Testing for time-varying Granger causality

Christopher F Baum1 Jesús Otero2 Stan Hurn3

1 Boston College, Chestnut Hill, MA, USA

2 Universidad del Rosario, Bogotá, Colombia
3 Queensland University of Technology, Brisbane, Australia

2021 Stata Economics Virtual Symposium

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 1 / 52
Introduction

Causal relationships in the econometric analysis of time series are typically

based on the concept of predictability and are established by testing for
Granger causality (Granger, 1969, 1988).

The popularity of Granger causality stems from the fact that it is identified
using reduced-form VAR models, applicable to a set of potentially jointly
determined variables.

Advantages:
No need for normalization
No need for guidance from economic theory

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 2 / 52
Introduction

Some important studies applying Granger causality in economics:

Money and income (Friedman and Kuttner, 1993; Swanson, 1998;

Shi, Hurn, and Phillips, 2020),
GDP and energy consumption (Lee, 2006; Arora and Shi, 2016),
CO2 emissions and economic growth (Grossman and Krueger, 1995),
economic growth and quality of health (Tapia Granados and Ionides,
2008),
oil prices and output (Hamilton, 1983).

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 3 / 52
Introduction

Standard econometric software packages for the estimation/analysis of

VAR models provide Granger causality tests.

However, just as with other aspects of structural stability, Granger

causality may be supported over one time frame, but may be fragile when
alternative periods are considered (see Thoma, 1994; Swanson, 1998;
Psaradakis, Ravn, and Sola, 2005).

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 4 / 52
Introduction

Phillips, Wu, and Yu (2011) and Phillips, Shi, and Yu (2014, 2015a,b)
derive theoretical results for testing for and date-stamping asset price
bubbles.

Based on these results, Shi, Phillips, and Hurn (2018) and Shi, Hurn, and
Phillips (2020) revisit the notion of time variation in testing for Granger
causality. They establish that it is possible to assess the stability of causal
relationships over time.

Shi, Phillips, and Hurn (2018) study the stationary VAR model,
Shi, Hurn, and Phillips (2020) study the lag-augmented VAR
(LA-VAR) model, which allows for non-stationary variables.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 5 / 52
Introduction

Although conceptually straightforward, these stability analysis algorithms

offer significant challenges in terms of implementation in software.

These methods produce large numbers of test statistics which must

be efficiently stored and displayed for analysis.
Analysis of the statistical significance of test results requires
bootstrapping in order to ensure correct inference.

This presentation illustrates how the analysis can be accomplished using a

new community-contributed Stata command, tvgc.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 6 / 52
Introduction

Organization of the talk:

briefly present the Granger causal framework,

address recursive techniques for assessing time variation,
present details of the tvgc command,
provide an example focusing on key US macro series.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 7 / 52
Granger causality

Consider without loss of generality the bivariate VAR(m) model given by

m m
(1) (1) (1)
X X
y1t = φ0 + φ1k y1 t−k + φ2k y2 t−k + ε1t (1)
k=1 k=1
m m
(2) (2) (2)
X X
y2t = φ0 + φ1k y1 t−k + φ2k y2 t−k + ε2t , (2)
k=1 k=1

where y1t and y2t , respectively, represent economic time series of interest.
Variable y1 is said to Granger cause variable y2 if the past values of y1
have predictive power for the current value of y2 , conditional on the past
returns of y2 .

The null hypotheses of no Granger causality from y1 to y2 involves testing

(2)
the joint significance of φ1k (k = 1, · · · , m) by means of a Wald test.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 8 / 52
Granger causality
Recasting the bivariate VAR(m) in matrix notation:
yt = Πxt + εt . (3)
where
yt = [ y1t y2t ]0 ,
0
xt = [ 1 yt−1 0
yt−2 0
· · · yt−k ]0
Π2×(2m+1) = [ Φ0 Φ1 · · · Φm ]
with
(2) 0
Φ0 = [ φ(1)
0 φ0 ]
and " #
(1) (1)
φ1k φ2k
Φk = (2) (2) for k = 1, . . . , m.
φ1k φ2k
The null of no Granger causality from variable y1 to y2 is R1→2 π = 0,
where R1→2 is the coefficient restriction matrix and π = vec(Π)
using row vectorization.
Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 9 / 52
Granger causality

