Chapter 11: Models using time series data

Chapter 11: Models using time series

data

Overview
This chapter introduces the application of regression analysis to time
series data, beginning with static models and then proceeding to dynamic
models with lagged variables used as explanatory variables. It is shown
that multicollinearity is likely to be a problem in models with unrestricted
lag structures and that this provides an incentive to use a parsimonious
lag structure, such as the Koyck distribution. Two models using the Koyck
distribution, the adaptive expectations model and the partial adjustment
model, are described, together with well-known applications to aggregate
consumption theory, Friedman’s permanent income hypothesis in the case
of the former and Brown’s habit persistence consumption function in the
case of the latter. The chapter concludes with a discussion of prediction
and stability tests in time series models.

Learning outcomes
After working through the corresponding chapter in the textbook, studying
the corresponding slideshows, and doing the starred exercises in the
textbook and the additional exercises in this guide, you should be able to:
• explain why multicollinearity is a common problem in time series
models, especially dynamic ones with lagged explanatory variables
• describe the properties of a model with a lagged dependent variable
(ADL(1,0) model)
• describe the assumptions underlying the adaptive expectations and
partial adjustment models
• explain the properties of OLS estimators of the parameters of ADL(1,0)
models
• explain how predetermined variables may be used as instruments in
the fitting of models using time series data
• explain in general terms the objectives of time series analysts and those
constructing VAR models

Additional exercises
A11.1
The output below shows the result of linear and logarithmic regressions of
expenditure on food on income, relative price, and population (measured
in thousands) using the Demand Functions data set, together with
the correlations among the variables. Provide an interpretation of the
regression coefficients and perform appropriate statistical tests.

225
20 Elements of econometrics


'HSHQGHQW9DULDEOH)22'
0HWKRG/HDVW6TXDUHV
6DPSOH
,QFOXGHGREVHUYDWLRQV

9DULDEOH&RHIILFLHQW6WG(UURUW6WDWLVWLF3URE

&
'3,
35(/)22'
323

5VTXDUHG0HDQGHSHQGHQWYDU
$GMXVWHG5VTXDUHG6'GHSHQGHQWYDU
6(RIUHJUHVVLRQ$NDLNHLQIRFULWHUL
6XPVTXDUHGUHVLG6FKZDU]FULWHULRQ
/RJOLNHOLKRRG)VWDWLVWLF
'XUELQ:DWVRQVWDW3URE)VWDWLVWLF



'HSHQGHQW9DULDEOH/*)22'
0HWKRG/HDVW6TXDUHV
6DPSOH
,QFOXGHGREVHUYDWLRQV

9DULDEOH&RHIILFLHQW6WG(UURUW6WDWLVWLF3URE

&
/*'3,
/*35)22'
/*323

5VTXDUHG0HDQGHSHQGHQWYDU
$GMXVWHG5VTXDUHG6'GHSHQGHQWYDU
6(RIUHJUHVVLRQ$NDLNHLQIRFULWHU
6XPVTXDUHGUHVLG6FKZDU]FULWHULRQ
/RJOLNHOLKRRG)VWDWLVWLF
'XUELQ:DWVRQVWDW3URE)VWDWLVWLF


&RUUHODWLRQ0DWUL[

/*)22'/*'3,/*35)22'/*323

/*)22'
/*'3,
/*35)22'
/*323


A11.2
Perform regressions parallel to those in Exercise A11.1 using your category
of expenditure and provide an interpretation of the coefficients.

A11.3
The output shows the result of a logarithmic regression of expenditure
on food per capita, on income per capita, both measured in US$ million,
and the relative price index for food. Provide an interpretation of the
coefficients, demonstrate that the specification is a restricted version of
the logarithmic regression in Exercise A11.1, and perform an F test of the
restriction.

226
 Chapter 11: Models using time series data


'HSHQGHQW9DULDEOH/*)22'3&
0HWKRG/HDVW6TXDUHV
6DPSOH
,QFOXGHGREVHUYDWLRQV

9DULDEOH&RHIILFLHQW6WG(UURUW6WDWLVWLF3URE

&
/*'3,3&
/*35)22'

5VTXDUHG0HDQGHSHQGHQWYDU
$GMXVWHG5VTXDUHG6'GHSHQGHQWYDU
6(RIUHJUHVVLRQ$NDLNHLQIRFULWHU
6XPVTXDUHGUHVLG6FKZDU]FULWHULRQ
/RJOLNHOLKRRG)VWDWLVWLF
'XUELQ:DWVRQVWDW3URE)VWDWLVWLF


A11.4
Perform a regression parallel to that in Exercise A11.3 using your category
of expenditure. Provide an interpretation of the coefficients, and perform
an F test of the restriction.

A11.5
The output shows the result of a logarithmic regression of expenditure on
food per capita, on income per capita, the relative price index for food,
and population. Provide an interpretation of the coefficients, demonstrate
that the specification is equivalent to that for the logarithmic regression in
Exercise A11.1, and use it to perform a t test of the restriction in Exercise
A11.3.


