random forests for ABC model choice and parameter estimation

ABC random forests for Bayesian testing and
parameter inference
Christian P. Robert
Universit´e Paris-Dauphine, Paris & University of Warwick, Coventry
Joint work with A. Estoup, J.M. Marin, P. Pudlo, L Raynal, & M. Ribatet

Outline
Approximate Bayesian computation
ABC for model choice
ABC model choice via random forests
ABC estimation via random forests

ABC basics
Exact ABC simulation of
approximate targets
Automated summary selection

Untractable likelihoods
Cases when the likelihood function
f (y|θ) is unavailable and when the
completion step
f (y|θ) =
Z
f (y, z|θ) dz
is impossible or too costly because of
the dimension of z
c MCMC cannot be implemented

The ABC method
Bayesian setting: target is π(θ)f (x|θ)
When likelihood f (x|θ) not in closed form, likelihood-free rejection
technique:
ABC algorithm
For an observation y ∼ f (y|θ), under the prior π(θ), keep jointly
simulating
θ ∼ π(θ) , z ∼ f (z|θ ) ,
until the auxiliary variable z is equal to the observed value, z = y.
[Tavar´e et al., 1997]

A as A...pproximative
When y is a continuous random variable, equality z = y is
replaced with a tolerance condition,
ρ(y, z)
where ρ is a distance
Output distributed from
π(θ) Pθ{ρ(y, z) < } ∝ π(θ|ρ(y, z) < )
[Pritchard et al., 1999]

ABC algorithm
Algorithm 1 Likelihood-free rejection sampler 2
for i = 1 to N do
repeat
generate θ from the prior distribution π(·)
generate z from the likelihood f (·|θ )
until ρ{η(z), η(y)}
set θi = θ
end for
where η(y) deﬁnes a (not necessarily suﬃcient) statistic

Output
The likelihood-free algorithm samples from the marginal in z of:
π (θ, z|y) =
π(θ)f (z|θ)IA ,y (z)
A ,y×Θ π(θ)f (z|θ)dzdθ
,
where A ,y = {z ∈ D|ρ(η(z), η(y)) < }.
The idea behind ABC is that the summary statistics coupled with a
small tolerance should provide a good approximation of the
posterior distribution:
π (θ|y) = π (θ, z|y)dz ≈ π(θ|η(y)) .

MA example
MA(q) model
xt = t +
q
i=1
ϑi t−i
Simple prior: uniform over the inverse [real and complex] roots in
Q(u) = 1 −
q
i=1
ϑi ui
under the identiﬁability conditions

MA example
MA(q) model
xt = t +
q
i=1
ϑi t−i
Simple prior: uniform prior over the identiﬁability zone, e.g.
triangle for MA(2)

MA example (2)
ABC algorithm thus made of
1. picking a new value (ϑ1, ϑ2) in the triangle
2. generating an iid sequence ( t)−q<t T
3. producing a simulated series (xt )1 t T
Distance: basic distance between the series
ρ((xt )1 t T , (xt)1 t T ) =
T
t=1
(xt − xt )2
or distance between summary statistics like the q autocorrelations
τj =
T
t=j+1
xtxt−j

Comparison of distance impact
Evaluation of the tolerance on the ABC sample against both
distances ( = 100%, 10%, 1%, 0.1%) for an MA(2) model

Comparison of distance impact
0.0 0.2 0.4 0.6 0.8
01234
θ1
−2.0 −1.0 0.0 0.5 1.0 1.5
0.00.51.01.5
θ2
Evaluation of the tolerance on the ABC sample against both
distances ( = 100%, 10%, 1%, 0.1%) for an MA(2) model

ABC advances
Simulating from the prior is often poor in eﬃciency
Either modify the proposal distribution on θ to increase the density
of x’s within the vicinity of y...
[Marjoram et al, 2003; Bortot et al., 2007, Beaumont et al., 2009]
...or by viewing the problem as a conditional density estimation
and by developing techniques to allow for larger
[Beaumont et al., 2002; Blum & Fran¸cois, 2009]
.....or even by including in the inferential framework [ABCµ]
[Ratmann et al., 2009]

