Evaluating performance of univariate, multivariate and VAR
approaches forecasting Swiss Gross Domestic Product using EViews
2016-100-EBC2086
Prof. Alain Heeq,
Martin Wischnath
16139285
Maastricht, 27th October 2016
Be Maastricht UniversityTable of Contents
1 Introduetion.
2: Dateien
3. Beonomic background ....
4 Univariate anal
4.1 Model specification
4.2 Model testing.
43° Volatility (GARCH-approach)...
Univariate forecast.
5.1 Building multivariate framework...
5.2 Cointegration Relationship ....
53 Simple conditional Model...
5A Conditional model forecast...
6 Multivariate VAR model ..
6.1 Estimation of VAR models...
6.2 Granger causality
6.3 Impulse responses.
64 VAR forecasting .
7 Forecast comparison
8 Conlusion.... 12
9 Appendices ses 13
9.1 Graphs...
9.2 TABLES vesen
10 Data Sources
11 Bibliography.1 Introduction
This paper tries to summarize our efforts on time series analysis during the course "Time
Series Modelling" by Prof. Alain Hecq. It should in particular accompany the research project
we conducted during the course, namely the model specification for uni- and multivariate data
analysis as well as the application and evaluation of different forecasting techniques.
‘Throughout our work we will make use of software package Eviews 9, which is specialised
‘on the type of time series work we will perform
For our research project, we start with three univariate time series of GDP data of France, the
United Kingdom and Switzerland and perform simple univariate analysis and construct a
multivariate model thereof. Additionally, we will build a vector auto regressive model (VAR)
of our three time series. We will also employ techniques to identify relationships between
time series, as we discussed throughout the course, in particular tests on cointegration and
Granger-causality. Finally, we will perform an out-of-sample forecast for the last 12 quarters
of the sample, namely from 2012 to 2014, to compare and asses the overall quality of our
models. To examine the quality of the forecasted time series we
mainly rely on Root
mean squared errors (RMSE) as an indicator for forecasting precision.
2 Data
‘The data used during this research was provided by the statistical agency of the European
Commission eurostat. We will use Gross Domestic Product data from the national accounts
database. The time frame we will examine is set from the first quarter of 1985 until the fourth
quarter of 2014 with a total number of 120 observations. For better comparison we will use
real prices as of 2010 and also use seasonally adjusted data. In the case of France we perform
Seasonal adjustment using the X13 method manually, a technique that is also employed by
eurostat. Taking the exponential attributes of GDP data into account we will co
use of log transformed level data, which is plotted in
inue with the
igure 1. We will perform all our
statistical tests on a 5% confidence level.
3. Economic background
For our model specification and forecasting purposes we will start off with Switzerland Gross
Domestic Product as the main variable, During our research we want to examine whether the
3non-EU member state Switzerland's economic state can be forecasted using selected European
countries, namely by using two major national economies France and the United Kingdor
This could be a real-world application performed by the Schnweizerische Nationalbank, for
instance.
Even though Switzerland is not a member state of the BU, it is surrounded by European
countries that are and thus itis likely that their economic prosperity has a large influence on
the Swiss economy and vice versa in today’s globalized in interdependent trade system.
‘To examine the interdependencies of Switzerland and selected central European countries and.
the EU in total we compute the correlation matrix of GDP growth rates of these countries in
Table 1. It turns out that different from the expected the Swiss economic growth is not highly
correlated with its neighbor state Germany or the Netherlands as another central European
state, Different from that, France has a high correletion of 0.90 with the economic growth in
the EU as a whole and can therefore be seen as good proxy for the EU in total.
In the European landscape the United Kingdom in particular is vastly different from the
majority of other countries conceming the political and economie system. The anglo-saxon
country employs weaker regulation and features a deregulated financial sector and overall less
labor protection. This is why we can assume higher dynamic in the economic activity picking
up global economic periods of growth as well as decline in a faster way which could be seen
as an indicator for future economic activity in the stronger regulated and thus less flexible
center of Europe. Switzerland in fact, could be seen as being similar in both econnomic in
political dimensions to Austria, Germany and the Benelux area, All of these countries feature
the same system of a developed social welfare system and also high regulation on labor
markets
Additionally, both the UK and Switzerland rely on a strong financial sector which tightens the
inderdependencies between the two countries. This finding gets underlined as the United
Kingdom's economic growth has relatively low correlation with other central European states
‘could be seen as an external variable for our forecasting purposes.