The heteroskedastic-consistent Wald statistic of the null hypothesis is

denoted by W1→2 and is defined as
h   i−1
W1→2 = T (R1→2 π̂)0 R1→2 V̂−1 Σ̂V̂−1 R01→2 (R1→2 π̂) ,

where V̂ = In ⊗ Q̂ and Q̂ = T −1 0 and Σ̂ = T −1 ˆ ˆ0

P P
t xt xt t ξt ξt with
ξˆt = ε̂t ⊗ xt and ε̂t = yt − Π̂xt .

The formulation of a test for Granger causality in a VAR(G ) system,

G = 2, . . . is straightforward.

This framework applies to testing for Granger causality in the context of a

VAR model estimated using stationary variables.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 10 / 52
Granger causality

To account for integrated variables, Toda and Yamamoto (1995) and

Dolado and Lütkepohl (1996) recommend estimating a LA-VAR model,
which is the original VAR(m) model augmented with d lags for the
possible maximum order of integration of the variables.

The resulting model is denoted VAR(m + d).

To test for Granger causality in the LA-VAR model, one proceeds just as
before.The coefficients associated to the additional d are not included in
the testing restrictions.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 11 / 52
Recursive testing algorithms

To allow for time variation in Granger causal orderings and to date-stamp

the timing of the changes, recursive estimation methods are required.

There are three algorithms that generate a sequence of test statistics:

the forward expanding (FE) window,

the rolling (RO) window,
the recursive evolving (RE) window.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 12 / 52
Recursive testing algorithms

Consider a sample of T + 1 observations {y0 , y1 , · · · , yT }, a number r such

that 0 < r < 1 and consider [Tr ] to denote the integer part of the product.

Then Tr1 ,r will be taken to denote a Wald test statistic computed over a
subsample starting at y[Tr1 ] and ending at y[Tr ] .

A schematic representation of the algorithms is given in what follows.

Each of the arrows is representative of a subsample over which the
relevant test statistic is computed.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 13 / 52
The forward expanding window

The FE algorithm (Thoma, 1994) is a standard forward recursion.

The Wald test statistic is computed first for a minimum window length,
τ0 = [Tr0 ] > 0, and the sample size then expands sequentially by one
observation until the final test statistic is computed using the entire
sample. The starting point of every subsample is the first data point.

This is what Stata’s prefix command rolling: produces with the

recursive option.

At the conclusion of the FE algorithm, a sequence of Wald test statistics,

Tr1 ,r with r1 = 0 and r ∈ [r0 , 1], is obtained.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 14 / 52
The forward expanding window

Sample interval [1,T ]

1 T
τ0 = [Tr0 ]

(a) Forward expanding window

Figure 1: Forward expanding window (based on Phillips, Shi, and Yu, 2015a)

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 15 / 52
The rolling window

In the RO algorithm (Swanson, 1998; Arora and Shi, 2016) a window of

size [Tw ] is rolled through the sample advancing one observation at a time
and a Wald test statistic is computed for each window.

This is what Stata’s prefix command rolling: produces by default.

The output from the RO algorithm is a sequence of test statistics Tr1 ,r

with r1 = r − w and r ∈ [r0 , 1], where each test statistic is computed from
a sample of the same size, [Tw ], with 0 < w < 1.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 16 / 52
The rolling window

Sample interval [1,T ]

1 T
τ0 = [Tr0 ]

τ0 = [Tr0 ]

τ0 = [Tr0 ]

(a) Rolling window

Figure 2: Rolling window (based on Phillips, Shi, and Yu, 2015a)

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 17 / 52
The recursive evolving window

In the RE algorithm, for a given observation of interest, the algorithm

computes a test statistic for every possible subsample of size r0 or larger
with the observation of interest providing the common end point of all the
subsamples.