ABC consistency
Recent studies on large sample properties of ABC posterior
distributions and ABC posterior means
[Liu & Fearnhead, 2016; Frazier et al., 2016]
Under regularity conditions on summary statistics,
incl. convergence at speed dT , characterisation of rate of posterior
concentration as a function of tolerance convergence
less stringent condition on tolerance decrease than for
asymptotic normality of posterior;
asymptotic normality of posterior mean does not require
asymptotic normality of posterior itself
Cases for limiting ABC distributions
1. dT T −→ +∞;
2. dT T −→ c;
3. dT T −→ 0
and limiting ABC mean convergent for 2
T = o(1/dT )
[Frazier et al., 2016]

Wilkinson’s exact BC
ABC approximation error (i.e. non-zero tolerance) replaced with
exact simulation from a controlled approximation to the target,
convolution of true posterior with kernel function
π (θ, z|y) =
π(θ)f (z|θ)K (y − z)
π(θ)f (z|θ)K (y − z)dzdθ
,
with K kernel parameterised by bandwidth .
[Wilkinson, 2008]
Theorem
The ABC algorithm based on the assumption of a randomised
observation y = ˜y + ξ, ξ ∼ K , and an acceptance probability of
K (y − z)/M
gives draws from the posterior distribution π(θ|y).

How exact a BC?
“Using to represent measurement error is
straightforward, whereas using to model the model
discrepancy is harder to conceptualize and not as
commonly used”
[Richard Wilkinson, 2008, 2013]

How exact a BC?
Pros
Pseudo-data from true model and observed data from noisy
model
Interesting perspective in that outcome is completely
controlled
Link with ABCµ and assuming y is observed with a
measurement error with density K
Relates to the theory of model approximation
[Kennedy & O’Hagan, 2001]
Cons
Requires K to be bounded by M
True approximation error never assessed
Requires a modiﬁcation of the standard ABC algorithm

Noisy ABC
Idea: Modify the data from the start
˜y = y0 + ζ1
with the same scale as ABC
[ see Fearnhead-Prangle ]
run ABC on ˜y
Then ABC produces an exact simulation from π(θ|˜y) = π(θ|˜y)
[Dean et al., 2011; Fearnhead and Prangle, 2012]

Consistent noisy ABC
Degrading the data improves the estimation performances:
Noisy ABC-MLE is asymptotically (in n) consistent
under further assumptions, the noisy ABC-MLE is
asymptotically normal
increase in variance of order −2
likely degradation in precision or computing time due to the
lack of summary statistic [curse of dimensionality]

Semi-automatic ABC
Fearnhead and Prangle (2010) study ABC and the selection of the
summary statistic in close proximity to Wilkinson’s proposal
ABC then considered from a purely inferential viewpoint and
calibrated for estimation purposes
Use of a randomised (or ‘noisy’) version of the summary statistics
˜η(y) = η(y) + τ
Derivation of a well-calibrated version of ABC, i.e. an algorithm
that gives proper predictions for the distribution associated with
this randomised summary statistic [calibration constraint: ABC
approximation with same posterior mean as the true randomised
posterior]
Optimality of the posterior expectation E[θ|y] of the parameter of
interest as summary statistics η(y)!

Bayesian model choice
Several models M1, M2, . . . are considered simultaneously for a
dataset y and the model index M is part of the inference.
Use of a prior distribution. π(M = m), plus a prior distribution on
the parameter conditional on the value m of the model index,
πm(θm)
Goal is to derive the posterior distribution of M, challenging
computational target when models are complex.

Generic ABC for model choice
Algorithm 2 Likelihood-free model choice sampler (ABC-MC)
for t = 1 to T do
repeat
Generate m from the prior π(M = m)
Generate θm from the prior πm(θm)
Generate z from the model fm(z|θm)
until ρ{η(z), η(y)} <
Set m(t) = m and θ(t)
= θm
end for
[Cornuet et al., DIYABC, 2009]

ABC estimates
Posterior probability π(M = m|y) approximated by the frequency
of acceptances from model m
1
T
T
t=1
Im(t)=m .
Issues with implementation:
should tolerances be the same for all models?
should summary statistics vary across models (incl. their
dimension)?
should the distance measure ρ vary as well?