Consequently, in this paper we try to build multiple models with the goal to examine the
influence of the EU, through France as a proxy, and the United Kingdom on Swiss GDP and
how their economic development can help forecasting the state of the Swiss economy.
Since we rely solely on the GDP and do not include other variables such as import or export
Statistic or indicators for economic prosperity we are aware that this can only represent a
simplified approach to real-world conditions,4 Univariate analysis,
4.1 Model specification
In our first approach on finding a suitable univariate model to correctly describe the data and
perform univariate forecasts we will employ an ARMA(p.q) approach.
Firstly, we will test on the order of integration of our time series, namely we will perform an
Augmented-Dickey-Fuller test (ADF) to test for the presence of a unit root. In our finding we
cannot reject the null-hypothesis of a unit root with an P-value of 0.324. As we have
confirmed to have a time series integrated of order one, we will continue using first
differences (GDP growth rates) of the data.
For the growth rates we perform a Wald coefficient test for the presence of seasonalities, We
find a P-value of 0.974 and thus cannot reject the null that the seasonality coefficients are
jointly insignificant. Thus, we will continue the univariate approach without the use of
seasonalities.
To find the best suitable ARMA model we take a look at the correlogram of the time series
and examine ACF and PACF values, We detect an AR(1) model as we see constant declining
correlations in ACF and only the first partial correlation being signi
int. Since the declining
ACF values are not undoubtedly to be identified we also consider an MA(2) model, which
tums out to have higher SIC (-7.445) and AIC (-7.538) values,
AY = a+ BAY + &%
Equiaion 1 Univariate AR(1) model
‘The finding of an AR(1)-model gets confirmed as we perform an automatic ARMA-model
selection following both Schwarz- and Akaike information criteria. The AR(1) model has the
lowest information criteria values in both cases, namely an $IC-value of -7.402 and an AIC
value of -7.476,
4.2. Model testing
‘To assess the overall quality of the selected AR-model we perform multiple misspecification
tests. To check for linearity we perform a RESET-test and receive a P-value of 0.0997 and
thus cannot reject the null of no misspecification on a 5% confidence level.Additionally, we test, whether we removed remaining autocorrelation completely. According
to the correlogram of the models residuals (Figure 5) we can conclude that the chosen model
performs well on removing autocorrelation as there are no significant ACP-values left.
‘Tests on heteroscedasticity and normality will most certainly be influenced by outliers present
in the data. According to the Boxplot we find three outliers: Beginning in 1999Q4 with a high
positive growth rate, due to the fact that the Swiss economy experienced a .boom’ during the
end of the 1990s ending a period of economic stagnation in the first half of the decade. This,
period from 1990 to 1997 characterized by low GDP overall change is also clearly identifiable
in Figure 6. The other two outliers are detected in 2008Q4 and the following quarter 2009Q1,
marking the impact of the global financial 1
and the following global recession period,
These outliers are also visible in the residuals of the univariate model (Figure 6). One can
notice that the periods of high growth or severe decline are in fact captured by the model,
even though high residuals remain, which is why we decided not to include Dummy-variables
to capture the outliers better as the main purpose of the model will be forecasting and the
models should remain simple and consistent.
Due to the aforementioned circumstances the null hypothesis of normel distributed residuals
gets rejected with a P-value of 0.00. The same is true for the test for heteroscedast
ity, due to
outliers and higher variance of GDP growth in the early 1980s. To address these issues we
will continue with the use of Huber-White robust standard errors, whenever possible, to
robustify against the observed heteroscedasticity.
4.3. Volatility (GARCH-approach)
We perform an additional ARCH-test for time varying volatility to further examine the
heteroscedasticity found in the data, The ARCH test for heterosceasticty gives a tvalue for
the lagged squared residual of 0.724, meaning that there is no significant volatility clustering
present,
4.4 Univariate forecast
We perform univariate out of sample forecasts for Swiss GDP in level for the period from
2012 to 2014, As visi
le in Table 3, the dynamic model has a RMSE-value of 835.55
compared to a RMSE of 377.04 in the static model. Intuitively, the static model which
forecasts while incorporating true realized values has a lower error.4.5 Model selection for explanatory variables
As we will need forecasting models for the proposed explanatory variables in the multivariate
model we will also identify suitable models for both French and British GDP. In both cases
‘we cannot reject null of a unit root with P-values in the case of France of 0.93 and for the
United Kingdom of 0.80. Accordingly, we will also work with growth rates of the French and
British GDP. respectively.
