + ψθ (θ0 ) + op ( ).
S s=1
T 2 j=1
T T T

Thus,
S
!
h i−1
1X s
b
A (θ0 ) = ψθ (θ0 ) A(θ0 ) − A (θ0 ) − A∞ (θ0 )
b
(C.1)
S s=1
 
i−1 S K
h
C(θ0 ) − 1 1
Cs (θ0 ) −
X X
C b (θ0 ) = ψθ (θ0 ) ψθ,θj (θ0 )Ab (θ0 )Abj (θ0 )
S s=1 2 j=1
" S
#
h i−1 1 X
− ψθ (θ0 ) A (θ0 ) + A∞,θ (θ0 ) Ab (θ0 ).
s b
(C.2)
S s=1 θ

Note that we have Ab∞ ∼ N while As →

d
N . To compute the weight for draw b, consider
 
S A
Ab (θ0 ) C b (θ0 )
θ0 + s √ + + op ( T1 )
Ab (θ0 ) C b (θ0 ) T
 
1 1X T
ψbb (θb ) = ψθ θ0 + √ + + op ( ) + √
T T T S s=1 T
   
A∞ θ0 + √T + T + op ( T )
Ab (θ0 ) C b (θ0 )
S C
b
C b (θ0 )
b 1 s
θ0 + A√(θT0 ) + T + op ( T1 )
1X 1
+ √ + + op ( )
T S s=1 T T
K
X Abj (θ0 ) 1 X Asθ (θ0 )
S
Ab∞,θ (θ0 ) 1 X Cs (θ0 ) X
S K
Cjb (θ0 )
= ψθ (θ0 ) + ψθ,θj (θ0 ) √ + √ + √ + + ψθ,θj (θ0 )
j=1
T S s=1 T T S s=1 T j=1
T

1 X X Aθ,θj (θ0 )Aj (θ0 ) X A∞,θ,θj (θ0 )Aj (θ0 ) 1 X X

S K s b K b b K K
Abk (θ0 )Abj (θ0 ) 1
+ + + ψθ,θj ,θk (θ0 ) + op ( ).
S s=1 j=1 T j=1
T 2 j=1 T T
k=1

Let:
 
i−1 S K
h 1
As (θ0 ) + Ab∞,θ (θ0 ) +
X X
Ab (θ0 ) = ψθ (θ0 )  ψθ,θj Abj (θ0 )
S s=1 θ j=1
 
i−1 1 X S K h S
h 1 i
Cs (θ0 ) + Asθ,θj (θ0 )Abj (θ0 ) + Ab∞,θ,θj (θ0 )Abj (θ0 ) 
X X
C b (θ0 ) = ψθ (θ0 )  ψθ,θj (θ0 )Cjb (θ0 ) +
S s=1 j=1
S s=1
 
h i−1 1 X K
+ ψθ (θ0 )  ψθ,θj ,θk (θ0 )Abk (θ0 )Abj (θ0 )
2
j,k=1
b b
a (θ0 ) = trace(A (θ0 )), ab2 (θ0 ) = trace(Ab (θ0 )2 ), cb (θ0 ) = trace(C b (θ0 )).

33
The determinant is
ab (θ0 ) ab2 (θ0 ) cb (θ0 )
 
−1 −1 1
ψbb (θb ) = ψθ (θ0 ) 1− √ − − + op ( ) .
T T T T

Hence
ab (θ0 ) ab2 (θ0 ) cb (θ0 )
 
−1 −1 1
ψbb (θb ) π(θb ) = ψθ (θ0 ) π(θ0 ) 1 − √ − − + op ( )
T T T T
 
b b K
X πθ,θj (θ0 ) Ab (θ0 )Abj (θ0 )
π θ (θ 0 ) A (θ 0 ) π θ (θ0 ) C (θ 0 ) 1 1
× 1 + √ + + + op ( )
π(θ0 ) T π(θ0 ) T 2 j=1 π(θ0 ) T T

ab (θ0 ) πθ (θ0 ) Ab (θ0 ) eb (θ0 )

