Utility of Space for Historical Studies Colin Flint 2002

32

The contribution of quantitative historical studies lies intheir ability to ground and contextualize historical problems(Monkkonen 1984). To do this, the historian locates the sub­ject in three dimensions=-vtemporal. socio-political-spatial,and cultural intellectual" (88). Alternatively, a geographicperspective sees Monkkoneri's three dimensions as beinginterwoven to produce a dynamic notion of context. Contextis constructed through social and political acts, and overlime the context mediates future social and political behav­ior. Culture is an integral part of this process as politicalquestions and responses are filtered and defined withinshared mean.ings and practices. For example, the particularand traditional work practices of miners in Netting­hamshire, England, created a region of more moderateunion activity that subsequently divided the National Unionof Mineworkers in the strike of 1984/5 (Griffiths and Jobn­ston 1991). In other words, the individual OJ group does notact either outside or even simply within a context. Contextis created and re-created by historical actions while simul­taneously mediating that behavior (Earle 1992).

A geographic approach to context inserts dynamism intothe definition offered by Anderton and Sellers (1989, 106)whereby "context is simply the background, or environs,relevant to a particular event. The event, the individual enti­ty experiencing the event, and the context can all be distin­guished or defined through their characteristics. Thus, coo­textual effects refer to the influences of contextcharacteristics on individuals or the events experienced byindividuals." The geographic approach concentrates moreon the interaction and mutual dependence of actor and con­text, or bow people and social movements construct contextand how context influences actions. In other words, the goalis to define and analyze context as "an enduring and regen­erative construct" (113).

Emphasis will be placed on describing a way of estimat­ing such contextual effects by the use of spatial statistics.However, I do not argue that the perspective and methodol­ogy are applicable to all situations. I agree that "there is no

S ince Edward M. Cook's (1980) appeal for the inclu­sion of geography in historical studies, HistoricalMethods bas published a number of articles illustrat­

ing the Tole of spatial structures in mediating human behav­ior (Taylor 1984; Hochberg and Miller 1992; Artz.rouni andKornlos 1996). However, those researchers do not theorizehow historical actions created and maintained spatial struc­tures in the first place. In addition, Douglas Anderton andDeborah Sellers (1989) provide a clear and careful discus­sion of the meaning of context and how it could be incor­porated into statistical models. My intention is to buildupon previous quantitative historical geography studies aswelJ as Anderton and Sellers's argument by highlightinghow geographers see context both as a force produced bythe actions of individuals and groups and one that shapessubsequent actions (Agnew 1987; Massey 1994; Sack 1997;Staeheli, Kodras, and Flint 1997). Once context has beendefined, 1 focus on some statistical modeling techniquessuitable for incorporating such spatial structures into quan­titative historical studies. In sum, with this article, 1 hope toillustrate the theoretical value and methodological necessi­ty of incorporating geographic notions of context into his­torical studies via spatial statistical models.

Abstract. The geographic perspectiveon human behavior empha­sizes the notion of context. Context suggests that human behavioris partly determined by the place in which individuals and groupsact. and in tum these actions re-create places. It follows that anelement of history is the contextual setting of activity. Spatial [3-

tistics are introduced as a tool for modeling context in historicalstudies. SpaLialdependence serves to operationalizc the specifici­ty of place and the diffusion of processes between places. Spatialheterogeneity serves to operationalize the idea that behavior is re­gionally specific rather than uniform across space. Both spatialdependence and spatial heterogeneity are included in spatial re­gression models. A spatial statistical analysis of the Nazi Partyvote in Bavaria between 1928and 1932 is adopted 10 illustrate theargument.Keywords: Nazi Party, geography, spatial statistics, Bavaria

COLIN FLINTDepartment of Geography

Pennsylvania State University

The Nazi Party in Geographic Context

The Theoretical and MethodologicalUtility of Space and Spatial Statistics

for Historical Studies

~USTORICAL METHODS. Winter 2002. Volume 35, Number I

particular moments in sucb intersecting social relations. netsof which have over time been constructed, laid down, inter­acted wiLhone another. decayed and renewed. Some of theserelations will be. as it were, contained within the place: oth­ers will stretch beyond it, tying any particular locality intowider relations and processes in which other places areimplicated too.

This definition emphasizes the historical dimension of place,as norms and institutions originating in earlier periods part­ly deternnine later activities. Thus, the dynamism of place iscentral to a contextual approach. Second, a geographicdimension, inseparable from the historic, can be seen in theimportance of the linkages between places in defining con­text. Hence, the relationships between places make up a

plays in the world economy, pinpoints the material concernsand condition of the inhabitants of a particular place. Sec­ond. locale. or the institutional setting of a place, identifiesthe social networks and organizations that transmit place­specific information and meaning. Finally, a sense of placeprovides a place-specific identity. In combination, problemsand demands posed by the political and economic dynam­ics of the world economy are digested and tackled within asetting of place-specific languages and institutions that thenproduce a mosaic of political responses. The key implica­tion for stati tical analysis is that social behavior acrossspace is not uniform but rather a function of place-specificpractices and traditions.

In a recent reflection on his earlier work, Agnew (1996,132) further defines the elements of place that "frame therange of possible political activities and actions for humanagents in particular places." In addition to reiterating andexpounding upon the original three components of place, heconsiders the impact of differential access to technology aswell as the use of geographic referents in the discourse ofpolitical parties. The most compelling additions to Agnew'sframework lie in the definition of interplace and cross-scaleconnections. First, be notes that "places are embedded interritorial Slates" (132), and so local pol itics may reflectlocal-central tensions. Second, "social class. ethnic andgender divisions and antagonisms have national and inter­national histories promulgated by political movements andinfluential commentators and leaders" (133). The signifi­cance of this second point is the introduction of a temporaldimension to illustrate that place-specific practices are tbeproduct of past events and organizations remembered andused in subsequent mobilizations. Agnew's conlributionforces us to recognize that places are the product of actionsover time, whereas history is partly the story of how placesadapt to new circumstances.

The discussion of interactions with other places andscales leads to the consideration that the nature of a place isthe combination of both local institutions and their connec­tions to the rest of the world. These historical and nodalaspects of place are captured by Massey's (1994, 120) def­inition of places:

Historical geographers. and some historians. have creat­ed an intellectual tradition examining the role of humanaction in shaping social and physical landscapes (e.g.,Meinig 1978; Cronon 19(1). in addition, historical geogra­phers have examined the role of historical context in (ram­ing political and social change (e.g., Bennett and Earle1983). Here, I attempt to define context in a way that can bemodeled by spatial statistics.To unpack the idea of context in a way that facilitates sta­

tistical analysis, one can use two related definitions ofplace, which is a suitable entrance into this discussionbecause of its relevance to political and social behavior.Place is the location of "immediate agency" and "meaning­ful identity" (Oakes 1997, 5 I0), and so it is the spatial set­ting most clearly implicated in the mediation of decisionsand acts. It is. therefore, a key geographic scale in under­standing the contextual setting of individuals (Johnston1991; Taylor 1981). But how is identity given meaning. andhow is agency mediated? To answer these questions. onemust have definitions of place that capture its unique com­ponents as well as the connections between other places andscales. Later. I will operationalize these abstract notions ofplace to facilitate statistical analysis.