This procedure is repeated taking the observation of interest to be every

point in the sample, subject only to the minimum window size. Thus every
observation in the sample byeond the first is associated with it a set of
Wald test statistics. Phillips, Shi, and Yu (2015b) propose that inference
be based on a sequence of supremum norms of these statistics.

The RE algorithm produces a sequence of test statistics Tr1 ,r with

r1 ∈ [0, r − r0 ] and r ∈ [r0 , 1] which are the sup norms of the Wald
statistics at each observation.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 18 / 52
The recursive evolving window

The RE algorithm encompasses both the FE and RO recursions as special

cases.

For each observation in turn, a sequence of test statistics is defined that

can be arranged in an upper triangular square matrix with column and row
dimensions equal to the largest number of usable observations.

the FE Wald statistic is the leading entry in each column,

the RO Wald statistic is located on the main diagonal,
the largest elements of each column are the RE statistics.

The information derived from these test statistics can be used over the full
sample or analyzed through the period in order to focus on the timing of
these time-varying phenomena via date stamping.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 19 / 52
The recursive evolving window

Sample interval [1,T ]

1 T
τ0 = [Tr0 ]

τ0 = [Tr0 ]

τ0 = [Tr0 ]

(a) Recursive evolving window

Figure 3: Sample sequences and window widths (based on Phillips, Shi, and Yu,
2015a)

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 20 / 52
Full sample analysis

If the null hypothesis of interest is whether a particular variable does not

Granger causes another variable at any time during the sample, with the
alternative that there is Granger causality at some time, a single test
statistic is required.

The maximal FE statistic is the largest element of the first row of the
upper triangular matrix of test statistics.
The maximal RO statistic is the largest element of the main diagonal
of the matrix.
The maximal RE statistic is the largest element of the entire upper
triangular matrix.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 21 / 52
Date stamping

Beyond these summary measures for the full sample, the sequence of FE,
RO and RE statistics can be graphed and compared with the bootstrap
percentiles derived by methods described in Shi, Hurn, and Phillips (2020),
section 3 and Shi, Phillips, and Hurn (2018), section 4.1.

These estimates can then be used to identify periods in which the

potential Granger-casual relationships vary significantly.

The estimated origination date of a change is determined as the first

instance at which the test statistic exceeds its critical value. Subsequent
changes are then identified in a similar fashion.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 22 / 52
The tvgc command
Syntax

The tvgc command tests whether the first variable in the varlist is
Granger-caused by the remaining variables.

Before using the tvgc command, it is necessary to tsset the data. The
varlist cannot contain gaps, but can contain time-series operators. tvgc
does not support the by: prefix. It may be applied to one unit of a panel
dataset.

tvgc varlist [if ] [in] [, prefix(string) p(integer ) d(string)

robust trend matrix window(integer ) boot(integer ) seed(integer )
sizecontrol(integer ) graph eps pdf notitle restab]

Ben Jann’s moremata package is required: ssc install moremata for

the latest version.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 23 / 52
The tvgc command
Options

The tvgc command supports the following options:

prefix can be used to provide a ‘stub’ with which variables created

in tvgc will be named. If this option is given, three Stata variables
will be created for the appropriate range of dates:
prefix forward varname, prefix rolling varname, and
prefix recursive varname. These variables contain the sequences of
these three test statistics over the sample period. The prefix option
must be specified to enable the graph option.
p sets the number of lags to be included in the VAR model, with a
default of 2.
d sets the number of lags to be included in the lag-augmented part of
the VAR model, with a default of 1. This option must be used when
there are integrated variables in the varlist. Set d=0 if no
augmented lags are needed.
Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 24 / 52
The tvgc command
Options

robust specifies that heteroskedasticity-robust test statistics should

be computed.

trend specifies that a linear trend should be included in the VAR.

matrix specifies that the T x T matrices of test statistics should be

returned. They are named r rhsvar for each of the test variables.

window specifies the number of observations to be included in the

rolling windows. If not specified, 20% of the sample is used.

boot sets the number of replications to perform the bootstrap

advocated by Phillips and Shi (2020). The default is 199.

