Advance Access publication May 6, 2008
Political Analysis (2008) 16:324–344
doi:10.1093/pan/mpn002
Two Sides of the Same Coin? Employing Granger
Causality Tests in a Time Series Cross-Section
Framework
M.V. Hood III
Department of Political Science, 104 Baldwin Hall, University of Georgia,
Athens, GA 30602-1615, USA
e-mail: [email protected]
Quentin Kidd
Department of Government, Christopher Newport University, 1 University Place,
Newport News, VA 23606, USA
e-mail: [email protected]
Irwin L. Morris
Department of Government and Politics, University of Maryland at College Park,
3140 Tydings Hall, College Park, MD 20742, USA
e-mail: [email protected] (corresponding author)
In this paper, we introduce a recently developed methodology for assessing the assumption of
causal homogeneity in a time series cross-section Granger framework. Following a description
of the procedure and the analytical contexts for which it is appropriate, we implement this new
approach to examine the transformation of the post-World War II party system in the South.
Specifically, we analyze the causal relationship between black mobilization and GOP growth in
the region. We find that black mobilization Granger caused Republican growth throughout the
South, whereas Republican growth Granger caused black mobilization only in the deep South.
We discuss the substantive significance of our results and conclude with guidelines for the
appropriate use of this procedure and suggestions for future extensions of the method.
1 Introduction
Introduced nearly 40 years ago, Granger models (see Granger 1969) of various types
remain the most popular methodology for evaluating the nature of the causal relationship
between two variables.1 Though originally designed for pairs of lengthy time series,
Author’s note: We would like to thank Christophe Hurlin and Baptiste Venet for their path-breaking work in this area and
especially for the helpful comments provided by Professor Hurlin. In addition, we would also like to thank Geoff Layman,
Mike Hanmer, and the participants of the American Politics Workshop at the University of Maryland, College Park for their
helpful advice. A previous version of the manuscript was presented at the 2006 Citadel Symposium on Southern Politics.
1
The use of Granger analysis in political science is extensive. Recent uses of Granger techniques range across the
discipline and focus on substantive questions as diverse as the causal relationships between democracy and
economic growth (Heo and Tan 2001; Baum and Lake 2003), partisanship and ideology in the American
electorate (Box-Steffensmeier and De Boef 2001), domestic politics and foreign policy (Moore and Lanoue
2003), school performance and the size of educational bureaucracies (Meier, Polinard, and Wrinkle 2000), and
terrorist incidents and the threat of terrorism (Enders and Sandler 2000).
The Author 2008. Published by Oxford University Press on behalf of the Society for Political Methodology.
All rights reserved. For Permissions, please email: [email protected]
324
Two Sides of the Same Coin?
325
Granger tests are increasingly used to evaluate causal relationships in time series crosssection (TSCS) data. The extension of the original Granger methodology to TSCS data has
the potential to improve upon the conventional Granger analysis for all the reasons that
TSCS analysis is generally preferable to cross-sectional or traditional time series analysis.
In general, TSCS models provide a number of improvements over the separate analyses of
time series data by cross-section. First, TSCS data allow for considerably more flexibility
modeling the behavior of cross-sectional units than conventional time series analysis (see
Greene 2000). Second, the TSCS framework allows for the analytical incorporation of
significantly more observations (and more degrees of freedom [df]) than would a comparable analysis of individual time series. Finally, and most significantly for our purposes,
TSCS Granger tests are significantly more efficient than conventional Granger tests (Hurlin
and Venet 2001). The potential benefits of TSCS Granger tests are, therefore, considerable.
However, just as TSCS Granger tests share the benefits of other types of TSCS analysis,
they also share an important potential flaw: the inappropriate assumption of causal homogeneity. As is often the case in TSCS analysis, this assumption is largely ignored. Failure
to adequately analyze the empirical foundations for this assumption could easily lead to
faulty substantive conclusions—inferring a causal relationship in all cross-sections when it
is only present in a subset of cross-sections or rejecting the presence of a causal relationship for an entire group of observations when a subset of the sample actually does manifest
the hypothesized causal relationship. As it becomes increasingly clear that (often untested)
assumptions about the specification of TSCS models may produce quite fragile results (see
Beck 2007; Wilson and Butler 2007), the development of a procedure for evaluating the
assumption of causal homogeneity in TSCS models would be a valuable methodological
innovation. Following a brief discussion of Granger analysis and its extension to TSCS
models, we describe just such a procedure—developed in Hurlin and Venet (2001)—for
evaluating the causal homogeneity between cross-sections in a TSCS framework. We then
present an application of this new methodology in which we examine the causal relationship
between black mobilization and Republican growth within the context of the Southern party
system. We conclude by identifying other potential applications of this methodology and
provide a set of guidelines for the appropriate use of this procedure in the research process.
2 Assessing Causality—The Granger Framework
Granger testing is a common method of investigating causal relationships (Granger 1969).
By estimating an equation in which y is regressed on k lagged values of y and k lagged
values of an additional variable x, we can evaluate the null hypothesis that x does not
Granger cause y.2 If one or more of the lagged values of x is significant, we are able to
reject the null hypothesis that x does not Granger cause y. In more straightforward terms,
we conclude that the evidence suggests that x causes y.
The Granger procedure is imperfect. Aside from the fact that no technique can fully
ascertain the nature of a causal relationship between two variables, the Granger procedure
is a simple bivariate model. It is possible, for example, that the relationship between x and
y is spurious (the correlation being caused by the impact of some other variable, z, on x
and y), and we cannot evaluate this possibility in the context of a simple bivariate model.
This renders the isolated use of Granger tests problematic. Conventional Granger tests are
also less than ideal for estimating the general causal relationship between two time series
in a variety of independent settings (or cross-sections). Estimating independent Granger
models for each of a set of cross-sections and then generalizing from frequently disparate
2
We define k as the number of lagged values of variables x and y.
326
M.V. Hood III, Quentin Kidd, and Irwin L. Morris
results is an inefficient—though common—procedure to test for the presence of a universal
causal relationship between variables. It does not, however, obviate the usefulness of
Granger tests as a first step toward uncovering the character of the causal relationship
between two variables. So, for example, in the literature on Southern politics, the causal
relationship between black mobilization and Republican growth (at the local, state, or
regional level) is unclear and under-analyzed. Granger tests would help clarify this relationship (as we demonstrate in an extended example below).
The development of multivariate vector autoregressive (VAR) Granger models was
a significant response to problems associated with bivariate Granger models. The rudimentary bivariate Granger procedure is easily extended to a VAR framework. Given a set of
variables, one can test for the existence of causal relationships between each pair of variables, and variables that are not Granger caused by any other variable in the system are
described as weakly exogenous (see Greene 2000). Suppose we were interested in the causal
interrelationship between economic growth, income inequality, and democratization. If we
could show that neither economic growth nor income inequality Granger caused democratization, then democratization would be weakly exogenous. Note that this is a relatively
stringent exogeneity condition; it precludes, for example, the possibility that a nonrecursive
relationship exists between two variables. This is not the only problem with Granger VAR
models. VAR time series are often short, and complicated multivariate VAR models quickly
overwhelm the available df in single time series.3 To address this problem, econometricians
have begun to modify Granger tests to incorporate panel and TSCS dynamics (see e.g.,
Holtz-Eakin, Newey, and Rosen 1988; Arellano and Bond 1991; Hurlin and Venet 2001;
Hurlin 2005).4 Within TSCS frameworks, Granger tests generate meaningful results with
significantly shorter time spans, incorporate significantly more observations, and produce
more efficient results than Granger tests in conventional contexts (Hurlin and Venet 2001).