In the analysis of the correlogram we find identical patterns of decreasing AC and a PAC only
significant for the first lag and thus will choose AR(I) models for both variable:
Kae = Gy + BAX, cu4 + ByOutlier; + £4,
AXn = Oz + PrbXre-1 t+ Ere
Equiation 2. Univariate AR(1) models for UK and France GDP growth
We test for misspecification in the AR model of French GDP growth (X,): We find
significant heteroscedasticity (P-value 0.000) due to an increased variance and outliers in the
1980s. Since there is heteroscedasticity present we will continue with the use of Huber-White
robust standard errors whenever possible. We do include a dummy variable to correct for a
significant outlier in Q4 2008 and QI 2009. Doing so allows us to reach better performance in
misspecification tests, accordingly we cannot reject the null hypothesis of normal distributed
outliers with a P-value of 0.538. The Ramsey-RESET test for linearity cannot reject the null
hypothesis of no misspecification with a P-value of 0.243,
For the British GDP we find significant heteroscedasticity (P-value 0.000) due to a varying
variances throughout the sample. Including dummies for the identified outliers does not help
in this case. For the same reason we reject the null hypothesis of normal distributed outliers
with a P-value of 0,001. Since there is heteroscedasticity present, accord
gto the White-test
for heteroscedasticity (P-value 0.000) we will also continue with the use of Huber-White
robust standard errors in this case.
5 Multivariate models
5.1 Building multivariate framework
We are now looking for ways to improve the univariate model by including additional
variables that should help forecasting the Swiss GDP due to a relationship out of economic
theory. We therefore build a ,simple conditional mode!’ that employs the Swiss GDP growth
as the dependent variable and a matrix of variables as explanatory independent ones.
¥5.2 Cointegration Relationship
To test for the necessity to include an Error-Correction-Model (ECM) into the model we test
for a cointegration relationship between the time series, namely whether there exists a long-
run relationship bewteen the data. For that purpose we will examine the residuals of a
COINTREG-model, perform the Engle-Granger test and the Johansen-test for the presence of
a cointegration relationship.
Figure 7 shows the correspondent residuals of the COINTREG estimation for a long-run
relationship of the three time series. The residuals do seem to have a unit-root and not be a
white-noise process, The Engle-Granger test performed on the data confirms that finding as it
cannot reject the null hypothesis of the three time series not being cointegrated with a P-value
of 0,995. Thus, we do not believe having a joint long-term relationship between the three
variables.
To confirm the finding we perform a Johansen cointegration test between the three series and
get a similar result as we cannot reject the null of None’ cointegration relationships between
the three in any direction with a P-value of 0.L11.
Economically, this finding states that — at least on a statistical level ~ there is no long-run
relationship between UK and France and Swiss GDP in the examined time period.
5.3 Simple conditional Model
We now construct a conditional model consisting of past values of the dependent variable,
current and possibly past values of the explanatory variables and according to our prior
finding without ineluding an ECM term.
AY, = a + BiAYeoa + BAX + ByAXoe + &
Equiation 3. Multivariate conditional model
To estimate the correct lag length to be inchided in the model we perform an automatic lag
determination procedure, which sclects the ‘best’ model by AIC or SIC criteria, Both
information criteria lead to a model without any lagged independent variables. The chosen
model performs well on misspecification tests, as it removes nearly all autocorrelation in the
correlogram (Figure 9), which gets confirmed by applying the LM test for serial correlation.
resulting a P-value of 0.697, thus we cannot reject the null hypothesis of no autocorrelation
We can also see normal distributed outliers (P-value 0.340) and no heteroscedasticity (P-value
80.380). The RESET-test confirms the finding with a P-value of 0.647, thus we cannot reject
the null of no misspecification.
According to our findings we achieve good performance in the aforementioned.
misspecification tests with a relatively simple model. This is in line with our parsimony
approach to achieve good forecasting performance with a relatively simple model paired with
preferably low costs and computation effort.
