ECON-C4210 - Econometrics II: Capstone

Lecture 10: Time series III

Otto Toivanen

Toivanen ECON-C4210 Lecture 10 1 / 37

Learning outcomes
• At the end of lecture 10, you know

1 what dynamic causal effects are

2 what the assumptions behind such a model are

3 how to estimate a dynamic causal model

4 how to interpet the results of a dynamic causal model

5 what Heteroskedasticity and Autoregression Consistent (HAC) standard errors are

6 what a Vector Autoregressive Regression (VAR) model is

7 how to estimate a VAR

8 how to estimate a VAR how to and interpret the results of a VAR model

Toivanen ECON-C4210 Lecture 10 2 / 37

Dynamic causal effect

• Think of macroeconomic data and questions.

• Example: The effect of a change in the ECB interest rate on inflation next month, in 6
months, in 12, in 24,... months.
• Often, we only have one ”subject”, say the country affected by ECB decisions.

• Difficult to think of an RCT.

Toivanen ECON-C4210 Lecture 10 3 / 37

Dynamic causal effect

• Substitute thought experiment: different treatments to the same (subject / observation

unit) at different points in time.
• Treatment = e.g. level of interest rate t periods ago.

• If the different times are drawn from the same distribution (i.e., Yy , Xt is stationary),
then we can estimate the dynamic causal effects from an OLS regression of Yt on Xt and
its lags.
• Such an estimator is called a distributed lag estimator.

Toivanen ECON-C4210 Lecture 10 4 / 37

The distributed lag model

Yt = β0 + β1 Xt + β2 Xt−1 ... + βt−p+1 Xt−p + ut

• Interpretation of β1 : The impact effect of change in Xt

• Notice that we are keeping the past constant!
• β2 : The 1-period dynamic multiplier. The effect of a change in Xt−1 on Yt , keeping
Xt , Xt−2 ,... Xt−p constant.
• β3 : The 2-period dynamic multiplier. The effect of a change in Xt−2 on Yt , keeping
Xt , Xt−1 , Xt−3 ... Xt−p constant.
Pj−1
• j-period cumulative dynamic multiplier = k=1 βk

• Example: 3-period cumulative dynamic multiplier = β1 + β2 + β3 + β4

...that is, the effect of all Xj for j = t, t − 1, t − 2, t − 3.

Toivanen ECON-C4210 Lecture 10 5 / 37
The distributed lag model: assumptions

Yt = β0 + β1 Xt + β2 Xt−1 ... + βt−p+1 Xt−p + ut

• Exogeneity: E[ut |Xt , Xt−1 , ...] = 0

• Strict exogeneity: E[ut |..., Xt+1 , Xt , Xt−1 , ...] = 0

• Notice that if X is strictly exogenous, then it is also exogenous.

• We will proceed with assuming exogeneity.

Toivanen ECON-C4210 Lecture 10 6 / 37

The distributed lag model: assumptions

A1 E[ut |Xt , Xt−1 , ...] = 0

A2 i) Y , X are stationary and ii) (Yt , Xt ) and (Yt−j , Xt−j ) become independent as j grows
large.

A3 (Yt , Xt ) have 8 nonzero finite moments; large outliers are unlikely.

A4 There is no perfect multicollinearity.

Toivanen ECON-C4210 Lecture 10 7 / 37

The distributed lag model: assumptions

• Assumptions 1 and 4 are familiar.

• Assumption 2 is different. The difference is due to us studying time series.

• Assumption 3 is familiar otherwise, but now need 8 finite moments. This is to ensure
”good” (=HAC) standard errors.

Toivanen ECON-C4210 Lecture 10 8 / 37

The distributed lag model: assumptions

• Assumption 2.i yields

1 internal validity
2 external validity
3 This is the time series equivalent of the ”identically distributed” of the cross-section i.i.d
assumption

• Assumption 2.ii
1 enables the use of a version of CLT
2 This is the time series equivalent of the ”independently distributed” of the cross-section i.i.d
assumption

• The intuition is that when the time periods are sufficiently separated, we have
independent experiments.

Toivanen ECON-C4210 Lecture 10 9 / 37

Heteroskedasticity and Autocorrelation Consistent (HAC) standard errors

• Under the above assumptions a distributed lag model, estimated via OLS, yields
consistent estimates of the coefficients βk .
• What about the standard errors? The variance of the sampling distribution is different as
Xt and ut may be autocorrelated (think of what the equivalent would mean in a
cross-section context).
• → need standard errors that are robust not only to heteroskedasticity but also to
autocorrelation.

