Explaining Parliamentary Party Unity QCA as a Research Tool for Comparative Legislative Studies

Erik Fritzsche 1 Explaining Parliamentary Party Unity QCA as a Research Tool for Comparative Legislative Studies Abstract: I want to find out which combinations of causal factors lead to parliamentary party unity. In order to solve the puzzle I do world-wide comparative research. This paper shows how such a research project can be put into action. Qualitative Comparative Analysis (QCA) and its multi value extension (MVQCA) provide helpful tools to tackle problems that usually arise out of comparative research. Applying QCA, it is possible to use in a transparent way even the poorest data sources and productive speculations regarding non-observations. Moreover, the researcher can identify different causal pathways to the phenomenon in question each with a combination of causal factors. In addition, one of the most important advantages of QCA is that it makes the dialogue between theory and data clearly visible to researchers. Therefore, it becomes likely to detect insufficiencies of research designs (e.g. regarding the considered variables, measurement, data quality, and case selection). I demonstrate step-by-step what (MV)QCA contributes to tackle the problematic nature of current comparative studies of parliaments analyzing the patterns of parliamentary party unity. After briefly introducing (MV)QCA, I point out how holistic single country studies and comparative causal analysis can be brought together with these methodological tools. I show in detail how QCA should be used in a correct manner and how the hermeneutic adventure takes place while analyzing data. This paper, therefore, makes explicit what comparative researchers do when they try to make most of their examined cases and what the added value of (MV)QCA is. Furthermore, it provides a step toward a more comprehensive theory of the causes of parliamentary party unity. 1. Introduction One of the key tasks of political science is to build theories of some range. In political science, when questions on causal patterns regarding political systems and institutions are to be answered, this is usually tried by comparison. However, comparing is difficult for a number of reasons. I want to present a way in which political scientists can design their comparative research. The QCA approach, that I use here, had a quite remarkable take off in the last years. Therefore, it is by no means new to use it. However, QCA has been used rarely the way Charles C. Ragin intended it in his first book on that topic in 1987 (see Ragin 1987). He wrote about a qualitative comparative research strategy. Nowadays, it is preferred to speak about a configurative approach (Rihoux/Ragin 2008). This terminology underlines the more technical aspects of QCA, i.e. seeking to detect necessary and sufficient conditions in a set of configurations that are represented by a set of observed conditions. These data analytic strategies make it possible to have an eye on the cases that are linked with analytic statements. It is therefore case-oriented research. However, the qualitative nature of oscillating between cases and data that Ragin described in 1987 (and as he did when he introduced the fuzzy set approach; see Ragin 2000) was hard to fulfill. My impression is, therefore, that a dialogue between theory and data could be more intense and explicit than it has been usually performed. In this paper I seek to demonstrate how such a qualitative comparison that oscillates between theory and data may look like. In order to show this, the paper here attempts to contribute to the illumination of the genuine research question of legislative studies: What causes party unity? 1 The author of this paper is a doctoral candidate at the Technical University of Dresden, Department of Political Science. Comments may be addressed to erik@fritzsche-net.de. 1

Party unity is an important phenomenon for legislative studies for at least two reasons. First, party unity is a precondition for high performance parliamentary types of government because it contributes to the stability of the government. Nevertheless, not every causal pathway that may lead to party unity is desirable (e.g. discipline by free choice versus corruption). Second, in presidential systems solid blocks of parties may induce a deadlock situation. On the other hand, if parties are very loose bodies, transaction costs for the realization of legislative projects may explode (cf. Mainwaring/Shugart 1997: 418-421). Therefore, my research phenomenon is located at the center of the types of government. It is essential for the functioning of both types of government. Consequently, it is worth asking what may cause different levels of party unity. There is quite a lot theoretical literature on the topic dealing with models that may help to perform comparative research (see Patzelt 2003, Bergman et al. 2003, Hazan 2003, Olson 2003, Carey 2007, Sieberer 2006, Jensen 2000, Bowler/Farrel/Katz 1999, Saalfeld 2005), some have been developed while a particular parliament was investigated (see Messerschmidt 2005, Saalfeld 1995, Skjæveland 1999, Valencia-Escamilla 2005). However, there is sparse comparative work available yet which is moreover either illustrative (Ozbudun 1970); restricted to certain regions (Mainwaring/Shugart 1997, Jensen 2000, Morgenstern 2004, Sieberer 2006); limited to statistical techniques and a subset of relevant variables (Poiré 2003, Carey 2007, Kailitz 2008); or exploring some hypotheses by taking into account only very few cases (Depauw 2003, Döring 2003, Carey 2003, Davidson-Schmich 2003 and 2006, Thames 2005 and 2007). Apart from that there are a lot of single case studies (see