5.4 Conditional model forecast
To perform forecasts on the conditional model we will also implement forecasted values for
the independent variables generated through the respective univariate models for France and
UK GDP growth from section 4.5.
We receive RMSE values for the static forecast of 473,33 and 574.08 for the dynamic
‘method. These errors are noticeably higher than the ones of the univariate model,
6 Multivariate VAR model
6.1 Estimation of VAR models
To better incorporate the explanatory variables chosen to forecast Swiss GDP we will
estimate a VAR model, which, differently from the simple conditional model, treats all
variables equally and forecasts them simultaneousely. We will estimate a VAR that
incorporates all data-scts, namely the GDP of Switzerland (Y), the GDP of France (X,) and
the United Kingdom (X,).
The fi
step of correctly specifying a VAR model is the correct determination of the number
of lags. According to the automatic lag determination that chooses models by information
critetia, a model with one lag is sufficient. Since in a VAR(1) model there is still a significant
amount of autocorrelation left, we decided to choose a two-lag model to achieve better
performance in removing autocorrelation, The resulting VAR(2) performs very well on the
correlogram (Figure 10).
We also test the resulting VAR-model for misspeci
ition. The null hypothesis of no serial
correlation cannot be rejected using the LM-test with a P-value of 0.67. We do however,
reject the null hypothesis of no heteroscedasticity. Additionally, on a 5% confidence level we
cannot reject the null hypothesis of normal distributed residuals (P-value of 0.097).Ye = (AY AX ep Xa, 0)!
AY = ce a AY Fa 2AK 2 Fa AK + WAX 2
$y DKe ea + MagAXne2 + Ere
Dae = ea + a Mea + AyD ag + ayaa Fa AN 2
H y5AXy 4 + O6AXo,c-2 + Ere
BNge = s+ ay, DYey + aypAVe 2 + dysAKi eo + ON
t ysAXz¢-4 + OypAXp cn2 + Ese
Equiation 4, VAR(2) approach forthe trivariate case
6.2 Granger causality
Analog to identifying a long-term relationship by testing for cointegration we will now
examine the dataset for short-term rel
nships by performing tests to identify Granger
causal relationships in the datasets. The identified Granger-causal relationships are displayed
in Table 2. We find that Switzerland’s GDP growth Granger-causes French GDP growth on a
5% significance level. The United Kingdom’s GDP growth Granger-causes both Swiss and
French GDP growth on a 10% and 5% significance level, respectively
Our findings support the hypothesis that the proposedly more dynamic "first-mover” anglo-
saxon economy can be used to forecast the central European countries’ economic activity,
6.3. Impulse responses
To further examine short-run interdependencies and ‘causal’ relationships between the
variables in the VAR we plot impulse responses of the three time series. The impulse
response function traces the effect of an exogenous shock of one variable and its effect on all
others. We tested for correlation in the residuals of the VAR and see a positive correlation of
the respective res
luals. To avoid a bias of correlated residuals we use Cholesky decomposed
impulse responses and try go abstract economic relationships from the computed impulse
responses.
Firstly, itis clearly vi
le that the growth of Swiss GDP reacts positively to a shoc!
British GDP for periods of lag two, three and four. The same is true for the growth rate of
French GDP and thus could confirm our hypothesis of an earlier response of UK GDP to
10global trends what could then be used to forecast French and Swiss GDP. Additionally, we
can detect a more immediate response of the growth rate of Swiss GDP to a shock in BU
(French) GDP growth, where we see significant impulses for the first two lags. The findings
also support the discovered Granger-causal relationships found in 5.2, namely the significant
relationship of the UK towards both France and Switzerland.
6.4 VAR forecasting
As we also perform forecasts using the VAR-model we receive RMSE-values for the static
model of 356.65 and for the dynamic model 452.76. The errors are lower than both the
univariate and conditional model's forecast, whether the differences in performance are
statistically significant will be evaluated in section 7.
7 Forecast comparison
Finally, we will evaluate the forecasting performance off all models outlined. Figures 12 and
13 show alll forecasts and compare them to the true outcome of Swiss GDP accompanied by
‘Table 3 which presents all forecasts with RMSE for both static and dynamic models.
Accordingly, the forecast with the lowest RMSE value for both static and dynamic forecasts
is clearly the VAR-approach. We will still compare it to the other forecasts through a test and
addi
nally add a combination of the VAR and univariate model.