Toivanen ECON-C4210 Lecture 10 10 / 37

Heteroskedasticity and Autocorrelation Consistent (HAC) standard errors

• In (well-behaving) cross-section data, the variance of β1 in a univariate regression is

σu2
var (βˆ1 ) =
T (σX2 )2

• Let’s consider a time-series with T = 2, with cov (ut , ut−1 ) 6= 0:

T
!
1 X
var ut =var [0.5(u1 + u2 )]
T t=1
=0.25[var (u1 ) + var (u2 ) + 2cov (u1 , u2 )]
=0.5(σu2 + ρ)

• Autocorrelation means that the usual variance formula for the error term does not apply.

Toivanen ECON-C4210 Lecture 10 11 / 37

Example: A (sort-of) Phillips curve

inflt = β0 + β1 δUEt + β2 δUEt−1 + ut

inflt = β0 + β1 δUEt + ut

Toivanen ECON-C4210 Lecture 10 12 / 37

Examples: A (sort-of) Phillips curve

Stata code
1 regr infl ue if time ind >= 13 & year <= 1965, robust estimates store est50s

Toivanen ECON-C4210 Lecture 10 13 / 37

Examples: A (sort-of) Phillips curve

Stata code
1 newey infl ue if time ind >= 13 & year <= 1965, lag(0) estimates store est50s

Toivanen ECON-C4210 Lecture 10 14 / 37

Sort-of Phillips curve
Table: Static Phillips curve

1949-1965 1966-1980 1981-2000 2000-2024m2

UE rate -0.000403∗ 0.000500∗∗ 0.000174 -0.000253∗
(0.000174) (0.000158) (0.000102) (0.000118)

Constant 0.00353∗∗∗ 0.00276∗∗ 0.00183∗∗ 0.00357∗∗∗

(0.000875) (0.000918) (0.000665) (0.000728)
Newey-West standard errors
1949-1965 1966-1980 1981-2000 2000-2024m2
UE rate -0.000403∗ 0.000500∗∗∗ 0.000174 -0.000253∗
(0.000175) (0.000124) (0.000126) (0.000111)

Constant 0.00353∗∗∗ 0.00276∗∗∗ 0.00183∗ 0.00357∗∗∗

(0.000963) (0.000702) (0.000772) (0.000685)
Observations 192 180 240 278
Toivanen ECON-C4210 Lecture 10 15 / 37
Sort-of Phillips curve
Table: Static Phillips curve

1949-1965 1966-1980 1981-2000 2000-2024m2

UE rate -0.000901 0.00238 0.000203 -0.00103∗∗∗
(0.000840) (0.00127) (0.000961) (0.000190)

L.UE rate 0.000508 -0.00190 -0.0000296 0.000827∗∗∗

(0.000837) (0.00127) (0.000965) (0.000183)
Newey-West standard errors
1949-1965 1966-1980 1981-2000 2000-2024m2
UE rate -0.000901 0.00238 0.000203 -0.00103∗∗∗
(0.00120) (0.00127) (0.00108) (0.000203)

L.UE rate 0.000508 -0.00190 -0.0000296 0.000827∗∗∗

(0.00114) (0.00126) (0.00105) (0.000193)
Toivanen ECON-C4210 Lecture 10 16 / 37
Estimation of the distributed lag model

Yt = β0 + β1 Xt + β2 Xt−1 ... + βt−p+1 Xt−p + ut

• You can estimate the model with OLS.

• You directly get the impact effect of X and the k-period dynamic multipliers & their
standard errors
• You do not get the cumulative multipliers directly, nor their standard errors.

Toivanen ECON-C4210 Lecture 10 17 / 37

Estimating the cumulative dynamic multipliers

Yt = β0 + β1 Xt + β2 Xt−1 ... + βt−p+1 Xt−p + ut

• Option 1: you estimate the above specification and then use post-estimation commands
to calculate the cumulative dynamic multipliers and their standard errors.

Toivanen ECON-C4210 Lecture 10 18 / 37

Cumulative dynamic multiplier

. newey infl ue L.ue, lag(3)

Regression with Newey–West standard errors Number of obs = 901

Maximum lag = 3 F( 2, 898) = 6.92
Prob > F = 0.0010

Newey–West
infl Coefficient std. err. t P>|t| [95% conf. interval]

ue
--. -.0007707 .0002407 -3.20 0.001 -.0012431 -.0002983
L1. .0008116 .0002206 3.68 0.000 .0003787 .0012445

_cons .0026148 .0005389 4.85 0.000 .0015571 .0036724

. scalar cdm = _b[ue] + _b[L.ue]

. scalar list cdm

cdm = .00004094

. testnl _b[ue] + _b[L.ue] = 0

(1) _b[ue] + _b[L.ue] = 0

chi2(1) = 0.20
Prob > chi2 = 0.6534

Toivanen ECON-C4210 Lecture 10 19 / 37

Estimating the cumulative dynamic multipliers

• Option 2: Transform the model to yield direct estimates of the cumulative dynamic
multipliers.
• Example: A 1-lag distributed lag model:

Yt =β0 + β1 Xt + β2 Xt−1 + ut
=β0 + β1 Xt − β1 Xt−1 + β2 Xt−1 + β1 Xt−1 + ut
=β0 + β1 (Xt − Xt−1 ) + (β2 + β1 )Xt−1 + ut
=β0 + β1 ∆Xt + (β2 + β1 )Xt−1 + ut

• Same trick works for higher order distributed lag models.