John Agnew (1987, 1996) bas consistently called for aconsideration of political behavior within a framework thatrealizes the mediating role of place. Agnew (1987, 5)defined three components of place that explain the geogra­phy of political behavior. First. location. or the role a place

The Theoretical Definition of Context

one correct method for analyzing a contextual effectbecause there are clearly distinct theoretical and method­ological notions of what a contextual effect is" (Andertonand Sellers 1989, 114). Instead, I offer a particular perspec­tive and methodology, a definition of context that can beexamined by a particular set of spatial statistics. Such anapproach leads to the understanding of how historicalactions were embedded in different spatial settings that, inturn, produced a mosaic of political and social responses.

This article is organized in the following manner. First,notions of context in the geographic literature are discussed,Second, a particular set of spatial statistical techniques thatcan be used to model the components of context that medi­ated historical behavior are described, and the theoreticaland methodological utility of these teclmiques is discussed.Third, these utilities are exemplified by an analysis of thegrowth of the Nazi Party's electorate in Bavaria between1928 and 1932, before the Nazis gained control in 1933. Inthis section, r elucidate how the techniques used in theexample lead to an understanding of historical processesbeing embedded within spatial structures. I conclude by dis­cussing the implications of the geographic approach for his­torical studies: the need to include the spatial selling whenanalyzing historical events.

33Winter 2002.Volume35.NUl11t~r I

identification and incorporation of spatial dependence andspatial beterogeneity, when appropriate, is necessary to en­ure unbiased and efficient estimates in regression analysis.Spatial dependence exists when the value of the depen­

dent variable in one spatial unit of analysis is partly a func­tion of the value of the same variable in neighboring units.The existence of spatial dependence may result from aprocess of diffusion that results in "Galton's problem,"which states that "certain traits in an area are often causednot by the same factors operating independently in eacharea but by diffusion processes" (O'Loughlin and Anselin1992, 17). In other words, an increase in electoral supportfor the Nazi Party in one place may have been a function ofincreased support in neighboring places and its consequentincreased propaganda activity and not simply the socioeco­nomic traits of the places themselves. The identification ofspatial dependence indicates where such processes of diffu­sion by political activity took place. The incorporation ofspatial dependence into statistical models includes suchagency as part of an explanation of political behavior. Inother words, the importance of the linkages between placesin defining context (Massey 1994) is incorporated into thestatistical modeling.

Spatial heterogeneity refers to a regional pattern in thedata that results in instability of parameters across the wholestudy (Anselin 1988, 9). Therefore, the slope of any regres­sion equation would not be constant when one comparesregions with the complete data set. In other words, spatialheterogeneity refers LO the regional differentiation of similarcharacteristics. For example, within Bavaria, Central andUpper Franconia was a culturally distinct region, being aProtestant enclave within a predominately Catholic area.The regional culture of Franconia was expected to produceregionally specific political behavior.Regional or provincialidentities and practices sbaped political behavior at a geo­graphic scale beneath the nation-state (Applegate 1990).Theidentification of spatial heterogeneity within the data indi­cates the presence of geographical variation in politicalbehavior,and its incorporation into statistical models illumi­nates the place-specific behavior of voter and party.

Spatial heterogeneity indicates the existence of regionalcontextual variation within the data. It may exist as a merenuisance, expressed as lack of constancy of the regressionerror variance. or it may have structural significance(O'Loughlin and Anse1in 1992. 27). Diagnostic tests forheteroskedasricity in ordinary least squares (OLS) modelsindicate the presence of structurally Significantheterogene­ity (Anselin 1992). Heteroskedasticity is the presence ofnonconstant variance of the random regression error over allthe observations. If heteroskedasticity is present. the OLSestimates are unbiased but inefficient, inference based onthe t and F statistics will be misleading, and the measures offit will be wrong (Anselin 1988, 120).Luc Anselin's Spacestat (1992. sect. 26, p. 8) software

compotes three tests for heteroskedasticity,but only two of

As social processes are embedded within spatial struc­tures, statistical techniques that model social behavior mustincorporate the spatiality of the data. The use of spatial sta­tistics to analyze data aggregated in geographic units (e.g.,precincts. counties, or states) is necessary for both theoreti­cal and methodological reasons. Theoretically. spatial sta­tistics allow for the inclusion of the two components of thecontext identified previously.First, one can capture linkagesbetween places by incorporating the diffusion of politicalstrength or mobilization.Toinvestigate these processes, oneincorporates the statistical concept of spatial dependenceinto the analysis. Second, the spatial condition of politicalbehavior leads to the expectation that people of similarsocioeconomic and cultural backgrounds do not necessarilybehave in the same way across different geographical con­texts (Agnew 1987; Brustein 1985). To investigate theseprocesses, the researcher incorporates the statistical conceptof spatial heterogeneity, thus allowing for the identificationand estimation of different relationships between the samevariables in different spatial settings. Methodologically, the

Methodology and Data

second key component of context thatmust be incorporatedinto statistical analysi .In summary. the relevant implications of these discu -

sions of place and political activity for the operationaliza­lion of statistical models are twofold. First, both Agnew(1987) and Massey (1994) point to a variety of processesand institutions that lead to an expectation of place-specificbehavior-or, in other words, the belief that people withsimilar socioeconomic and cultural characteristics are like­ly to behave differently within unique contextual settings.Thus, the geographic approach supports Monkkonen's(1984) argument that quantitative historical studies shouldinvestigate the context of historical problems rather thanuniversal laws. In this particular case, we are identifying thecomposition of the Nazi Party-Nazi being the shortenedversion of Nationalsozialistische Deutsche Arbeiterpartie.the National Socialist GermanWorkers' Party (NSDAP)­electorate in different spatial settings rather than casting ablanket explanation of the NSDAP vote across the whole ofGermany. Second. part of the explanation for place-specificbehavior lies in the network of connections between places.Therefore. historical quantitative studies should includeinterlinkages between places in the modeling of contextualeffects.

Both of these theoretically generated expectations maybe incorporated into the analysis of political behavior viathe spatial statistical analysis of aggregate data. Spatial het­erogeneity allows for the identification of place-specificbehavior, and spatial dependence identifies the interactionbetween places as space was being transformed by partyactivity. In the following section, I discuss spatial statisticaltechniques that can be used to model these implications.