Back to sufficiency
‘Sufficient statistics for individual models are unlikely to
be very informative for the model probability.’
[Scott Sisson, Jan. 31, 2011, X.’Og]
If η1(x) sufficient statistic for model m = 1 and parameter θ1 and
η2(x) sufficient statistic for model m = 2 and parameter θ2,
(η1(x), η2(x)) is not always sufficient for (m, θm)
c Potential loss of information at the testing level

Limiting behaviour of B12 (T → ∞)
ABC approximation
B12(y) =
T
t=1 Imt =1 Iρ{η(zt ),η(y)}
T
t=1 Imt =2 Iρ{η(zt ),η(y)}
,
where the (mt, zt)’s are simulated from the (joint) prior
As T go to inﬁnity, limit
B12(y) =
Iρ{η(z),η(y)} π1(θ1)f1(z|θ1) dz dθ1
Iρ{η(z),η(y)} π2(θ2)f2(z|θ2) dz dθ2
=
Iρ{η,η(y)} π1(θ1)f η
1 (η|θ1) dη dθ1
Iρ{η,η(y)} π2(θ2)f η
2 (η|θ2) dη dθ2
,
where f η
1 (η|θ1) and f η
2 (η|θ2) distributions of η(z)

Limiting behaviour of B12 ( → 0)
When goes to zero,
Bη
12(y) =
π1(θ1)f η
1 (η(y)|θ1) dθ1
π2(θ2)f η
2 (η(y)|θ2) dθ2
,
c Bayes factor based on the sole observation of η(y)

Limiting behaviour of B12 (under sufficiency)
If η(y) sufficient statistic for both models,
fi (y|θi ) = gi (y)f η
i (η(y)|θi )
Thus
B12(y) =
Θ1
π(θ1)g1(y)f η
1 (η(y)|θ1) dθ1
Θ2
π(θ2)g2(y)f η
2 (η(y)|θ2) dθ2
=
g1(y) π1(θ1)f η
1 (η(y)|θ1) dθ1
g2(y) π2(θ2)f η
2 (η(y)|θ2) dθ2
=
g1(y)
g2(y)
Bη
12(y) .
[Didelot, Everitt, Johansen & Lawson, 2011]
c No discrepancy only when cross-model sufficiency

Poisson/geometric example
Sample
x = (x1, . . . , xn)
from either a Poisson P(λ) or from a geometric G(p) Then
S =
n
i=1
yi = η(x)
suﬃcient statistic for either model but not simultaneously
Discrepancy ratio
g1(x)
g2(x)
=
S!n−S / i yi !
1 n+S−1
S

Poisson/geometric discrepancy
Range of B12(x) versus Bη
12(x) B12(x): The values produced have
nothing in common.

Formal recovery
Creating an encompassing exponential family
f (x|θ1, θ2, α1, α2) ∝ exp{θT
1 η1(x) + θT
1 η1(x) + α1t1(x) + α2t2(x)}
leads to a suﬃcient statistic (η1(x), η2(x), t1(x), t2(x))

Formal recovery
f (x|θ1, θ2, α1, α2) ∝ exp{θT
1 η1(x) + θT
1 η1(x) + α1t1(x) + α2t2(x)}
In the Poisson/geometric case, if i xi ! is added to S, no
discrepancy

Formal recovery
f (x|θ1, θ2, α1, α2) ∝ exp{θT
1 η1(x) + θT
1 η1(x) + α1t1(x) + α2t2(x)}
Only applies in genuine suﬃciency settings...
c Inability to evaluate loss brought by summary statistics

MA(q) divergence
1 2
0.00.20.40.60.81.0
1 2
0.00.20.40.60.81.0
1 2
0.00.20.40.60.81.0
1 2
0.00.20.40.60.81.0
Evolution [against ] of ABC Bayes factor, in terms of frequencies of
visits to models MA(1) (left) and MA(2) (right) when equal to
10, 1, .1, .01% quantiles on insuﬃcient autocovariance distances. Sample
of 50 points from a MA(2) with θ1 = 0.6, θ2 = 0.2. True Bayes factor
equal to 17.71.