Employing conventional Granger tests with TSCS data, however, raises two important
inferential issues, both dealing with the potential heterogeneity of the individual crosssections. One form of cross-sectional variation is due to distinctive intercepts, and this type
of variation may be addressed with a fixed effects model.5 The other, arguably more
problematic type of heterogeneity—causal variation across units—requires a more complex analytical response.6 Until recently, this type of heterogeneity was largely ignored
3
These problems are discussed in some detail in Kmenta (1997) and Greene (2000).
Some substantive examples of this methodology include Podrecca and Carmeci (2001), Weinhold and Nair
(2001), Davis and Hu (2004), Hurlin and Venet (2004), and Erdil and Yetkiner (2005). Though political scientists
commonly distinguish between panel data (more cross-sections than time periods) and TSCS data (more time
points than cross-sections), economists often do not. The procedure presented below is most appropriate for
TSCS data, so we use the TSCS terminology. However, it is also appropriate for panel data with a manageable
number of cross-sections (i , 50 in most cases) and sufficiently long time series for modeling temporal dynamics
(t . 10).
5
For example, suppose black mobilization has the same causal impact on Republican growth throughout the
South. That would not preclude variation in Republican growth across the Southern states due to other idiosyncratic factors outside the model. Fixed effects models are one way to address this type of cross-sectional
variation.
6
This causal heterogeneity should not be confused with the ‘‘unobserved heterogeneity’’ common in censored
TSCS data or in TSCS data for which sample selection bias is a potential problem. In those situations, one or
more unobservable variables (unobservable for practical or theoretical reasons) is correlated with (or is potentially correlated with) an observed independent variable. This correlation may lead to spurious results, attributing
a causal effect to an observed variable because the unobserved variable cannot be incorporated into the model. In
this case, there is no expectation of causal heterogeneity; if all appropriate variables were observable, the same
causal model would be appropriate for all observations—that is, causal homogeneity. The various solutions to
‘‘unobserved heterogeneity’’ (see, e.g., Heckman 1981; Vella and Verbeek 1998, 1999; Wooldridge 2000, 2002,
2005) are based upon this substantive assumption. Our concern is with the appropriateness of the assumption of
causal homogeneity in the first place.
4
Two Sides of the Same Coin?
327
(with unknown results) in TSCS VAR models. According to Erdil and Yetkiner (2005),
there are two distinctive literatures dealing with TSCS or panel VAR models. One, the
literature based on early work by Hsiao (1986) and Holtz-Eakin, Newey, and Rosen (1988)
largely ignores this type of heterogeneity. So, for example, tests for weak exogeneity based
on this literature did not (and do not) effectively account for the possibility of causal
variation among the cross-sections; so it was not possible to know for which set of
cross-sections the exogeneity result actually applied.7
Another strain of this literature, which was based on recent work by Hurlin and Venet
(2001), explicitly addresses this type of heterogeneity. Hurlin and Venet (2001) outline
a procedure for evaluating the character of the causal processes (homogenous versus
heterogenous) within a TSCS framework.8 Their methodology is an extension of the basic
Granger model. In this conventional context, series x may be said to cause series y if and
only if the expectation of y given the history of x, E(yjxt-k), is different from the unconditional expectation of y, E(y).9 That is, if
Eðyjytk Þ 5 Eðyjytk ; xtk Þ;
ð1Þ
then x has no effect on—and thus does not cause—y. However, if
Eðyjytk Þ 6¼ Eðyjytk ; xtk Þ;
ð2Þ
we would conclude that x does cause y. The expected value of y is different given x, so this
implies that x causes y.
Extension of Hurlin and Venet (2001) on this basic model to the TSCS context is
straightforward (but not trivial). Their analytical results are based on a model of the
following type,
yi;t 5 ai þ
p
X
cðkÞ yi;tk þ
k51
p
X
ðkÞ
bi xi;tk þ vi;t
ð3Þ
k50
for each of the cross-sections i and for all t in [1,T]:10 Here we have a TSCS model where
the regressors are (1) lagged values of the dependent variable (yi,tk) subset by crosssections i and (2) lagged values of the independent variable (xi,tk) also subset by crosssection. The error term is represented by vi,t. cðkÞ are the autoregressive coefficients, and
ðkÞ
bi are the regression coefficients. The fixed effects are represented by ai. The number of
time periods is indicated by p. To retain sufficient degrees of freedom for model estimation, Hurlin and Venet assume the following constraints:
1. cðkÞ are constant and
ðkÞ
2. bi are assumed constant for all k 2 [1, p].
7
This may be partly attributable to the preponderance of large i, small t data sets in applied econometrics. Panels
with hundreds of cross-sections and 10 or fewer time points are very common. Given that data structure, tests for
causal heterogeneity—even the tests developed by Hurlin and Venet—would be difficult to implement. However,
in political science—and increasingly in economics—TSCS data sets with significantly fewer cross-sections and
significantly more time points are quite common. It is these types of data sets for which the tests for causal
heterogeneity are particularly appropriate.
8
Technically, this is also a TSCS VAR model, but it is limited—as currently developed by Hurlin and Venet—to
the evaluation of bivariate relationships.
9
The independent variable is denoted by x and the dependent variable by y. The time period is indicated by t, and
the number of lags is indicated by k.
10
Note that the residuals satisfy the conventional assumptions. See Hurlin and Venet (2001) for further details.
328
M.V. Hood III, Quentin Kidd, and Irwin L. Morris
The first assumption precludes variation in the autoregressive coefficient from crosssection to cross-section. The second assumption precludes variation in the regression
coefficients from time period to time period.11 In addition, whereas the autoregressive
slope coefficients are identical for all cross-sections, the regression coefficients are allowed to vary across individual cross-sections. Within the TSCS framework, there are
three possible causal scenarios:
1. An identical causal relationship exists between x and y in all cross-sections.12
2. No causal relationship exists between x and y in any cross-section.
3. There is a causal relationship between x and y in some subset of n cross-sections, but the
nature of the relationship is not constant across cross-sections.13 Hurlin and Venet
subdivide this scenario into two categories: (a) causal relationships of different types
across cross-sections and (b) at least one cross-section shows no evidence of any causal
relationship.
The formalization Hurlin and Venet’s formalization of the first scenario in the TSCS Granger
framework—that there is no evidence that x causes y in any of the cross-sections—implies that
Eðyi;t jyi;t1 ; yi;t2 ; . . . yi;0 ; ai Þ 5 Eðyi;t jyi;t1 ; yi;t2 ; . . . yi;0 ; xi;t ; xi;t1 ; xi;t2 ; . . . xi;0 ; ai Þ;
ð4Þ
for every cross-sectional unit in the panel. Eðyi;t jyi;t1 ; yi;t2 ; . . . yi;0 ; xi;1 ; xi;t1 ;
xi;t2 ; . . . xi;0 ; ai Þ is the best linear predictor of yi,t given the past values of y for each
cross-section and the past values of x for each cross-section.14 According to equation
(4), the expected value of y given past values of y is equal to the expected value of y given
both past values of y and past values of x for each cross-section i. This is a straightforward
extension of the null hypothesis for the standard Granger model; knowing the previous
values of the independent variable x provides no new information about the current value
of the dependent variable y that is not already contained in the previous values of y. This
equation merely states that this is true for each and every cross-section in the TSCS.