HISTORlCAL METHODS34

where y is a vector or observations on the dependent vari­able. X is a matrix of explanatory variables. including thetemporal- patial lag. B is the vector of the regression coef­ficients. and E is an error term. However, the existence ofspatial dependence often persisted despite the inclusion ofthe temporal-spatial lag. If this was the case, the estimatesof the regression coefficients would be biased and inconsis­tent (Anselin 1988, 59). If spatial dependence remained inthe presence of the temporal-spatial lag. it was replaced bythe spatially lagged dependent variable.

Spatial dependence can exist in two forms (Anselin 1988,11-13). In its substantive form, spatial dependence is inter­preted as spatial contagion. The behavior in one Kreis waspartly explained by similar behavior in neigh boling Kreise,discussed earlier as "Galton's problem." Methodologically,one incorporates substantive spatial dependence into the re­gression equation by adding the spatially lagged dependentvariable. Forma1Jy. this may be expressed by the equation

y =pWY + XB + E.

where the notation is the same as lor the usual regressionequation and p is a spatial autoregressive coefficient and WY

y=X6+E.

ficients between the two spatial regimes indicates that het­erogeneity existed at the subregional scale. In other words,different political behavior existed within the subregions, orcontextual settings, deftned by the spatial regimes.

Spatial dependence was directly incorporated into themodels in one of two ways depending on the diagnostictests reported in the initial models. The average value of theNazi vote in neighboring Kreise (counties) in the first of thetwo elections defining a particular period of change,referred to hereinafter as the temporal-spatial lag, wasincorporated into the initial OLS model. [f this variable waseither significant or insignificant but spatial dependenceexisted, then the temporal-spatial lag was dropped andreplaced by tbe spatially lagged dependent variable. Thespatially lagged dependent variable was the average valueof the dependent variable in neighboring Kreise, Ln boththese cases, the neighbors of each Kreis (county) weredefined by first-order contiguity. If the temporal-spatial lagwas positive in sign and statistically Significant, it indicatesthat the change in the Nazi Party vote in one Kreis was part­ly a function of the size of support in the neighboring Kreisein the first of two elections in that particular period ofchange. If the spatially lagged dependent variable was pos­itive in sign and statistically significant, it indicates that thechange in the Nazi vote in a Kreis was partly a function ofthe change in the Nazi Party vote in neighboring Kreise.The inclusion of either of these variables models the role ofthe interlinkages between places in defining the contextualsetting of the voter.

The OLS model may be represented formally by theusual regression equation

where b is a matrix of the estimates of the regression coeffi­cients and var(b) is the corresponding (asymptotic) variancematrix (Anselin 1992, sect. 32. p. 2). A 'ignjficant value forthe Chow test measuring the stability of the regression coef-

these are reported at anyone time. The first reported test iseither the Lagrange multiplier test developed by Breusch andPagan (1979) or the studentized version developed by Koen­ker and Bassett (1982). When the errors are normal. theBreusch-Pagan test is reported, and when the errors are non­normal, the Koenker-Bassett test is reported. When there islittle prior information about the form of heteroskedasticiry,the White (1980) test is the most valid for heteroskedasticity.When there are sufficient degrees of freedom. Spacestat(Anselin 1992) reports the White test in addition to either theBreusch-Pagan or Koenker-Bassen test.

With regard to the Nazi Party vote in Bavaria, if thesediagnostic tests are significant, subregional patterns of vot­ing behavior were present. In other words, political behav­ior was not uniform across the mosaic of place-specific con­texts that constituted Bavaria. After the identification ofheterogeneity, I used previous historical studies of thegrowth of tbe Nazi Party in Bavaria to identify subregionsthat were likely to display voting behavior different fromthe remainder of the region. J used the identification ofthese subregions, called spatial regimes, to estimate siruc­tural change models (Anselin 1992, sect. 32. p. I). To esti­mate a structural change model, I identified cases within theparticular subregion by the use of a dummy variable, andthe structural change estimation reported separate regres­sion coefficients for the two sets of cases, those in the ub­region and those in the remainder of the region. The struc­tural change model is represented by the equations

Yi = ai + XiBij + £i for d = 0 andYj = aj +Xjl3ij + Ej for d = 1,

where both the constant terms (ui[jD and the slope terms(Bilj I) take on different values (O'Loughlin and Anselin1992, 3 L). To diagnose whether the structural change esti­marion captures the heterogeneity within the region. testsfor heteroskedasticity should be insignificant.

In addition, a Chow test was reported for the model as awhole and for each of the explanatory variables. The Chowstatistic (Chow 1960) tests the stability of regression coeffi­cients. It is distributed as an F variate with K. N - MKdegrees of freedom (with Mas tbe number of regimes). Thetest is a test of the null hypothesis

HI):g'b = O.

where b is a vector of all the regression coefficients (includ­ing the constant terms) and g' is a K by 2K matrix (lk I-Ik],with Ik as a K by K identity matrix (Anselin 1992, sect. 32,p. 2). The corresponding Wald test may be expressed as theequation

W = (g'b){g'[var(h)j-I g)-I (g'b).

35Wimer 2001, Volume 35. Number I

where tr stands for the lrace matrix operator, M = I -X(X'X)_IX', Y is a vector of observations on the dependentvariable, e is a vector of OLS residuals, W is the spatialweights matrix. S2 the maximum likelihood estimate oferror variance, and B is the vector of OLS coefficients(Anselin 1992. sect. 26. p. 13).

Finally, the existence of nonnormality in the data re­quired the use of estimation methods that were robust in thepresence of nonnormality. The jackknife resampling meth­od consists of repeated estimations on the data set fromwhich one observation is dropped at a lime and yieldsresults that are robust to nonnormality and hetero-

LMLAG= (e"Wy/s2)2 / [(WX8)'MWX8/s2

+ tr[W'W + W2]).

KR = (g'Z'Zg)/(a'a/hN)'

where g stands for the coefficient vector in the auxiliaryregression and a is the resulting residuaJ vector (Anselin1992, sect. 26, p. 12).The final test is the Lagrange multiplier lest for substan­

tive spatial dependence. Valid only under the assumption ofnormality, it is asymptotic in nature and may be expressedin the following equation:

where tr represents the matrix trace operator, e is a vector ofthe OLS residuals, i1 = e'elN is the maximum likelihoodestimate for error variance, and W is the spatial weightsmatrix (O'Loughl in and Anselin 1992, 29).The third test is a specification robust procedure. the

Kelejian-Robinson (1992) test for spatial error dependence.In contrast to the previous two tests, this one does notrequire normality in the error terms and requires less infor­mation about the precise form of the spatial weights matrix.Because it requires a large sample size, it must be interpret­ed with caution for smaJl data sets. The statistic is calculat­ed from an auxiliary regression of cross products of residu­als and cross products of the explanatory variables (matrixZ). The cross products are calculated for all the pairs ofobservations for which a nonzero correlation is postulated,with eacb pair being entered just once. for a total of hNpairs. The Kelejian-Robinson statistic may be expressed inthe following equation:

LM£RR = {e'We / S2} 2/ tr rW'W + W]2,

model and the elements of the weights matrix. To test thesignificance of this statistic, it is converted to a standardizedz value and compared with a normal distribution. This testis the least useful and reliable of the four because it cannotdiscriminate between spatiaJ error dependence and substan­tive spatiaJ dependence and it is sensitive to non normalityand heteroskedasticity (Anselin and Rey (991).