'HSHQGHQW9DULDEOH/*)22'3&
0HWKRG/HDVW6TXDUHV
6DPSOH
,QFOXGHGREVHUYDWLRQV

9DULDEOH&RHIILFLHQW6WG(UURUW6WDWLVWLF3URE

&
/*'3,3&
/*35)22'
/*323

5VTXDUHG0HDQGHSHQGHQWYDU
$GMXVWHG5VTXDUHG6'GHSHQGHQWYDU
6(RIUHJUHVVLRQ$NDLNHLQIRFULWHU
6XPVTXDUHGUHVLG6FKZDU]FULWHULRQ
/RJOLNHOLKRRG)VWDWLVWLF
'XUELQ:DWVRQVWDW3URE)VWDWLVWLF


A11.6
Perform a regression parallel to that in Exercise A11.5 using your category
of expenditure, and perform a t test of the restriction implicit in the
specification in Exercise A11.4.

227
20 Elements of econometrics

A11.7
In Exercise 11.9 you fitted the model
LGCAT = b 1 + b 2 LGDPI + b 3 LGDPI (− 1) + b 4 LGPRCAT + b 5 LGPRCAT (− 1) + u
where CAT stands for your category of expenditure.
• Show that (b 2 + b 3 ) and (b 4 + b 5 ) are theoretically the long-run
(equilibrium) income and price elasticities.
• Reparameterise the model and fit it to obtain direct estimates of these
long-run elasticities and their standard errors.
• Confirm that the estimates are equal to the sum of the individual short-
run elasticities found in Exercise 11.9.
• Compare the standard errors with those found in Exercise 11.9 and
state your conclusions.

A11.8
In a certain bond market, the demand for bonds, Bt, in period t is negatively
related to the expected interest rate, ite+1 , in period t + 1:

Bt = b 1 + b 2 ite+1 + u t (1)

where ut is a disturbance term not subject to autocorrelation. The expected

interest rate is determined by an adaptive expectations process:
( )
ite+1 − ite = λ it − ite (2)

where it is the actual rate of interest in period t. A researcher uses the

following model to fit the relationship:
Bt = γ 1 + γ 2 it + γ 3 Bt −1 + vt (3)

where vt is a disturbance term.

• Show how this model may be derived from the demand function and the
adaptive expectations process.
• Explain why inconsistent estimates of the parameters will be obtained
if equation (3) is fitted using ordinary least squares (OLS). (A
mathematical proof is not required. Do not attempt to derive expressions
for the bias.)
• Describe a method for fitting the model that would yield consistent
estimates.
• Suppose that ut was subject to the first-order autoregressive process:

u t = ρu t −1 + ε t

where εt is not subject to autocorrelation. How would this affect your

answer to the second part of this question?
• Suppose that the true relationship was actually
Bt = b1 + b 2it + ut (1*)

with ut not subject to autocorrelation, and the model is fitted by

regressing Bt on it and Bt–1, as in equation (3), using OLS. How would
this affect the regression results?
• How plausible do you think an adaptive expectations process is for
modelling expectations in a bond market?

228
 Chapter 11: Models using time series data

A11.9
The output shows the result of a logarithmic regression of expenditure
on food on income, relative price, population, and lagged expenditure on
food using the Demand Functions data set. Provide an interpretation of
the regression coefficients, paying attention to both short-run and long-run
dynamics, and perform appropriate statistical tests.


'HSHQGHQW9DULDEOH/*)22'
0HWKRG/HDVW6TXDUHV
6DPSOHDGMXVWHG
,QFOXGHGREVHUYDWLRQVDIWHUDGMXVWLQJHQGSRLQWV

9DULDEOH&RHIILFLHQW6WG(UURUW6WDWLVWLF3URE

&
/*'3,
/*35)22'
/*323
/*)22'

5VTXDUHG0HDQGHSHQGHQWYDU
$GMXVWHG5VTXDUHG6'GHSHQGHQWYDU
6(RIUHJUHVVLRQ$NDLNHLQIRFULWHU
6XPVTXDUHGUHVLG6FKZDU]FULWHULRQ
/RJOLNHOLKRRG)VWDWLVWLF
'XUELQ:DWVRQVWDW3URE)VWDWLVWLF


A11.10
Perform a regression parallel to that in Exercise A11.9 using your category
of expenditure. Provide an interpretation of the coefficients, and perform
appropriate statistical tests.

A11.11
In his classic study Distributed Lags and Investment Analysis (1954), Koyck
investigated the relationship between investment in railcars and the
volume of freight carried on the US railroads using data for the period
1884–1939. Assuming that the desired stock of railcars in year t depended
on the volume of freight in year t–1 and year t–2 and a time trend, and
assuming that investment in railcars was subject to a partial adjustment
process, he fitted the following regression equation using OLS (standard
errors and constant term not reported):
Iˆ = 0.077 F + 0.017F – 0.0033t – 0.110K
t t–1 t–2 t–1
R2 = 0.85
where It = Kt – Kt–1 is investment in railcars in year t (thousands), Kt is the
stock of railcars at the end of year t (thousands), and Ft is the volume of
freight handled in year t (ton-miles).
Provide an interpretation of the equation and describe the dynamic
process implied by it. (Note: It is best to substitute Kt – Kt–1 for It in the
regression and treat it as a dynamic relationship determining Kt.)

