Professional Documents
Culture Documents
infrastructure investment
and economic
development
A two-way causal relationship between One of the most difficult questions in exploring the relationship between
tebcommunications infrastructure in- telecommunications and economic development is determining a causal connec-
vestment and economic development, tion: While we observe that telecommunications development and general
eetabiished for the US economy in pre-
vkus anaiysis, Is tested at the more economic development often proceed together, is it telecommunications invest-
iocaiized state and sub-state level and ment that promotes economic development, or economic development that
for two specific sub-categories of teie- creates demand for more telecommunications services?’
communications infrastructure invest-
ment- central office equipment, and Researchers have investigated the relationship between telecommunica-
cable and wire. For time series of these tions investment and economic development for almost 30 years2
two sub-categoriee of teiecommunica-
tfons investment complied for the Com- Studies have found that highly developed national economies are
monwealth of Pennsylvania and two correlated with highly developed telecommunications infrastructure.
countries within Pennsylvania this However, it was only recently that statistical techniques sufficiently
analysis tests two causal hypotheses
consistent with the US national-level rigorous to disentangle the relationships began to be applied to these
analysis: First, the level of economic issues. The evolution of this research, employing state-of-the-art statis-
actlvity at any point in time is a reliable
tical techniques, has now confirmed at the national economic level the
pmdictor (‘cause’) of the amount of
telecommunications investment at a ia- existence of a feedback process in which telecommunications invest-
tef point In time. Second, the amount of ment enhances economic activity and growth, while economic activity
t&communfcations investment at any and growth stimulate demands for telecommunications infrastructure
poht in time is a reliable predictor
(‘cm~se’) of the level of economic activ- investment.3
ity at a later point in time. What has not been answered, prior to the present study, is whether or
not this interdependent relationship between economic activity and
Francis J. Cronin is Managing Director of
DRiIMcGraw-Hill, 24 Hartwell Avenue, telecommunications infrastructure investment at the national level also
Lexington, MA 02173, USA (Tel: + 1 617 holds at the state and sub-state levels - and if so, how it occurs. Does it
660 6424; fax: +1 617 660 6463). Edwin result from more highly developed economies needing, and being able
B. Parker is Presidentof Parker Telecom-
munications, PO Box 402, Gleneden to afford, more communications? Is telecommunications investment a
Beach, OR 97366, USA (Tel: +l 503 764 productive stimulus contributing to economic growth, or merely a
3056; fax: +l 503 764 3059). ElisahethK. consequence of that growth? Or is it both? Does the relationship
Colieran and Mark A. Gold are Senior
Associateswith DRI. between telecommunications investment and the economy vary be-
tween the national and state level, or between the state and sub-state
This study was originally performed by level? Does the relationship vary across sub-state areas (urban versus
DRYMcGraw-Hillin coniunctionwith Par-
ker Telecommunicationsfor a consortium rural, high-income versus low-income)? And finally, does the rela-
of Pennsylvaniatelephone companies. tionship between investment and telecommunications infrastructure
Data
office equipment and in cable and wire. To control for inflation, values
have been adjusted to 1977 constant dollars.
At the county level, our original intent was to conduct the analyses on
data provided by LECs for a diverse set of many counties. This
approach would have allowed us to test whether the relationship
between economic activity and telecommunications investment varies,
based on a community’s characteristics (industrial composition, popula-
tion density, socioeconomic characteristics, etc). However, data limita-
tions stemming from the perpetual property records accounting system
maintained by most of the LECs prevented us from compiling sufficient-
ly long historical records of investment activity by local area. As a result
our analysis is limited to the information provided by GTE in York and
Somerset counties, for which decades of historical investment data for
the individual local exchanges are available. To control for inflation,
values have been adjusted to 1977 constant dollars.
Additionally, DRI’s sub-state analysis could not extend below the
county level as a result of the statistical requirements that a substantial
amount of history be available for each input to the tests. Data on
economic activity below the county level have not been compiled in a
consistent time series prior to 1988. For this reason, analysis at the
exchange level was not possible unless the exchange borders were
closely aligned with a county. A limited set of counties, including
Somerset and York, exhibited this characteristic. DRI’s sub-state
analysis focused on these counties.