The second test is the Lagrange multiplier test for spatialerror dependence that may be stated as the following equa­tion:

where W is the weights matrix and e is the vector of theresiduals (O'Loughlin and Anselin 1992, 29). The distribu­tion of this statistic depends on the specification of the

e= We + x,

where the notation is the same as above with We being aspatial lag of the errors and x is a "well-behaved" error termwith mean of zero and variance matrix .1.21 (Anselin 1992,sect. 29, p. I). The presence of both the spatial lag and thewell-behaved error term creates a problem of simultaneity.In ligbt of this problem, a maximum likelihood procedurethat includes the estimation of a nonlinear likelihood func­tion must be executed (Anselm 1988,59). Lf the spatial er­ror dependence is ignored, the OLS estimates would be un­biased but could result in misleading inference if thevariance estimates are not adjusted. because the OLS vari­ance expressions do not account for the dependence amongthe errors.

The difficulty in distinguishing between spatial errordependence and substantive spatial dependence stems fromthe similarity of the two expressions representing eitherform of dependence. The similarity of the two expressionsleads to the common factor hypothesis. that the product ofthe spatial autoregressive coefficient with the regressioncoefficients equals the negative of the coefficients of thespatially lagged dependent variables (Anselin ] 988,225-30). If the common factor hypothesis holds, then itsuggests that the spatial error model is the correct specifi­cation. if it does not hold, then it implies that either the sub­stantive spatial dependence model is the correct specifica­tion or may point to specification errors. such as the wrongspatial weights matrix or missing explanatory variables.

Four tests for the presence of spatial dependence, eithererror or substantive, were used. The first test is an applica­tion of the Moran's J rest to the residuals of the OLS regres­sion to test for spatial autocorrelation in the errors of a re­gression model. This test may be expressed in the equation

1=e"Wele'e

y = X8 + s and

is the spatial lag, the average of the value of the dependentvariable in neighboring Kreise (Anselin 1992. sect. 27. p. 1).

In addition (0 the substantive interpretation. spatialdependence may also have to be controlled for as a nui­sance. This form of spatial dependence is known as spatialerror dependence as it is associated with model specifica­tion errors that are not restricted to one Kreis but spill acrossthe spatial units of observation. The spatial error model isusuaJly expressed as a spatial autoregressive process in theerror term. In other words, the usual assumptions ofhomoskedastic and uncorrelated errors no longer hold. Toestimate regression coefficients in the presence of spatialerror dependence. a spatial autoregressive model is estimat­ed that may be stated in the following equations:

HISTORICAL METHODS36

The aving grace is that the available aggregate data arevery comprehensive as long as it is clear that inferencesabout individual behavior are not being made. Rather, oneexamines voting behavior aggregated at the scale of theKreis. The use of aggregate data is less of a problem for anontology that focuses on the contexts of individual behaviorrather than on individuals divorced from the structures thaimediate their everyday life.' The statistical analysis ofaggregate voting and socioeconomic data identifies thesocial composition of particular contexts that providedexperiences that nurtured a vote for the Nazis.

Both the census and election files were obtained from thearchive Wahl- und Sozialdaten del' Kreise lind Gemeindetides Deutschen Retches, 1920-1933 at the Central Archiveof the University of Cologne." The Cologne data were dis­aggregated at a scale of over six thousand geographic unitsthat included Kreise. villages. and neighborhoods withincities. All the usual socioeconomic measures except income(not collected in Germany then or now) are to be found inthis file: those relating to the 1925 census data were used tocalculate the explanatory variables adopted in the modeling.I selected the geographic units within Bavaria from this fileand aggregated the voting and socioeconomic data at thecounty scale. Areal boundaries of the Kreise were obtainedfrom an Office of Strategic Service. (aSS) Map, No. 6289(1944). 1 digitally coded the Kreise borders from this mapinto a Geographic Information System (GIS) to allow forspatial analysis of the data.' The incorporation of spatialanalysis with a GIS was necessary to allow for the con­struction of a contiguity or spatial weights matrix. The dig­ital coding of the spatial units within a GIS allowed for thedevelopment of matrices that were based on first-order con­tiguity to define the immediate neighbors of the Kreise.These matrices were necessary to include the interactionbetween neigbboring Kreise in the structural-spatial regres­sion models that are discussed below.

The estimation of models that include spatial dependen­cy and heterogeneity and the relevant socioeconomic vari­ables i a process thai requires the evaluation of diagnostictests and the significance of the explanatory variables. Theinitial OLS model estimated in the Spacestat softwareincluded the same even variables plus additional explana­tory variables for each separate model.' The consistent vari­ables were suggested by previous analyses and informed bytheoretical frameworks commonly used to explain the NaziParty vote. Walter D. Burnham's (1972) theory of politicalconfessional ism argues that Catholics and the industrialproletariat would not have been attracted to the Nazi Partybecause of their respective allegiances to the Center Partyand the Social Democrats and Communists. The mass theo­ry (Arendt 1958; Kornhauser 1959) posits that it was thealienated and isolated (surrogately measured by unemploy­ment and electoral turnout) who were attracted to the NaziParty's totalitarian structure. Finally. the class theory(Lipset 1960) argues that the economic policies of the Nazi

with p and 8 replaced by their instrumental estimates(Anselin 1992. sect. 28, p. 3). Then, pseudo-error terms arecreated by random sampling (with replacement) from thevector e. Estimates for p and B in the resampled data areobtained by means of instrumental variables, adopting thespatial lag WY (Anselin 1992. sect. 28. p. 4). This proce­dure is repealed a large number of times (999 in the follow­ing example) to generate a frequency distribution for the pand 1\ estimates. The reported bootstrap estimation is themean of this frequency (see Anselin 1988.91-96). One usesthe jackknife and boot trap techniques when appropriaterather than transforming the dependent variable to createnormality.

It is of methodological importance to identify and incor­porate the spatiality within the data. because the existenceof both heterogeneity and dependence violates the assump­tions of multiple regression. Spatial dependency violatesthe assumption of serial and spatial independence of thevariables and of the residuals (Ansel in 1988, 11-14). Withregard to spatial heterogeneity, the presence of het­eroskedasticity produces unbiased but inefficient OLS esti­mates so that inference based on the I and F statistics willbe misleading and the measures of fit will be wrong.