MA(q) divergence
1 2
0.00.20.40.60.81.0
1 2
0.00.20.40.60.81.0
1 2
0.00.20.40.60.81.0
1 2
0.00.20.40.60.81.0
Evolution [against ] of ABC Bayes factor, in terms of frequencies of
visits to models MA(1) (left) and MA(2) (right) when equal to
10, 1, .1, .01% quantiles on insuﬃcient autocovariance distances. Sample
of 50 points from a MA(1) model with θ1 = 0.6. True Bayes factor B21
equal to .004.

A stylised problem
Central question to the validation of ABC for model choice:
When is a Bayes factor based on an insuﬃcient statistic T(y)
consistent?
Note/warnin: c drawn on T(y) through BT
12(y) necessarily diﬀers
from c drawn on y through B12(y)
[Marin, Pillai, X, & Rousseau, JRSS B, 2013]

A benchmark if toy example
Comparison suggested by referee of PNAS paper [thanks!]:
[X, Cornuet, Marin, & Pillai, Aug. 2011]
Model M1: y ∼ N(θ1, 1) opposed
to model M2: y ∼ L(θ2, 1/
√
2), Laplace distribution with mean θ2
and scale parameter 1/
√
2 (variance one).
Four possible statistics
1. sample mean y (suﬃcient for M1 if not M2);
2. sample median med(y) (insuﬃcient);
3. sample variance var(y) (ancillary);
4. median absolute deviation mad(y) = med(|y − med(y)|);

A benchmark if toy example
Comparison suggested by referee of PNAS paper [thanks!]:
[X, Cornuet, Marin, & Pillai, Aug. 2011]
Model M1: y ∼ N(θ1, 1) opposed
to model M2: y ∼ L(θ2, 1/
√
2), Laplace distribution with mean θ2
and scale parameter 1/
√
2 (variance one).
q
q
q
q
q
q
q
q
q
q
q
Gauss Laplace
0.00.10.20.30.40.50.60.7
n=100
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
Gauss Laplace
0.00.20.40.60.81.0
n=100

Framework
Starting from sample
y = (y1, . . . , yn)
the observed sample, not necessarily iid with true distribution
y ∼ Pn
Summary statistics
T(y) = Tn
= (T1(y), T2(y), · · · , Td (y)) ∈ Rd
with true distribution Tn
∼ Gn.

Assumptions
A collection of asymptotic “standard” assumptions:
[A1] is a standard central limit theorem under the true model with
asymptotic mean µ0
[A2] controls the large deviations of the estimator Tn
from the
model mean µ(θ)
[A3] is the standard prior mass condition found in Bayesian
asymptotics (di eﬀective dimension of the parameter)
[A4] restricts the behaviour of the model density against the true
density
[Think CLT!]

Asymptotic marginals
Asymptotically, under [A1]–[A4]
mi (t) =
Θi
gi (t|θi ) πi (θi ) dθi
is such that
(i) if inf{|µi (θi ) − µ0|; θi ∈ Θi } = 0,
Cl vd−di
n mi (Tn
) Cuvd−di
n
and
(ii) if inf{|µi (θi ) − µ0|; θi ∈ Θi } > 0
mi (Tn
) = oPn [vd−τi
n + vd−αi
n ].

Between-model consistency
Consequence of above is that asymptotic behaviour of the Bayes
factor is driven by the asymptotic mean value µ(θ) of Tn
under
both models. And only by this mean value!

under
Indeed, if
inf{|µ0 − µ2(θ2)|; θ2 ∈ Θ2} = inf{|µ0 − µ1(θ1)|; θ1 ∈ Θ1} = 0
then
Cl v
−(d1−d2)
n m1(Tn
) m2(Tn
) Cuv
−(d1−d2)
n ,
where Cl , Cu = OPn (1), irrespective of the true model.
c Only depends on the diﬀerence d1 − d2: no consistency

under
Else, if
inf{|µ0 − µ2(θ2)|; θ2 ∈ Θ2} > inf{|µ0 − µ1(θ1)|; θ1 ∈ Θ1} = 0
then
m1(Tn
)
m2(Tn
)
Cu min v
−(d1−α2)
n , v
−(d1−τ2)
n