Toivanen ECON-C4210 Lecture 10 20 / 37

Cumulative dynamic multiplier

. gen diff = ue - L.ue

(1 missing value generated)

. newey infl diff L.ue, lag(3)

Regression with Newey–West standard errors Number of obs = 901

Maximum lag = 3 F( 2, 898) = 6.92
Prob > F = 0.0010

Newey–West
infl Coefficient std. err. t P>|t| [95% conf. interval]

diff -.0007707 .0002407 -3.20 0.001 -.0012431 -.0002983

ue
L1. .0000409 .0000912 0.45 0.653 -.000138 .0002199

_cons .0026148 .0005389 4.85 0.000 .0015571 .0036724

Toivanen ECON-C4210 Lecture 10 21 / 37

Distributed lag model with autocorrelation

Yt = β0 + β1 Xt + β2 Xt−1 ... + βt−p+1 Xt−p + ut

• Now assume that

ut = ũt + φut−1
in other words, ut is autocorrelated (AR(1)).
• Besides using HAC standard errors, we could consider transforming the regression
equation so that the error term in the transformed equation is i.i.d.
• As an example, let’s work with the following distributed lag model:
Yt = β0 + β1 Xt + β2 Xt−1 + ut

Toivanen ECON-C4210 Lecture 10 22 / 37

Estimating the cumulative dynamic multipliers

• We can proceed as follows:

Yt =β0 + β1 Xt + β2 Xt−1 + ut
Yt − φYt−1 =β0 + β1 Xt + β2 Xt−1 + ut − φYt−1
=β0 + β1 Xt + β2 Xt−1 + ut
− φ (β0 + β1 Xt−1 + β2 Xt−2 + ut−1 )
| {z }
=Yt−1

=(β0 − φβ0 ) + β1 Xt + (β2 − φβ1 )Xt−1 − φβ2 Xt−2 + ũt

+ φut−1 − φut−1
Yt =δ0 + φYt−1 + δ1 Xt + δ2 Xt−1 + δ3 Xt−2 + ũt

• The outcome is an ADL(1,2) model with i.i.d error term.

Toivanen ECON-C4210 Lecture 10 23 / 37

Vector autoregressive (VAR) models

• What if you want to model two or more series simultaneously?

• What if they interact, i.e., the past values of Y affect current value of X and vice versa?

• VAR models are designed for these situations. Below an example:

Yt =β10 + β11 Yt−1 + β12 Yt−2 + γ11 Xt−1 + γ12 Xt−2 + u1t
Xt =β20 + β21 Xt−1 + β22 Xt−2 + γ21 Yt−1 + γ22 Yt−2 + u2t

Toivanen ECON-C4210 Lecture 10 24 / 37

Vector autoregressive (VAR) models

• Without any constraints on the coefficients, one could estimate the VAR equation by
equation.
• Often, theory suggests some constraints on the coefficients, e.g., in macroeconomic
models.
• Example of VAR: US Phillips curve.

• We use data from 2000m1 - 2015m12.

Toivanen ECON-C4210 Lecture 10 25 / 37

Plotting the Phillips curve

Stata code
1 twoway s c a t t e r i n f l ue i f y e a r <= 1965 | | s c a t t e r i n f l ue i f y e a r > 1960 & y e a r <= 1980 | | ///
2 s c a t t e r i n f l ue i f y e a r > 1980 & y e a r <= 2000 | | s c a t t e r i n f l ue i f y e a r > 2000 , ///
3 t i t l e ( ”US P h i l l i p s c u r v e ” ) ///
4 s u b t i t l e ( ” 1949m1 − 2024m2” ) ///
5 l e g e n d ( l a b ( 1 ” 1949 − 1965 ” ) l a b ( 2 ” 1966 − 1980 ” ) l a b ( 3 ” 1981 − 2000 ” ) l a b ( 4 ” 2001− ” 2024m2 ) ) ///
6 graphregion ( color ( white )) bgcolor ( white )
7
8 graph expor t ” L I I 1 0 4 p h i l l i p s . pdf ” , r e p l a c e