 
−1 1
= ψθ (θ0 ) π(θ0 ) 1 − √ + √ + + op ( )
T π(θ0 ) T T T
PK πθ,θj (θ0 ) b
where eb (θ0 ) = −ab (θ0 ) ππ(θ
θ (θ0 )
0)
Ab (θ0 ) − ab2 (θ0 ) − cb (θ0 ) + πθ (θ0 ) b
π(θ0 ) C (θ0 ) + 1
2 j=1
b
π(θ0 ) A (θ0 )Aj (θ0 ). The
normalized weights are
−1
ψbb (θb ) π(θb )
b b
w (θ ) = P −1
B
c=1 ψbc (θc ) π(θc )
ab (θ0 ) πθ (θ0 ) Ab (θ0 ) eb (θ0 )
  
1 1 a(θ0 ) πθ (θ0 ) A(θ0 ) e(θ0 ) 1
= 1− √ + √ + + op ( ) 1+ √ − √ − + op ( ) .
B T π(θ0 ) T T T T π(θ0 ) T T T

PB b b
The posterior mean θSLT = b=1 wb (θb )θb with θb = θ0 + A√(θT0 ) + C (θ0 )
T + op ( T1 ). After some simplification,

B B
A(θ0 ) C(θ0 ) πθ (θ0 ) 1 X (Ab (θ0 ) − A(θ0 ))Ab (θ0 ) 1 X (ab (θ0 ) − a(θ0 ))Ab (θ0 )
θSLT = θ0 + √ + + −
T T π(θ0 ) B T B T
B=1 b=1
πθ (θ0 ) a(θ0 )A(θ0 )θ0 a2 (θ0 )θ0 πθ (θ0 ) θ0 1
+2 − −[ A(θ0 )]2 + op ( )
π(θ0 ) T T π(θ0 ) T T
B
A(θ0 ) C(θ0 ) πθ (θ0 ) 1 X (Ab (θ0 ) − A(θ0 ))Ab (θ0 ) 1
= θ0 + √ + + + C M (θ0 ) + op ( )
T T π(θ 0 ) B T T
B=1

where terms in A and C are defined from (C.1) and (C.2).

34
D.1 Results For The Example in Section 6.1
The data generating process is yt = m0 + σ0 et , et ∼ iid N (0, 1). As a matter of notation, a hat is used to
denote the mode, a bar denotes the mean, superscript
PS PTs denotes aPspecific draw and a subscript S to denote
1 s 1 S s
average over S draws. For example, eS = ST s=1 t=1 et = S s=1 e .

1
PT
MLE: Define e = T t=1 et . b = m0 + σ0 e ∼ N (0, σ02 /T ). For the variance
Then the mean estimator is m
0 −1 0
estimator, eb = y − m
b = σ0 (e − e) = σ0 M e, M = IT − 1(1 1) 1 is an idempotent matrix with T − 1 degrees
of freedom. Hence σ 2
bM L =e b0 eb/T ∼ σ02 χ2T −1 .

BC: Expressed in terms of sufficient statistics (m,

b σb2 ), the joint density of y is
PT
b 2 b2
 
1 T /2 t=1 (m − m) −T σ
p(y; m, σ 2 ) = ( ) exp − × .
2πσ 2 2σ 2 2σ 2
R∞
The flat prior is π(m, σ 2 ) ∝ 1. The marginal posterior distribution for σ 2 is p(σ 2 |y) = −∞
p(y|m, σ 2 )dm.
R∞ T
√
Using the result that −∞ exp(− 2σ b 2 )dm = 2πσ 2 , we have
2 (m − m)

b2
 
2 2 −(T −1)/2 2 2T − 3 Tσ
p(σ |y) ∝ (2πσ ) b /2σ ) ∼ invΓ
exp(−T σ , .
2 2
β
The mean of an invΓ(α, β) is α−1 . b2 T T−5 .
Hence the BC posterior is σ 2BC = E(σ 2 |y) = σ