A11.12
Two researchers agree that a model consists of the following relationships:
Yt = α1 + α2Xt + ut (1)
Xt = β1 + β2Yt–1 + vt (2)
Zt = γ1 + γ2Yt + γ3Xt + γ4Qt + wt (3)

229
20 Elements of econometrics

where ut, vt, and wt, are disturbance terms that are drawn from fixed
distributions with zero mean. It may be assumed that they are distributed
independently of Qt and of each other and that they are not subject to
autocorrelation. All the parameters may be assumed to be positive and it
may be assumed that α2β2 < 1.
• One researcher asserts that consistent estimates will be obtained if (2)
is fitted using OLS and (1) is fitted using IV, with Yt–1 as an instrument
for Xt. Determine whether this is true.
• The other researcher asserts that consistent estimates will be obtained
if both (1) and (2) are fitted using OLS, and that the estimate of β2 will
be more efficient than that obtained using IV. Determine whether this is
true.

Answers to the starred exercises in the textbook

11.6
Year Y K L Year Y K L
1899 100 100 100 1911 153 216 145
1900 101 107 105 1912 177 226 152
1901 112 114 110 1913 184 236 154
1902 122 122 118 1914 169 244 149
1903 124 131 123 1915 189 266 154
1904 122 138 116 1916 225 298 182
1905 143 149 125 1917 227 335 196
1906 152 163 133 1918 223 366 200
1907 151 176 138 1919 218 387 193
1908 126 185 121 1920 231 407 193
1909 155 198 140 1921 179 417 147
1910 159 208 144 1922 240 431 161

Source: Cobb and Douglas (1928)

The table gives the data used by Cobb and Douglas (1928) to fit the
original Cobb–Douglas production function:

Yt = b 1 K tb Lbt v t
2 3

Yt, Kt, and Lt, being index number series for real output, real capital input,
and real labour input, respectively, for the manufacturing sector of the
United States for the period 1899–1922 (1899=100). The model was
linearised by taking logarithms of both sides and the following regressions
was run (standard errors in parentheses):
logY ˆ = –0.18 + 0.23 log K + 0.81 log L R2 = 0.96
(0.43) (0.06)         (0.15)
Provide an interpretation of the regression coefficients.
Answer:
The elasticities of output with respect to capital and labor are 0.23 and
0.81, respectively, both coefficients being significantly different from zero
at very high significance levels. The fact that the sum of the elasticities
is close to one suggests that there may be constant returns to scale.
Regressing output per worker on capital per worker, one has

230
 Chapter 11: Models using time series data

ˆ
Y K
log
= 0.01 + 0.25 log R2 = 0.63
L L
(0.02) (0.04)
The smaller standard error of the slope coefficient suggests a gain in
efficiency. Fitting a reparameterised version of the unrestricted model
ˆ
Y K
log
= –0.18 + 0.23 log + 0.04 log L R2 = 0.64
L L
(0.43) (0.06) (0.09)
we find that the restriction is not rejected.

11.7
The Cobb–Douglas model in Exercise 11.6 makes no allowance for the
possibility that output may be increasing as a consequence of technical
progress, independently of K and L. Technical progress is difficult to
quantify and a common way of allowing for it in a model is to include an
exponential time trend:

Yt = β 1 K tβ 2 Lβt 3 e ρ t v t

where ρ is the rate of technical progress and t is a time trend defined to be

1 in the first year, 2 in the second, etc. The correlations between log K, log
L and t are shown in the table. Comment on the regression results.
ˆ
logY = 2.81 – 0.53 log K + 0.91 log L + 0.047 t R2 = 0.97
(1.38) (0.34) (0.14) (0.021)

&RUUHODWLRQ

/*./*/7,0(

/*.
/*/
7,0(


Answer:
The elasticity of output with respect to labour is higher than before, now
implausibly high given that, under constant returns to scale, it should
measure the share of wages in output. The elasticity with respect to capital
is negative and nonsensical. The coefficient of time indicates an annual
exponential growth rate of 4.7 per cent, holding K and L constant. This
is unrealistically high for the period in question. The implausibility of the
results, especially those relating to capital and time (correlation 0.997),
may be attributed to multicollinearity.

11.16
The output below shows the result of fitting the model
LGFOOD = β1 + β2λLGDPI + β2λ(1 – λ)LGDPI(–1)
+ β2λ(1 – λ)2LGDPI(–2) + β3LGPRFOOD + u
using the data on expenditure on food in the Demand Functions data
set. LGFOOD and LGPRFOOD are the logarithms of expenditure on food
and the relative price index series for food. C(1), C(2), C(3), and C(4)
are estimates of β1, β2, λ and β3, respectively. Explain how the regression
equation could be interpreted as an adaptive expectations model and
discuss the dynamics implicit in it, both short-run and long-run. Should
the specification have included further lagged values of LGDPI?
231
20 Elements of econometrics


'HSHQGHQW9DULDEOH/*)22'
0HWKRG/HDVW6TXDUHV
6DPSOHDGMXVWHG
,QFOXGHGREVHUYDWLRQVDIWHUDGMXVWLQJHQGSRLQWV
&RQYHUJHQFHDFKLHYHGDIWHULWHUDWLRQV
/*)22' && & /*'3,& & & /*'3,&
 & &A /*'3,& & &A /*'3,
& /*35)22'

