FACULTY OF ECONOMICS AND BUSINESS

CAMPUS BRUSSEL
Master of Business Engineering

Statistical Modelling
Distributed lag models

Studenmund (2013). Using econometrics: A practical guide (6th.

edition). Edinburgh: Pearson Education Limited.
Koop, G. (2005). Analysis of economic data (second Edition). West
Sussex: John Wiley & Sons.

1
Overview part time series models
 Regression with time lags: Distributed lag models
 Univariate time series analysis
 Regression with time series variables

2
Time series data versus cross-sectional data
 In econometrics, it is important to distinguish between cross-sectional
data and time-series data.
o Cross-sectional data: data on a number of economic units (e.g. consumers,
households, firms, etc.) at a particular point in time.
o Time-series data: data collected over time on one economic unit.
 Time series observations are usually correlated, whereas a simple
random sample of cross-sectional observations are typically independent.
o E.g. the household income of two families are usually not related, whereas
the household income of a family in a certain year is related to the
household income of the same family in the previous year.
 Time series data have a natural ordering according to time whereas
cross-sectional data have no particular ordering.

3
Time series data versus cross-sectional data
 We can distinguish between univariate time series models and time series
regression models
o Univariate time series model: model a time series of one variable
o Time series regression: model the relation between two variables measured
at a set of equidistant time points.
 Time series data often contain dynamic relations between variables: a
change in a variable can have an impact on the same variable, or other
variables in one or more future time periods.
o E.g. When the income tax rate is increased, consumers have less disposable
income, they will spend less money, which reduces profits of suppliers, etc.
The effects of tax increase do not occur immediately, i.e. they are
distributed over different future time periods.
o E.g. A firm’s decision to purchase new manufacturing machinery will not
immediately affect production. It takes time to purchase and install the
equipment, train workers, etc.

4
Modelling dynamic relations
 Suppose we want to model the inflation rate (in a period of 50 years)
using the interest rate in these years
 The dynamic relation between the two variables can be modeled in three
ways:
o A distributed lag model specifies that the current rate of inflation depends
on the current interest rate, and also on past values of the interest rate. The
model also implies that a change in the current interest rate will have an
impact on the inflation now and in future periods.
o An autoregressive distributed lag model specifies that the current rate of
inflation depends on past values of the inflation rate, and on current and past
values of the interest rate.
o A serial correlation (or autocorrelation) model assumes that the error in the
current inflation rate depends on the error values of past inflation rates.

 In this chapter we further describe distributed lag models.

5
Example: The effect of safety training on accidents
 Many companies provide safety training to workers in an effort to reduce
losses due to accidents. To evaluate the training program a company may
be interested to study the relation between the hours of training provided to
each worker and the losses due to accidents.
 We have time series data collected on a monthly basis during 5 years (i.e.
60 observations) for the following variables:
o : losses due to accidents (in $ per month)
o : safety training provided to each worker (hours per month).
 As it is likely that safety training provided in previous periods also affects
the current accident rate (and losses), we can use a distributed lag model.
 The model specifies that current losses depend on present safety training
and on safety training up to 4 months ago.

6
Distributed lag model
 The model can be formulated as follows:

 Notation:
o are the values of and in the current time period
o indicates the value of in the previous time period
o represents the value of two periods ago
o…
o represents the value of four periods ago
o is assumed to be a Normally distributed error term as in standard
regression models
 The variables , etc. are called lagged variables.
o represents the variable lagged periods

7
Example: The effect of safety training on accidents
 The figure below shows (for the first 10 records of the data set) the
dependent variable (Losses), the explanatory variable (training) and the
training variable lagged 1 to 4 periods (training_lag1,.., training_lag4).
 When we use OLS to fit the distributed lag model, the first four records
will be omitted because they contain missing values.

8
Example: The effect of safety training on accidents
SPSS output:

9
Example: The effect of safety training on accidents
 What can the company conclude about the effectiveness of safety programs?
 Assuming that no other changes in the company’s training policy are made,
increasing the safety training of each worker by one hour in a certain month is
associated with
o An immediate predicted average reduction of losses due to accidents of $131.99 (p<.01).
o A predicted average reduction of losses due to accidents of $449.86 one month later
(p<.01).
o A predicted average reduction of losses due to accidents of $422.52 two months later
(p<.01).
o A predicted average reduction of losses due to accidents of $187.10 three months later
(p<.01).
o A predicted average reduction of losses due to accidents of $27.77 four months later
(p=.536).
 Note that the effect of training_lag4 is not significant. This means that safety
training four months ago has no effect on current losses due to accidents. This
suggests that the company should hold safety trainings every 4 months.

10
Example: The effect of safety training on accidents
 The effect of safety training on losses due to accidents has a hump-shaped
pattern over time. The immediate effect is rather small ($132), then it
increases to more than $400 for the next two months, then it falls to $187
after 3 months, then it drops to about 0.
 Increasing safety training in a given period will not only reduce predicted
average losses immediately, but also for the next 3 months:
 The combined benefit of increasing the safety training of each worker
with one hour equals
$131.99+$449.86+$422.52+$187.10+$27.77=$1219.24 in the month in
which the training is carried out and in the subsequent 4 months.
 Note: the conclusions are only correct if the assumptions of the model are
satisfied. Besides the Gauss-Markov assumptions the model assumes that
o Lag length is correctly specified (i.e., , etc. are not significant)
o and are stationary.