The theoretical implication of neglecting the spatiality ofthe data in statistical rrodeling is that place-specific con­texts would be ignored. In the case of the Nazi Party. theinteraction between counties that underlies spatial depen­dency informs the linkages between places a<;a componentof context and illustrates how the agency of political partiescreated spaces of power (Flint 1998). In tum, these newspaces of power changed existing institutions and practicesto create a changing place-specific geography of politicalbehavior. The outcomes of these processes manifest them­selves in the form of spatial heterogeneity. The followingdiscussion of the spatial statistical analysis of the Nazi Partyvote in Bavaria illustrates how spatial statistical modelsmay be interpreted 10 illuminate the role of politicalprocesses in constructing places.

Those who choose to study elections held in Weimar Ger­many are restricted by the available data. No survey datathat would allow for individual-level analysis are availableand, therefore. one is restricted to the use of aggregate data.

e=y-pWY-XB.

skedasticity (Mackinnon and White 1985). However. thejackknife estimate is not robust to spatial dependence; whennonnormality and spatial dependence are present, the boot­strap estimation technique should be adopted (Anselm1988. 92-96). The bootstrap estimation uses the random­ness present in artificially created resampled data sets as thebasis for statistical inference. Estimates are calculated froma large number of data sets, obtained by random samplingwith replacement. First. instrumental variables are used toprovide an estimate for the error vector in the form of theresiduals:

37Wimer 2002. Volume 35. Number I

stag election of 1928, and Gaue (new organizational dis­tricts) were created to organize local branches into regionsalmost identical to Reichstag electoral districts (Orlow1969, 139). In Bavaria, the eight sub-Gaue were used toorganize saturation campaigns leading up to the Reichstagelection of 1930 (Pridharn j 973. 97). In March L930, tbeNazi Party introduced its agrarian program in an attempt toattract the rural vote. The agricultural crisis of 1930 coin­cided with the Nazis' reorientation to agricultural issues, thereorientation of the German National Party (DNVP) towardthe interest of big business in 1928, and an internal crisis inthe Bavarian Peasants' and Middle-Class League (BBMB),an economic interest party. During 1929 and 1930. the NaziParty (NSDAP) increased the number of meetings held inrural areas, as it believed that political meetings were themost effective way of gaining the support of previous non­voters (138). In addition. the Nazis put up candidates in dis­trict council elections in Middle and Upper Franconia toincrease their profile (121).

The content of the Nazis' message leading up to theReichstag election of 1930 varied from area to area depend­ing on the nature of the electorate (Pridham 1973). InProtestant Upper Franconia, the party stressed a conserva­tive message; in Catholic areas, the Nazi Party argued thatChristianity was compatible with anti-Semitism and that theCatholic parties had lost credibility because of their coop­eration with the Weimar government (137). The Nazis'share of the vote in the Reichstag election of 1930 wasLightly lower in Bavaria than the national average of 18percent. Within Bavaria, the Nazis' support was strongest inthe Protestant regions of MiddLe and Upper Franconia,gaining 23 percent of the VOte.Modest support, of about 11percent. occurred in Catholic Lower Bavaria and the UpperPalatinate (140). The NSDAP was especially strong in areasof mixed religious belief. such as the towns of Mernrningen,Neu-Ulm, and Nordlingen in Catholic south Bavaria (140).However. the Nazi Party also did well in some stronglyCatholic towns.

The ability of the Nazis to project a different message indifferent areas, combined with the spatial pattern of reli­gious affiliation, leads to an expectation of heterogeneity inthe nature of the party's support and the need to estimatestructural change models. In Bavaria, locational and institu­tional aspects of context (Agnew 1987) were interacting toform an environment beneficial to the Nazi Party. TheNSDAP was creating a more efficient organization anddefining its message in relation to the economic context; inaddition, its political competitors were becoming weaker.The organizational changes that the party initiated werecoupled with the dissemination of a new message designedto appeal to the electorate in Light of the economic crisis.Such activity and opportunity lead to an expectation for thediffusion of Nazi Party support during this period ofchange. This diffusion would be manifest in the modelingtaking the form of a positive and Significant spatially lagged

The growth of the Nazi Party vote in Bavaria betweenMay 1928 and July 1932 illustrates how the geographicalcontext of the voter produced place-specific behavior andhow the Nazi Party generated support by making institu­tionalized connections between places. Across the whole ofGermany, the Nazi Party increased the size of it'>electoralsupport from 2.6 percent in the Reichstag (parliament) elec­tion of May 1928 to 18.3 percent in September 1930 and upto 37.3 percent in July 1932. The trend of increased supportwas similar within Bavaria, with the actual percentage ofthe vote increasing from 6.8 percent in 1928 to a size justunder the national average in the otber two elections (Prid­ham 1973).

The Reichstag election of September 1930 marked thebreakthrough for the Nazi Party. Nationally, the party reor­ganized itself after its national failure in the previous Reich-

Example: The Nazi Party Vote in Bavaria,May 1928 to July 1932

Parry attracted the self-employed middle class. The sevenvariables measured (I) the percentage of the populationwho were Protestant; (2) the percentage of the work forcewho were manual industrial workers; (3) the percentage ofthe work force who were blue-collar workers in trade andtransport; (4) the percentage of the work force who wereself-employed; (5) the percentage of the work force whowere unemployed; (6) the electoral turnout as a percentageof eligible voters for the second of the two elections withinthe period of change in the Nazi vole: and (7) the temporal­spatial lag discussed below.

These variables test the dominant theoretical frameworksused to explain the Nazi Party vote. [n addition, the variablemeasuring the percentage of the work force who were blue­collar workers in trade and transport was found to be animportant explanatory variable in previous analysis thatmodeled the support for the Nazi Party using a regionalapproach. For this reason, it was included in the analysis(0'Loughlin, Flint, and Anselin 1994). This variable teststhe tendency for a section of the working class, includingartisans, that most closely empathized and identi.fied withthe middle class to support the Nazi Party (Ault andBrustein 1998). The additional explanatory variables weresuggested by a stepwise regression analysis that did notinclude the temporal-spatial lag. The inclusion of theseadditional variables was necessary to counter the critique ofelectoral geography that contextual influences identified bygeographers are a function of poorly specified models thatdo not include all the relevant explanatory variables (McAl­lister 1987). In addition to evaluating the significance ofthese variables, diagnostic tests reported the type of spatial­ity in the data. Both the diagnostic tests and significance ofthe variables were interpreted to estimate subsequent mod­els that contained only the significant variables while alsoincorporating the correct type of spatiality.