‘Through averaging with weights gained through a regression on the real values we receive
two datasets of combined forecasts, For the dynamic approach we gain weights w, of 0.86 for
the univariate model and 0.14 for the VAR model, Analog we get static weights for the static
approach with 0.84 for the univariate and 0.16 for the VAR model.
We run the Diebold-Mariano test as a regression of the forecasting differences on a constant
using HAC standard errors. The results of this test are
ible in Table 4, which states the
correspondent t-stati
ic and P-value under the null hypothesis that the two forecasts are
identical. For the dynamic forecasts the VAR approach alone still delivers better performance
than the VAR combined with the univariate model. The Diebold-Mariano test confirms a
significant difference between the two. However, no such significant difference can be found
for the static combined against the static univariate forecasts. Since the RMSE are also
comparable, we must assume they are identical,8 Conlusion
In our work we specified a univariate, multivariate conditional and a VAR model to exeeute
various tests on the given time series and eventually evaluate the forecasting performance of
the respective models as well as a combined forecasting approach.
We started off with a correlati
little about how the economies are connected with one another. We then tested for a long-run
matrix between the three growth rates that tells us relatively
relationship between the three countries and rejected such hypothesis. However, we did find
short-run Granger-causal relationships that showed how the UK Granger-causes both France
and Switzerland. We conducted a finer examination of this issue as we plotted impulse
responses which uncovered how the UK influences both the EU (France) and Switzerland
after a certain time lag.
For forecasting purposes this relationship proved to be useful as the two-lag VAR-model
showed the best dynamic forecasting performance and is, supported by Diebold-Mariano
tests, signi
Si
be differentiated from the combined forecast. Since the univariate model can be estimated
mntly better than the other forecasting methods.
1, the simple univariate model shows a good performance, in the static ease it even cannot
with relatively low costs and computing power we want to highlight that it performed
extraordinary well.
As for the economic interpretation of these findings we can conclude that-with all the
aforementioned limitations present— there exist short- and medium-run interdependen:
between both the EU, proxied through France, and Switzerland as well as the UK and
Sei
etland in economic performance.
Finally, the employed software package Eviews proved to be well suited for this type of
research as it implements almost all tests and procedures natively.
129° Appendix
9.1 Graphs:
2
ot
o
02
Figure 2. Line plot of S
6 bs ‘30 'o2 ‘od “os “os Oo Od Od Oe Of WO 12 14
— France GOP log
— Swizerland GOP log
— UK GP fog)
Figure 1. Line plot of data sample from 1985 (QU) to 2014 (4)
‘Switzeriand GDP growth rate
Hath! i
Fe
‘oe bet “oo! “oe! “od Ge Ge 0 0d “od Oo oo WO 2
erland GDP growth rates from 1985 (QI) to 2014 (Qa)oll IM bia ch -
|} “th
oo es oo ad ode Be oo ad bv be be Wd
Figure 3, Line plot of French and British GDP growth rates from 1985 (Q1) to 2014 (Q4)
apne nite OLOGEUIT,
Ee nar tc 9
Sie eesti
ance oan carla any oe Ato i
ae) ‘esd Gost —Gtesar como
‘iinagoe Shir Seow
esti cn
Figure 4. Eviews output of univariate model
Sample: 198501 201108
Ines obsenatons: 107
Ost
bor. or” aaasa
5 c003 ‘ono? oaded
6 2006 ona ons
1-404 008. 124
8402” 0s 1271
3 0008 ocor 42472
Doss oasy tera
003. dott 403
Oxo. ot. dane
01a. 043. 526596
208. “008” ses
006i 005i seme
oro oa? 72652
Figure 5. Cortelogeamm of univariate model
1402
03
08
04
00
03
6688 9092 84 a6 8 00 02 ad
Residual —— Actual —— Fitted
08
Figure 6. Residuals of univariate AR(1) estimation
810
= tr
a
~08:
6 ‘aa ‘so ‘G2 “ow oe on bo wo od 08 08
— Residual — Actual — Fitea
10
"2 4
Figure 7. Residuals of multivariate COINTREG estimation
15Dependent Vr: (LOGS)
Motes lest Squses
Sara acted) 190509 209104
Ince abeeritins 106 ator aahsiens
ts etrstedari conten dan ore & cove
vrata Coticenk Si Ener _ataie Pros
DiLesiswT:1y) _oaa01"e —ogreeae — geToars oame
DLOGIUK)” —Daaeaes —aeesTY | aaTPOTe 040
DILOGR) §——oaraano—xeBez0 3457059 OUD
e Doma dovorie aaa ovat
Feauares 443088 Moen dopandener C008
Aajotes R-squorad 042885 SD eepencent vat tots
82 otrageeuon 04806. Akio rite ‘76n0
Sumsqisedwas Duets Semen aren ‘tee?