11
Selection of lag length
 To select the lag length, we can use the following procedure
o fit a model with the maximum possible lag length that seems
reasonable
o test against H A :  qmax  0
o if is rejected we select as lag length, otherwise we drop ,
refit the model and test .
o Keep on dropping the lag order by one and re-estimating the model until the
coefficient of the highest lag is significantly different from 0.
 Applying this procedure in the example, we drop the variable
training_lag4 and refit the model.
 We then select lag length 3 because after removing training_lag4 all
regression coefficients are significant.

12
Example: The effect of safety training on accidents
Final model using lag length :

13
Exercise 1: The effect of education spending on
economic growth
 Development economists are often interested in the effect of education
spending on economic growth. They suspect that the positive effect of
raising education levels may take 5 to 10 years to manifest themselves in
higher growth rates.
 We consider annual data for a country from 1910 to 1995 on the variables
o = GDP growth (measured as % change per year)
o = Education spending (measured as $ per child under age 16)
 We fit distributed lag models with maximum lag length 10, and
subsequently drop lagged variables on the highest lag until we reject the
null hypothesis that the coefficient on the highest lag is 0.
 Using this strategy, we select lag length 6 for the final model.

14
Exercise 1: The effect of educational spending on GDP
For the model with lag length 6 we obtain the following output:

15
Exercise 1: The effect of educational spending on GDP

16
Exercise 1: The effect of educational spending on GDP
 Explain which distributed lag models were fitted to select the lag length
of educational spending.
 Interpret the output of the selected distributed lag model
o Discuss the goodness-of-fit of the model: compute
o Interpret the regression coefficients
o Discuss the combined effect of an increase in educational spending on the
current period and on future periods.

17
Solution: The effect of educational spending on GDP
 The following table describes the distributed lag models that were fitted
to select the lag length of educational spending. The table also shows the p-
value of the coefficient for the highest lag of educational spending. We
select because in this model is rejected.
Maximum lag Model p-value when testing
included vs H A :  qmax  0

10 .41

9 .26

8 .19

7 .14

6 <.001

18
Solution: The effect of educational spending on GDP
SPSS output Goodness-of-fit of the model

19
Solution: The effect of educational spending on GDP

The model explains 72% of the variation in GDP around its mean

After controlling for overfitting the model explains 69.4% of the variation
in GDP around its mean


The correlation between observed and predicted GDP equals .849

20
Solution: The effect of educational spending on GDP

The observed GDP growth values deviate from the true regression plane
by approximately .372%, on average.

21
Solution: The effect of educational spending on GDP
Assuming that no other changes to the educational spending program
are made, increasing the educational spending of each child by one dollar
in a given year is associated with
o An immediate predicted average increase of .009% of GDP growth (p=.36)
o A predicted average increase of .002% of GDP growth in the next year (p=.89)
o A predicted average increase of .000% of GDP growth two years later (p=.975)
o A predicted average increase of .031% of GDP growth three years later (p<.01)
o A predicted average increase of .040% of GDP growth four years later (p<.001)
o A predicted average increase of .059% of GDP growth five years later (p<.001)
o A predicted average increase of .047% of GDP growth six years later (p<.001).
 We see that increasing educational spending of each child younger than
16 by one dollar does not affect GDP growth in the current year, nor does it
affect GDP growth in the next two years. However, this increase has a
significant positive effect on GDP growth three, four, five and six years
later.
22
Solution: The effect of educational spending on GDP
 The total predicted average increase in GDP growth due to increasing the
educational spending of each child younger than 16 with one dollar equals
.009%+.002%+.000%+.031%+.040%+.059%+.047%=.188% in the year in
which the educational spending was increased and in the subsequent 6
years.

23
Exercise 2: Okun’s law
 In the economic model known as Okun’s law, the change in the
unemployment rate from one period to the next depends on the rate of
growth of output in the economy.
 represents the unemployment rate in quarter
 indicates the change in the unemployment rate from
one quarter to the next.

 indicates the growth of output in the

economy (i.e. percentage change in gross domestic product from one
quarter to the next).
 We use quarterly US data from 1985 to 2009 (98 time points).

24
Exercise 2: Okun’s law

 As changes in growth are likely to have a distributed-lag effect on the

change in the unemployment rate we use a distributed lag model to model
the relation between and

25
Exercise 2: Okun’s law
 We start with a distributed lag model with lag length :

 SPSS output of estimated coefficients:

26
Exercise 2: Okun’s law
 As the coefficient of is not significantly different from 0 (and also
has the wrong sign), we remove this predictor and fit a model with lag
length :

27
Exercise 2: Okun’s law
 Interpret the coefficients of the distributed-lag model.
 Compute the total effect of a change in output growth on the change of
the unemployment rate.

28
Solution
 Interpret the coefficients of the distributed-lag model.
 The final model reads:

 A one percent increase in the growth of the current quarter leads to a

predicted average immediate decrease in the change of the unemployment
rate of .20%.
 It follows from the final model that

 Hence, a one percent increase in the growth of the current quarter leads to
a predicted average decrease of .165% in the change of the unemployment
rate from the current quarter to the next quarter.
It also follows from the final model that

29
Solution
 Hence, a one percent increase in the growth of the current quarter leads to
a predicted average decrease of .07% in the change of the unemployment
rate from the next quarter to the following quarter.

 Compute the total effect of a change in output growth on the change

of the unemployment rate.
A one percent increase in the growth of the current quarter leads to a
predicted average total decrease of .202+.165+.070=.437% in the change of
the unemployment rate for the current quarter and the next two quarters.

30