Toivanen ECON-C4210 Lecture 10 26 / 37

Plotting the Phillips curve

US Phillips curve
1949m1 - 2024m2
.02

.01

1949 - 1965
1966 - 1980
inflation

0 1981 - 2000
2001-
2024m2

-.01

-.02
0 5 10 15
UE rate

Toivanen ECON-C4210 Lecture 10 27 / 37

Inflation
.02

.01
inflation

-.01

-.02
0 200 400 600 800 1000
time

Toivanen ECON-C4210 Lecture 10 28 / 37

Unemployment
15

10
UE rate

0
0 200 400 600 800 1000
time

Toivanen ECON-C4210 Lecture 10 29 / 37

Change in unemployment
10

5
UE rate, D

-5
0 200 400 600 800 1000
time

Toivanen ECON-C4210 Lecture 10 30 / 37

VAR on inflation and UE

Stata code
1
2 forvalues t = 1/10{
3 dfuller ue i f t i m e i n d >= 1 3 , l a g s ( ‘ t ’ )
4 }
5 forvalues t = 1/10{
6 dfuller d . ue i f t i m e i n d >= 1 3 , l a g s ( ‘ t ’ )
7 }
8 forvalues t = 1/10{
9 dfuller i n f l i f t i m e i n d >= 1 3 , l a g s ( ‘ t ’ )
10 }

Toivanen ECON-C4210 Lecture 10 31 / 37

DF tests

. forvalues t = 1/10{
2. dfuller d.ue if time_ind >= 13, lags(`t')
3. }

Augmented Dickey–Fuller test for unit root

Variable: D.ue Number of obs = 890

Number of lags = 1

H0: Random walk without drift, d = 0

Dickey–Fuller
Test critical value
statistic 1% 5% 10%

Z(t) -21.738 -3.430 -2.860 -2.570

MacKinnon approximate p-value for Z(t) = 0.0000.

Toivanen ECON-C4210 Lecture 10 32 / 37

DF tests

. forvalues t = 1/10{
2. dfuller infl if time_ind >= 13, lags(`t')
3. }

Augmented Dickey–Fuller test for unit root

Variable: infl Number of obs = 890

Number of lags = 1

H0: Random walk without drift, d = 0

Dickey–Fuller
Test critical value
statistic 1% 5% 10%

Z(t) -12.934 -3.430 -2.860 -2.570

MacKinnon approximate p-value for Z(t) = 0.0000.

Toivanen ECON-C4210 Lecture 10 33 / 37

VAR estimation results

. var infl d.ue if time_ind >= 13

Vector autoregression

Sample: 13 thru 902 Number of obs = 890

Log likelihood = 3486.821 AIC = -7.813081
FPE = 1.39e-06 HQIC = -7.792505
Det(Sigma_ml) = 1.36e-06 SBIC = -7.759247

Equation Parms RMSE R-sq chi2 P>chi2

infl 5 .002824 0.3670 515.974 0.0000

D_ue 5 .416157 0.0077 6.920339 0.1402

Coefficient Std. err. z P>|z| [95% conf. interval]

infl
infl
L1. .5831457 .0334794 17.42 0.000 .5175274 .648764
L2. .0345948 .0334775 1.03 0.301 -.03102 .1002096

ue
LD. .0001338 .0002278 0.59 0.557 -.0003127 .0005803
L2D. .0003456 .0002261 1.53 0.126 -.0000974 .0007887

_cons .0011203 .0001277 8.77 0.000 .0008701 .0013706

D_ue
infl
L1. -5.43021 4.932801 -1.10 0.271 -15.09832 4.237903
L2. 7.239596 4.932536 1.47 0.142 -2.427996 16.90719

ue
LD. .0475497 .033565 1.42 0.157 -.0182365 .1133359
L2D. -.0497456 .0333062 -1.49 0.135 -.1150245 .0155333

_cons -.008321 .0188126 -0.44 0.658 -.0451931 .0285511

Toivanen ECON-C4210 Lecture 10 34 / 37
Granger causality

• Do lagged values of X predict Yt , or the other way round?

• H0: the coefficients on all the lagged values of X in the Y regression are jointly
insignificant.
• Granger causality 6= causality.

• Does it make sense to assume the future does not affect the past?

Toivanen ECON-C4210 Lecture 10 35 / 37

VAR on inflation and UE

Stata code
1 v a r i n f l d . ue i f t i m e i n d >= 13
2 vargranger

Toivanen ECON-C4210 Lecture 10 36 / 37

Testing Granger causality

. vargranger

Granger causality Wald tests

Equation Excluded chi2 df Prob > chi2

infl D.ue 2.7756 2 0.250

infl ALL 2.7756 2 0.250

D_ue infl 2.2265 2 0.328

D_ue ALL 2.2265 2 0.328

Toivanen ECON-C4210 Lecture 10 37 / 37