seed sets the initial seed for random number generation in

bootstrapping.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 25 / 52
The tvgc command
Options

sizecontrol specifies the number of observations to be included in

the bootstrap computations in order to control the empirical size,
with a default of 12.

graph specifies that the timeseries of the three test statistics should
be graphed along with their 90% and 95% critical values. The graphs
will be saved with names specified by the prefix() option as
prefix forward, prefix rolling and prefix rolling. The eps and
pdf options specify the format in which the graphs are saved. The
notitle option suppresses the graph titles.

restab specifies that a LATEX table containing the test statistics and
their 95th and 99th percentile values should be written to
restab.tex. The file will be replaced if it exists. When including this
fragment in a LATEX document, use the booktabs package.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 26 / 52
The tvgc command
Implementation

Although reference is made to the rolling-window algorithms provided by

Stata’s rolling: prefix, all computations are performed in Mata rather
than the ado-file language in order to produce results with acceptable
speed. Even without bootstrapping, computation of the upper triangular
matrix of test statistics requires (T )(T − 1)/2 estimates of the VAR or
LA-VAR model for each right-hand variable where T is the number of
usable observations.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 27 / 52
Empirical application
The tvgc command is illustrated using a three-variable VAR specification
for monthly US data.

logarithm of industrial production (ln i),

unemployment rate (u),
logarithm of the price of crude oil (ln o).

These variables are a subset of those used by Hamilton (1983) to study

the relationship between oil and the macroeconomy using Granger
causality tests.

The sample period is January 1959 to December 2019, with 732

observations accessed from FRED. The task of downloading these series is
simplified with the Stata command freduse; see Drukker (2006). The
Stata dataset used in this illustration can be downloaded using the
command bcuse us outoil; install bcuse from SSC if needed.
Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 28 / 52
Empirical application
5.0 12.0

4.5 10.0

8.0
4.0

6.0
3.5

4.0
3.0
1959 1965 1971 1977 1983 1989 1995 2001 2007 2013 2019 1959 1965 1971 1977 1983 1989 1995 2001 2007 2013 2019

(a) ln i (b) u
5.0

4.0

3.0

2.0

1.0
1959 1965 1971 1977 1983 1989 1995 2001 2007 2013 2019

(c) ln o

Figure 4: Variables in levels

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 29 / 52
Empirical application

It is apparent that both ln i and ln o are trending.

The order of integration of the variables is assessed with the ADFmax and
DFGLS unit root tests of Leybourne (1995) and Elliott, Rothenberg, and
Stock (1996).

Using the Stata commands adfmaxur (Otero and Baum, 2018) and ersur
(Otero and Baum, 2017), the results suggest:
presence of a unit root in ln i and ln o,
stationarity of the unemployment rate

As there are I(1) variables in the VAR model, our analysis proceeds in the
context of a LA-VAR model where d = 1.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 30 / 52
Empirical application

The optimal lag order of the VAR model is chosen using the Stata
command varsoc.

The command is applied to a model that includes a linear trend, which

enters as an exogenous variable. The maximum number of lags is set to 12
as the data are monthly.

The Schwartz lag-order selection statistic recommends p = 2 lags, while

Akaike favors p = 6 lags. The more parsimonious choice of p = 2 is
adopted here.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 31 / 52
Empirical application

GC ?
Letting x −−→ y to denote that the direction of Granger causality being
tested runs from x to y , the following relationships are tested:

GC ? GC ?
u −−→ ln i and ln o −−→ ln i;

GC ? GC ?
ln i −−→ u and ln o −−→ u;

GC ? GC ?
ln i −−→ ln o and u −−→ ln o.

The required commands are respectively:

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 32 / 52
Empirical application

. tvgc li u lo, trend win(72) sizecontrol(12) p(2) d(1) robust

. tvgc u li lo, trend win(72) sizecontrol(12) p(2) d(1) robust

. tvgc lo li u, trend win(72) sizecontrol(12) p(2) d(1) robust

In all three cases the chosen options indicate

presence of a linear trend as an exogenous variable, trend,
initial estimation window of 72 observations, win(72),
size of the tests controlled over a one-year period, sizecontrol(12)
two lags in the VAR, p(2),
one lag in the lag-augmented part of the VAR, d(1),
tests robust to heteroskedasticity, robust.