Alternatively,
Eðyi;t jyi;t1 ; yi;t2 ; . . . yi;0 ; ai Þ 6¼ Eðyi;t jyi;t1 ; yi;t2 ; . . . yi;0 ; xi;t1 ; xi;t2 ; . . . xi;0 ; ai Þ
ð5Þ
"i 2 ð1; 2; . . . NÞ;
implies the existence of a causal relationship for every cross-sectional unit in the data. In
contrast to equation (4), this inequality indicates that the expected value of y depends upon
both past values of y and past values of x for each cross-section i. Thus, we can reject the
TSCS version of the Granger null hypothesis; knowing the previous values of the independent variable x provides new information about the current value of the dependent
11
Note that regressions coefficients are allowed to vary across lag lengths.
Situations in which a causal relationship is present in each cross-section but the character of the causal relationship is variable are included in Scenario 3.
13
Heterogeneity arising from level differences between cross-sections is addressed by including unit-specific
(fixed) effect parameters. This is an imperfect strategy. If we were willing to assume the absence of true causal
heterogeneity—attributing all cross-sectional heterogeneity to sampling error—then an array of alternatives to
fixed effects would be available to us (for a thorough discussion of these options, see Wooldridge 2002). The
development of alternatives to fixed effects would be a welcome—but still as yet unrealized—extension of the
procedure presented here.
14
Hurlin and Venet (2001) explicitly define the best linear predictor for the instantaneous case (includes the
current value of x as a regressor). Our definition of the best linear predictor for the more conventional Granger
case (does not include the current value of x as a regressor) is a straightforward extension.
12
Two Sides of the Same Coin?
329
variable y that is not already contained in the previous values of y. This equation also
indicates that the inequality holds for all cross-sections in the TSCS. Note that each
individual linear predictor is assumed constant, which indicates that the nature of the
causal relationship is identical across all cross-sections.15
The third scenario allows for the presence of at least one causal relationship (up to
a maximum of N not necessarily identical causal relationships) but does not require the
existence of N causal relationships—heterogeneous causality. So, heterogeneous causality
requires that equation (5) holds for at least one cross-section i in the data. The final case
refers to that situation in which equation (4) holds for at least one cross-section i in the
data. This would indicate that for at least one cross-section, there is no causal relationship.
This exhausts the theoretically possible outcomes associated with TSCS Granger tests. We
now move to a more detailed description of the implementation of the Hurlin and Venet
procedure for empirically distinguishing between these various cases.16
3 A Method for Causality Testing in TSCS Settings
The methodology of Hurlin and Venet for evaluating causal heterogeneity in TSCS data
comprise a set of nested F-tests. F-tests are standard tools for assessing the significance of
parameter restrictions in linear models; more specifically, they provide a means for estimating the likelihood that one or more parameter restrictions significantly affect a model’s
goodness of fit.17 The procedure described below is a series of particular F-tests. These
tests are easily implemented using the constrained regression technique in Stata or in
similar programs such as E-Views. Before outlining the specific hypotheses and requisite
tests, we outline the basic process below:
1. conduct TSCS nonstationarity tests for each variable to be examined,18
2. create a dummy variable to represent each cross-section in the sample,
3. create a set of slope parameters by multiplying each unit-specific variable by lags of the
independent variable,
4. specify the necessary equations with the proper constraints, excluding the constant term,
5. for each model, record the sum of squared residuals or save it as a new variable, and
6. calculate the required test statistics and determine significance.
15
See original constraints on autoregressive coefficients and regression coefficients.
Effectively distinguishing between substantive differences between cross-sections and differences resulting
from sampling error depends upon a sufficiently long-time series. For this reason, these procedures are most
appropriate for TSCS data and are not appropriate for panels with particularly short time series (t , 10).
17
Note that the F-test tends to be the most conservative of a set of analogous tests (i.e., Wald, LM, and LR tests)
and is the least sensitive (of these tests) to small sample deviations from normally distributed disturbances (see
Greene 2000).
18
These models operate under the assumption that all series to be tested do not contain a unit root. The crucial
requirement is covariance stationarity, and none of the results require more than weakly stationary series.
Popular statistical software packages (such as Stata) include standard tests to assess stationarity in TSCS data.
Two examples would be the Levin, Lin, Chu Test or the Im, Pesaran, and Shin Test. See Baltagi (2005) for
a discussion of unit root tests for TSCS data. A potentially useful extension of this methodology to a context
including partially integrated series remains undeveloped. Although some procedures do exist for evaluating the
level of integration of partially integrated series (i.e., the Robinson log periodogram estimator [roblpr in Stata]),
it is not clear how to most effectively account for time series that manifest different levels of partial integration—a likely possibility in TSCS data sets in which there is a serious question about the comparability of the
cross-sections (for further discussion of diagnostic tests for fractional integration, see Baum and Wiggins 2001).
16
330
M.V. Hood III, Quentin Kidd, and Irwin L. Morris
A flowchart outlining the nested testing procedure described in the previous section is
presented in Fig. 1.19 The first test to be conducted is designed to test the proposition that x
does not cause y for any cross-section in our sample. Rejection of this hypothesis implies
that for at least one (and possibly more) cross-section, x Granger causes y.20 Formally, the
first hypothesis with cross-sections denoted as i is as follows:
H1:
For all i, x does not cause y.
We assess this first hypothesis by constructing a test statistic, hereafter referred to as F1,
comparing the sum of squared residuals from a set of restricted models (RSS2) to the sum
of squared residuals produced by a set of baseline (unrestricted) models (RSS1). As in the
traditional Granger causality test, the unrestricted model includes lags of yi,tk, lagged
values of the independent variable (xi,tk), and the fixed effects themselves (ai) to predict
current values of yi,t. Lagged values of the dependent variable are constrained to be equal
(ci,t1 5 ci,tk) for all models presented. In the unrestricted model, subsequent lags withinTSCS slope coefficients are also set to be equal (bi,t1 5 bi,tk). In the restricted model
slope coefficients and lags are constrained to 0 (bi,t1 5 0), leaving only the unit specific
effects and the various lags of the dependent variable 0 to predict current values of y.21 The
test statistic to determine the presence of causality (F1) is calculated as:
F1 5
ðRSS2 RSS1 Þ=ðNpÞ
;
RSS1 =½NT Nð1 þ pÞ p
where N is the number of cross-sections, p is the number of lags, and T is the number of
time periods. Interpretation of the statistic relies on the F-distribution with Np, NT N(1 þ
p) p df. An insignificant statistic for this test indicates that x does not cause y in any
cross-section, and thus, the testing process ends here. A significant test statistic indicates
that for at least one (and possibly all) of the cross-sections, there exists a causal relationship and we proceed to a test for homogenous causality.
Having determined that x does cause y for i 1, we now proceed to determine whether
a common (or homogenous) causal process exits for all i. The collective nature of the
causal process is determined by testing the following hypothesis:
H2:
For all i, x causes y.
In order to test H2, we calculate another test statistic, hereafter referred to as F2.