Checking for adequate statistics
Run a practical check of the relevance (or non-relevance) of Tn
null hypothesis that both models are compatible with the statistic
Tn
H0 : inf{|µ2(θ2) − µ0|; θ2 ∈ Θ2} = 0
against
H1 : inf{|µ2(θ2) − µ0|; θ2 ∈ Θ2} > 0
testing procedure provides estimates of mean of Tn
under each
model and checks for equality

Checking in practice
Under each model Mi , generate ABC sample θi,l , l = 1, · · · , L
For each θi,l , generate yi,l ∼ Fi,n(·|ψi,l ), derive Tn
(yi,l ) and
compute
^µi =
1
L
L
l=1
Tn
(yi,l ), i = 1, 2 .
Conditionally on Tn
(y),
√
L { ^µi − Eπ
[µi (θi )|Tn
(y)]} N(0, Vi ),
Test for a common mean
H0 : ^µ1 ∼ N(µ0, V1) , ^µ2 ∼ N(µ0, V2)
against the alternative of diﬀerent means
H1 : ^µi ∼ N(µi , Vi ), with µ1 = µ2 .

Toy example: Laplace versus Gauss
qqqqqqqqqqqqqqq
qqqqqqqqqq
q
qq
q
q
Gauss Laplace Gauss Laplace
010203040
Normalised χ2 without and with mad

Random forests
ABC with random forests
Illustrations

Leaning towards machine learning
Main notions:
ABC-MC seen as learning about which model is most
appropriate from a huge (reference) table
exploiting a large number of summary statistics not an issue
for machine learning methods intended to estimate eﬃcient
combinations
abandoning (temporarily?) the idea of estimating posterior
probabilities of the models, poorly approximated by machine
learning methods, and replacing those by posterior predictive
expected loss
[Cornuet et al., 2016]

Random forests
Technique that stemmed from Leo Breiman’s bagging (or
bootstrap aggregating) machine learning algorithm for both
classification and regression
[Breiman, 1996]
Improved classification performances by averaging over
classification schemes of randomly generated training sets, creating
a “forest” of (CART) decision trees, inspired by Amit and Geman
(1997) ensemble learning
[Breiman, 2001]

Growing the forest
Breiman’s solution for inducing random features in the trees of the
forest:
boostrap resampling of the dataset and
random subset-ing [of size
√
t] of the covariates driving the
classiﬁcation at every node of each tree
Covariate xτ that drives the node separation
xτ cτ
and the separation bound cτ chosen by minimising entropy or Gini
index

Breiman and Cutler’s algorithm
Algorithm 3 Random forests
for t = 1 to T do
//*T is the number of trees*//
Draw a bootstrap sample of size nboot
Grow an unpruned decision tree
for b = 1 to B do
//*B is the number of nodes*//
Select ntry of the predictors at random
Determine the best split from among those predictors
end for
end for
Predict new data by aggregating the predictions of the T trees
[ c Tae-Kyun Kim & Bjorn Stenger, 2009]

Subsampling
Due to both large datasets [practical] and theoretical
recommendation from G´erard Biau [private communication], from
independence between trees to convergence issues, boostrap
sample of much smaller size than original data size
N = o(n)
Each CART tree stops when number of observations per node is 1:
no culling of the branches

Idea: Starting with
possibly large collection of summary statistics (s1i , . . . , spi )
(from scientiﬁc theory input to available statistical softwares,
to machine-learning alternatives)
ABC reference table involving model index, parameter values
and summary statistics for the associated simulated
pseudo-data
run R randomforest to infer M from (s1i , . . . , spi )

Idea: Starting with
pseudo-data
at each step O(
√
p) indices sampled at random and most
discriminating statistic selected, by minimising entropy Gini loss

Idea: Starting with
pseudo-data
Average of the trees is resulting summary statistics, highly
non-linear predictor of the model index