&RHIILFLHQW6WG(UURUW6WDWLVWLF3URE

&
&
&
&

5VTXDUHG0HDQGHSHQGHQWYDU
$GMXVWHG5VTXDUHG6'GHSHQGHQWYDU
6(RIUHJUHVVLRQ$NDLNHLQIRFULWHU
6XPVTXDUHGUHVLG6FKZDU]FULWHULRQ
/RJOLNHOLKRRG'XUELQ:DWVRQVWDW


Answer:
There is a discrepancy between the theoretical specification, which has
two lagged values, and the regression specification, which has three.
Fortunately, in this case it makes little difference. Here is the output for the
regression with two lags:


'HSHQGHQW9DULDEOH/*)22'
0HWKRG/HDVW6TXDUHV
6DPSOHDGMXVWHG
,QFOXGHGREVHUYDWLRQVDIWHUDGMXVWPHQWV
&RQYHUJHQFHDFKLHYHGDIWHULWHUDWLRQV
/*)22' && & /*'3,& & & /*'3,& &
 &A /*'3,& /*35)22'

&RHIILFLHQW6WG(UURUW6WDWLVWLF3URE

&
&
&
&

5VTXDUHG0HDQGHSHQGHQWYDU
$GMXVWHG5VTXDUHG6'GHSHQGHQWYDU
6(RIUHJUHVVLRQ$NDLNHLQIRFULWHU
6XPVTXDUHGUHVLG6FKZDU]FULWHULRQ
/RJOLNHOLKRRG+DQQDQ4XLQQFULWH
)VWDWLVWLF'XUELQ:DWVRQVWDW
3URE)VWDWLVWLF


Suppose that the model is

LGFOOD = β1 + β2LGDPIe + β3LGPRFOOD + u
where LGDPIe is expected LGDPI at time t + 1, and that expectations for
income are subject to the adaptive expectations process
LGDPIe – LGDPIe = λ(LGDPI – LGDPIe).
The adaptive expectations process may be rewritten
LGDPIe = λLGDPI + (1 – λ)LGDPIe .
Lagging this equation one period and substituting, one has.
LGDPIe = λLGDPI + λ(1 – λ)LGDPI(–1) + (1 – λ)2LGDPIe(–1).

232
 Chapter 11: Models using time series data

Lagging a second time and substituting, one has

LGDPIe = λLGDPI + λ(1 – λ)LGDPI(–1) + λ(1 – λ)2LGDPI(–2)
+ (1 – λ)3LGDPIe(–2).
Substituting this into the model, one has the regression specification as
stated in the question. The actual regression in the textbook includes a
further lagged term.
The output implies that the estimate of the long-run income elasticity, β2, is
0.50 (original and revised output). The estimate of λ, the speed of adjustment
of expectations, is 0.92 (0.94 in the revised output). Hence the estimate
of the short-run income elasticity, β2λ, is 0.46 (0.47 in the revised output).
The price side of the model has been assumed to be static. The estimate of
the price elasticity is –0.09 (–0.08 in the revised output). For the theoretical
specification, the coefficient of the dropped unobservable term is β2(1 – λ)3.
Given the estimates of β2 and λ, its estimate is 0.0003. Hence we are justified
in neglecting it. For the revised output, its estimate is even lower, 0.0001.

11.18
A researcher is fitting the following supply and demand model for a
certain commodity, using a sample of time series observations:
4GW E   E  3W  X GW 
4 VW D   D  3W  X VW 
where Qdt is the amount demanded at time t, Qst is the amount supplied,
Pt is the market clearing price, and udt and ust are disturbance terms that
are not necessarily independent of each other.. It may be assumed that the
market clears and so Qdt = Qst.
• What can be said about the identification of (a) the demand equation,
(b) the supply equation?
• What difference would it make if supply at time t was determined
instead by price at time t − 1 ? That is,
4 VW D   D  3W   X VW 
• What difference would it make if it could be assumed that udt is
distributed independently of ust?
Answer:
The reduced form equation for Pt is


3W E  D   X GW  X VW  .
D  E
Pt is not independent of the disturbance term in either equation and so
OLS would yield inconsistent estimates.
Provided that udt is not subject to autocorrelation, Pt–1 could be used as
an instrument in the demand equation. Provided that ust is not subject to
autocorrelation, OLS could be used to fit the second equation. It makes no
difference whether or not udt is distributed independently of ust.
The first equation could, alternatively, be fitted using OLS, with the
variables switched. From the second equation, Pt–1 determines Qt, and
then, given Qt, the demand equation determines Pt:

3W 4W  E   X GW .
E

The reciprocal of the slope coefficient provides a consistent estimator of β2.

233
20 Elements of econometrics

Answers to the additional exercises

A11.1
The linear regression indicates that expenditure on food increases by
$0.032 billion for every extra $ billion of disposable personal income (in
other words, by 3.2 cents out of the marginal dollar), that it increases
by $0.403 billion for every point increase in the price index, and that it
increases by $0.001 billion for every additional thousand population. The
income coefficient is significant at the 1 per cent level (ignoring problems
to be discussed in Chapter 12). The positive price coefficient makes no
sense (remember that the dependent variable is measured in real terms).
The intercept has no plausible interpretation.
The logarithmic regression indicates that the income elasticity is 0.59
and highly significant, and the price elasticity is –0.12, not significant.
The negative elasticity for population is not plausible. One would expect
expenditure on food to increase in line with population, controlling for
other factors, and hence, as a first approximation, the elasticity should be
equal to 1. However, an increase in population, keeping income constant,
would lead to a reduction in income per capita and hence to a negative
income effect. Given that the income elasticity is less than 1, one would
still expect a positive elasticity overall for population. At least the estimate
is not significantly different from zero. In view of the high correlation,
0.995, between LGDPI and LGPOP, the negative estimate may well be a
result of multicollinearity.