Acceptance of H2 (denoted by an insignificant test statistic) indicates that a common
causal process is manifest for all cross-sections in our sample. At this juncture, further
testing is unnecessary as x is said to Granger cause y for all TSCS cross-sections. Rejection
of H2, denoted by a significant test statistic, would indicate that for at least one or more
TSCS cross-sections, x does not Granger cause y. The F2 test statistic is calculated using
the sum of squared residuals from the unrestricted model described above (RSS1) along
with the sum of squared residuals (RSS3) from a restricted model in which the slope terms
are constrained to be equal for each cross-section in the sample (bt1 5 btk). Calculation
of the F2 test statistics is as follows:
19
Nomenclature and equations come directly from Hurlin and Venet (2001).
It is important to note that rejection of this hypothesis does not necessarily imply the presence of a homogenous
causal process for the entire TSCS data set (or that x causes y for all cross-sections).
21
Slope coefficients for x can be either constrained to zero or excluded from the model altogether.
20
331
Two Sides of the Same Coin?
F1: Not
Significant
H1: Test for the
Presence of a Causal
Relationship
Result: Causal
Relationship Not
Present for any CrossSection
F1: Significant
H2: Test to Determine
the Nature of
Causality
F2: Not
Significant
Result: Causal
Relationship Present
for all Cross-Sections
F2: Significant
H3a/b: Test for the
Presence of a Causal
Relationship for a
Specific Cross-Section
i or Subset of CrossSections j
F3a/b: Not
Significant
Result: Causal
Relationship Not
Present for Specific
Cross-Section i or
Subset of CrossSections j
F3a/b: Significant
Result: Causal
Relationship Present
for Specific CrossSection i or Subset of
Cross-Sections j
Fig. 1 Granger causality testing for TSCS data.
F2 5
ðRSS3 RSS1 Þ=½pðN 1Þ
:
RSS1 =½NT Nð1 þ pÞ p
Again, the test statistic is interpreted using an F-distribution with Np, NT N(1 þ p) p
df.
If H2 is rejected, a third hypothesis can be utilized in order to determine for which
cross-section i, x causes y or:
332
H3a:
M.V. Hood III, Quentin Kidd, and Irwin L. Morris
For i, x does not cause y.
For each cross-section i, H3a is conducted using the unrestricted sum of squared residuals estimated previously (RSS1) in addition to the sum of squared residuals (RSS2,i)
from a model in which the slope coefficient for the cross-section in question is constrained
to 0 or excluded from the model equation (bi,tk5 0). The statistic to test H3a for crosssection i is calculated as follows:
F3a 5
ðRSS2;i RSS1 Þ=p
:
RSS1 =½NT Nð1 þ 2pÞ þ p
If the F3a test statistic is significant using a F-distribution with Np, NT N(1 þ p) p df,
then we can reject H3a, indicating that x does cause y for cross-section i. If the F3a test
statistic is insignificant, then x does not cause y for cross-section i. Testing of H3a is
conducted for each TSCS cross-section across the requisite number of time periods
(t k) in order to determine which cross-sections contributed to the previous finding of
causality (denoted by rejection of H1).
When theory suggests grouping cross-sections in a specific manner, a second test
statistic can be calculated to examine the causal nature for some subset of panel members
j, or:
H3b: For j, x does not cause y.
In this case, the slope coefficients for the subset of cross-section members j in question
are constrained to 0 (bj,tk 5 0). The sum of squared residuals from this restricted equation
(RSS2,j) is again compared to those from the unrestricted model (RSS1) to produce the F3b
test statistic:
F3b 5
ðRSS2;j RSS1 Þ=ðnnc pÞ
;
RSS1 =½NT Nð1 þ pÞ nc p
where nnc is the number of TSCS cross-sections for which b is constrained to 0 and nc is
the number of TSCS cross-sections for which b is not constrained to 0. At this juncture,
a significant F3b test statistic allows for the rejection of H3b for subgroup j, inferring that x
Granger causes y in this subset of cross-sections. We now apply this TSCS Granger procedure to test one recent theory of political development in the post-World War II South.
4 Substantive Application: The Transformation of the Southern Party System
Though dramatic, the rise of Southern Republicanism and the electoral mobilization of
Southern blacks were not unexpected. Key’s (1949) work alludes to this transformation,
and by the end of the 1970s, both dynamics were the foci for a considerable body of
scholarship (for a description of this literature, see Black and Black 2002). Undoubtedly,
national political dynamics played an important role in the formative stages of the
development of Southern Republicanism (see Black and Black 1987, 1992, 2002;
Carmines and Stimson 1989), but national level forces cannot fully explain the subregional
variation in GOP growth or the electoral mobilization of Southern blacks.22
Explanations of Republican growth have focused primarily on demographic and
economic factors such as immigration, the transformation and growth of the Southern
22
Hood, Kidd, and Morris (2004); Nadeau and Stanley (1993); Rhodes (2000); and Shafer and Johnston (2001)
provide useful overviews of this literature.
Two Sides of the Same Coin?
333
economy, the growth in religious conservatism, and racial context. Recently, Hood, Kidd,
and Morris (2004) have added a new wrinkle to the literature by suggesting that black
mobilization drove Republican growth in the region. We argue that as blacks mobilized and
became Democrats, it became increasingly difficult for white conservatives to maintain
control of the party. In increasing numbers, white conservatives left for the relatively more
attractive Republican Party (see Heard 1952). As conservative whites left the Democratic
Party, the opportunities for Southern blacks in the Democratic Party increased. With the
Democratic Party an increasingly attractive option, southern blacks mobilized in greater
numbers. In short, the usefulness of the Republican Party increased over time for white
conservative voters as the attractiveness of the Democratic Party decreased. We also argue
that this mobilization dynamic is most likely to play out in those Southern states with the
largest black populations—especially those with black populations large enough (if mobilized) to control the Democratic Party in an increasingly Republican environment.23
To investigate the potentially endogenous relationship between GOP growth and black
mobilization in the South, we turn to a TSCS Granger analysis to examine the extent to
which: (1) GOP growth Granger causes black mobilization and (2) black mobilization
Granger causes GOP growth for the Southern states from 1960 to 2004. We provide a fuller
description of our data in the next section.
4.1 Variable Operationalization
For this study, the former states of the Confederacy serve as our unit of analysis, producing
a total of 11 cross-sections. For each Southern state, we have collected biennial data over
a 44-year time period, from 1960 through 2004, for two variables.24 The first of these
variables, Black Mobilization, taps into the potential influence that the political mobilization of blacks may have produced in regard to politics in the region. Black Mobilization is
calculated at the state level as the number of black registered voters divided by the total
number of registered voters, or:
Number of Blacks Registered to Vote25
:
Total Number of Registered Voters
Operationalized as it is, our measure of black electoral strength places blacks within the
context of the existing electorate—a much more precise method for estimating the potential influence of blacks as an electoral presence than alternative indicators (i.e., the percentage of blacks registered to vote).
The second indicator of interest, GOP Strength, is measured at the state level utilizing
an index developed by David (1972). General election vote percentages for Republican
candidates in gubernatorial, senate, and congressional elections were utilized to create
a composite state-level index of GOP strength.26 Following the construction of each GOP
23
These would be the deep South states or Alabama, Georgia, Louisiana, Mississippi, and South Carolina.
Our time series consist of 2-year election cycles (i.e., 1960, 1962, . . . 2004).