HISTORICAL METHODS38

tions and political discontent, In addition, it appears thatrural areas were more likely to support the Nazi Party thanurban areas. Sucb a result would be consistent with the in­ception and dissemination of the Nazi Parry's agrarian pro­gram. The positive value of the variable measuring the per­centage of the work force who were white-collar workers inagriculture and the negative sign of the variable measuringthe percentage who were blue-collar domestic employees, aprobable surrogate measure for an urban population, sup­ports the interpretation of the diffusion of Nazi Party sup­port into rural areas during this period of change. However,the difficulties of working with aggregate data become clearwhen one tries to interpret the positive sign of the variablemeasuring the percentage of the work force employed in tbetrade-and-transport sector, which would seem to be measur­ing a primarily urban form of employment.The following period of change-September 1930 to

July 1932-marked the Nazi Party's sustained surge inelectoral popularity, During that time. the NSDAP defineditself as a catchall protest party to gain the votes of sup­porters of a mixture of economic interest parties. InBavaria, the Nazi Party's rural campaign only became fullymobilized after the Reichstag election of 1930, and the ruralpropaganda machine was slower to develop in the BavarianGaue tban in other parts of Germany (Pridharn 1973,226).rn 1931, the Bavarian Galle published a newspaper aimedexclusively at peasants, but the party lacked speakerstrained in rural issues (231). The Nazi Party promoted anantisrate as well as an anti-Jewish message. promising tocancel debts to Jewish cattle traders (232). In the Reichstagelection of July 1932, the Nazi Party became the largestparty bur stiU failed to obtain an absolute majority, Facedwith electoral weariness and a conservative national gov­ernmeru, the Nazi Party clothed the election in terms of ajudgment on the Weimar Republic as a whole (281).

The modeling of the change in the Nazi Party votebetween 1930 and 1932 (table 2) illustrates two importantfeatures of a spatial statistical approach. First, the hetero­geneity within the initial models required the estimation ofa structural change model with spatial regimes defined bythe Protestant enclave of Central and Upper Franconia.Though places with a relatively high Protestant populationwere more likely to support the Nazi Party across the wholeof Bavaria, there were subregional differences in the sup­port from places experiencing unemployment and places ofmanual industrial employment and unemployment. In otherwords, the regional context of Central and Upper Franconiaproduced support from places with no particular socioeco­nomic characteristics, that is. working class or middle class.On the other hand, in the rest of Bavaria, voters in placesexperiencing economic distress supported tbe Nazi Party.unless the unemployed were manuaJ industrial workers.

The reported diagnostic tests support these conclusions(see table 3). The significant values of the Chow-Wald teston the two regimes as a whole as well as the individual vari-

Note: R' = .63: Sq. corr. = .S I.'p< .OS; .'p< .01.

0.143.960.010.050.151.031.30

U.44**t5.32··0.07'"-0.19"·0.69**

-2.43-2.81-

ConstantSpatially lagged dependent variableProtestant populationElectoral turnoutTrade-and-transport workersBlue-collar domestic workersWhite-collar fann workers

SDCoefficientVariable

TABLE 1Nazi Party Electorate in Bavaria, May 1928-

September 1930: Spatial-Lag Bootstrap Estimation

variable, Theoretically, it would represent the role of theinterlinkages between places in defining the context of thevoter's decision as well as the agency of the Nazi Party cre­ating new spaces of political strength.

The results of the bootstrap estimation of the change inthe Nazi Party vote in Bavaria between 1928 and 1930 arepresented in table 1. The nonnonnality of the data requiredthe bootstrap estimation technique that provided for robustestimates but did not facilitate the reporting of diagnosticstatistics. The model reported was the outcome of a processstarting with the estimation of an OLS model in the Space­stat (Anselin 1992) software consisting of a choice of vari­ables informed by theory as well as an earlier stepwise esti­marion. By interpreting the diagnostics of this model andlooking at the statistical significance of the independentvariables, the model reported in table I was deemed themost appropriate. At the top of the table, the pseudo-z?and squared correlation between the observed and predict­ed values are reported as measures of goodness of fit.

The significance and positive sign of the spatially laggeddependent variable support the expectation that the NaziParty was able to diffuse its electoral strength acrossBavaria during this period of change. The spatially laggeddependent variable indicates that the change in the size ofthe Nazi Party vote in one Kreis was partly a function of thechange in neighboring Kreise. In other words, increasedsupport in a particular county arose from the increased suc­cess, profile. and activity of the Nazi Party in neighboringareas. Following Masse} (1994), the statistical modelingincorporates the interlinkages between places as part of thecontextual setting explaining voting behavior.

The interpretation of the socioeconomic independent var­iables in table 1 reveals both the strengths and weaknessesof using aggregate data. The independent variables do notinform us about the nature of individuals who voted for theNazi Party. Rather, they suggest that predominately Protes­tant places with relatively low electoral turnout in the 1930Reichstag election nurtured support for the Nazi Party,which may be interpreted as a context of Protestant institu-

39Winter 2002. Volume 35. Number I

In summary, the two models discussed exemplify themethodological utility of spatial statistics. First, the inclu­sion of the spatially lagged dependent variable negates theproblem of a lack of independence between the observa­tions. Second, the estimation of a structural change modelremoves the heteroskedasticity from the data and thus pro­duces efficient OLS estimates. The estimation of differentspatial regimes illustrates how different contextual settingsinteracted with the agency of the Nazi Parry to producepJace-specific political behavior. Finally, the spatial errorestimation controlled for heteroskedastic and correlatederrors and so allowed for more accurate inference.

The models also illustrate how the theoretical notion ofsocially constructed contexts ca.n be operationalized via

64(J.626

6.066.19

8II

Likelihood ratio testWaId test

.0165.77

Test (J11 Cl)lIl11/(}/1 factor hypothesis

Likelihood ratio test

.790.07

Diagnosticsfor spatial dependence

Breusch-Pagan te~t

Diagnosticsfor irefem.lk(da.ltidr_r

.000

.652

.000

.000

.000

Stability of individual coefficients

17.080.2014.0313.571359

CONSTANTPROTMANtNDUNEMPMANUNEMP

.00026.485Chow- WaJd test

I'dfTest

TABLE 3Regression Diagnostics: Test on Structural Instability for Two

Regimes Defined by Central and Upper Franconia

Value

ables show differences in the regression slopes between thetwo subregions for all the variables except Protestant. With­in Bavaria, place-specific contexts were constructed throughthe interaction of Nazi Party activity with previous nets ofsocial relations to produce a regionalization in the composi­lion of the NSDAP's electorate. Following Agnew (1987)and Massey (1994), the identification and interpretation ofspatial regimes illustrate how the combination of institution­alized practices created place-specific social behavior.