Outcome of ABC-RF
Random forest predicts a (MAP) model index, from the observed
dataset: The predictor provided by the forest is “suﬃcient” to
select the most likely model but not to derive associated posterior
probability
exploit entire forest by computing how many trees lead to
picking each of the models under comparison but variability
too high to be trusted
frequency of trees associated with majority model is no proper
substitute to the true posterior probability
And usual ABC-MC approximation equally highly variable and
hard to assess

Posterior predictive expected losses
We suggest replacing unstable approximation of
P(M = m|xo)
with xo observed sample and m model index, by average of the
selection errors across all models given the data xo,
P( ^M(X) = M|xo)
where pair (M, X) generated from the predictive
f (x|θ)π(θ, M|xo)dθ
and ^M(x) denotes the random forest model (MAP) predictor

Posterior predictive expected losses
Arguments:
Bayesian estimate of the posterior error
integrates error over most likely part of the parameter space
gives an averaged error rather than the posterior probability of
the null hypothesis
easily computed: Given ABC subsample of parameters from
reference table, simulate pseudo-samples associated with
those and derive error frequency

toy: MA(1) vs. MA(2)
Comparing an MA(1) and an MA(2) models:
xt = t − ϑ1 t−1[−ϑ2 t−2]
Earlier illustration using ﬁrst two autocorrelations as S(x)
[Marin et al., Stat. & Comp., 2011]
Result #1: values of p(m|x) [obtained by numerical integration]
and p(m|S(x)) [obtained by mixing ABC outcome and density
estimation] highly diﬀer!

Diﬀerence between the posterior probability of MA(2) given either
x or S(x). Blue stands for data from MA(1), orange for data from
MA(2)

xt = t − ϑ1 t−1[−ϑ2 t−2]
Earlier illustration using two autocorrelations as S(x)
Result #2: Embedded models, with simulations from MA(1)
within those from MA(2), hence linear classiﬁcation poor

Simulations of S(x) under MA(1) (blue) and MA(2) (orange)

xt = t − ϑ1 t−1[−ϑ2 t−2]
Earlier illustration using two autocorrelations as S(x)
Result #3: On such a small dimension problem, random forests
should come second to k-nn ou kernel discriminant analyses

classiﬁcation prior
method error rate (in %)
LDA 27.43
Logist. reg. 28.34
SVM (library e1071) 17.17
“na¨ıve” Bayes (with G marg.) 19.52
“na¨ıve” Bayes (with NP marg.) 18.25
ABC k-nn (k = 100) 17.23
ABC k-nn (k = 50) 16.97
Local log. reg. (k = 1000) 16.82
Random Forest 17.04
Kernel disc. ana. (KDA) 16.95
True MAP 12.36

Comments
unlimited aggregation of arbitrary summary statistics
recovery of discriminant statistics when available
automated implementation with reduced calibration
self-evaluation by posterior predictive error
soon to be included within DIYABC

Two basic issues with ABC
ABC compares numerous simulated dataset to the observed one
Two major diﬃculties:
to decrease approximation error (or tolerance ) and hence
ensure reliability of ABC, total number of simulations very
large;
calibration of ABC (tolerance, distance, summary statistics,
post-processing, &tc) critical and hard to automatise

classiﬁcation of summaries by random forests
Given a large collection of summary statistics, rather than selecting
a subset and excluding the others, estimate each parameter of
interest by a machine learning tool like random forests
RF can handle thousands of predictors
ignore useless components
fast estimation method with good local properties
automatised method with few calibration steps
substitute to Fearnhead and Prangle (2012) preliminary
estimation of ^θ(yobs)
includes a natural (classiﬁcation) distance measure that avoids
choice of both distance and tolerance
[Marin et al., 2016]

random forests as non-parametric regression
CART means Classiﬁcation and Regression Trees
For regression purposes, i.e., to predict y as f (x), similar binary
trees in random forests
1. at each tree node, split data into two daughter nodes
2. split variable and bound chosen to minimise heterogeneity
criterion
3. stop splitting when enough homogeneity in current branch
4. predicted values at terminal nodes (or leaves) are average
response variable y for all observations in ﬁnal leaf