To produce the output shown here, the options boot(499),

seed(123), prefix, graph and eps were also used.
Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 33 / 52
Empirical application

The results for the full sample show that we fail to reject the null
hypothesis of no Granger causality from income and unemployment to the
price of oil when applying the FE window.

In all other cases there is evidence of Granger causality when the test
statistics are compared with the 95th percentiles of the empirical
distribution of the bootstrap test statistics.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 34 / 52
Empirical application

Table 1: Wald tests of Granger causality

Direction of causality Max Wald FE Max Wald RO Max Wald RE

GC ?
u−
−−→ ln i 20.524 31.073 38.806
(10.283) (10.355) (10.775)
[15.751] [15.110] [16.131]
GC ?
ln o −
−−→ ln i 12.037 28.322 31.689
(8.709) (8.970) (9.324)
[12.459] [13.526] [14.389]
GC ?
ln i −
−−→u 70.205 68.762 75.290
(10.360) (10.469) (10.544)
[15.850] [17.025] [17.892]
GC ?
ln o −
−−→u 46.355 42.252 64.877
(9.673) (9.807) (10.118)
[13.607] [13.607] [13.962]
GC ?
ln i −
−−→ ln o 4.349 25.639 30.328
(7.913) (8.565) (9.344)
[14.964] [14.956] [14.964]
GC ?
u−
−−→ ln o 3.440 17.229 17.253
(9.333) (9.333) (10.121)
[15.417] [14.775] [15.417]

The 95th and 99th percentiles of the empirical distribution of the bootstrap statistics are in parentheses and brackets,
respectively.
Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 35 / 52
Empirical application

Figures 5 to 10 display the time-varying Granger causality test results. In

general, these plots all support the conclusion that Granger causal
relationships are extremely dynamic and that the patterns of causation
found in the data depend on the type of recursive algorithm used.

These plots display the 90th and 95th percentiles of the empirical
distribution of the bootstrap statistics, to be compared with the sequence
of FE, RO, and RE test statistics.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 36 / 52
Empirical application
20 30

15

20

10

10

0 0
1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1 1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1

GC ? GC ?
(a) Forward: u −−→ ln i (b) Rolling: u −−→ ln i

40

30

20

10

0
1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1

GC ?
(c) Recursive: u −−→ ln i

Figure 5: Time-varying Granger causality tests between ln i and u

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 37 / 52
Empirical application
80 80

60 60

40 40

20 20

0 0
1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1 1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1

GC ? GC ?
(a) Forward: ln i −−→ u (b) Rolling: ln i −−→ u

80

60

40

20

0
1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1

GC ?
(c) Recursive: ln i −−→ u

Figure 6: Time-varying Granger causality tests between ln i and u

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 38 / 52
Empirical application
8
25

20
6

15

10

2
5

0 0
1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1 1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1

GC ? GC ?
(a) Forward: ln i −−→ ln o (b) Rolling: ln i −−→ ln o

30

20

10

0
1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1

GC ?
(c) Recursive: ln i −−→ ln o

Figure 7: Time-varying Granger causality tests between ln i and ln o

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 39 / 52
Empirical application
15 30

10 20

5 10

0 0
1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1 1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1

GC ? GC ?
(a) Forward: ln o −−→ ln i (b) Rolling: ln o −−→ ln i

30

20

10

0
1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1

GC ?
(c) Recursive: ln o −−→ ln i

Figure 8: Time-varying Granger causality tests between ln i and ln o

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 40 / 52
Empirical application
10 20

8
15

10

5
2

0 0
1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1 1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1

GC ? GC ?
(a) Forward: u −−→ ln o (b) Rolling: u −−→ ln o

20

15

10

0
1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1

GC ?
(c) Recursive: u −−→ ln o

Figure 9: Time-varying Granger causality tests between u and ln o

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 41 / 52
Empirical application
50
40

40

30

30

20

20

10
10

0 0
1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1 1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1