SMD: The estimator equates the auxiliary statistics computed from the sample with
PS thes average of the
1
statistics over simulations. Given σ, the mean estimator m
b S solves m
b =m
bS + σS s=1 e . Since we use
σ2 σ2
sufficient statistics, m
b is the ML estimator. Thus, m b S ∼ N (m, T0 + ST ). Since yts − y st = σ(est − es ), the
1 S T s 2
bS2 is the σ 2 that solves σ
b2 = σ 2 ( ST s
P P
variance estimator σ s=1 t=1 (et − e ) ) Hence

b2
σ χ2T −1 /T
bS2 =
σ 1
P P s s 2
= σ 2
2 = σ 2 FT −1,S(T −1) .
ST s (b
e
t t − e ) χ S(T −1) /(ST )

d2 2 2 (T −1)
The mean of a Fd1 ,d2 random variable is d2 −2 . Hence E(b
σSM D ) = σ S(T −1)−2 .

LT: The LT is defined as
2 !

b2 − σ 2
T σ
σ 2 ) ∝ 1σ2 ≥0 exp −
pLT (σ 2 |b
2 σ4
2b
which implies
σ4
 
2 2 2b
2
σ |b
σ ∼LT N σ
b , truncated to [0, +∞[.
T
φ( a−µ
σ )
For X ∼ N (µ, σ 2 ) we have E(X|X > a) = µ + 1−Φ( a−µ
σ (Mills-Ratio). Hence:
σ )

2
φ( √0−bσ ) r p !
b2
2/T σ p 2 φ(− T /2)
ELT (σ 2 |b
σ2 ) = σ
b2 + 2 b2 = σ
2/T σ b2 1+ .
1 − Φ( √0−bσ
p
2
) T 1 − Φ(− T /2)
2/T σ
b

q √
2 φ(− √T /2)
Let κLT = T 1−Φ(− T /2) . We have ELT (σ 2 |b
σ2 ) = σ
b2 (1 + κLT ) . The expectation of the estimator is

T −1
E ELT (σ 2 |b
σ2 ) = σ2


(1 + κLT )
T

35
from which we deduce the bias of the estimator
 
T −1 1
E ELT (σ 2 |b
σ2 ) − σ2 = σ2


κLT − .
T T
2
The variance of the estimator is 2σ 4 TT−1
2 (1 + κLT ) and the Mean-Squared Error (MSE)
 2 !
4 T −1 2 T −1 1
σ 2 2 (1 + κLT ) + κLT −
T T T

which is the squared bias of MLE plus terms that involve the Mills-Ratio (due to the truncation).

SLT: The SLT is defined as

 
2
2    2 
2 2 χS(T −1) 2 χ2S(T −1) 2
 T σ b −σ ST   T [ χS(T −1) ]2 σ
2
b / ST − σ 
σ ) ∝ 1σ2 ≥0 exp −
2 2
pSLT (σ |b = 1 2 ≥0 exp −
ST
   
σ
 2 2bσ4 2 2bσ4
 
  

where
2 T
1X1X s χ2S(T −1)
bS2 = σ 2
σ (et − es )2 = σ 2 .
S s=1 T t=1 ST
This yields the slightly more complicated formula
!
2 2 s 2
χ2S(T −1) 2b
σ4 ST
σ |b
σ , (e )s=1,...,S ∼ N σ
b / , [ ]2
ST T χ2S(T −1)

and the posterior mean becomes

 
b2 ST /χ2 −1)
σ
φ − r 2σb 4 S(T
ST

T ( 2 )2
ST χ
S(T −1)
p ST
ESLT (σ 2 |b
σ2 ) = σ
b2 + 2/T b2
σ
χ2S(T −1) χ2S(T −1)
 
b2 ST /χ2S(T −1)
σ
1 − Φ − r 2σb 4 ST

( )2
T χ2
S(T −1)
 p 
ST φ − T /2 p ST
b2
=σ +  2/T 2 b2 .
σ
χ2S(T −1)
 p
1 − Φ − T /2 χS(T −1)

√
φ(− T /2) p
Let κSLT = √ 2/T χ2 ST = κLT χ2 ST (random). We can compute
1−Φ(− T /2) S(T −1) S(T −1)

