Adesanya dissagregation of data corrected

Mathematical Theory and Modeling www.iiste.org
ISSN 2224-5804 (Paper) ISSN 2225-0522 (Online)
Vol.3, No.1, 2013

Disaggregation Comparison Analysis of Economic Time Series
Data
T.O. OLATAYO and K.K. ADESANYA

Department of Mathematical Sciences,Olabisi Onabanjo University, Ago-Iwoye, Ogun State,
Nigeria. Otimtoy@yahaoo.com and adesanyakehindekazeem@yahoo.com.

Abstract - Econometrics modeling often implies the use of a number of time series data,
some of which could be available at lower frequency and therefore, it could be convenient to
disaggregate these data to high frequency form instead of estimating with a significant loss of
information. The main aim of temporal disaggregation is to derive an estimate of the
underlying high frequency (HF) observation of an observed low frequency (LF) time series.
The method adopted by Chow-Lin, Fernadez, Litterman, (static model), and Santo
Silvacardoso (dynamic model) were used to make comparison in disaggregation economic
analysis of time series data. The parameters employed in this study are Autoregressive test,
Correlation and standard Deviation.
Result of analysis in low frequency form (Annual) confirmed that Chow-Lin has the
correlation value of 0.9914, Fernandez has the correlation value 0.9914, Litterman has the
correlation value 0.9701 and Santo Silvacardoso has the correlation value of 0.9914. Result of
analysis in high frequency (monthly) confirmed that Chow-Lin has the correlation value of
0.9899 and Standard Deviation of 212850.48, Fernadez has the correlation value of 0.9899
and Standard Deviation of 78553.54, Litterman has the correlation value of 0.9997 and
Standard Deviation of 789109.18 while Santo Silvacardoso has the correlation value of
0.9898 and Standard Deviation of 2337.24.
The performance indicators of disaggregated values for Chow-Lin, Fernandez,
Litterman being a static model and Santo Silvacardso being a dynamic model, annual and
monthly data confirms that the results of analysis are very good with high correlation figures
while the ability of the estimated monthly data captured the true dynamic of the series. Santo
Silvacardoso being a dynamic model preformed better with minimum standard deviation
while Litterman technique being a classic and static model preformed poorly in
disaggregating to high frequency form.

Keywords: Disaggregation, Low frequency data, High Frequency Data, Static Model,
Dynamic model.

119

Vol.3, No.1, 2013

INTRODUCTION

Temporal disaggregation methods play an important role for the estimation of
short term economic indicators. The need for temporal disaggregation can stem from a
number of reasons, due to the high costs involved in collecting the statistical
information needed for estimating national accounts, could decide to conduct large
sample surveys only annually.
Consequently, quarterly (or even monthly) national accounts could be obtained
through an indirect approach, which is by using related quarterly (or monthly) time
series as indicators of the short term dynamics of the annual aggregates. Econometric
modeling often implies the use of a number of time series, some of which could be
available only at lower frequencies, and therefore, it would be convenient to
disaggregate these data instead of estimating, with a significant loss of information, the
complete model at lower frequency level. Chow and Lin (1971).
Different strategies have been developed to get an estimate of the
autoregressive parameter from the lower frequency data is the most applied procedures
are those proposed by Chow and Lin (1971), Bourney and Laroque (1979), other
authors have proposed atternatives restriction on the DGP (Data Generation Process)
of the disturbance series in the High Frequency regression model. Fernadez (1981)
proposes a random walk model for the disturbances that avoids the estimation of
parameter at the High Frequency level. Litterman (1983) refines the Fernadez solution
by introducing Markov process to take account of serial correlation in the residuals.
Moauro and Savio (2002) encompasses the three solution, generalizing the restrictions
in the class of ARIMA (Auto Regressive Integrated Moving Average) processes.
Recently, some authors have proposed techniques based on dynamic regression
models in the identification of the relationship linking the series to be estimated and
the related indicators. Aadland (2000).
In this study, efforts will be geared towards disaggregation of low frequency
time series data into high frequency time series data, through the comparison of static

120

Vol.3, No.1, 2013

and dynamic models as propounded by Chow Lin, Litterman, Fernandez and Santo
Silvarcardoso.