A11.2
OLS logarithmic regressions
LGDPI LGP LGPOP R2
coef. s.e. coef. s.e. coef. s.e.
ADM –1.43 0.20 –0.28 0.10 6.88 0.61 0.975
BOOK –0.29 0.28 –1.18 0.21 4.94 0.82 0.977
BUSI 0.36 0.19 –0.11 0.27 2.79 0.51 0.993
CLOT 0.71 0.10 –0.70 0.05 0.15 0.36 0.998
DENT 1.23 0.14 –0.95 0.09 0.26 0.54 0.995
DOC 0.97 0.14 0.26 0.13 –0.27 0.52 0.993
FLOW 0.46 0.32 0.16 0.33 3.07 1.21 0.987
FOOD 0.59 0.08 –0.12 0.08 –0.29 0.26 0.992
FURN 0.36 0.28 –0.48 0.26 1.66 1.12 0.985
GAS 1.27 0.24 –0.24 0.06 –2.81 0.74 0.788
GASO 1.46 0.16 –0.10 0.04 –2.35 0.49 0.982
HOUS 0.91 0.08 –0.54 0.06 0.38 0.25 0.999
LEGL 1.17 0.16 –0.08 0.13 –1.50 0.54 0.976
MAGS 1.05 0.22 –0.73 0.44 –0.82 0.54 0.970
MASS –1.92 0.22 –0.57 0.14 6.14 0.65 0.785
OPHT 0.30 0.45 0.28 0.59 3.68 1.40 0.965
RELG 0.56 0.13 –0.99 0.23 2.72 0.41 0.996
TELE 0.91 0.13 –0.61 0.11 1.79 0.49 0.998
TOB 0.54 0.17 –0.42 0.04 –1.21 0.57 0.883
TOYS 0.59 0.10 –0.54 0.06 2.57 0.39 0.999

234
 Chapter 11: Models using time series data

The price elasticities mostly lie in the range 0 to –1, as they should,
and therefore seem plausible. However the very high correlation
between income and population, 0.995, has given rise to a problem of
multicollinearity and as a consequence the estimates of their elasticities
are very erratic. Some of the income elasticities look plausible, but that
may be pure chance, for many are unrealistically high, or negative when
obviously they should be positive. The population elasticities are even less
convincing.

Correlations between prices, income and population

LGP, LGDPI LGP, LGPOP LGP, LGDPI LGP, LGPOP
ADM 0.61 0.61 GASO 0.05 0.03
BOOK 0.88 0.87 HOUS 0.49 0.55
BUSI 0.98 0.97 LEGL 0.99 0.99
CLOT –0.94 –0.96 MAGS 0.99 0.98
DENT 0.94 0.96 MASS 0.90 0.89
DOC 0.98 0.98 OPHT –0.68 –0.67
FLOW –0.93 –0.95 RELG 0.92 0.92
FOOD –0.60 –0.64 TELE –0.98 –0.99
FURN –0.95 –0.97 TOB 0.83 0.86
GAS 0.77 0.76 TOYS –0.97 –0.98

A11.3
The regression indicates that the income elasticity is 0.40 and the price
elasticity 0.21, the former very highly significant, the latter significant at
the 1 per cent level using a one-sided test. If the specification is

FOOD DPI
log = b 1 + b 2 log + b 3 log PRELFOOD + u
POP POP

it may be rewritten

log FOOD = β 1 + β 2 log DPI + β 3 log PRELFOOD
+ (1 − β 2 ) log POP + u.

This is a restricted form of the specification in Exercise A11.2:

log FOOD =b 1 + b 2 log DPI + b 3 log PRELFOODb 3
+ b 4 log POP + u

with β4 = 1 – β2. We can test the restriction by comparing RSS for the two
regressions:

    
)   
  

The critical value of F(1,40) at the 0.1 per cent level is 12.61. The critical
value for F(1,41) must be slightly lower. Thus we reject the restriction.
Since the restricted version is misspecified, our interpretation of the
coefficients of this regression and the t tests are invalidated.

235
20 Elements of econometrics

A11.4
Given that the critical values of F(1,41) at the 5 and 1 per cent levels are
4.08 and 7.31 respectively, the results of the F test may be summarised as
follows:
Restriction not rejected: CLOT, DENT, DOC, FURN, HOUS
Restriction rejected at the 5 per cent level: MAGS
Restriction rejected at the 1 per cent level: ADM, BOOK, BUSI, FLOW,
FOOD, GAS, GASO, LEGL, MASS, OPHT, RELG, TELE, TOB, TOYS
However, for reasons that will become apparent in the next chapter, these
findings must be regarded as provisional.