The second key component of the model is the Significantvalue of lambda and Lhe other diagnostic tests. The diag­nostic tests of the initial models had indicated the suitabili­ty of not only a structural change model but also a spatialerror estimation to con IToI for the spatial autocorrelationacross the error terms. The inclusion of the spatial auiore­gressive coefficient (lambda) controls for the spatial errorautocorrelation and so produces unbiased and efficient esti­mares (Anselin 1992, sect. 29, p. 2). The statisticallyinsignificant value of the Breusch-Pagan lest shows that thespatial regimes chosen for the structural change model badremoved all the heteroskedasticiry from the data. Finally,the tests for spatial error dependence and the common fac­tor hypothesis illustrate that the spatial autocorrelation pres­ent was not substantive but rather was contained in theerrors, confirming the choice of the spatial error estimation.At the top of the table, the values of R2 are complementedby the addition of measures of goodness of fit based onmaximum. likelihood estimates. The R2 is inaccurate in spa­tial regression models, and so these other measures arereported to allow for a comparison of models (Anselin1992, sect 26, p. 4). LIK is the logarithm of the joint like­lihood of the maximum likelihood parameter estimates. TheAkaike (198 l) information criteria (ArC) and the Schwartzinformation criteria (SC) are different forms of correctionfor tile joint likelihood's tendency to increase when addi­tional variables are added [0 the model (Anselio 1988,246). In a comparison of models, the one with the best fit isthe one with the lowest value for an information criterion.

Note: R' = .86: Sq. corr, = .83: UK (logarithmof joint likelihood of maximum likelihood parameter estimates) =-420.-t3: AIC (A~aika informauon criteria) = 8(>0.85:SC (Schwartz iinfonnaticn criteria) = 890.69.*1' < .05:"1' < .01.

Central and UpperFranconia Other counties

Variable Coefti.cient SD Coefficient SD

Constant nAO"" 1.56 42.59*'" 5.89Protestant population 0.29** 0.02 027** 0.04Manual industrial employment -0.25** 0.08 -O.89M 0.15Unemployment -{l.18 D.25 2.85*$ 0.78Manual unemployment -0.03 0.06 -0.95*" 0.24). 0.27" 0.11 0.27* 0.11

TABLE 2Na7j Party Electorate in Bavaria. September 1930-July 1932:

Structural-Change Spatial Error Estimation

HISTORICAL METHODS40

Agnew. J.A. 1987. Place and politics: The geographical mediation of stateand society. Boston: Allen & Unwin.

--. 1996. Mapping politics: How context counts in electoral geogra­phy. Political Geography 15: 129-46.

Akaike, H. 19S I. Likelihood of a model and information criteria. Journalof Econometrics 16:3-14.

Anderton. D. L.. and D. E. Sellers. 1989. A brief review of contextual­effect models and measurement. Historical Methods 22: 106-15.

Anselin. L. 1988. Spatial econometrics: Methods and models. Dordrecht,Holland: Kluwer Academic Publishers.

---. 1992. Spacestat: A program/or statistical analysis of spat illI data.Santa Barbara: University of California. National Center for Geograph­ic Information & Analysis.

Anselin. L., and S. Rey. 1991. Properties of tests for spatial dependence inlinear regression models. Geographical Analysis 23: 112-31.

Applegate. C. 1990. A nation of provincials: The German idea of Heimat.Berkeley: University of Califomia Press.

Archer. J. C.. and P. J. Taylor. 1981. Section and party. Chichester: JohnWiley.

Arendt. H. 1958. The origins of totalitarianism. Cleveland: World Pub­lishing.

Artzrouni. M .. and J. Komlos. 1996. The formation of the European statesystem: A spatial "predatory" model. Historical Methods 29: 126-34.

Ault, B.. and w. Brustein, 1998. Joining the Nazi Party: Explaining thepolitical geography of NSDAP membership. 1925-1933. AmericanBehavioral Scientist 41: 1304-23.

Benneu, S .. and C. Earle. 1983. Socialism in America: A geographicalinterpretation of its failure. Political Geography Quarterly 2: 31-55.

Breusch, T. S., and A. R. Pagan. 1979. A simple lest for heteroskedasticltyand random coefficient variation. Econometrica 47: 1287-94.

Brustein. W. 1985. Class conflict and class collaboration in regional rebel­lions. 1500-1700. Theory and Society 14: 445-68.

Burnham. W. D. 1972. Political immunization and political confessional­ism: The United States and Weimar Germany. Journal of Interdiscipli­nary History 3: 1-30.

Chow, G. C. 1960. TesLSof equality between sets of coefficients in two lin­ear regressions. Econometrica 28: 591-{)05.

Cook. E. M. Jr. 1980. Geography and history: Spatial approaches to earlyAmerican history. Historical Methods 13: 19-28.

REFERENCES

l. Gary King's (1997) considerable methodological advancements inthe ability to make ecological inferences when using aggregate data con­tinue to deny the ontological necessity of considering individual behaviorwithin structural constraints.

2. The election file includes returns for the Reichstag elections between1920 and 1933, and the census file contains socioeconomic data collectedfrom a variety of sources and at a number of different times. Jiirgen Fal­ter and Wolf Gruner (1981) have described how this data set was revisedduring the 19705 not only to correct data errors, which were mainly a resultof punching errors. but also to compensate for internal political boundarychanges within Weimar Gennany to make the territorial units within thedata set as consistent and coherent as possible. The changes in internalpotitical borders in Weimar Germany were a product of major changes inadministrative units between 1919 and 1933 that were partly a result of theincorporation of suburbs into urban areas and the reform of local govern­ment. Falter and Gruner believe that the sources of the data set are therespective volumes of the Statistik des Deutschen Reiches (Statistics of [heGerman Reich) issued by the Statistisches Reichsamt (State StatisticalOffice) in Berlin.

3. Rusty Dodson, David Fogel. and Steve Kirin of the Department ofGeography. University of California-Santa Barbara, coded the originalmap into a GIS. Subsequently, Michael Shin of the Department of Geog­raphy and Regional Studies, University of Miami, and r revised this mapto incorporate the boundary changes and reduce the Dumber of KreisuniisOD the map to 743.

4. The Spacestat software developed and distributed by Luc Anselin(1992) allows for the estimation of OLS models with spatial diagnostics aswell as a host of descriptive spatial statistics and spatial econometric models.

NOTES

Context is a dynamic phenomenon and one that is insep­arable from individual and group actions. Through a casestudy, we have tried to show that places frame. and are inturn shaped by, the actions of individuals and groups. Con­text determines the amount and content of informationbeing received by individuals and groups. the institutional­ized norms and practices that will filler and organize thatinformation, and the range of possible responses (Thrift1983). In turn, such responses will change the nature of aplace and its role in Framing the next round of challengesand responses. Thus. a dynamic and geographically sensi­tive notion of context is essential for understanding histori­cal actions.