Fritzsche, forthcoming). This paper aims at the development of a more general theory on the causes of party unity by comparing parliamentary party groups (PPGs) from all over the world. In order to do that, a research design will be presented that may help to uncover the complex causal patterns behind this phenomenon. I will outline how the data was captured and which data was collected. The central part of the paper, however, is data analysis. In this stage of the presented research the dialogue between theory and data will become explicit and it will be shown which expansions to the resign design of this paper seem to be promising. One best understands the paper as the presentation of a relatively extensive pretest or a first walk on the hermeneutic spiral it will not finish at the top of that spiral, but it will show which routes are likely to be promising. As usual with methodological contributions, the journey is the reward. Before starting, a brief technical introduction to QCA is given that refers to the central aspects of a hermeneutical use of this research tool. 2. The QCA-Approach: A Brief Technical Introduction I first want to introduce briefly some technical aspects of QCA as presented by Ragin 1987. Before data analysis the researcher has to do the following steps. - As usual among comparative research designs, the researcher has to set up some hypotheses regarding the phenomenon of interest, he has to identify the corresponding variables and he has to find solutions for the respective measurement. - The obtained data will be referred to as raw data. - QCA uses binary data, MVQCA (see Cronqvist 2005) may also use multi-value scales. If data is not binary, the researcher has to dichotomize his data. This can be done using theoretical arguments, by constructing indices, or by using thresholds for metric data. The respective variables are called conditions, the values of the dependent variable are called outcome. Of course, this dichotomization is a reduction of information. However, this is the cost of QCA. 2

- Then a truth table can be presented. Every row of this truth table contains a combination of truth values of the respective conditions i.e. a row does not necessarily imply one case (as is the case in the raw data set). The number of observed cases with a certain combination of present/absent conditions (= a row) are shown in a special column which is labeled number of instances. A combination of truth values of the conditions and the output the researcher wants to explain (e.g. high/medium/low party unity) is called a configuration. The described truth table is after its compilation out of the raw data set the starting point for an analysis performed with the algebra of George Boole. Using the Boolean algebra, it is possible to minimize the truth table. Minimization is the process by which the information contained in the truth table will be compressed to a relatively simple logical expression: a Boolean equation. - There are several configurations with the outcome that the researcher wants to analyze. For example, the researcher wants to know what may cause a high level of party unity, where high level of party unity is the value of the outcome variable party unity if it has the value 1 (= high level of party unity is present). - The minimization process is strictly comparative (or we may say experimental in its logic). The logical rule to minimize a condition, i.e. to state it is irrelevant with regard to the outcome the researcher whishes to explain, is as follows: If two Boolean expressions differ in only one causal condition yet produce the same outcome, then the causal condition that distinguishes the two expressions can be considered irrelevant and can be removed to create a simpler, combined expression (Ragin 1987: 93). There are two facts to note, here. First, the comparison is made with two Boolean expressions, i.e. combinations, and not cases. However, combinations imply cases. Secondly, only those combinations are used for minimization whose outcome is the one we are interested in (e.g. high levels of party unity and the corresponding value 1 of the outcome variable), and hence whose output code is the same. - After the first step of identifying possible minimizations, starting with the combinations of the truth table with the outcome in question, a second step with the resulting minimized combinations of the first step will follow. This will be continued until no further minimizations are possible. The resulting expressions are called prime implicants. A prime implicant is a Boolean expression that cannot be minimized again. Each of the prime implicants implies the combinations that were used to derive it (and the cases that are implied by these combinations, i.e. which have been observed with these combinations of causal factors). The analyst uses a prime implicant chart to reduce some occurring redundancy, i.e. that different prime implicants imply the same combination. - The result of such a minimization is a reduced Boolean equation, i.e. an expression that implies all the combinations with the outcome in question. It typically consists of prime implicants and it is written that way: OUTCOME IN QUESTION = A(1)B(1)C(1) + B(1)C(0) + D(0), where the conditions A(1), B(1) and C(1) are linked with each other by Boolean multiplication, i.e. the logical and ; and where the terms A(1)B(1)C(1) and B(1)C(0) are linked with the Boolean addition, i.e. the logical or. A 1 in 3