Illustration
conditional expectation f (x) and well-speciﬁed dataset

Illustration
single regression tree

Illustration
ten regression trees obtained by bagging (Bootstrap AGGregatING)

Illustration
average of 100 regression trees

bagging reduces learning variance
When growing forest with many trees,
grow each tree on an independent bootstrap sample
at each node, select m variables at random out of all M
possible variables
Find the best dichotomous split on the selected m variables
predictor function estimated by averaging trees
Improve on CART with respect to accuracy and stability

prediction error
A given simulation (ysim, xsim) in the training table is not used in
about 1/3 of the trees (“out-of-bag” case)
Average predictions ^Foob(xsim) of these trees to give out-of-bag
predictor of ysim

Related methods
adjusted local linear: Beaumont et al. (2002) Approximate Bayesian
computation in population genetics, Genetics
ridge regression: Blum et al. (2013) A Comparative Review of
Dimension Reduction Methods in Approximate Bayesian Computation,
Statistical Science
linear discriminant analysis: Estoup et al. (2012) Estimation of
demo-genetic model probabilities with Approximate Bayesian
Computation using linear discriminant analysis on summary statistics,
Molecular Ecology Resources
adjusted neural networks: Blum and Fran¸cois (2010) Non-linear
regression models for Approximate Bayesian Computation, Statistics and
Computing

ABC parameter estimation (ODOF)
One dimension = one forest (ODOF) methodology
parametric statistical model:
{f (y; θ): y ∈ Y, θ ∈ Θ}, Y ⊆ Rn
, Θ ⊆ Rp
with intractable density f (·; θ)
plus prior distribution π(θ)
Inference on quantity of interest
ψ(θ) ∈ R
(posterior means, variances, quantiles or covariances)

common reference table
Given η: Y → Rk a collection of summary statistics
produce reference table (RT) used as learning dataset for
multiple random forests
meaning, for 1 t N
1. simulate θ(t)
∼ π(θ)
2. simulate ˜yt = (˜y1,t, . . . , ˜yn,t) ∼ f (y; θ(t)
)
3. compute η(˜yt) = {η1(˜yt), . . . , ηk (˜yt)}

ABC posterior expectations
Recall that θ = (θ1, . . . , θd ) ∈ Rd
For each θj , construct a separate RF regression with predictors
variables equal to summary statistics η(y) = {η1(y), . . . , ηk(y)}
If Lb(η(y∗)) denotes leaf index of b-th tree associated with η(y∗)
—leaf reached through path of binary choices in tree—, with |Lb|
response variables
E(θj | η(y∗)) =
1
B
B
b=1
1
|Lb(η(y∗))|
t:η(yt )∈Lb(η(y∗))
θ
(t)
j
is our ABC estimate

ABC posterior expectations
For each θj , construct a separate RF regression with predictors
variables equal to summary statistics η(y) = {η1(y), . . . , ηk(y)}
If Lb(η(y∗)) denotes leaf index of b-th tree associated with η(y∗)
—leaf reached through path of binary choices in tree—, with |Lb|
response variables
E(θj | η(y∗)) =
1
B
B
b=1
1
|Lb(η(y∗))|
t:η(yt )∈Lb(η(y∗))
θ
(t)
j
is our ABC estimate

ABC posterior quantile estimate
Random forests also available for quantile regression
[Meinshausen, 2006, JMLR]
Since
^E(θj | η(y∗
)) =
N
t=1
wt(η(y∗
))θ
(t)
j
with
wt(η(y∗
)) =
1
B
B
b=1
ILb(η(y∗))(η(yt))
|Lb(η(y∗))|
natural estimate of the cdf of θj is
^F(u | η(y∗
)) =
N
t=1
wt(η(y∗
))I{θ
(t)
j u}
.