In addition, the key components of the geographic notionof context may be operationaLizedto allow for a modelingof aggregate data sensitive to place-specific settings. Inter­actions between places may be incorporated into statisticalmodels via the concept of spatial dependence. The incorpo­ration of spatial heterogeneity identifies how the spatial cir­cumstances of the individual produces place-specificbehavior. These two concepts and the processes they repre­sent are, of course. inseparable.When they model historicalbehavior by means of aggregate data. actors need not be iso­lated from the everyday contexts that mediate their activi­ties and decisions.The implications of the inclusion of context in historical

studies address issues of ontology and the division of acad­emic inquiry along disciplinary fractures. First, an under­standing of the mediating role of context diminishes theexplanatory power of an ontology that focuses on individu­als and groups at the expense of their geographic setting.Understanding is not simply a matter of noting the particu­lar institutions. attitudes. and power politics that set limitsto particular actions. Instead, the mutual process of con­structing contexts and historical projects needs to beemphasized. Finally, the dynamic nature of the creation ofunique places, and the techniquescapable of modeling themillustrated in this article, show the necessity and possiblechannels of a fruitful dialogue between geograpbers andhistorians.

spatial statistics. In the first model, the role of dynamicinterlinkages between places (Agnew 1987;Massey 1994)was incorporated into the model by the i.nclusionof the spa­tially lagged dependent variable. In other words, spatialdependence was included as part of the explanation of theincrease in the Nazi Parry vote. The Nazi Party's ability tobuild support was partly a function of its ability to spread itsmessage across space. In the second model, spatial hetero­geneity in the electorate's behavior was illustrated by theestimation of different models for separate spatial regimes.Thus, place-specific political behavior, or the effect of dif­ferent contextual settings. was directly incorporated into theanalysis.

Conclusion

41Wimer 2002. Volume 35. NUID:Jer I

McAllister, I. 1987. Social context, turnout, and the vote: Australian andBritish comparisons. Political Geography Quarterly 6: 17-30.

Melnig; D.W. 1978.The continuous shaping of America:Aprospectus forgeographers and historians. American Historical Review 83: 1186-1205.

Monkkonen, E. H. 1.984.The challenge of quantitative history.HistoricalMethods 17: 86-94.

Oakes, T. 1997. Place and the paradox of modernity.Annals of the Associ­ation of American Geographers 87: 509-31.

Office of Strategic Services. 1944. German COl/lily (Kreise) Boundaries.Map No. 6289.

O'Loughlin.T. 1993.Geo-economic competition in the Pacific Rim: Thepolitical geography of Japanese and US exports, 1966-1988. Transac­tions, Institute of British Geographers N.S. 18: 438-59.

O'Loughlin, 1., and L.Anselin. 1992.Geography of international conflictand cooperation: Theory and methods. InThe new geopolitics, edited byM. D. Ward, 11-38. Philadelphia: Gordon and Breach.

O'Loughlin, 1.. C. Flint, and L.Anselin. 1994. The political geographyof the Nazi vole: Context, confession and class in the 1930 Reichstagelection. Annals of the Association of American Geographers 84:351-80.

Orlow. D. 1969. The history of rile Nazi Parry. 1919-1933. Vol. 1. Pitts­burgh: Pittsburgh University Press.

Pridham, G. 1973.Hitler's rise 10 power-The Nazi movement in Bavaria,1923-1933. London: Hart-Davis, MacGibbon.

Sack. R. D. 1997. Homo geographicus: Aframework for action, awareness,and moral concern. Baltimore,Md.: Johos Hopkins University Press.

Staeheli. LA., 1. E. Kodras, and C. Flint. 1997. Introduction: [n Statedevolution in America: implications for a diverse society, edited byL. A. Staeheli, J. E. Kodras, and C. Flint, xii-xxxiii. Thousand Oaks,Calif.: Sage Publications.

Taylor,H. 1984. The use of maps in the study of the black ghetto-forma­tion process: Cincinnati, 1802-1910. Historical Methods 17:44-58.

Taylor. P. J. 1981. Geographical scales within the world-economyapproach. Review 5: J-U.

Thrift, N. 1983. On the determination of social action in time and space.Environment and Planning D: Society and Space I: 23-57.

While. H. 1980.A heteroskedastic-consistenr covariance matrix estimatorand a direct lest for heteroskedasticity,Econometrica 48: 817-38.

Cronen. W, 1991. Nature's metropolis: Chicago and the Great WesT.NewYork.:W. W. Norton.

Earle. C, 1992. Geographical inquiry and American historical problems.Stanford: Stanford University Press.

Falter, J. W. 1990. The first German volkspanei: The social foundations ofthe NSDAP. 1.n Elections, ponies, and political traditions: Social foun­dations of German parties and party systems, edited by K. Rohe,53-81.Munich: Berg.

Falter, J. W. 1991.Hitters Wahler. Munich: C.H. Beck,Falter, J. W.. and W. D. Gruner. 1981.Minor and major flaws of a widely

used data set: The ICPSR "German WeimarRepublik:data, 1919-1933"under scrutiny."Historical Social Research 20: 4-26.

FLint,C. 1998, Forming electorates, forging spaces: The Nazi Party voteand the social construction of space..American Behavioral Scientist 41:1282-1303.

Griffiths, M. J., and R. J. Johnston. 1991.What's in a place? An approacbLO the concept of place, as illustrated by the British National Union ofMlneworkers' strike, 1984-85. Antlpode 23: 185-213.

Hochberg, L. J., and D. W. Miller. 1992.Constructing a central-place bier­archy from a commercial directory.Historical Methods 25: 8(}"94,

Johnston. R. J. 1991.A question of place. Oxford: Blackwell.Johnston, R. J.. and C. Pattie. 1988. Changing voter allegiances in GreatBritain. Regional Studies 22: 241-75.

Kelejian, H., and D. P. Robinson. 1992. Spatial aurocorrelation: A newcomputationally simple test with an application to per capita county pol­icy expenditures. Regional Science and Urban Economics 22: 317-31.

King, G. 1997. A solution to the ecological inference problem: Recon­structing individual behavior from aggregOle data. Princeton. N.J.:Princeton University Press.

Koenker, R.. and G. Bassett. 1982. Robust tests for hereroskedasticitybased on regression quantiles. Econometrica 50: 43-{i1.

Kornhauser,W. 1959.The politics of mass society. NewYork: Free Press.Lipset, S. M. 1960. Political man: The social bases of politics. Garden

City. N.Y.: Doubleday.Massey, D. 1994. Space. place. and gender. Minneapolis: University ofMinnesota Press.

MacKinnon, J. G., and H.White. 1985.Some hereroskedasuciry-consisiemcovariance matrix estimators with improved finite sample properties.Journal of Econometrics 29: 305-25.

HlSTORTCAL METHODS42