parentheses indicate the presence; a 0 in parentheses indicate the absence of the respective condition A, B, C or D. 2 - The equation has to be read as follows: If the output in question is present, then the presence of A and B and C or the presence of B and the absence of C or the absence of d is given. Note that the result is not only a multivariate expression, it is an equifinal one, as well: It shows that different causal pathways may lead to the outcome we wish to explain. The calculation can be delegated to computer programs. Here, I use TOSMANA from Lasse Cronqvist. 3 What is the added value of such a data analytical tool, and which problems may occur? To show some answers to that question, I will demonstrate this in the next sections of the paper by presenting a research design and doing data analysis as a hermeneutical enterprise. 3. The Causes of Party Unity Applying QCA Due to limited knowledge about the research object a qualitative research process is less constrained and essentially open with regard to the concepts and measurement, the selected cases, the obtained data and the possible inferences than research processes on topics about which we know relatively much. However, qualitative research has not to be chaotic. Here, I present the examined cases, the dependent variable and concepts, hypotheses, measurements and data on parliamentary party unity. The obtained data matrix will be analyzed with the help of the above mentioned algorithms of QCA based on Boolean algebra that make multivariate and configurative analysis possible. The solutions will be interpreted with grip to the cases and the theoretical statements I outline at the beginning of this section. This helps to get insights regarding the research design: Did variables and measurements prove their worth or is there some need to adjust them or to add new conditions? Did the case selection turn out to be promising, and in which direction should I expand the analyzed empirical material? As I show, there are some inadequacies of my research design that ask for adjustment and that could hardly be known in advance of this study. This underlines that the search for answers regarding the research question is a hermeneutical enterprise, and that the answers, here, are very tentative. The good news is, however, that it is clear what can be learned from the present research effort for my future research activities on the question raised here. 3.1 Case Selection Often, parliaments are the units of analysis among studies on party unity. That is plausible if the PPGs within a parliament seem to be relatively similar. Nevertheless, there are PPGs within a parliament differing so tremendously that it does not seem to be a just treatment to choose parliaments as the units of analysis. Therefore, I chose individual PPGs to be the units of analysis here. The following table shows the examined PPGs. However, often the obtained data showed that PPGs within a parliament are equal. Where this has been the case, the respective PPGs of the same parliament were summed up in a line of the dataset and were labeled with the same case id. For example, the case id DB comprises all PPGs of the German Bundestag that have been investigated in my research, i.e. CDU/CSU, SPD, FDP. In contrast to the cases of the Italian PPGs: The DC and PSI had a different configuration of 2 One may also use upper/lower case notation if a dichotomous data set is used. 3 Cronqvist, Lasse (2006a): Tosmana - Tool for Small-N analysis. Version 1.254. Marburg. Download: http://www.tosmana.net. 4

conditions than the PCI. Therefore, I constructed two case ids for these cases: ICDCPSI and ICPCI. [insert table 1 here] The case selection was made relatively pragmatic. To gain empirical variance in my dataset I chose to collect data on Western and Eastern European, Scandinavian as well as North and Latin America PPGs. Both the workload that can be met within the research process and the available literature regarding the cases and variables I chose to collect data on forced me to be pragmatic. However, one needs to begin with a promising set of cases which is open for reassessment. Later on, I show how in the process of data analysis possible directions of the expansion of cases to analyze uncover. How were the time periods of the cases determined? I tried to construct the time periods in a way that the respective cases were relatively stable with regard to the investigated conditions and party unity. For example, my statement concerning the Mexican PPGs is limited to a relatively small time period from 1998 to 2000. This is the case due to the government of the PRI that ended in 2000 with the last PRI president, Ernesto Zedillo. After 2000, patterns of party unity changed (see Weldon 2002). Furthermore, the literature for the time from 1998 to 2000 seemed suitable to the research effort here. For an overview of the time periods that constitute the respective cases see table 1 above. 3.1 Model, Measurement and Data 3.1.1 The Phenomenon in Question: Party Unity (PU) The first theoretical challenge I have to deal with is the concept of party unity. By party unity I mean a united voting of the deputies of a PPG. Here, three things are important. First, I do not analyze the unity of a party as a whole, but of the parliamentary part of a party which is not necessarily the same. That is very plausible, because of the second important argument: Usually, researchers restrict the phenomenon of unity to the voting behavior; however, it would be plausible to stretch the concept to other political actions such as media statements. Of course, it is by far the easiest way to deal with the problem if the concept is narrowed to the voting pattern, not least because voting is the most concrete indicator of the unity of a party which can be observed as a unity of a PPG only, not the party as a whole. If one would like to deal with party unity as the unity of the party as a whole, than one must turn away from voting behavior and one needs to analyze other behavioral patterns. For comparative work this is not possible because we lack data and theoretical tools for the measurement of such a broad concept of party unity, especially if the measurement tool should be able to produce comparable data. Third, I do not use the words party cohesion or party discipline because they imply not only the empirical phenomenon (united voting), rather they additionally imply different causal patterns: on the one hand shared beliefs, preferences, opinions and perceptions (party cohesion), on the other hand the anticipation and the actual use of power, (negative) sanctions and rewards (party discipline) (cf. Hazan 2003, Carey 2007, Morgenstern 2004). 