Tests of a restriction
RSSU RSSR F t
ADM 0.125375 0.480709 116.20 10.78
BOOK 0.223664 0.461853 43.66 6.61
BUSI 0.084516 0.167580 40.30 6.35
CLOT 0.021326 0.021454 0.25 –0.50
DENT 0.033275 0.034481 1.49 1.22
DOC 0.068759 0.069726 0.58 –0.76
FLOW 0.220256 0.262910 7.94 2.82
FOOD 0.016936 0.023232 15.24 –3.90
FURN 0.157153 0.162677 1.44 1.20
GAS 0.185578 0.300890 25.48 –5.05
GASO 0.078334 0.139278 31.90 –5.65
HOUS 0.011270 0.012106 3.04 1.74
LEGL 0.082628 0.102698 9.96 –3.16
MAGS 0.096620 0.106906 4.36 –2.09
MASS 0.143775 0.330813 53.34 7.30
OPHT 0.663413 0.822672 9.84 3.14
RELG 0.053785 0.135532 62.32 7.89
TELE 0.054519 0.080728 19.71 4.44
TOB 0.062452 0.087652 16.54 –4.07
TOYS 0.031269 0.071656 52.96 7.28

A11.5
If the specification is
FOOD DPI
log = b 1 + b 2 log + b 3 log PRELFOOD + γ 1 POP + u ,
POP POP
it may be rewritten
log FOOD = β 1 + β 2 log DPI + β 3 log PRELFOOD
+(1 − β 2 + γ 1 ) log POP + u.

This is equivalent to the specification in Exercise A11.1:

log FOOD = β 1 + β 2 log DPI + β 3 log PRELFOOD
+ β 4 log POP + u

with β4 = 1 – β2 + γ1. Note that this is not a restriction. (1) – (3) are just
different ways of writing the unrestricted model.

236
 Chapter 11: Models using time series data

A t test of H0: γ1 = 0 is equivalent to a t test of H0: β4 = 1 – β2, that is, that

the restriction in Exercise A11.3 is valid. The t statistic for LGPOP in the
regression is –3.90, and hence again we reject the restriction. Note that
the test is equivalent to the F test. –3.90 is the square root of 15.24, the F
statistic, and it can be shown that the critical value of t is the square root
of the critical value of F.

A11.6
The t statistics for all the categories of expenditure are supplied in the
table in the answer to Exercise A11.4. Of course they are equal to the
square root of the F statistic, and their critical values are the square roots
of the critical values of F, so the conclusions are identical and, like those of
the F test, should be treated as provisional.

A11.7
• Show that (b 2 + b 3 ) and (b 4 + b 5 ) are theoretically the long-run
(equilibrium) income and price elasticities.
In equilibrium, LGCAT = LGCAT , LGDPI = LGDPI (− 1) = LGDPI and
LGPRCAT = LGPRCAT (− 1) = LGPRCAT . Hence, ignoring the transient
effect of the disturbance term,

LGCAT = β1 + β 2 LGDPI + β 3 LGDPI + β 4 LGPRCAT + β 5 LGPRCAT
= β1 + (β 2 + β 3 )LGDPI + (β 4 + β 5 )LGPRCAT .

Thus the long-run equilibrium income and price elasticities are

θ = b 2 + b 3 and φ = b 4 + b 5 , respectively.
• Reparameterise the model and fit it to obtain direct estimates of these
long-run elasticities and their standard errors.
We will reparameterise the model to obtain direct estimates of θ and
φ and their standard errors. Write b 3 = θ − b 2 and φ = β 4 + β 5 and
substitute for β3 and β5 in the model. We obtain

LGCAT = b1 + b 2 LGDPI + (θ − b 2 )LGDPI (− 1) + b 4 LGPRCAT + (φ − b 4 )LGPRCAT (− 1) + u

= b1 + b 2 (LGDPI − LGDPI (− 1)) + θLGDPI (− 1)
+ b 4 (LGPRCAT − LGPRCAT (− 1)) + φLGPRCAT (− 1) + u
= b1 + b 2 DLGDPI + θLGDPI (− 1) + b 4 DLGPRCAT + φLGPRCAT (− 1) + u

where DLGDPI = LGDPI – LGDPI(–1) and DLGPRCAT = LGPRCAT –

LGPRCAT(–1).
The output for HOUS is shown below. DLGPRCAT has been abbreviated
as DLGP.

237
20 Elements of econometrics


'HSHQGHQW9DULDEOH/*+286
0HWKRG/HDVW6TXDUHV
6DPSOHDGMXVWHG
,QFOXGHGREVHUYDWLRQVDIWHUDGMXVWLQJHQGSRLQWV

9DULDEOH&RHIILFLHQW6WG(UURUW6WDWLVWLF3URE

&
'/*'3,
/*'3,
'/*3
/*35+286

5VTXDUHG0HDQGHSHQGHQWYDU
$GMXVWHG5VTXDUHG6'GHSHQGHQWYDU
6(RIUHJUHVVLRQ$NDLNHLQIRFULWHU
6XPVTXDUHGUHVLG6FKZDU]FULWHULRQ
/RJOLNHOLKRRG)VWDWLVWLF
'XUELQ:DWVRQVWDW3URE)VWDWLVWLF


• Confirm that the estimates are equal to the sum of the individual short-
run elasticities found in Exercise 11.9.
The estimates of the long-run income and price elasticities are 1.01 and
–0.45, respectively. The output below is for the model in its original
form, where the coefficients are all short-run elasticities. It may be seen
that, for both income and price, the sum of the estimates of the short-
run elasticities is indeed equal to the estimate of the long-run elasticity
in the reparameterised specification.