S(T − 1) T −1
E ESLT (σ 2 |b
σ2 ) = σ2 + σ2


E(κSLT )
S(T − 1) − 2 T
and the bias
2 T −1
E ESLT (σ 2 |b
σ2 ) − σ2 = σ2 + σ2


E(κSLT )
S(T − 1) − 2 T
which is the bias of SMD and the Mills-Ratio term that comes from taking the mean of the truncated normal
rather than the mode. The variance is similar to the LT and the SMD
1 T −1 S
2σ 4 κ1 + 2σ 4 V(κSLT ) + 4σ 4 2
Cov(κSLT , 2 ).
T −1 T χS(T −1)

36
The extra term is due to κSLT being random. We could simplify further noting that κSLT = κLT χ2 ST ,
S(T −1)
2 2
E(κSLT ) = κLT S(TST 2 S T
−1)−2 , V(κSLT ) = κLT (S(T −1)−2)2 (S(T −1)−4) and Cov(κSLT , χ2
S
) = κLT S 2 T V(1/χ2S(T −1) ) =
S(T −1)
2
S T
κLT (S(T −1)−2) 2 (S(T −1)−4) .

The MSE is
 2
2 T −1 1 T −1 S
σ4 + E(κSLT ) + 2σ 4 κ1 + 2σ 4 V(κSLT ) + 4σ 4 Cov(κSLT , 2 )
S(T − 1) − 2 T T −1 T2 χS(T −1)
(T − 1)2 4σ 4
 
2 1 T −1
= 2σ 4 2
+ κ1 + 2
E(κ2SLT + E(κSLT )
[S(T − 1) − 2] T −1 T S(T − 1) − 2 T
| {z }
MSE of SMD
T −1 S
+2σ 4 V(κSLT ) + 4σ 4 Cov(κSLT , 2 ).
T2 χS(T −1)

RS: The auxiliary statistic for each draw of simulated data is matched to the sample auxiliary statistic.
Thus, m
b =P mb + σ b eb . Thus conditional on m b and σ 2,b , mb = m b − σ b eb ∼ N (0, σ 2,b /T ). For the variance,
2,b 2,b b b 2
σ
b =σ t (et − e ) /T . Hence

2
b2 b2
P  
σ t (et − e) /T T − 1 Tσ
σ 2,b = P b b 2
= σ 2
P b b 2
∼ invΓ ,
t (et − e ) /T t (et − e ) /T
2 2
 
b2
σ 2 ) ∼ invΓ T −3
Note that pBC (σ 2 |b 2 , Tσ
2 under a flat prior, the Jacobian adjusts to the posterior to match
the true posterior. To compute the posterior mean, we need to compute the Jacobian of the transformation:
2,s 2
|ψθ |−1 = ∂σ 7
σ 2 . Since σ
∂b
2,b
= P (eTbσb−eb )2 , |ψθ |−1 = P (eTb −eb )2 .
t t t t
Under the prior p(σ 2,s ) ∝ 1, the posterior mean without the Jacobian transformation is
B P
2 1 X 2 (et − e)2 /T B→∞ 2 T
σ =σ Pt b b 2
−→ σ
T −3
b
B t (et − e ) /T
b=1

The posterior mean after adjusting for the Jacobian transformation is

PB
σ 2,b · P (eTb −eb )2 P T 2
P
b ( t (ebt −eb )2 )
1
P b 2
b=1 b (z )
σ 2RS = PB
t t
= σ
b 2
P P b − eb )2 /T
= T σ
b 2 B
1
P b
(e bz
2,b
b=1 1/σ b=1 t t B

p p
where 1/z b = t (ebt − eb )2 . As B → ∞, B1 b (z b )2 −→E[(z b )2 ] and S1 b z b −→E[z b ]. Now z b ∼ invχ2T −1
P P P
1
with mean T −3 and variance (T −3)22 (T −5) giving E[(z b )2 ] = (T −3)(T
1 2
b2 T T−5 =
−5) . Hence as B → ∞, σ RS,R = σ
2
σ BC .