4 4 It is very interesting to study concepts of political actors in order to learn about the perception of the political world which the users of such concepts have: For example, in German political culture the word Fraktionszwang ( party force ) exists; the users (which are even citizens with a very well educated background) have the idea that the united voting behavior is forced by party leaders who can easily push their weak-willed deputies into their direction. The nub is that deputies use the term themselves, however, interpreting the term force much laxer than the citizens do. By force they mean the general political circumstances that cause some (self-) discipline. Consequently, if a deputy says that he has to comply with the party line due to the party 5

With (MV)QCA it is not possible to analyze interval or ratio scaled data. Therefore, the measurement of the concept party unity contains, first, three measurement categories that help to order the empirical cases on an ordinal scale: Party Unity = PU = 2, for a high level of party unity 1, for a moderate level or middle level of party unity 0, for a (relatively) low level of party unity Second, I need a rule how the empirical cases can be classified into the three categories. The most relevant information I have used were the Rice-Scores of the cases I analyzed, which indicate the voting behavior on roll call votes. These scores were calculated by the Rice-Index (see Rice 1928) which is commonly computed as follows: RICEij = %Yes %No for PPG i in vote j. 5 It needs to be pointed out, that Rice-Scores are calculated differently by different researchers with regard to the included votes, the treatment of absentees and abstention, or the weighing of votes. This may be adequate in order to pay attention to certain voting rules. 6 However, these differences in the calculation of the Rice-Index induce doubts with regard to the comparability of the corresponding scores. Here, scores of Rice-Indices were used whose measurement seem to be relatively coherent even though they are not exactly congruent. Table 1, that was introduced in the section on case selection above, shows the scores that were used here. However, these Rice-Scores were used to order each case into one of the categories of the PUscale. Therefore, the validity of my three fold PU-scale is higher than the validity of the information that is claimed to be represented in the Rice-Scores, which are surely not as precise as their values seem to suggest. Of course, the PU-scale is rougher than the Rice- Scores: But roughness and validity may work hand in hand if fine grained measurement is not (yet) validly possible. How were the cases ordered into the classification of the PU-scale? I start the calibration (for its logic see Ragin 2008) of the PU-scale by classifying the PPGs with the highest Rice- Scores, i.e. those of the Belgian Chamber of Deputies, into the category for a high level of party unity (PU = 2). The PPGs with the lowest level, i.e. those of US House of Representatives were classified into the category for a low level of party unity (PU = 0). With these cases and their PU-classification two cornerstones were obtained to classify the remaining cases, i.e. to classify these empirical relatives (voting behavior) into a set of force, meaning that he has to discipline himself in order to help the party as a whole, citizens understand that the respective deputy has some weakness of character and that he is not fighting for his convictions. This causes a lot of trouble in the relationship between voters and deputies and is harming to a certain extent the legitimacy of the parliamentary system. Actually, this loss of legitimacy has no empirical foundation; it is simply a misunderstanding that would be avoided if one communicator, here the deputy, would use words that describe reality more aptly. For more empirical descriptions of such patterns in Germany see Patzelt 1999 and Patzelt 2003. 5 The Rice-Index has values between 0 and 100 (or between 0 and 1, respectively). His interpretation is rather easy. For example, a value of 0 means that 50 % of PPG voted for a given motion and the other 50 % voted against it. Consequently, the respective PPG is totally divided. A value of 100 (or 1, respectively) indicates that there are no rebels. 6 For example, if the majority of deputies is necessary to accept a motion (as in the Russian State Duma) then every absentee should be treated as a nay because he apparently does not act according to the line of his PPG. 6