'HSHQGHQW9DULDEOH/*+286
0HWKRG/HDVW6TXDUHV
6DPSOHDGMXVWHG
,QFOXGHGREVHUYDWLRQVDIWHUDGMXVWLQJHQGSRLQWV

9DULDEOH&RHIILFLHQW6WG(UURUW6WDWLVWLF3URE

&
/*'3,
/*'3,
/*35+286
/*35+286

5VTXDUHG0HDQGHSHQGHQWYDU
$GMXVWHG5VTXDUHG6'GHSHQGHQWYDU
6(RIUHJUHVVLRQ$NDLNHLQIRFULWHU
6XPVTXDUHGUHVLG6FKZDU]FULWHULRQ
/RJOLNHOLKRRG)VWDWLVWLF
'XUELQ:DWVRQVWDW3URE)VWDWLVWLF


• Compare the standard errors with those found in Exercise 11.9 and state
your conclusions.
The standard errors of the long-run elasticities in the reparameterised
version are much smaller than those of the short-run elasticities in the
original specification, and the t statistics accordingly much greater. Our
conclusion is that it is possible to obtain relatively precise estimates of
the long-run impact of income and price, even though multicollinearity
prevents us from deriving precise short-run estimates.

238
 Chapter 11: Models using time series data

A11.8
• Show how this model may be derived from the demand function and the
adaptive expectations process.
The adaptive expectations process may be rewritten

ite+1 = λit + (1 − λ )ite .

Substituting this into (1), one obtains

Bt = b1 + b 2 λit + b 2 (1 − λ )ite + u t .

We note that if we lag (1) by one time period,

Bt −1 = b1 + b 2 ite + u t −1 .

Hence

b 2 ite = Bt −1 − b1 − u t −1 .

Substituting this into the second equation above, one has

Bt = b1λ + b 2 λit + (1 − λ )Bt −1 + u t − (1 − λ )u t −1 .

This is equation (3) in the question, with γ1 = β1λ, γ2 = β2λ, γ3 = 1 – λ,

and vt = ut – (1 – λ)ut–1.
• Explain why inconsistent estimates of the parameters will be obtained if
equation (3) is fitted using ordinary least squares (OLS). (A mathematical
proof is not required. Do not attempt to derive expressions for the bias.)
In equation (3), the regressor Bt–1 is partly determined by ut–1. The
disturbance term vt also has a component ut–1. Hence the requirement
that the regressors and the disturbance term be distributed
independently of each other is violated. The violation will lead to
inconsistent estimates because the regressor and the disturbance term
are contemporaneously correlated.
• Describe a method for fitting the model that would yield consistent
estimates.
If the first equation in this exercise is true for time period t + 1, it is
true for time period t:

ite = λit −1 + (1 − λ )ite−1 .

Substituting into the second equation in (a), we now have

Bt = b 1 + b 2 λit + b 2 λ (1 − λ )it −1 + (1 − λ ) ite−1 + u t .
2

Continuing to lag and substitute, we have

Bt = β1 + β 2 λit + β 2 λ (1 − λ )it −1 + ... + β 2 λ (1 − λ ) it − s +1 + (1 − λ ) ite− s +1 + ut .
s −1 s

For s large enough, (1 − λ ) will be so small that we can drop the

s

unobservable term ite− s +1 with negligible omitted variable bias. The
disturbance term is distributed independently of the regressors and
hence we obtain consistent estimates of the parameters. The model
should be fitted using a nonlinear estimation technique that takes
account of the restrictions implicit in the specification.

239
20 Elements of econometrics

• Suppose that ut were subject to the first-order autoregressive process:

u t = ρu t −1 + ε t

where εt is not subject to autocorrelation. How would this affect your

answer to the second part of this question?
vt is now given by
vt = ut – (1 – λ)ut–1 = ρut–1 + εt – (1 – λ)ut–1 = εt – (1 – ρ – λ)ut–1.
Since ρ and λ may reasonably be assumed to lie between 0 and 1, it is
possible that their sum is approximately equal to 1, in which case vt is
approximately equal to the innovation εt. If this is the case, there would
be no violation of the regression assumption described in the second
part of this question and one could use OLS to fit (3) after all.
• Suppose that the true relationship was actually
Bt = b 1 + b 2 i t + u t (1*)

with ut not subject to autocorrelation, and the model is fitted by regressing

Bt on it and Bt–1, as in equation (3), using OLS. How would this affect the
regression results?
The estimators of the coefficients will be inefficient in that Bt–1 is a
redundant variable. The inclusion of Bt–1 will also give rise to finite
sample bias that would disappear in large samples.
• How plausible do you think an adaptive expectations process is for
modelling expectations in a bond market?
The adaptive expectations model is implausible since the expectations
process would change as soon as those traders taking advantage of
their knowledge of it started earning profits.