Interpolation was used to fill in the gaps between missing years for both the number of blacks who were
registered to vote and for the total number of registered voters in each Southern state. [Data sources: VEP
News (Various Years); Statistical Abstract (Various Years); Current Population Reports: P-20 Series on Voting
and Registration (Various Years)].
26
Estimates from 1960 through 1970 are obtained from David’s work, whereas estimates for the remaining years
are calculated by the authors. [Data sources: David (1972); Guide to U.S. Elections (1994); America Votes
(Various Years)].
24
25
334
M.V. Hood III, Quentin Kidd, and Irwin L. Morris
state index, a 10-year (5-time point) moving average was applied to smooth any sharp
variations present in each series.27 The David Index of Party Strength was the method of
choice for Lamis (1988) in his detailed study of party change in the South. Specifically,
GOP Strength is calculated as:
ð% Republican Vote½Senate Election þ % Republican Vote½Gubernatorial Election
þ % Republican Vote½Average Republican Congressional VoteÞ=3:28
These two series are plotted by state from 1960 through 2004 in Fig. 2, with GOP Strength
represented by black squares and Black Mobilization represented by white triangles. As
evidenced by this set of 11 plots there is a great deal of variation, both temporal and crosssectional, for each of these series. The minimum and maximum values for GOP Strength
are 0.01 (Mississippi-1960) and 0.57 (Tennessee-2004), respectively. For Black Mobilization, the minimum value is 0.04 found in 1960 Mississippi, whereas the maximum value of
0.38 is also associated with Mississippi in 1994.
4.2 Testing for Nonstationarity
Before proceeding with the TSCS Granger tests outlined above, we need to establish that
both pooled time series are stationary (do not contain a unit-root). We utilize two different
tests designed to detect the presence of unit-roots specifically in TSCS data. Table 1
presents test statistics from the Levin, Lin, and Chu and the Im, Pesaran, and Shin procedures, both of which indicate that nonstationarity is not an issue for either series.29
4.3 Causality Testing
The first step in untangling the causal process between the two variables of interest is to
test H1. In our case, we want to know if:
Granger causes
/ Black Mobilization
ði;tÞ ;
GOP Strengthði;tkÞ
and if
Granger causes
/ GOP Strength
Black Mobilizationði;tkÞ
ði;tÞ ;
27
Comparisons between our measure of Republican Party strength and actual party registration data from
Louisiana and Florida from 1950 to 2000 (the only two Southern states that did track party registration during
the time of our study) indicate a high level of congruity (r 5 0.94 for LA and 0.94 for FL) [data available from
the authors upon request].
28
Special transformations had to be made for Louisiana for each election following the 1978 institution of an open
primary system. We used the following method to calculate our index of GOP party strength for 1978 through
2004: 1. If there was only one election (open primary): GOP 5 percent of total republican vote (including votes
won by other Republican candidates in the primary). 2. If there was both a primary and a general election, and a.
The general election contained both a Republican and a Democrat: GOP 5 percent of total vote won by
Republican Candidate. b. The general election contained two Democratic candidates: GOP 5 0%, c. The
general election contained two Republican candidates: GOP 5 100%.
29
We also conducted unit-root tests for the raw (unsmoothed) GOP Strength series. The Levin, Lin, and Chu (test
statistic: 4.60; p , .001) and the Im, Pesaran, and Shin (test statistic: 2.77; p , .001) tests both indicate this
series is also stationary.
335
Two Sides of the Same Coin?
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
19601962196419661968197019721974197619781980198219841986198819901992199419961998200020022004
19601962196419661968197019721974197619781980198219841986198819901992199419961998200020022004
Alabama
Arkansas
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
19601962196419661968197019721974197619781980198219841986198819901992199419961998200020022004
19601962196419661968197019721974197619781980198219841986198819901992199419961998200020022004
Georgia
Florida
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
1960 1962 196419661968197019721974197619781980198219841986198819901992199419961998200020022004
19601962196419661968197019721974197619781980198219841986198819901992199419961998200020022004
Louisiana
Mississippi
GOP Strength
Black Mobilization
Fig. 2 GOP Strength and Black Mobilization by state, 1960–2004.
for the 11 states in our sample. In addition, if a causal relationship is detected, we need to
determine its consistency across the states in our sample. We begin by testing H1, which is:
H1: For all states, GOP Strength (Black Mobilization) does not cause Black Mobilization
(GOP Strength).
In order to test H1, we calculated the F1 test statistics using the sum of squared residuals
from the unrestricted models:
GOP Strengthi;t 5 ci;t1 GOP Strengthi;t1 . . . ci;tk GOP Strengthi;tk
þ bi;t1 Black Mobilizationi;t1 . . . bi;tk Black Mobilizationi;tk þ ai
þ ei;t1 . . . ei;tk ;
ð9Þ
336
M.V. Hood III, Quentin Kidd, and Irwin L. Morris
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
19601962196419661968197019721974197619781980198219841986198819901992199419961998200020022004
19601962196419661968197019721974197619781980198219841986198819901992199419961998200020022004
North Carolina
South Carolina
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
19601962196419661968197019721974197619781980198219841986198819901992199419961998200020022004
19601962196419661968197019721974197619781980198219841986198819901992199419961998200020022004
Texas
Tennessee
0.6
0.5
0.4
0.3
0.2
0.1
0
19601962196419661968197019721974197619781980198219841986198819901992199419961998200020022004
Virginia
GOP Strength
Black Mobilization
Fig. 2 Continued.
Black Mobilizationi;t 5 ci;t1 Black Mobilizationi;t1 . . . ci;tk Black Mobilizationi;tk
þ bi;t1 GOP Strengthi;t1 . . . bi;tk GOP Strengthi;tk þ ai þ ei;t1 . . . ei;tk ;
ð10Þ
where the slope coefficients for subsequent lags within states are constrained to be equal
(i.e., ci,t1 5 ci,tk " k and bi,t1 5 bi,tk " k) and a second model identical to the one
specified in equation (9) (equation 10), where the slope coefficients and subsequent lags
for all the states in our sample are constrained to 0 (i.e., bi,t1 . . . bi,tk 5 0). Again,
a significant F1 test statistic indicates that for at least one (and possibly all) of the states in
our analysis, Black Mobilization Granger causes GOP Strength or GOP Strength causes
Black Mobilization.
337
Two Sides of the Same Coin?
Table 1 TSCS unit-root tests
Test
Test statistics
Levin, Lin, and Chu
GOP Strength
Black Mobilization
Im, Pesaran, and Shin
GOP Strength
Black Mobilization
2.93***
4.06***
1.99**
2.23***
Note. For both tests, a significant test statistic allows for the rejection of the null
hypothesis that the series is nonstationary.*p , .10; **p , .05; ***p , .01.
The F1 test statistics are presented in Table 2. The first test of H1 analyzes whether GOP
Strength, for the members of our TSCS collectively, Granger causes Black Mobilization.
The F1 test statistic is statistically significant at one lag t1, allowing us to reject H1. So
for at least one state (and possibly all), there is statistical evidence that GOP Strength
Granger causes Black Mobilization. The F1 test statistics, however, are insignificant at two
and three lags, an indication that there is little evidence to support the claim that this
relationship exists beyond a single lag period.