MATERIALS AND METHODS

The disaggregation of low frequency data (annual) to high frequency data
(monthly), thereafter using both the low and high frequency data. Different
strategies that have been developed to get an estimate of the autoregressive
parameter from the lower frequency data, the method adopted by Chow-Lin,
Fenandez, Litterman and Santo Silvacardoso were used to examine the
performance indicators of estimate of private consumption expenditure.
(a) The Chow-Lin Model
The Chow-Lin (1971) disaggregation method is based on the assumption that yt
can be represented by a linear regression model with first order autoregressive errors;

with and )
The model is thus a particular case with scalar system matrices Z=1, T= , H=1.

As far the initial coordinates are concerned as αt is a stationary zero mean AR(1)
process, it is assumed that the process applies since time immemorial, giving α1 ~
which amounts to setting α =0, W1=0, and H1=(1- )-1/2
If some elements of xt are non stationary, the CL model postulates full
cointegration between them and the series yt.
Deterministic components (such as a linear trend) are handled by including
appropriate regressors in the set xt e.g. by setting xt, and writing ;

with the first two elements of B being denoted µ and alternatively, they can be
accommodated in the transistor equation, which becomes
αt = αt-1 + M + gt + εt.

121

Vol.3, No.1, 2013

The state space form corresponding to this case features Wt = (1.t.01) for t > 1,
whereas W1 = [ (1-О)-1(1-2О)/(1-О)2, 01]. The first two elements of the vector of
exogenous regressive xt are zero, since M and g do not enter the measurement
equation.
(b) The Litterman and Fernadez models
According to the litterman (1983) model, the temporally disaggregated process
is a regression model with ARIMA disturbances.
yt = x1t β + µt , Δµt = ОΔµt-1 + εt.
Litterman explicitly assumes that the vt process has started off at time t=0 with
µ = Δµo, (litterman 1983). This is usually inadequate, unless the set of indicator
o

include constant (which would capture the effect of initial value); the inclusion of a
linear trends amounts to allowing for non-zero drift in the ARIMA process.
The Fernandez (1981) model arises in the particular case when О=0 and thus µt

is a random walk.
The state space representation is obtained by defining the state vector and
system matrices as follows;
Αt = µ Z1= (1,1). T = 1 1 H= 0
t-1
Δµt 0 О 1

The Litterman initialization implies µ = µ + Оµ +
1 0 o ε1 = εt which is implemented
casting; 0
α1 = 0, W1= 0, H1 = 1

Alternatively including µ-1 in the vector β as its first element, in which case xt
features a zero element in first positive, and assuring that the stationary process has
started in the indefinite past, the initial conditions are:

1 01 1 0
W1 = 0 01 H1 = +
O 1
This follows from writing:

122

Vol.3, No.1, 2013

and taking Δµo ~ N(0, δ2/(1-Ø2), ε1~N(0,δ2). The diffuse nature arises from the non-
stationary of the model.
It should be noticed that in this second setup we cannot include a constant in xt,
since this effect is captured by µ-1.
Finally, the ARIMA process can be extended to include a constant and a trend
in Δµt = Δµt-1 m + gt + εt; the parameters m and g are incorporated in the vector β and
the matrices W1 and Wt are easily extended; for instance, if β = (µ-1, m, g, β12)1 where
β2 corresponds to the regression effects affecting only the measurement equation.
(c) Santo Silvacardoso Model
Santo Silvacardoso considered a slightly different representative of the
underlying high frequency data which emplicitly takes into account the presence of the
tagged dependent variable;
(1-OL) yt.µ = xt.µ βt.+ εt.µ,
t=1 µ = 2, … s
The solution for yh is given by t=2,… T µ = 1, 2 … s

Yn1 A1A1 T12 -1
A1Zn + Ci ( C1 (AiAi)-1 Ci)-1 yi – C1n
=
where Yn2 T21 A2A2 A2Zn + C21 (C2 (A12A2)-1Ci)-1 (y12 – C2 (A12 A2)-1 A2Z*n