tg etn SISE907 Harman Qunnertor “7 T2Pt
Festa 2708513. uronic stat 2azosrs
Prob tlie) oo0000. wa tate 2ieeo3
Provalésbste) ono
Figure 8, Eviews output of multivariate mode!
Dow: 10/476 Time: 20:02
Sample: 198501 201104
Incas cbseratons: 108
2justod fo 3 dynamicregressors
PettalCoreiaton AZ PAC _OSut Prob
1 -001..001., 00373 9847
‘0078 0079 07200 98606
0077 0049 13908,
00:1 0009 14083,
0033 0022 15319
0.4143 0139 33873,
0.012 oot 39045,
201...004.. 3.9387
0.20 0099 50307
002, 001.. $5861 1048
1-000. 902. 538883. 0.890,
1 "0033 ‘0002 sia4s 0.925
1--018.,-0.43.. 8.0250 caee
fi
0060 0085 s47s7 (a6)
0166 0.170 11930 cae
0078 0101 12696 6605,
“Probabilites maynotbe valor his ecuaton specication
Figure 9. Comelogramm of multivariate modelFigure 11. VAR impulse response plot120,000.
118,000
116,000.
114,000
112,000
10,000)
108,000
106,000
104,000
Cw ee
a
‘Unieviate (@ynamic)
Muttsaiteeyramic}
— VAR tayeamie}
— switeerand
— Combination namie)
Figure 12. Graph of out-of-sample forecast and comparison of dynamic models
129,000
‘18,000
116.00.
114,000
112,000
110,000 4
108,000 /
100,000.)
104,000.
Meow wae we
208 2010 aot ota ate
= Canbinaionsetatie)
—— sutzetand
— VaR Gato)
Mutivariate ratios
Unie statiy
Figure 15, Graph of out-of-sample forecast and comparison of static models
189.2 Tables
UK Switzerland Netherlands _ Germs France FU28
UK 1,000 0590 0.568 ssi 0.556 o.r26
Switwrland 0540 L000) 0.560 sae 0.036 0.678
Netherlands 056i 0.569 1.000, 0.622 0.730 0.792
Germany 0.551 0.588 0.622 1.000 0.758 ass
Krance 0.656) 0.696 (030 0.365 L000 2.900
E28 0.726 0.678 0.792 0.385 0900 1.000
GDP growth rate with selected EU states and EL28
‘Table 1. Correlation matrix of S
Franoape | _— 708" [swirind ODP
re |e eee
2
‘Table 2, Granger causal relationships between GDP growth rates (chi-sq-statistic,* significant on 10% level, **
significant on 5% level)
"RMSE dynamic
"Univariate
Multivariate
VAR 452.76
Comb (least sq) «778.15
Table 3, Overview of RMSE of all forecasting models
Diebold-Mariano-Test stat Prob
—_ dynamic —_ -
Univariate vs. Multivariate 162i 0.00
-— Combys. VARS BUDS
"Comb vs. Multivariate 1332S.
‘Comb vs. Univariate “8.01 0.00,
— static —
Univariate vs, Multivariate 3.18 000
~— Combys. VARS OST
~ Comb vs. Multivariate 338 000
‘Comb vs. Univariate -038 OST
‘Table 4. Forecasting comparison using Diebold-Mariano tess
1910 Data Sources
European Union: curostat Statistical Service. National Accounts. (2016). Gross Domestic
Product of selected EU member states and accociates. Retrieved from:
hupi//ec.europa.cu/eurostat/data/database
2011 Bibliography
Diebold, F. X. (2015). Forecasting in Economics, Business, Finance and Beyond. University
of Pennsylvania,
Wooldridge, J. M. (2013). Introductory econometrics a modern approach (5. ed.). Mason,
Ohio u.a.: South-Westem Cengage Learning,
21