A11.9
The regression indicates that the short-run income, price, and population
elasticities for expenditure on food are 0.14, –0.10, and –0.05, respectively,
and that the speed of adjustment is (1 – 0.73) = 0.27. Dividing by 0.27,
the long-run elasticities are 0.52, –0.37, and –0.19, respectively. The
income and price elasticities seem plausible. The negative population
elasticity makes no sense, but it is small and insignificant. The estimates of
the short-run income and price elasticities are likewise not significant, but
this is not surprising given that the point estimates are so small.

A11.10
The table gives the result of the specification with a lagged dependent
variable for all the categories of expenditure.

240
 Chapter 11: Models using time series data

OLS logarithmic regression

long-run
LGDPI LGP LGPOP LGCAT(–1)
effects
coef. s.e. coef. s.e. coef. s.e. coef. s.e. DPI P
ADM –0.38 0.18 –0.10 0.06 2.03 0.74 0.68 0.09 –1.18 –0.33
BOOK –0.36 0.20 –0.21 0.22 2.07 0.74 0.75 0.12 –1.46 –1.05
BUSI 0.10 0.13 0.03 0.18 0.78 0.45 0.72 0.11 0.33 0.09
CLOT 0.44 0.10 –0.40 0.07 0.01 0.32 0.43 0.09 0.77 –0.70
DENT 0.71 0.18 –0.46 0.16 –0.13 0.51 0.47 0.13 1.34 –0.87
DOC 0.23 0.14 –0.11 0.10 0.21 0.35 0.78 0.10 1.04 –0.52
FLOW 0.20 0.24 –0.31 0.27 0.07 0.98 0.75 0.11 0.81 –1.25
FOOD 0.14 0.09 –0.10 0.06 –0.05 0.19 0.73 0.11 0.53 –0.35
FURN 0.07 0.22 –0.07 0.22 0.82 0.91 0.68 0.12 0.21 –0.23
GAS 0.10 0.17 –0.06 0.03 –0.13 0.45 0.76 0.08 0.42 –0.26
GASO 0.32 0.11 –0.10 0.02 –0.59 0.25 0.80 0.06 1.56 –0.47
HOUS 0.30 0.05 –0.09 0.04 –0.13 0.10 0.73 0.05 1.11 –0.32
LEGL 0.40 0.14 0.10 0.09 –0.90 0.36 0.68 0.09 1.23 0.30
MAGS 0.57 0.21 –0.48 0.37 –0.56 0.44 0.55 0.12 1.27 –1.08
MASS –0.28 0.29 –0.23 0.11 1.08 0.89 0.75 0.12 –1.14 –0.93
OPHT 0.30 0.24 –0.28 0.33 –0.45 0.85 0.88 0.09 2.48 –2.25
RELG 0.34 0.09 –0.71 0.17 1.25 0.38 0.51 0.09 0.68 –1.44
TELE 0.15 0.14 0.00 0.12 0.68 0.37 0.81 0.12 0.77 0.02
TOB 0.12 0.14 –0.12 0.05 –0.31 0.43 0.71 0.11 0.43 –0.43
TOYS 0.31 0.11 –0.27 0.08 1.44 0.47 0.47 0.12 0.58 –0.51

A11.11
Given the information in the question, the model may be written
K t* = β1 + β2Ft–1 + β3Ft–2 + β4t + ut
Kt – Kt–1 = It = λ( K t* – Kt–1).
Hence
It = λβ1 + λβ2Ft–1 + λβ3Ft–2 + λβ4t – λKt–1 + λut.
From the fitted equation,
l = 0.110
0.077
b2 = = 0.70
0.110
0.017
b3 = = 0.15
0.110
−0.0033
b4 = = –0.030.
0.110

Hence the short-run effect of an increase of 1 million ton-miles of freight

is to increase investment in railcars by 7,000 one year later and 1,500 two
years later. It does not make much sense to talk of a short-run effect of a
time trend.
In the long-run equilibrium, neglecting the effects of the disturbance term,
Kt and K t* are both equal to the equilibrium value K and Ft–1 and Ft–2 are
both equal to their equilibrium value F . Hence, using the first equation,
K = β1 + (β2 + β3) F + β4t.

241
20 Elements of econometrics

Thus an increase of one million ton-miles of freight will increase the stock
of railcars by 940 and the time trend will be responsible for a secular
decline of 33 railcars per year.

A11.12
• One researcher asserts that consistent estimates will be obtained if (2) is
fitted using OLS and (1) is fitted using IV, with Yt–1 as an instrument for
Xt. Determine whether this is true.
(2) may indeed be fitted using OLS. Strictly speaking, there may be
an element of bias in finite samples because of noncontemporaneous
correlation between vt and future values of Yt–1.
We could indeed use Yt–1 as an instrument for Xt in (1) because Yt–1 is a
determinant of Xt but is not (contemporaneously) correlated with ut.
• The other researcher asserts that consistent estimates will be obtained if
both (1) and (2) are fitted using OLS, and that the estimate of β2 will be
more efficient than that obtained using IV. Determine whether this is true.
This assertion is also correct. Xt is not correlated with ut, and OLS
estimators are more efficient than IV estimators when both are
consistent. Strictly speaking, there may be an element of bias in finite
samples because of noncontemporaneous correlation between ut and
future values of Xt.

242