Derivation of the Bias Reducing Prior The bias of the MLE estimator has E(b
σ ) = σ 2 − T1 σ 2 and
σ 2 ) = 2σ 4 ( T1 − T12 ). Since the auxiliary parameters coincide with the parameters of interest,
variance V (b
√ √
∇θ ψ(θ) and ∇θθ0 ψ(θ) = 0. For Z ∼ N (0, 1), A(v; σ 2 ) = 2σ 2 (1− T1 )Z, Thus ∂σ2 A(v; σ 2 ) = 2(1− T1 )Z, as =
7
b2,b (σ 2,b ) = σ
This holds because σ σ 2,b /dσ 2,b |−1 = |dσ 2,b /db
b2 so that |db σ 2 |.

37
√ 1
2σ 2 (1 − T )(Z − Z s ). The terms in the asymptotic expansion are therefore

1 2 s 1
∂σ2 A(v s ; σ 2 )as = ) Z (Z − Z s ) ⇒ E(∂σ2 A(v s ; σ 2 )as ) = −σ 2 2(1 − )2
2σ 2 (1 −
T T
s 4 1 2
V (a ) = 4σ (1 − )
T
s s0 1 2 4
cov(a , a ) = 2(1 − ) σ
T
1 S − 1 0 1  1 S − 1 σ2 S
(1 − )V (as ) + cov(as , as ) = σ 4 (1 − )2 4(1 − ) + 2 =
S S T S S 3(S − 1)

Noting that |∂σb2 σ 2,b | ∝ σ 2,b , it is analytically simpler in this example to solve for the weights directly, ie.
w(σ 2 ) = π(σ 2 )|∂σb2 σ 2,b | rather than the bias reducing prior π itself. Thus the bias reducing prior satisfies

−2σ 2 (1 − T1 )2 1 2
∂σ2 w(σ 2 ) =  =− 2 .
σ 4(1 − S ) + 2 S−1
1

σ 4 (1 − T1 )2 4(1 − S1 ) + 2 S−1
S S

Taking the integral on both sides we get:

1 1
log(w(σ 2 )) ∝ − log(σ 2 ) ⇒ w(σ 2 ) ∝ ⇒ π(σ 2 ) ∝ 4
σ2 σ
which is the Jeffreys prior if there is no re-weighting and the square of the Jeffreys prior when we use the
Jacobian to re-weight. Since the estimator for the mean was unbiased, π(m) ∝ 1 is the prior for m.
The posterior mean under the Bias Reducing Prior π(σ 2,s ) = 1/σ 4,s is the same as the posterior without
weights but using the Jeffreys prior π(σ 2,s ) = 1/σ 2,s :
PS 2,s
PT
s=1 σ (1/σ 2,s ) S t=1 (et − e)2 /T
σ 2RS = PS = PS 2
= σ PS PT s 2
≡σ 2
bSM D.
2,s 2,s s
s=1 1/σ s=1 1/σ s=1 t=1 (et − e ) /(ST )

38
D.2 Further Results for Dynamic Panel Model with Fixed Effects

Table 4: Dynamic Panel ρ = 0.9, β = 1, σ 2 = 2

Mean over 1000 replications

MLE LT SLT SMD ABC RS Bootstrap

Mean 0.751 0.751 0.895 0.898 0.889 0.899 0.751

ρb : SD 0.030 0.030 0.026 0.025 0.025 0.025 0.059
Bias -0.149 -0.149 -0.005 -0.002 -0.011 -0.001 -0.149

Mean 0.934 0.934 0.998 1.000 0.996 1.000 0.935

βb : SD 0.070 0.071 0.074 0.073 0.073 0.073 0.139
Bias -0.066 -0.066 -0.002 0.000 -0.004 0.000 -0.065

Mean 1.857 1.865 1.972 1.989 2.054 2.097 1.858

b2 :
σ SD 0.135 0.141 0.145 0.145 0.151 0.153 0.269
Bias -0.143 -0.135 -0.028 -0.011 0.054 0.097 -0.142

S – – 500 500 1 1 500

B – 500 500 – 500 500 –

See note to Table 3.

39
References
Arellano, M. and Bonhomme, S. 2009, Robust Priors in Nonlinear Panel Data Models, Econometrica
77(2), 489–536.

Bao, Y. and Ullah, A. 2007, The Second-Order Bias and Mean-Squared Error of Estimators in
Time Series Models, Journal of Econometrics 140(2), 650–669.