The second half of Table 2 details the F1 test statistics used to examine the hypothesis
that Black Mobilization Granger causes GOP Strength. At one lag, the F1 test statistic is
significant, indicating again that this causal process is at work in one or more of the states
in our sample. Again, there appears to be no evidence that this relationship exists in
subsequent time periods (i.e., t2, t3). In summary, we can reject H1 for the relationships
under study and proceed next to determine the nature of the causal process across the 11
states of the South.
The following hypothesis is designed to examine the homogeneity of the causal relationship uncovered in the preceding step:
H2: For all states, GOP Strength (Black Mobilization) causes Black Mobilization (GOP
Strength).
In order to test H2, we calculate another set of test statistics, hereafter referred to as F2,
using the sum of squared residuals from the unrestricted model specified above and those
from a model where the slope coefficients are constrained to be equal for all states and lag
Table 2 F1 test statistics
Lags
GOP Strength Granger causes Black Mobilization
t1
t2
t3
Black Mobilization Granger causes GOP Strength
t1
t2
t3
F1
3.214***
0.532
0.668
1.652*
1.197
0.979
Note. Critical values: F1: based on an F-distribution with Np,
NTN(1þp)p df (Hurlin and Venet 2001).*p , .10; **p , .05; ***p,.01.
338
M.V. Hood III, Quentin Kidd, and Irwin L. Morris
Table 3 F2 test statistics
Lags
GOP Strength Granger causes Black Mobilization
t1
t2
t3
Black Mobilization Granger causes GOP Strength
t1
t2
t3
F2
1.966**
0.493
0.594
0.748
0.559
0.619
Note. Critical values: F2: based on an F-distribution with Np, NTN(1 þ p) p df (Hurlin and Venet 2001).*p , .10; **p , .05; ***p , .01.
periods (i.e., b1,t1 5 b2,t2 5 bn,tk). Failure to reject H2 (insignificant test statistic)
indicates that the causal process is homogenous for all 11 states in our sample. At this
juncture, if H2 is confirmed, further testing is unnecessary as Black Mobilization is said to
Granger cause GOP Strength (or GOP Strength is said to Granger cause Black Mobilization) for all the states in our sample. On the other hand, rejecting H2 (significant test
statistic) would indicate that for at least one or more states n 1, Black Mobilization does
not Granger cause GOP Strength (or GOP Strength does not Granger cause Black Mobilization).
The results of our tests for H2 are located in Table 3. Again, the bidirectional relationship between Black Mobilization and GOP Strength is examined for up to three lag
periods. In the case of GOP Strength Granger causing Black Mobilization, H2 is rejected
at one lag. Therefore, we must conclude that the causal process in this case is heterogeneous or does not exist across all the states in our sample. Reversing the equation, however,
we find that H2 is accepted in the case of Black Mobilization Granger causing GOP
Strength. So, for this relationship, the causal process appears to be homogenous across
all 11 states, making further examination unnecessary.
In order to determine which states in our sample contribute to the causal finding leading
from GOP Strength to Black Mobilization, we must turn to a third hypothesis or H3a which
can be formally stated as:
H3a:
For state i, GOP Strength does not cause Black Mobilization.
Rejection of H3a indicates the presence of a causal relationship for the specific state
under consideration. The F3a test statistics are calculated using the sum of squared residuals from the unrestricted model in equation (1) and a second model in which the slope
coefficients and subsequent lags for the state under testing are constrained to 0 (i.e., F3a:
bi,t1 . . . bi,tk 5 0). We present state-by-state results in the upper portion of Table 4 for
one lag period t1.30 The test results indicate that for Alabama, Georgia, Louisiana,
Mississippi, North Carolina, and South Carolina the relationship GOP Strength Granger
causes Black Mobilization appears to hold. For the remaining five states in our study there
is insufficient statistical evidence to reject H3a.
What one quickly notices about these two sets of states is that they divide nicely into
two well-identified geographic subregions: deep South and rim South. The sole exception
30
There is little evidence that this causal relationship exists beyond the first lag period (see Table 2). We therefore
restrict tests of H3a to t1.
339
Two Sides of the Same Coin?
Table 4 F3 test statistics: GOP Strength Granger causes Black
Mobilization(t-1)
State
F3a
Alabama
Arkansas
Florida
Georgia
Louisiana
Mississippi
North Carolina
South Carolina
Tennessee
Texas
Virginia
7.500***
0.190
0.002
13.932***
10.008***
16.660***
4.752***
3.736***
0.550
0.121
1.38
Subregion
Deep
Rim
F3b
6.464***
1.147
Note. Critical values: F3: based on an F-distribution with Np, NT N(1 þ p)
p df (Hurlin and Venet 2001). Deep South: AL, GA, LA, MS, SC. Rim
South: AR, FL, NC, TN, TX, VA.*p , .10; **p , .05; ***p , .01
is North Carolina, typically considered a rim South state. The bottom portion of Table 4
tests H3b by grouping states into their respective subregion. By convention, we classify
North Carolina as a rim state along with Arkansas, Florida, Tennessee, Texas, and Virginia
with the remaining states comprising the deep South. In this example, H3b is specified as:
H3b: For subregion j, GOP Strength does not cause Black Mobilization.
In order to determine whether GOP Strength ganger causes Black Mobilization for each
subregion, we derive a set of test statistics (F3b) using the sum of squared residuals from
equation (1) and from a second model in which the slope coefficients for the states
comprising the subregion are constrained to 0 (i.e., for the deep South: F3b: bj,t1 5 0).31
As indicated in Table 4, H3b is rejected collectively for the deep South states and accepted
for the rim states. Thus, testing by subregion, we may conclude that at t1, GOP
Strength Granger causes Black Mobilization in the deep South, but not in the rim South.
5 Discussion of Results
In summary, we find that Black Mobilization Granger causes GOP Strength and that GOP
Strength Granger causes Black Mobilization. The former causal process can be characterized as homogenous for the 11 states in our sample. The later process, conversely, is
heterogeneous, existing in the deep South (and North Carolina), but not in the rim states.
In the case of the rim South, the causal process is one-sided and can be characterized as:
Granger causes
/ GOP Strength
Black MobilizationðRim;t1Þ
31
In this case, we only calculate the F3b test statistic to one lag t1.
ðRim;tÞ :
340
M.V. Hood III, Quentin Kidd, and Irwin L. Morris
Fig. 3 Mean Black population by state, 1960–2000.
In the case of the deep South states, the process is reciprocal:
Granger causes
/ GOP Strength
Black MobilizationðDeep;t1Þ
ðDeep;tÞ ;
and
Granger Causes
/ Black Mobilization
GOP StrengthðDeep;t1Þ
ðDeep;tÞ :
Not only is there a geographical differentiation associated with these subregions, but
a substantive one as well. Although the size of the black population is much higher in
the South compared to other regions of the United States, within the South the highest
concentrations of black citizens are located in the deep South states. This fact becomes
clear looking at Fig. 3, which arrays the 11 states in the region according to the relative size
of their average black populations from 1960 through 2000. The average black population
in the region varies from a high of 37% in Mississippi to a low of 12% in Texas.
The horizontal line in Fig. 3 separates states based on the significance of their F3a test
statistics in Table 4. These test statistics were significant for those states above the horizontal line and insignificant for the reminder located below this division. This line also
suggests a potential threshold for the noted counter-mobilization effect to manifest itself.
In states with an average black population over 20%, expansion of the Republican Party
produced increased levels of mobilization within the black electorate. Below this level, it is
possible that a necessary critical mass does not exist to trigger the noted counter-mobilization reaction found in the deep South states.