Statistical background
13J. Geweke, ‘Inference and causality in Causality literature
economic time series models’, in Z. Gri-
liches and M.D. Intriligator, eds, Handbook The nature and estimation of causal relationships is covered extensively
of Economics, Vol- 2, North Holland- in economic literature.r3 A survey of broad philosophical issues and
Elsevier. Amsterdam. 1964, Ch 19. perspectives may be found in the Journal of Econometrics. l4 A subset of
“See, & A. Zellner; ‘Causality and caus-
al law in economics’, Journal of Eco- this literature, pioneered by Granger,” concerns the application of
nometrics. Vol 39. No l/2, Seotember/ certain forms of time series analysis to establish evidence of predictabil-
October 1966, pp 7-21; and-C.W:J. Gran- ity that would be consistent with causation. This avenue of thought has
ger, ‘Some recent developments in a con-
cept of causality’, Journalof Econometrics, developed through time and is effectively surveyed by Geweke.16 The
Vol 39, No l/2, September/October 1966, statistical procedures involved in applying ‘Grange? causality or similar
pp 19*212. tests now constitute a subject that appears regularly in applied eco-
I%. Granger, ‘Economic processes in-
volving feedback’, information and Con- nometrics textbooks.i7
trd, Vol 6, 1963, pp 2646; C. Granger, Specific mathematical specifications are required to depict the
‘Investigating causal relations by eco- assumed relationships between the phenomena being examined. Two
nometric models and cross spectral
methods’, Econometrica, July 1969, pp standard statistical tests have been judged appropriate for testing the
424436. direction of causality in small samples. These are the Granger test and
“Geweke, op cif, Ref 13. the Modified Sims test.18 Granger defines causality as: x causes y if, and
“See. for examole. E. Berndt. The frac-
tice of Econothethcs, Addison-Wesley, only if, y is better predicted by using the past history of x than by not
Reading, MA, 1991, Ch 6. doing so, with the past of y being used in either case. The Modified Sims
‘*For a detailed discussion of the prop- test is very similar to Granger’s. The Modified Sims test finds x causes y
erties of causality tests when used with
small samples, as in the present study, see if, and only if, a specification including earlier values of x and earlier,
Guilkey and Salemi. They conclude that current and future values of y is a better predictor of the time series of x
both the Granger and Modified Sims proc- than a specification including earlier values of x and earlier and current
edures provide accurate estimates with
smell samples. 0. Guilkey and M. Salemi, values of y (see the Appendix for a complete discussion of these
‘Small sample properties-of three tests for causality tests).
Granger-causal ordering in a bivariate Of course, we would prefer to specify the ‘true’ relationship. Since, in
stochastic system’, Review of Economics
and Statistics, November 1962, pp 666 general, we do not have such knowledge, we employ standard alterna-
660. tive specifications (ie Granger and Modified Sims). DRI employes the
Granger test and the Modified Sims test because they are the standard
tests used in causality analysis and both possess acceptable small-sample
properties. It has been noted by researchers that these two tests, while
designed to examine the same phenomena, have slightly different
representations of the time sequencing of the relationships and, there-
fore, may generate similar or different results. In fact, when examining
a relationship between two variables, one test may indicate a causal
relationship during a given time frame or for a specific geographic area
while the other test does not; for a different time frame or geographic
area the latter test might indicate a causal relationship while the first test
does not. This does not invalidate the tests; rather it indicates that the
‘true’ relationship is more closely approximated by one test for a given
time interval or geographic area and by the second test for other time
intervals or areas.
We report results for both tests in order to increase the robustness of
our research. It is not necessary that both tests nor all lag structures
indicate a causal relationship - only that one of the tests and one of the
lag structures give test results greater than the critical value chosen.
Estimation issues
In conducting causality tests, it is important to address several estima-
tion issues that can have important implications for the results of the
analysis. First, it is important to determine the appropriate relationship
between the variables across time. For example, does a change in
investment in COE lead to a change in economic activity in the
following year, or does the effect show up two, three or four years later?
This time dimension is referred to as the lag structure of the variable.
(Should the test determine that COE is related to economic activity two
years later, economic activity is said to have a two-year lag structure.)
Statistical tests exist to determine the lag structure that should not be
considered in the analysis, based on possible biases in the behaviour of
the variables. I9 In the current study these tests indicate that at least a
two-year lag structure would be appropriate to remove any biases. Our
analysis uses two, three and four-year lag structures. While it is not
necessary to test more than one lag structure, the current analysis
includes these longer time frames to capture any latent impacts between
two variables that may take time to materialize.