T12 = A1A2 -C1 (C1(A1A1)-1C1) C1(AiA1)-1 AiA2
T21 = A2A1 -C21 (C2 (A2A2)-1 C2(A2A2)-1 A12 A1
The method of samtos silva and Cardoso. For notation convenience, let t= s(t-1)
+ µ be the interm running on the periods and re-write as follows;
Yt = Oyr-1 + Xr β + εT = 1,….n
t-1 t-1
Yt = ( Σ Oi xi-1) β + ØT yo + ( Σ Oi εT-1)
i=0 i=0

RESULTS
The evaluation of the results were made with respect to both Low and High
Frequencies data, with the results that examined ‘economic reasonableness’ of the annual
and estimated monthly regression model and, eventually, by comparing different estimates. A
number of estimates have been calculated according to different specifications, both in the
original data and in log-transformed form.
Summary of the results obtained in disaggregation comparison of economic time
series data of annual Nigeria GDP from CBN (1981-2009):

123

Vol.3, No.1, 2013

Table 1: Unit roots tests for annual Nigeria GDP (1981-2009) :low frequency
Intercept and trend
statistic None Only intercept both
Levels
ADF(1) 0.144685 0.123473 0.041476
(0.023166) (0.027735) (0.045681)
PP 0.144685 0.123473 0.041476
(0.023166) (0.027735) (0.045681)
First difference
ADF(1) 0.034904 -0.071382 -0.448537
(0.119209) (0.145472) (0.255566)

PP -0.282981 -0.427759 -0.995286
(0.136712) (0.163720) (0.238537)
Log-levels
ADF(1) 0.015485 -0.004487 -0.229935
(0.002449) (0.016359) (0.130912)
PP 0.015485 -0.004487 -0.229935
(0.002449) (0.016359) (0.130912)
First difference in logs
ADF(1) -0.303820 -0.835713 -0.828735
(0.140703) (0.199238) (0.202855)
PP -0.303820 -0.835713 -0.828735
(0.140703) (0.199238) (0.202855)
Mackinnon 5% critical values for rejection of hypothesis of a unit root in
parentheses. PP test statistics have been calculated using 3 lags truncation for Bartlett
Kernel (Newey and West, 1994). From the table of low frequency of GDP, the study
reported that ADF and PP at level with their respective values greater than (-3.34)
Mackinnon 5% , thus unit root is present, therefore annual GDP is cointegrated, hence
no error correction model. At first difference, ADF and PP values are each greater than
5% Mackinnon hence there is present of unit root, therefore there exist cointegration.
The study reported at log level both ADF and PP results shows that the GDP
cointegrated at both lag 1 and lag 3 for ADF and PP respectively since their respective
values each is greater than Mackinnon 5% . Likewise, we found out that ADF and PP
indicate cointegration since their values each is greater than Mackinnon 5%. The
distributed lagged model specified for their relationships were stable for control of
action and prediction.

124

Vol.3, No.1, 2013

Table 2 Unit roots tests for annual Nigeria PCE (1981-2009) :low frequency
Intercept and trend
Statistic None Only both
intercept
Levels
ADF(1) 0.132804 0.106901 -0.007913
(0.037072) (0.045731) (0.079247)
PP 0.132804 0.106901 -0.007913
(0.037072) (0.045731) (0.079247)
First difference
ADF(1) -0.443656 -0.872442 -1.202197
(0.228404) (0.197424) (0.202624)
PP -0.688948 -0.872442 -1.202197
(0.187493) (0.197424) (0.202624)
Log-levels
ADF(1) 0.016026 -0.024149 -0.477070
(0.005752) (0.036345) (0.176411)
PP 0.016026 -0.024149 -0.477070
(0.005752) (0.036345) (0.176411)
ADF(1) -0.988757 -1.320202 -1.321400
(0.196094) (0.189513) (0.192816)
PP -0.988757 -1.320202 -1.321400
(0.196094) (0.189513) (0.192816)