As Key noted decades ago:
The range of the Negro population—from 49.2 percent in Mississippi to 14.4 percent in Texas—
suggests that even ‘‘the South’’ is by no means homogenous and that if the Negro influences the
politics of the South, there ought to be wide variations in political practices from state to state
(1949:10–11).
Two Sides of the Same Coin?
341
This variation has traditionally been thought of in quantitative terms. The example
utilized suggests that a qualitative distinctiveness also exists. These results, taken in
tandem, suggest the transformation of Southern politics during the last half of the twentieth century can be viewed from a theoretical perspective that attributes the same logic of
action to both blacks and conservative whites. This is a simple and straightforward feedback loop (in over half of the Southern states), but it is also a novel characterization of the
twin pillars of the transformation of Southern politics: Republican growth and black
mobilization.
With few exceptions, research on Republican growth in the South has focused on
other types of causal explanations including economic dynamics, religious or cultural
orientations, migration patterns, or the geographic concentration of blacks (i.e., the
black-belt hypothesis), while largely overlooking more overtly political explanations.
Similarly, existing work on black mobilization in the South rarely focuses explicitly on
the party dynamics highlighted in this example.
In an important respect, our results are consistent with the black empowerment literature (Browning, Marshall, and Tabb 1984; Bobo and Gilliam 1990; Harris, SinclairChapman, McKenzie 2005). However, instead of focusing on the outcome produced by
black mobilization—the election of black officials (in the South)—and then viewing that
outcome as an inducement to further black mobilization, one might explain our results
with a more fundamental political force: the change in the benefits for blacks of Democratic Party membership and activism (and, implicitly, electoral mobilization). Clearly,
Republican growth (and the exodus of white conservatives from the Democratic Party)
opened up significant opportunities for blacks.32 Our research suggests that Southern
blacks saw these opportunities—where they were greatest—and took advantage of them.
6 Future Directions
There is considerable evidence of a nonrecursive relationship between the two factors of
interest; however, this two-way causal flow is not homogenous across all the states in our
sample. Again, for the rim South states there is no endogeneity issue, as Black Mobilization was found to Granger cause GOP Strength, with no evidence for the converse. For
those states located in the deep South, the TSCS Granger tests did point to evidence of
a nonrecursive relationship between these two variables of interest.
The logical next step would involve specifying a set of models within a multivariate
framework in order to determine if the causal relationships uncovered stand up to the
addition of control variables. In formulating multivariate explanatory models, the best
course of action would call for separate analyses, one for deep South states and another
for the rim South states. For the former, instrumental variables will need to be employed to
alleviate known problems associated with the use of endogenous regressors, whereas for
the latter a more traditional TSCS regression framework can be utilized (for a discussion of
panel models using instrumental variables, see Baltagi 2005).
7 Conclusion
This manuscript examines a new methodology for evaluating the causal homogeneity
assumption in TSCS data via the use specially designed Granger causality tests. The TSCS
32
For example, in 2004, blacks comprised 58% of the turnout in the South Carolina Democratic Primary and 47%
in the Georgia Democratic Primary (Bullock and Gaddie 2005a; 2005b).
342
M.V. Hood III, Quentin Kidd, and Irwin L. Morris
Granger tests presented allow one to control for, and detect, the possibility that the causal
process, if present, is not homogenous for members of the cross-section under study (i.e.,
heterogenous causality). As the utilization of TSCS data for the exploration of political
phenomena has grown exponentially over the last decade, this diagnostic device should
enjoy widespread application.
Policy studies exploring the relationship between concealed weapons and crime over
time and across some geographic unit are one example where these specialized Granger
tests might be employed. The study of Kovandzic and Marvell (2003) attempts to determine the effect that concealed weapon permits produce on the rate of various types of
violent crime in Florida. The pair find little evidence that the rate of concealed weapon
permits within a county act to depress violent crime rates. It is possible, however, that
a nonrecursive relationship may exist in which increases in the violent crime rate may
cause an increase in citizens seeking to obtain concealed weapon permits. Likewise, the
finding that permits are not related to a decrease (or increase) in violent crime rates may
not be valid across all the Florida counties sampled. It is possible that a relationship
between these two factors may be heterogeneous, present for some subset of counties
but not for others.
Shaw’s (1999) examination of campaign-specific effects on presidential voting patterns
provides another example where this methodology might be employed. The Republican
share of the two-party vote, by state, is estimated on a weekly basis using candidate
appearances and campaign advertising as explanatory factors. The included campaign
effects were found to be positively related to a candidate’s vote share within a state.
The possibility also exists, however, that campaigns may adjust to a changing political
environment. In this case, vote share (or estimated vote share) may cause the frequency of
advertising or appearances to shift. Likewise, one might envision a situation where campaigns affect vote share, but not in a uniform manner across all states. States with higher
numbers of independents may be more apt to be influenced by campaign effects, compared
to states with lopsided partisan divisions.
Although time-series, cross-sectional research has become prevalent in political science scholarship, diagnostic tests and other extensions to deal with known TSCS issues
have sometimes been slow to develop. Recent research indicates that one known problem—the questionable assumption of causal heterogeneity—is far more prevalent than
previously realized. We agree with the conclusion of Wilson and Butler that there is a clear
‘‘need for extensive sensitivity testing as part of the research process,’’ (2007: 119) especially where TSCS data are involved. The modified Granger tests presented in this manuscript offer one possible tool to help researchers evaluate the extent of causal
heterogeneity within a TSCS data set. Such information, in turn, can then be used to
specify more accurate multivariate time-series cross-sectional models.
References
America Votes. 1996, 1998, 2000, 2002, 2004. Washington, DC: Congressional Quarterly Press.
Arrellano, Manuel, and Stephen Bond. 1991. Some tests of specification for panel data: MonteCarlo evidence
and an application to unemployment equations. Review of Economic Studies 58:277–97.
Baltagi, Badi H. 2005. Econometric analysis of panel data. 3rd ed. Hoboken, NJ: John Wiley & Sons.
Baum, Christopher F., and Vince Wiggins. 2001. Tests for Long Memory in a Time Series. Stata Technical
Bulletin Reprints, Volume 10. College Station, TX: Stata Press.
Baum, Matthew A., and David A. Lake. 2003. The political economy of growth: Democracy and human capital.
American Journal of Political Science 47:333–47.
Beck, Nathaniel. 2007. From statistical nuisances to serious modeling: Changing how we think about the
analysis of time-series-cross-section data. Political Analysis 15(2):97–100.
Two Sides of the Same Coin?
343
Black, Earl, and Merle Black. 1987. Politics and society in the South. Cambridge, MA: Harvard University
Press.
Black, Earl, and Merle Black. 1992. The vital south. Cambridge, MA: Harvard University Press.
Black, Earl, and Merle Black. 2002. The rise of Southern Republicans. Cambridge, MA: Harvard University Press.
Bobo, Lawrence, and Franklin D. Gilliam. 1990. Race, sociopolitical participation, and black political empowerment. American Political Science Review 84:377–97.
Box-Steffensmeier, Janet, and Suzanna De Boef. 2001. Macropartisanship and macroideology in the sophisticated electorate. Journal of Politics 63:232–48.
Browning, Rufus P., Dale Rogers Marshall, and David H. Tabb. 1984. Protest is not enough: The struggle of
Blacks and Hispanics for equality in urban politics. Los Angeles: University of California Press.