numerical relatives (the PU-scale). To cut a long story short, undoubtedly united are the PPGs of the German Bundestag, Denmark, Great Britain, and the French PPGs. Brazil s PPGs were ordered to the category of low party unity although their Rice-Score are a bit higher than those of the caucuses of the US House. However, Brazil has constantly high rates of party switching indicating the little unity of their PPGs. An exception is the PT, whose Rice-Score is dramatically high. It has been categorized as high party unity. A medium level of party unity can be observed for Argentine and Russian PPGs. Their Rice-Scores are significantly higher than in the US or in Brazil. Mexican PPGs need to be differentiated: While the PAN and PRD are united on a medium level, the PRI is clearly more united (however, note that this is a statement, that is limited to time period investigated here). Czech PPGs are significantly less united than the rest of the parliamentary systems, so they were classified to the medium level unity. For the Italian Camera dei Deputati before 1993 (the period I want to cover with my cases) Rice-Scores are unavailable due to secret voting that was used for all motions. But in order to analyze the respective PPGs, I searched the corresponding literature for some hints on the level of party unity and the consequences of rebellious behavior. This search bases on the assumption that the more rebellious PPGs are the more fatal are the respective consequences and the more should be written about these consequence in the respective literature on that case. The Democrazia Christiana (DC) had some crisis within their PPGs, that led to defeats regarding their motions that should have passed when the PPGs would have been united. In some cases that led to the collapse of a government the DC helped to form. Therefore, it would not be convincing to put the DC in a group with a high level of party unity because defections on motions that are essential for the government lead to dramatic consequences in a parliamentary type of government. The Partito Socialisto Italiano (PSI) seems to be more united but collapsed her own government, as well. The relatively disunited PSI is a cause for the Italian consociativismo, i.e. the search for votes by the government among members of the Communist opposition in order to find a majority that does not stand among others due to PSI rebels. I classified both DC and PSI into the medium category. The Italian PCI is significantly more united and I put it in the high level category. All in all, this is, of course, a very rough measurement that is primarily based on expert judgments. However, I think that such expert judgments are appropriate in this case and help to investigate some interesting PPGs that would otherwise not have been taken into account. 3.1.2 The Impact of the Type of Government My model with the independent variables is based on rationality assumptions that are typically assumed when legislators are analyzed: They act rationally and they are re-election-, policy- and office-seeking individuals. I use this simple rationality assumption together with the characteristics and the shape of institutions in which these legislators act in order to derive the hypotheses on possible impacts of these institutional elements on party unity. According to the common sense among legislative specialists, the type of government should have the most important impact on party unity. For both majority and opposition parties in a parliamentary type of government the incentives to act cohesively are tremendous. Majority parties have to bring their government into office and need to keep it there. Effective government and, therefore, competitiveness can hardly be expected for disunited parties in a parliamentary type of government. Likewise, for minority parties it is a necessary condition to act in unison if they want to roll the majority in case of rebelling members and to present to 7

the voters a cohesive policy alternative to the governmental PPGs. On contrary, in a presidential type of government the incentive of keeping a government in office or to put the government to flight is simply not given due to the two chains of delegation. We should, therefore, expect that in parliamentary systems party unity is significantly higher than in presidential ones. I simply measured whether the government is depending on the confidence of the parliament or not. This is fully adequate for my research context apart from the controversies of regime classification. If this is the case, the respective type of government is called parliamentary and the TYPEGOV-Variable is set 1; if not it is called presidential type of government and the variable is set 0. Table 2 presents the data. [insert table 2 here] For most of my cases the classification according to the vote-of-confidence-criterion was not problematic, at all. Only the Russian State Duma and the French Assemblée Nationale were a bit problematic because both parliaments are placed within hybrid types of government. However, in France the criterion in question is given, and the French government is depending on the Assemblée Nationale (besides the president). The Russian State Duma may use the vote of no confidence and there is a compulsory for a vote of investiture but the president alone has the right to propose the new prime minister and he may dissolute the parliament after his candidate failed to meet the majority threshold two times. Thus, whether the government depends on the confidence of the legislators will depend on their handling of the president s dissolution option. If the legislators find that the dissolution is no political option for the president or that the dissolution is not entailed with electoral losses or any other costs, they are strong vis-à-vis the government. However, this should be by no means the regular expectation. Instead, dissolution is risky what legislators anticipate. Hence, the Russian government is relatively independent from the State Duma s confidence and needs to be classified as a (more or less) presidential type of government. 