Kernel (Newey and West, 1994). From the table of low frequency of PCI, the study
reported that ADF and PP at level with their respective values greater than (-3.34)
Mackinnon 5%, thus unit root is present, therefore annual PCI is cointegrated, hence
no error correction model. At first difference, ADF and PP values are each greater than
5% Mackinnon hence there is present of unit root, therefore there exist cointegration.
The study reported at log level both ADF and PP results shows that the PCI
cointegrated at both lag 1 and lag 3 for ADF and PP respectively since their respective
values each is greater than Mackinnon 5%. Likewise, we found out that ADF and PP
indicate cointegration since their values each is greater than Mackinnon 5%. The
distributed lagged model specified for their relationships were stable for control of
action and prediction.

125

Vol.3, No.1, 2013

Annual Nigerian National Account
The annual series aggregated and the chosen indicator is estimated above. As
confirmed by the unit roots tests (table 1 and 2 ). Moreover, the residual based ADF
test τe (table 3) is coherent with the hypothesis of cointegration.
Table 3 : Residual-based cointegration tests: ADF(1) on Nigeria national accounts
τe τα
Levels -2.632536 -2.553634
Log levels -2.700094 -2.644761
5% asymptotic -3.34 -3.78

Table 4: estimates of the auxiliary annual regression on Nigeria national account
(PCE)
Variants Α β Φ
1 48.686 0.670884 0.601003 -0.405550
(19.625) (0.164094) (0.129655) (0.202708)
2 0.684426 0.593367 -0.412778
(0.151929) (0.123654) (0.202077)
3 0.414404 0.392369 0.597102 0.032398
(0.141003) (0.058722) (0.060532) (0.190810)
4 0.367033 0.651491 0.151401
(0.066069) (0.065554) (0.187903)

Table 4 contains parameters’ estimates for dynamic models in both levels and
logarithms, and precisely according to variants 1, 2, 3 and 4 (that is, model in levels
with or without intercept, and model in logs with intercept or without intercept, which
in this last case turns out to be significant). Concentrating on the estimates obtained
through variants (1 and 2) with 3 and 4, we find that the Low Frequency estimated
values are very similar.

126

Vol.3, No.1, 2013

Table 5 Unit roots tests for monthly Nigeria GDP (1981-2009) :high frequency
Intercept and trend
Statistic None Only both
intercept
Levels
ADF(4) 0.000239 0.000194 -3.12E-05
(7.57E-05) (8.16E-05) (0.000125)
PP 0.009858 0.008258 0.002204
(0.000620) (0.000744) (0.001295)
First difference
ADF(4) -0.016152 -0.021480 -0.031404
(0.003631) (0.004167) (0.005053)
PP -0.008279 -0.010343 -0.006392
(0.007090) (0.008190) (0.010017)
Log-levels
ADF(4) 0.000353 -0.000939 -0.047182
(0.000218) (0.001156) (0.008042)
PP 0.001249 0.000787 -0.070416
(0.000458) (0.002457) (0.014402)
ADF(4) -0.242826 -0.252969 -0.252892
(0.030019) (0.030444) (0.030541)
PP -0.182021 -0.186895 -0.186789
(0.030246) (0.030558) (0.030660)

Kernel (Newey and West, 1994).
From the table of high frequency of GDP, the study reported that ADF and PP
at level with their respective values greater than (-3.34) Mackinnon 5% , thus unit root
is present, therefore annual GDP is cointegrated, hence no error correction model. At
first difference, ADF and PP values are each greater than 5% Mackinnon hence there
is present of unit root, therefore there exist cointegration. The study reported at log
level both ADF and PP results shows that the GDP cointegrated at both lag 1 and lag 3
for ADF and PP respectively since their respective values each is greater than
Mackinnon 5% . Likewise, we found out that ADF and PP indicate cointegration since

127

Vol.3, No.1, 2013

their values each is greater than Mackinnon 5%. The distributed lagged model
specified for their relationships were stable for control of action and prediction.