GC ? GC ?
(a) Forward: ln o −−→ u (b) Rolling: ln o −−→ u

60

40

20

0
1960m1 1970m1 1980m1 1990m1 2000m1 2010m1 2020m1

GC ?
(c) Recursive: ln o −−→ u

Figure 10: Time-varying Granger causality tests between u and ln o

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 42 / 52
Summary of findings

Estimation using the FE and RE windows indicate that during most of the
study period there is evidence of Granger causality from unemployment to
income; see Figures 5(a) and 5(c), and vice versa, see Figures 6(a) and
6(c). These results provide strong support for the intuition that these two
measures of economic activity are closely related.
FE estimation also shows that the price of oil Granger-causes income in
the late 1960s and early 1970s; see Figure 8(a).

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 43 / 52
Summary of findings

In contrast, strong evidence of Granger causality from income to the price

of oil is apparent in the 1980s and 2000s with the RO and RE windows:
see Figures 7(b) and 7(c), respectively.

The fact that the FE window fails to pick up the opening of this causal
channel confirms a well-known problem with the FE algorithm: namely,
that it is not sensitive to changes late in the sample period.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 44 / 52
Summary of findings

A particularly strong illustration of the effects of the first oil shock of

October 1973 is to be found in Figures 10(a), 10(b) and 10(c) showing
Granger causality from the oil price to unemployment.All the algorithms
identify a period of strong causality which starts at the time of the first oil
shock and lasts until the second oil shock in 1979.

Interestingly, although the causal channel from the oil price to

unemployment is active at times in the latter half of the sample period,
the channel is not evident during the great recession of 2008–2009.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 45 / 52
Concluding remarks

Evaluation of Granger-causal relationships among macroeconomic

aggregates is an important component of macroeconometric modeling, as
it is key to formally assessing the temporal stability of these relationships.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 46 / 52
Concluding remarks

Evaluation of Granger-causal relationships among macroeconomic

aggregates is an important component of macroeconometric modeling, as
it is key to formally assessing the temporal stability of these relationships.
We present the tvgc command to test for time-varying Granger causality.
The command produces full-sample test statistics as well as date-stamping
of the periods in which there are significant findings of Granger-causal
relationships.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 46 / 52
Concluding remarks

Evaluation of Granger-causal relationships among macroeconomic

aggregates is an important component of macroeconometric modeling, as
it is key to formally assessing the temporal stability of these relationships.
We present the tvgc command to test for time-varying Granger causality.
The command produces full-sample test statistics as well as date-stamping
of the periods in which there are significant findings of Granger-causal
relationships.
Using US monthly data on industrial production, unemployment and oil
prices, we find support for the conclusion that causal relationships can
change dramatically over any given sample period.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 46 / 52
Concluding remarks

Evaluation of Granger-causal relationships among macroeconomic

aggregates is an important component of macroeconometric modeling, as
it is key to formally assessing the temporal stability of these relationships.
We present the tvgc command to test for time-varying Granger causality.
The command produces full-sample test statistics as well as date-stamping
of the periods in which there are significant findings of Granger-causal
relationships.
Using US monthly data on industrial production, unemployment and oil
prices, we find support for the conclusion that causal relationships can
change dramatically over any given sample period.
Arbitrarily choosing the sample period over which to conduct causality
tests may lead to misleading inference compared with a strategy that
allows data-driven identification of change points.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 46 / 52
Concluding remarks

An extended version of this presentation, with additional details on

inference, is available from Baum, Hurn, and Otero (2021).

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 47 / 52
References I

Arora, V. and S. Shi (2016). Energy consumption and economic growth in

the United States. Applied Economics 48 (39), 3763–3773.
Baum, C. F., S. Hurn, and J. Otero (2021). The dynamics of U.S.
industrial production: A time-varying Granger causality perspective.
Econometrics and Statistics, in press.
Dolado, J. J. and H. Lütkepohl (1996). Making Wald tests work for
cointegrated VAR systems. Econometric Reviews 15 (4), 369–386.
Drukker, D. M. (2006). Importing Federal Reserve economic data. The
Stata Journal 6 (3), 384–386.
Elliott, G., T. J. Rothenberg, and J. H. Stock (1996). Efficient tests for an
autoregressive unit root. Econometrica 64 (4), 813–836.
Friedman, B. M. and K. N. Kuttner (1993). Another look at the evidence
on money-income causality. Journal of Econometrics 57 (1-3), 189–203.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 48 / 52
References II