‘% determine the appropriate lag struc- A second issue to be addressed in the estimation process is whether
ture, LaGrangian Multiplier tests can be the existence of trends in the variables will cloud the findings of the
run on the residuals of estimated autore- analysis. In other words, as both the level of economic activity and the
gressive processes for different variable
series. For a discussion of the LaGrangian level of telecommunications investment have increased over time it is
Multiplier (LM) test, see T.C. Mills, Time possible that this upward trend could mask the true relationship between
Series Techniques for Economists, Cam- the change in one variable and the related change in the other variable.
bridge University Press, Cambridge, UK,
1990, pp 147-l 50. DRI conducted statistical tests to determine whether trends in the
*% identify the presence of trends (unit variables exist, and the appropriate solution.*’ These tests uncovered
roots), the Dickey-Fuller test can be ap- evidence of trends in certain variables which were removed by using, for
plied. The Dickey-Fuller test applies data
to a specified autoregressive process to each year, the change in the variable from the previous year (referred to
ascertain whether a trend exists in the as the first difference of the variable) rather than the actual levels.
data; if a trend exists, the estimation of The technical specifications and tests used in the analyses are
parameters using ordinary least squares
will probably be biased. For more informa- described more fully in the Appendix.
tion on the Dickey-Fuller test, see D. Dick- Another important issue is the selection of the level of statistical
ey and W. Fuller, ‘Distribution of the esti- significance at which the hypothesis that telecommunications invest-
mator for autoregressive time series with a
unit root’, S&tis&a~ Association, 1979, pp ment causes economic growth can be satisfied. The causality tests are
427431. actually constructed to test the reverse hypothesis, ie that there is no
‘99% confidence level; b95% confidence level; Three-yearlag F value (df = 3,20) F value (df = 1.19)
‘90% confidence level. ‘W stands for ‘degrees Income causes investment NS NS
of freedom’, a measure of the number of uncon- Investment causes income 6.4@ 6~35~
strained cases used in testing statistical signifi- Four-yearlag F value (df = 4,18) F value (df = 1 ,I 8)
cance. The ‘F value’ is the statistic used to Income causes investment NS 3.41=
measurethe significanceof the relationship.‘NS’ Investment causes inwme 4.18’ 4.15c
stands for not significant.
Findings
State level
This section discusses the relationships between two significant compo-
nents of telecommunications investment and economic activity at the
state level. The two investment variables are net investment in central
office equipment and net investment in cable and wire. Table 1 presents
the results of the analyses of investment in central office equipment.
As the table shows, for the two- and three-year lag periods we find no
significant effect of changes in total personal income in Pennsylvania on
central office equipment investment. One of the two tests (the Modified
Sims) shows a significant effect with a four-year lag period. This result is
“Since numerous factors (eg data report- consistent with the long investment planning cycles used in the telecom-
ing errors, definitional problems and speci-
fication errors) can affect the statistical munications industry for new central office equipment investment.
results - ie obscuring a relationship that in Installation and replacement of telephone switching equipment have
fact exists - we cannot ‘technically’ con- been more typically determined by multiyear investment budgeting
clude that no relationship exists, only that
we fail to reject the null hypothesis of no cycles than by quick response to market demands. Furthermore, the
relationship. longer time lag between economic activity and investment may reflect
Table 2. The relationship between cable and wire telecommunicetionr plant investment and
total personal income In Pennsylvania.
County level
The intent of the county-level analysis is to determine whether the
“8. Egan and L.D. Taylor, Capital Budget-
statistically significant causal relationship between telecommunications
ing for Technology Adoption in Telecom- investment and economic growth identified in previous studies at the
munications: The Case of Fiber, Center for national level, and in this study at the state level, can be observed at
Telecommunications and Information Stu-
dies, Research Working Paper Series,
more localized geographic locations. Additionally, if such relationships
Columbia Universitv, New York, April can be determined at the county level, it would be useful to explore
1989; comments by Lee M. Bauman,’ in potential differences in the impact of telecommunications development
‘Pacific Telesis - surviving and striving’,
Telephony, 9 September 1991, p 28.
on counties varying in size and industrial composition. However, the
=Ff. Cohen, The Economic impact of scope of the analyses is limited to the two counties - Somerset and York
Broadband Communications on the US - for which data could be collected for the appropriate time frame.