Beaumont, M., Zhang, W. and Balding, D. 2002, Approximate Bayesian Computation in Population
Genetics, Genetics 162, 2025–2035.

Bester, A. and Hansen, C. 2006, Bias Reduction for Bayesian and Frequentist Estimators, Mimeo,
University of Chicago.

Blum, M., Nunes, M., Prangle, D. and Sisson, A. 2013, A Comparative Review of Dimension
Reduction Methods in Approximate Bayesian Computation, Statistical Science 28(2), 189–208.

Cabrera, J. and Fernholz, L. 1999, Target Estimation for Bias and Mean Square Error Reduction,
Annals of Statistics 27, 1080–1104.

Cabrera, J. and Hu, I. 2001, Algorithms for Target Estimation Using Stochastic Approximation,
InterStat 2(4), 1–18.

Calvet, L. and Czellar, V. 2015, Accurate Methods for Approximate Bayesian Computation Filter-
ing, Journal of Financial Econometrics.

Chernozhukov, V. and Hong, H. 2003, An MCMC Approach to Classical Estimation, Journal of

Econometrics 115:2, 293–346.

Creel, M. and Kristensen, D. 2013, Indirect Likelihood Inference, mimeo, UCL.

Dean, T., Singh, S., Jasra, A. and Peters, G. 2011, Parameter Estimation for Hidden Markov
Models with Intractable Likelihoods, arXiv:1103.5399.

Diggle, P. and Gratton, J. 1984, Monte Carlop Methods of Inference for Implicit Statistical Methods,
Journal of the Royal Statistical Association Series B 46, 193–227.

Drovandi, C., Pettitt, A. and Faaddy, M. 2011, Approximate Bayesian Computation using Indirect
Inference, Journal of the Royal Statistical Society, Series C 60(3), 503–524.

Drovandi, C., Pettitt, A. and Lee, A. 2015, Bayesian Indirect Inference Using a Parametric Auxiliary
Model, Statistical Science 30(1), 72–95.

Duffie, D. and Singleton, K. 1993, Simulated Moments Estimation of Markov Models of Asset
Prices, Econometrica 61, 929–952.

Forneron, J. J. and Ng, S. 2015, Reverse Importance Sampling by Optimization, mimeo, Columbia
University.

Gallant, R. and Tauchen, G. 1996, Which Moments to Match, Econometric Theory 12, 657–681.

Gao, J. and Hong, H. 2014, A Computational Implementation of GMM, SSRN Working Paper
2503199.

40
Gourieroux, C. and Monfort, A. 1996, Simulation-Based Econometric Methods, Oxford University
Press.

Gourieroux, C., Monfort, A. and Renault, E. 1993, Indirect Inference, Journal of Applied Econo-
metrics 85, 85–118.

Gourieroux, C., Renault, E. and Touzi, N. 1999, Calibration by Simulation for Small Sample Bias
Correction,, in R. Mariano, T. Schuermann and M. Weeks (eds), Simulation-based Inference in
Econometrics: Methods and Applications, Cambridge University Press.

Gourieroux, G., Phillips, P. and Yu, J. 2010, Indirect Inference of Dynamic Panel Models, Journal
of Econometrics 157(1), 68–77.

Hansen, L. P. 1982, Large Sample Properties of Generalized Method of Moments Estimators,

Econometrica 50, 1029–1054.

Hansen, L. P. and Singleton, K. J. 1982, Generalized Instrumental Variables Estimation of Nonlinear

Rational Expectations Models, Econometrica 50, 1269–1296.

Heggland, K. and Frigessi, A. 2004, Estimating Functions in Indirect Inference, Journal of the
Royal Statistical Association Series B 66, 447–462.

Hsiao, C. 2003, Analysis of Panel Data, Cambridge University Press.

Jacquier, E., Johannes, M. and Polson, N. 2007, MCMC Maximum Likelihood for Latent State
Models, Journal of Econometrics 137(2), 615–640.

Jiang, W. and Turnbull, B. 2004, The Indirect Method: Inference Based on Intermediate Statistics-
A Synthesis and Examples, Statistical Science 19(2), 239–263.