Bullock, III, Charles S., and Ronald Keith Gaddie. 2005a. An assessment of voting rights progress in Georgia.
Report prepared for the Project on Fair Representation, American Enterprise Institute. http://www.aei.org/
publications.
Bullock, III, Charles S., and Ronald Keith Gaddie. 2005b. An Assessment of voting rights progress in South
Carolina. Report prepared for the Project on Fair Representation, American Enterprise Institute. http://
www.aei.org/publications.
Carmines, Edward, and James Stimson. 1989. Issue evolution: Race and the transformation of American
politics. Princeton: Princeton University Press.
Current Population Reports: P-20 Series on Voting and Registration. U.S. Department of Commerce. Bureau of
the Census. [1980-Table 5; 1982-Table 16, 1984-Table 2; 1986-Table 4; 1988-Table 2; 1990-Table 4; 1992Table 4; 1994-Table 4a (electronic version); 1996-Table 4 (electronic version); 1998-Table 4 (electronic
version); 2000-Table 4a (electronic version); 2002-Table 4a (electronic version); 2004-Table 4a (electronic
version)]. Washington, DC: Government Printing Office.
David, Paul T. 1972. Party strength in the United States, 1872–1970. Charlottesville, VA: University of Virginia
Press.
Davis, E. Philip, and Yuwei Hu. 2004. Is there a link between pension-fund assets and economic growth—A
cross-country study. Working Paper. The Pensions Institute, City University: London.
Enders, Walter, and Todd Sandler. 2000. Is transnational terrorism becoming more threatening? A time-series
investigation. Journal of Conflict Resolution 44:307–32.
Erdil, Erkan, and I. Hakan Yetkiner. 2005. A panel data approach for income-health causality. In The economics
of health reforms, ed. John N. Yfantopoulos, 701–24. Athens: Atiner Publications.
Granger, Clive W. J. 1969. Investigating causal relations by econometric models and cross-spectral methods.
Econometrica 37:424–38.
Greene, William H. 2000. Econometric analysis. 4th ed. Englewood Cliffs, NJ: Prentice-Hall.
Guide to U.S. Elections. 3rd ed. 1994. Washington, DC: Congressional Quarterly Press.
Harris, Frederick C., Valeria Sinclair-Chapman, and Brian D. McKenzie. 2005. Macrodynamics of Black
Political Participation in the Post-Civil Rights Era. Journal of Politics. 67:1143–63.
Heard, Alexander. 1952. A two-party South? Chapel Hill: The University of North Carolina Press.
Heckman, James J. 1981. The incidental parameters problem and the problem of initial conditions in estimating
a discrete time-discrete data stochastic process. In Structural Analysis of Discrete Panel Data with Econometric Applications, ed. C. F. Manski and D. McFadden, 179–95. Cambridge, MA: MIT Press.
Heo, Uk, and Alexander C. Tan. 2001. Democracy and economic growth: A causal analysis. Comparative
Politics 33:463–73.
Holtz-Eakin Douglas, Whitney Newey, and Harvey S. Rosen. 1988. Estimating vector autoregressions with
panel data. Econometrica 56:1371–96.
Hood III, M. V., Quentin Kidd, and Irwin Morris. 2004. The reintroduction of the Elephas Maximus to the Southern
United States: The rise of Republican State Parties, 1960 to 2000. American Politics Research 31:68–101.
Hsiao, Cheng. 1986. Analysis of panel data. Cambridge: Cambridge University Press.
Hurlin, Christophe. 2005. Testing for Granger causality in heterogeneous panel data models [English Title].
Revue Economique 56:1–11.
Hurlin, Christophe, and Baptiste Venet. 2001. Granger causality tests in panel data models with fixed coefficients. Working Paper Eurisco 2001–09, University of Paris Dauphine.
Hurlin, Christophe, and Baptiste Venet. 2004. Financial development and growth: A re-examination using a panel
Granger test. Working Paper. University of Orléans, University of Paris Dauphine.
Kmenta, Jan. 1997. Elements of econometrics. 2nd ed. Ann Arbor: University of Michigan Press.
Key, V. O. 1949. Southern politics in state and nation. New York: Vintage.
Kovandzic, Tomislav, and Thomas B. Marvell. 2003. Right-to-carry concealed handguns and violent crime:
Crime control through gun decontrol? Criminology and Public Policy 2:363–96.
344
M.V. Hood III, Quentin Kidd, and Irwin L. Morris
Lamis, Alexander P. 1988. The two-party South. Oxford: Oxford University Press.
Meier, Kenneth J., L. Polinard, and Robert D. Wrinkle. 2000. Bureaucracy and organizational performance:
Causality arguments about public schools. American Journal of Political Science. 44:590–602.
Moore, Will H., and David J. Lanoue. 2003. Domestic politics and U.S. Foreign Policy: A study of cold war
conflict behavior. Journal of Politics 65:376–96.
Nadeau, Richard, and Harold W. Stanley. 1993. Class polarization in partisanship among native Southern
Whites, 1952–90. American Journal of Political Science 37:900–19.
Podrecca, Elena, and Gaetano Carmeci. 2001. Fixed investment and economic growth: New results on causality.
Applied Economics 33:177–82.
Rhodes, Terrel L. 2000. Republicans in the South: Voting for the State House, Voting for the White House.
Westport, CT: Praeger.
Shafer Byron E., and R. G. C. Johnston. 2001. The transformation of Southern politics revisited: The House of
Representatives as a window. British Journal of Political Science. 31:601–25.
Shaw, Daron R. 1999. The effect of TV ads and candidate appearances on statewide presidential votes, 1988–96.
American Political Science Review 93:345–61.
Statistical Abstract of the U.S. U.S. Department of Commerce. United States Bureau of the Census. [1976-Table
747; 1979-Table 840; 1980-Table 849]. Washington, DC: U.S. Government Printing Office.
Vella, Francis, and Marno Verbeek. 1998. Whose wages do unions raise? A dynamic model of unionism and
wage rate determination for young men. Journal of Applied Econometrics 13:163–83.
Vella, Francis, and Marno Verbeek. 1999. Two-step estimation of panel data models with censored endogenous
variables and selection bias. Journal of Econometrics 90:239–63.
VEP [Voter Education Project] News. [1968-Vol.2, no. 4; 1969-Vol. 3, no. 12; 1970-Vol. 4, nos. 1,2]. Atlanta,
GA: Southern Regional Council.
Weinhold, Diana, and Usha Nair. 2001. Causality tests for cross country panels: A new look at FDI and
economic growth in less developed countries. Oxford Bulletin of Economics and Statistics 63:153–72.
Wilson, Sven E., and Daniel M. Butler. 2007. A lot more to do: The sensitivity of time-series-cross-section
analyses to simple alternative specifications. Political Analysis 15(2):101–23.
Wooldridge, Jeffrey M. 2000. A framework for estimating dynamic, unobserved effects panel data models with
possible feedback to future explanatory variables. Economics Letters 68:245–50.
Wooldridge, Jeffrey M. 2002. Econometric analysis of cross section and panel data. Cambridge, MA: MIT
Press.
Wooldridge, Jeffrey M. 2005. Simple solutions to the initial conditions problem for dynamic, nonlinear panel
data models with unobserved heterogeneity. Journal of Applied Econometrics 20:39–54.