128

Vol.3, No.1, 2013

Table 6: Unit roots tests for monthly Nigeria PCE (1981-2009) :high frequency
Intercept and trend
Statistic None Only intercept both
Levels
ADF(4) 0.000383 0.000300 -9.59E-05
(0.000122) (0.000141) (0.000245)
PP 0.010299 0.008759 0.002948
(0.000990) (0.001219) (0.002222)
First difference
ADF(4) -0.032360 -0.037469 -0.043962
(0.004980) (0.005318) (0.005725)
PP -0.017307 -0.020109 -0.022043
(0.009975) (0.010740) (0.011679)
Log-levels
ADF(4) 8.58E-05 -5.18E-05 -0.000876
(3.10E-05) (0.000136) (0.000525)
PP 0.001616 3.52E-05 -0.003513
(0.000229) (0.001064) (0.003967)
ADF(4) -0.047802 -0.055155 -0.055161
(0.006439) (0.006844) (0.006855)
PP -0.025820 -0.029747 -0.029693
(0.012155) (0.013053) (0.013073)

parentheses PP test statistics have been calculated using 4 lags truncation for Bartlett
Kernel (Newey and West, 1994).
From the table of high frequency of PCE, the study reported that ADF and PP at
level with their respective values greater than (-3.34) Mackinnon 5% , thus unit root is
present, therefore annual PCE is cointegrated, hence no error correction model. At first
difference, ADF and PP values are each greater than 5% Mackinnon hence there is
present of unit root, therefore there exist cointegration. The study reported at log level
both ADF and PP results shows that the PCE cointegrated at both lag 1 and lag 3 for
ADF and PP respectively since their respective values each is greater than Mackinnon
5% . Likewise, we found out that ADF and PP indicate cointegration since their values

129

Vol.3, No.1, 2013

each is greater than Mackinnon 5%. The distributed lagged model specified for their
relationships were stable for control of action and prediction.

Monthly Disaggregation of Annual Nigerian National Account
The series disaggregated and the chosen indicators are estimated above. As
confirmed by the unit roots tests (table 4.4, 4.5 and 4.6a), both series are I(1).
Moreover, the residual based ADF test τe (table 4.6b) is coherent with the hypothesis
of cointegration.

Table 7 : Residual-based cointegration tests :ADF(4) on Nigeria national accounts
Τe Τα
Levels -3.363091 -3.364926
Log levels -3.450114 -3.483024
5% asymptotic -3.34 -3.78
*Davidson and Mackinnon(1993), Table 20.2 p.722

Table 8: estimates of the auxiliary monthly regression on Nigeria national
account (PCE)
variants Α Β Φ
1 17.074 0.029475 0.987704 -4.68E-05
(5.875) (0.007142) (0.005033) (0.003571)
2 0.033687 0.985958 1.19E-05
(0.007069) (0.005051) (0.000103)
3 0.013367 0.022680 0.980698 0.477684
(0.013570) (0.011506) (0.010481) (0.025889)
4 0.018026 0.986979 0.357109
(0.010490) (0.008317) (0.023430)

Table 8 contains parameters’ estimates for dynamic models in both levels and
logarithms, and precisely according to variants 1, 2, 3 and 4 (that is, model in levels

130

Vol.3, No.1, 2013

with or without intercept, and model in logs with intercept or without intercept, which
in this last case turns out to be significant). Concentrating on the estimates obtained
through variants 2, 3 and 4, we find that the HF estimated values are very similar.

Table 9: Disaggregation comparison indicators of private consumption
Expenditure

PCE Chow-Lin Fernandez Litterman SSC
Annual % changes
Correlation 0.9914 0.9914 0.9701 0.9914
Monthly % changes
Correlation 0.9899 0.9899 0.9997 0.9898
Standard dev. 212850.48 78553.54 789109.18 2337.24

Estimated (for Chow-Lin, Fernandez, Litterman (Static models) and Santo
Silva-Cardoso(Dynamic model), annual and monthly confirms that the results are
surely very good with high correlation figures while the ability of the estimated
monthly data capture the ‘true’ dynamics of the series. Santo Silva Cardoso being a
dynamic model performed better with minimum standard deviation while Litterman
technique a classical and static model performed poorly from the disaggregation of
Monthly national account data.