Granger, C. W. J. (1969). Investigating causal relations by econometric

models and cross-spectral methods. Econometrica 37, 424–438.
Granger, C. W. J. (1988). Some recent development in a concept of
causality. Journal of Econometrics 39 (1-2), 199–211.
Grossman, G. M. and A. B. Krueger (1995). Economic growth and the
environment. Quarterly Journal of Economics 110 (2), 353–377.
Hamilton, J. D. (1983). Oil and the macroeconomy since World War II.
Journal of Political Economy 91 (2), 228–248.
Lee, C.-C. (2006). The causality relationship between energy consumption
and GDP in G-11 countries revisited. Energy Policy 34 (9), 1086–1093.
Leybourne, S. (1995). Testing for unit roots using forward and reverse
Dickey-Fuller regressions. Oxford Bulletin of Economics and
Statistics 57 (4), 559–571.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 49 / 52
References III
Otero, J. and C. F. Baum (2017). Response surface models for the Elliott,
Rothenberg, and Stock unit-root test. The Stata Journal 17 (4),
985–1002.
Otero, J. and C. F. Baum (2018). Unit-root tests based on forward and
reverse Dickey-Fuller regressions. The Stata Journal 18 (1), 22–28.
Phillips, P. C. B. and S. Shi (2020). Real time monitoring of asset
markets: bubbles and crises. In H. D. Vinod and C. R. Rao (Eds.),
Handbook of Statistics: Financial, Macro and Micro Econometrics
Using R, Volume 42, pp. 61–80. Amsterdam: Elsevier.
Phillips, P. C. B., S. Shi, and J. Yu (2014). Specification sensitivity in
right-tailed unit root testing for explosive behaviour. Oxford Bulletin of
Economics and Statistics 76 (3), 315–333.
Phillips, P. C. B., S. Shi, and J. Yu (2015a). Testing for multiple bubbles:
Historical episodes of exuberance and collapse in the S&P 500.
International Economic Review 56 (4), 1043–1077.
Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 50 / 52
References IV

Phillips, P. C. B., S. Shi, and J. Yu (2015b). Testing for multiple bubbles:

Limit theory ofreal-time detectors. International Economic
Review 56 (4), 1079–1134.
Phillips, P. C. B., Y. Wu, and J. Yu (2011). Explosive behavior in the
1990s NASDAQ: When did exuberance escalate asset values?
International Economic Review 52 (1), 201–226.
Psaradakis, Z., M. O. Ravn, and M. Sola (2005). Markov switching
causality and the money-output relationship. Journal of Applied
Econometrics 20 (5), 665–683.
Shi, S., S. Hurn, and P. C. B. Phillips (2020). Causal change detection in
possibly integrated systems: Revisiting the money-income relationship.
Journal of Financial Econometrics 18 (1), 158–180.
Shi, S., P. C. B. Phillips, and S. Hurn (2018). Change detection and the
causal impact of the yield curve. Journal of Time Series Analysis 39 (6),
966–987.
Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 51 / 52
References V

Swanson, N. R. (1998). Money and output viewed through a rolling

window. Journal of Monetary Economics 41 (3), 455–474.
Tapia Granados, J. A. and E. L. Ionides (2008). The reversal of the
relation between economic growth and health progress: Sweden in the
19th and 20th centuries. Journal of Health Economics 27 (3), 544–563.
Thoma, M. A. (1994). Subsample instability and asymmetries in
money-income causality. Journal of Econometrics 64 (1-2), 279–306.
Toda, H. Y. and T. Yamamoto (1995). Statistical inference in vector
autoregressions with possibly integrated processes. Journal of
Econometrics 66 (1-2), 225–250.

Baum, Otero, Hurn Testing for time-varying Granger causality 2021 Stata Symposium 52 / 52