Economy and on US Competitiveness,
Economic Strategy Institute, Washington,
Somerset and York are the state’s seventh and ninth largest counties
DC, 1992. (respectively), based on a square mile measure. These counties differ in
Source: US Department of Commerce, Bureau Resident population (‘000) 11 862.2 11 777.0 11 905.5 12 050.1
of Economic Analysis, and DRINcGraw-Hill. Per capita income ($900) 9.9 13.6 18.7 19.6
Table 6. The relatlonshlp between central office te4ecommunkatlons plant lnveetment and total
employment In York County
test, but not the Granger test, indicates that the conclusion that
investment in outside plant causes economic activity can be accepted at
the 90% confidence level. The causal relationship of economic growth
to investment in outside plant is not detected at the county level, even
though the Modified Sims test demonstrates such a relationship at the
state level.
The statistical tests for Somerset County for both COE (see Table 8)
and outside plant (see Table 9) fail to detect significant relationships in
either direction for all lag structures of COE and for the two- and
three-year lag structures for outside plant investment. However, for the
four-year lag structure the Granger test finds that investment in outside
plant causes economic activity at the 90% confidence level. This finding
concerning the causal relationship of outside plant to economic growth
in Somerset County provides some support for the results in York
County.
The positive impacts of central office equipment on employment in
York County and outside plant investment on employment in both York
and Somerset Counties are particularly important findings. These
results are consistent with the significant negative correlation between
unemployment and telecommunications infrastructure found by Parker
and Hudson in rural counties in Oregon and Washington. The Parker
and Hudson findings were comtemporaneous and therefore could not,
by themselves, indicate causal direction. The York County data pre-
sented here show that investment in both central office equipment and
outside plant causes changes in employment activity for the three lag
structures tested, while the Somerset County analysis finds that invest-
ment in outside plant causes economic growth.
The finding of a significant result for outside plant investment in both
counties, but not in the state as a whole, is consistent with a finding of
the 1980 Hardy study that the greatest economic benefits were obtained
Table 8. The relationship between central oftioe tekommunloationa plant Investment and total
employment in Somerset County.
Tabk 9. Tha ralatkxnahlp bahvaan outside office talacommunlcationa invaatmant and total
employment in %meraat County.
Appendix
Specifications and tests used in the analysis of causality
Standard regression analysis is one Both the Granger and the Modified any of a range of tests (eg an F test on
technique economists use to deter- Sims tests seek to determine evidence the joint significance of the lagged A
mine the statistical relationship be- of the causal role of A on B through an variables), then one concludes that A
tween two data series. Within this examination of comparative time ‘Granger causes’ B.
structure, however, causal rela- series representations. In the case of The Modified Sims test is a bit less
tionships are presumed. In order to the Granger test, an autoregressive intuitive. Again one considers
investigate, on a statistical basis, evi- representation of B is specified. The whether A causes B by considering
dence of causal ordering, special representation must include at least alternative time series representa-
methods are required. Therefore eco- enough lagged variables to reduce the tions. Here, however, one analyses a
nomists use other techniques to deter- resulting series of residuals to ‘white time series representation of the
mine causality. These tests seek to noise’ (ie no serial correlation). Furth- potentially causal variable A. The
determine the evidence of a causal ermore, B must be represented as a ‘base case’ representation involves an
relationship between two data series stationary autoregressive process. If autoregressive representation of A
by assessing whether the inclusion of the level B cannot be represented as a combined with lagged and contempor-
past information about one data series stationary process, an autoregressive aneous values of the variable B. As in
enhances the predictability of the representation of a first difference of the Granger test, this representation
second. B is considered, etc, until a stationary must be stationary and enough lagged
Statistical tests of causality were autoregressive representation may be variables must be included to render
proposed by Granger,25 building on made. A resulting autoregressive rep- the residuals series as white noise.
earlier work by Weiner.26 A thorough resentation is then considered against This base representation is compared
review of theoretical and practical another representation for the predic- to an alternative representation which
issues is presented by Geweke.” The tion of B. This representation includes includes not only the autoregressive
analysis of causality performed in this all the autoregressive terms of the first terms of A and lagged and contempor-
study involves the application of a but includes also lagged values of the aneous values of B but forward values
Granger test and a similar test prop- variable A. If the inclusion of the prior of B as well. A is said to ‘cause’ B if
osed by Sims,‘* later modified by history of A enhances the predictabil- the inclusion of the forward values of
Geweke. ity of the variable B, as indicated by B increases the predictability (as mea-