131

Vol.3, No.1, 2013

DISCUSSION OF RESULTS
Qualitative analysis very often has to rely on data whose observation frequency
is systematically lower than desired. However, as it would require enormous resources
to actually observe this process, most countries use annual estimates of economic
activity as the basis of their statistics. In contrast, many other variables such as money
stock and interest rates are available at a far higher frequency (and can often also be
observed more accurately). Nevertheless, researchers , policy makers and public, all
have genuine interest in high frequency information on low frequency data for
efficient and timely decision making. Therefore, statistical offices all around the world
work on providing temporarily disaggregated data to serve this aim.

CONCLUSION

The performance indicators of disaggregated estimates of private consumption
expenditure estimates (for chow-lin, Fernandez, litterman) being a static model and
santo silvacardoso being a dynamic model, annual and monthly data confirms that the
results of analysis are very good with high correlation figures while the ability of the
estimated monthly data capture the true dynamic of the series. Santo silvacardoso
being a dynamic model performed better with minimum standard deviation while
litterman technique being a classical and static model performed poorly from the
disaggregating of monthly national account data.

REFERENCES
Aadland D.M. (2000) Distribution and Interpolation Using Transformed Data,
Journal of Applied Statistics.
Bourney and Laroque (1979): An analysis of transformations (with discussion),
Journal of the Royal Statistical Society B 26, 211-246.
Chow G., Lin A.L. (1971): Best Linear Unbiased Interpolation, Distribution and
Extrapolation of Time Series by Related Series. The Review of
Economics and Statistics.

132

Vol.3, No.1, 2013

Di Fonzo T. (2003), Temporal disaggregation of economic time series: towards
a dynamic extension, European Commission (Eurostat) Working Papers
and Studies, Theme 1, General Statistics.
Fernandez, P.E.B. (1981), A methodological note on the estimation of time
series, the Review of Economics and Statistics, 63, 3, 471-478.
Litterman, R.B. (1983), A random walk, Markov model for the distribution of
time series, Journal of Business and Economic Statistics, 1, 2, pp. 169-
173.
Moauro F. and Savio G. (2002): Temporal Disaggregation Using multivariate
Structural Time Series Models. Fortcoming in the Econometric Journal.
Salazar, E.L., Smith, R.J. and Weale, M. (1997). Interpolation Using a Dynamic
Regression Model: Specification and Monte Carlo Properties, NIESR
Discussion Paper n. 126.

133

This academic article was published by The International Institute for Science,
Technology and Education (IISTE). The IISTE is a pioneer in the Open Access
Publishing service based in the U.S. and Europe. The aim of the institute is
Accelerating Global Knowledge Sharing.

More information about the publisher can be found in the IISTE’s homepage:
http://www.iiste.org

CALL FOR PAPERS

The IISTE is currently hosting more than 30 peer-reviewed academic journals and
collaborating with academic institutions around the world. There’s no deadline for
submission. Prospective authors of IISTE journals can find the submission
instruction on the following page: http://www.iiste.org/Journals/

The IISTE editorial team promises to the review and publish all the qualified
submissions in a fast manner. All the journals articles are available online to the
readers all over the world without financial, legal, or technical barriers other than
those inseparable from gaining access to the internet itself. Printed version of the
journals is also available upon request of readers and authors.

IISTE Knowledge Sharing Partners

EBSCO, Index Copernicus, Ulrich's Periodicals Directory, JournalTOCS, PKP Open
Archives Harvester, Bielefeld Academic Search Engine, Elektronische
Zeitschriftenbibliothek EZB, Open J-Gate, OCLC WorldCat, Universe Digtial
Library , NewJour, Google Scholar

Adesanya dissagregation of data corrected

More Related Content

What's hot (18)

Viewers also liked (20)

Similar to Adesanya dissagregation of data corrected (20)

More from Alexander Decker (20)

Adesanya dissagregation of data corrected