Kass, R., Tierney, L. and Kadane, J. 1990, The Validity of Posterior Expansion Based on Lapalce’s
Method, in R. K. S. Gleisser and L. Wasserman (eds), Bayesian and Likelihood Mehtods in
Statistics and Econometrics, Elsevier Science Publishers, North Holland.

Kirkpatrick, S., Gellatt, C. and Vecchi, M. 1983, Optimization by Simulated Annealing, Science
220, 671–680.

Kormiltsina, A. and Nekipelov, D. 2014, Consistent Variance of the Laplace Type Estimators,
SMU, mimeo.

Lise, J., Meghir, C. and Robin, J. M. 2015, Matching, Sorting, and Wages, Review of Economic
Dynamics. Cowles Foundation Working Paper 1886.

Marin, J. M., Pudio, P., Robert, C. and Ryder, R. 2012, Approximate Bayesian Computation
Methods, Statistical Computations 22, 1167–1180.

Marjoram, P., Molitor, J., Plagnol, V. and Tavare, S. 2003, Markov Chain Monte Carlo Without
Likelihoods, Procedings of the National Academy of Science 100(26), 15324–15328.

Meeds, E. and Welling, M. 2015, Optimization Monte Carlo: Efficient and Embarrassingly Parallel
Likelihood-Free Inference, arXiv:1506:03693v1.

41
Michaelides, A. and Ng, S. 2000, Estimating the Rational Expectations Model of Speculative
Storage: A Monte Carlo Comparison of Three Simulation Estimators, Journal of Econometrics
96:2, 231–266.

Nekipelov, D. and Kormilitsina, A. 2015, Approximation Properties of Laplace-Type Estimators,

in N. Balke, F. Canova, F. Milani and M. Wynne (eds), DSGE Models in Macroeconomics:
Estimation, Evaluation, and New Developments, Vol. 28, pp. 291–318.

Newey, W. and Smith, R. 2004a, Higher Order Properties of GMM and Generalized Empirical
Likelihood Estimators, Econometrica 72, 219–225.

Newey, W. and Smith, R. 2004b, Higher Order Properties of GMM and Generalized Empirical
Likeliood Estimators, Econometrica 71:1, 219–255.

Nickl, R. and Potscher, B. 2010, Efficient Simulation-Based Minimum Distance Estimation and
Indirect Inference, Mathematical Methods of Statistics 19(4), 327–364.

Pagan, A. and Ullah, A. 1999, Nonparametric Econometrics, Vol. Themes in Modern Econometrics,
Cambridge University Press.

Pritchard, J., Seielstad, M., Perez-Lezman, A. and Feldman, M. 1996, Population Growth of Human
Y chromosomes: A Study of Y Chromosome MicroSatellites, Molecular Biology and Evolution
16(12), 1791–1798.

Rilstone, P., Srivastara, K. and Ullah, A. 1996, The Second-Order Bias and Mean Squared Error
of Nonlinear Estimators, Journal of Econometrics 75, 369–385.

Robert, C. and Casella, G. 2004, Monte Carlo Statistical Methods, Textbooks in Statistics, second
edn, Springer.

Sisson, S. and Fan, Y. 2011, Likelihood Free Markov Chain Monte Carlo, in S. Brooks, A. Celman,
G. Jones and X.-L. Meng (eds), Handbook of Markov Chain Monte Carlo, Vol. Chapter 12,
pp. 313–335. arXiv:10001.2058v1.

Smith, A. 1993, Estimating Nonlinear Time Series Models Using Simulated Vector Autoregressions,
Journal of Applied Econometrics 8, S63–S84.

Smith, A. 2008, Indirect Inference, in S. Durlauf and L. Blume (eds), The New Palgrave Dictionary
of Economics, Vol. 2, Palgrave Mcmillian.

Tavare, S., Balding, J., Griffiths, C. and Donnelly, P. 1997, Inferring Coalescence Times From DNA
Sequence Data, Genetics 145, 505–518.

Tierney, L. and Kadane, J. 1986, Accurate Approximations for Posterior Moments and Marginal
Densities, Journal of the American Statistical Association 81, 82–86.

42