ABC posterior quantile estimate
Since
^E(θj | η(y∗
)) =
N
t=1
wt(η(y∗
))θ
(t)
j
with
wt(η(y∗
)) =
1
B
B
b=1
ILb(η(y∗))(η(yt))
|Lb(η(y∗))|
natural estimate of the cdf of θj is
^F(u | η(y∗
)) =
N
t=1
wt(η(y∗
))I{θ
(t)
j u}
.
ABC posterior quantiles + credible intervals given by ^F−1

ABC variances
Even though approximation of Var(θj | η(y∗)) available based on
^F, choice of alternative and slightly more involved version
In a given tree b in a random forest, existence of out-of-baf entries,
i.e., not sampled in associated bootstrap subsample
Use of out-of-bag simulations to produce estimate of E{θj | η(yt)},
˜θj
(t)
,
Apply weights ωt(η(y∗)) to out-of-bag residuals:
Var(θj | η(y∗
)) =
N
t=1
ωt(η(y∗
)) (θ
(t)
j − ˜θj
(t) 2

ABC covariances
For estimating Cov(θj , θ | η(y∗)), construction of a speciﬁc
random forest
product of out-of-bag errors for θj and θ
θ
(t)
j − ˜θj
(t)
θ
(t)
− ˜θ
(t)
with again predictors variables the summary statistics
η(y) = {η1(y), . . . , ηk(y)}

Reference table of N = 10, 000 Gaussian replicates
Independent Gaussian test set of size Npred = 100
k = 53 summary statistics: the sample mean, the sample
variance and the sample median absolute deviation, and 50
independent pure-noise variables (uniform [0,1])

−2 −1 0 1 2
−2−1012
ψ1
ψ~
1
0.5 1.0 1.5 2.0 2.5
0.51.01.52.02.5
ψ2
ψ~
2
0.05 0.10 0.15 0.20 0.25 0.30 0.35
0.050.150.250.35
ψ3
ψ~
3
0.0 0.2 0.4 0.6 0.80.00.20.40.60.8
ψ4
ψ~
4
Scatterplot of the theoretical values with their corresponding
estimates

−4 −3 −2 −1 0 1 2
−4−3−2−1012
Q0.025(θ1 | y)
Q
~
0.025(θ1|y)
−1 0 1 2 3 4
−101234
Q0.975(θ1 | y)
Q
~
0.975(θ1|y)
0.2 0.4 0.6 0.8 1.0 1.2
0.20.40.60.81.01.2
Q0.025(θ2 | y)
Q
~
0.025(θ2|y)
1 2 3 4 512345
Q0.975(θ2 | y)
Q
~
0.975(θ2|y)
Scatterplot of the theoretical values of 2.5% and 97.5% posterior
quantiles for θ1 and θ2 with their corresponding estimates

ODOF adj local linear adj ridge adj neural net
ψ1(y) = E(θ1 | y) 0.21 0.42 0.38 0.42
ψ2(y) = E(θ2 | y) 0.11 0.20 0.26 0.22
ψ3(y) = Var(θ1 | y) 0.47 0.66 0.75 0.48
ψ4(y) = Var(θ2 | y) 0.46 0.85 0.73 0.98
Q0.025(θ1|y) 0.69 0.55 0.78 0.53
Q0.025(θ2|y) 0.06 0.45 0.68 1.02
Q0.975(θ1|y) 0.48 0.55 0.79 0.50
Q0.975(θ2|y) 0.18 0.23 0.23 0.38
Comparison of normalized mean absolute errors

True ODOF loc linear ridge Neural net
0.00.10.20.30.40.5
Var
~
(θ1|y)
True ODOF loc linear ridge Neural net
0.00.20.40.60.81.0
Var
~
(θ2|y)
Boxplot comparison of Var(θ1 | y), Var(θ2 | y) with the true
values, ODOF and usual ABC methods

Comments
ABC RF methods mostly insensitive both to strong correlations
between the summary statistics and to the presence of noisy
variables.
implies less number of simulations and no calibration
Next steps: adaptive schemes, deep learning, inclusion in DIYABC

random forests for ABC model choice and parameter estimation

More Related Content

What's hot (20)

Viewers also liked (9)

Similar to random forests for ABC model choice and parameter estimation (20)

More from Christian Robert (17)

Recently uploaded (20)

random forests for ABC model choice and parameter estimation