3.1.3 The Candidate Selection Methods Another promising explanatory variable is the method of candidate selection because the selection of possible rebels should make them more disciplined. I use the differentiation of candidate selection methods into two dimensions: (regional) centrality, i.e. where the selection takes place (local, regional, or national level), and exclusivity, i.e. who selects (a party leader, a small clique of leaders or a nominee committee, delegates, closed primaries, open primaries) (see Hazan/Rahat 2006). From these two dimensions I derived some hypotheses on possible effects on the phenomenon in question (cf. Hazan/Rahat 2001). Party unity should be strengthened when the candidate selection is central and exclusive, e.g. a small clique of national party leaders makes the decisions on the candidates punishing the rebels. When candidates are selected on regional levels this should harm party unity because the members of a PPG may be pushed to act according to the interests of the regional level, and there is no guarantee that the regional interests of the PPG members are congruent with the interests of the PPG or their leadership (that is rather unlikely, actually). This effect should be even stronger if the selection method on the regional level is inclusive (delegates, party members in closed primaries) because this constitutes a need for the deputy to build on a personal profile and should make him more susceptible to rebel. I studied the literature about my cases and tried to make a plausible assessment of the nature of the selection process according to these two dimensions: centrality and exclusivity. If more than one selectorate has been involved in the selection process, I was able to find one selectorate that is plausibly the most powerful. In most cases the authors I relied on noticed informal procedures and did not concentrate on formal statutes alone. However, in some cases data was not sufficient to 8

distinguish the cases on my inclusivity or exclusivity scales. I therefore put some categories together and got a more valid distribution (which is better to handle with QCA, as well). The data organized in this way is shown in table 3. [insert table 3 here] It becomes visible here that these ordinal measurement scales are necessary: For higher scales comparative data on candidate selection methods is totally unavailable. Consequently, only the qualitative measurements presented in table 3 can be made I documented the measurement and all decisions that are connected with them carefully by setting up measurement protocols. They have two functions. First, there I stated why a certain case has a certain measurement value for a given variable and which data sources were used to make the respective decision. Second, these measurement protocols enable the researcher to write down some remarks on the Gestalt of the case, i.e. some information on the shape of the case, that go beyond the pure numerical value of the measurement. In consequence, this information enables the case-oriented view while analyzing the numerical or Boolean solutions that are calculated in the stage of data analysis. So to speak, these protocols constitute sketches of the examined cases, that help to remember how the case is shaped in order to bring this knowledge to the stage of data analysis. 7 Of course, such sketches are necessary if the analyst wants to have an holistic view on his cases and if the dialogue between data and theory is intended. 3.1.4 The Electoral System Another promising condition that may help to explain the levels of party unity is the electoral connection, i.e. the electoral system. The hypothesis is that the more incentives are present to cultivate a personal vote the more challenging this should be for party unity. As with inclusive candidate selection methods deputies need to individualize themselves and build a profile in order to gain the maximum proportion of a personal vote. This should lead to more rebelling behavior in the assembly. To test such an effect I used the typology Paul Mitchell developed in order to separate electoral systems according to their incentives to cultivate a personal vote (Mitchell 2000). So I distinguished the examined PPGs into such that operate under conditions of three separate kinds of electoral systems: candidate-centered, partycentered, and intermediate systems. 8 Table 4 contains the data of my measurement with Mitchell s typology. [insert table 4 here] As can be seen from the data, the observed variance is quite good. That will help to gain insights from the effect of this condition. Again, I wrote down all relevant information on the sources and reasons of my measurement decisions in the measurement protocols. 7 However, I cannot present the measurement protocols here, because they would have summed up to approximately 100 to 150 printed pages. The reader may get an impression of these protocols by visiting my web site on which I present all these protocols (in German). See http://www.erik-fritzsche.de/research.htm. 8 I also used the more detailed typology developed by Carey and Shugart (Carey/Shugart 1995) to measure the possible effects of the electoral system on cultivating personal votes. I then cross tabulated the obtained data with the measurement of Mitchell s typology. Both classifications seem valid, but Mitchell s typology is more parsimonious (but still valid) which is better with regard to the required parsimoniousness of the QCA data input. 9

3.2 The Hermeneutic Adventure Takes Place Data Analysis with QCA 3.2.1 Bivariate Patterns The first analytical step is to look for bivariate relationships. Looking at the various crosstabs, the remarkable statement has to be that there cannot be found an obtrusive finding. However, one finding is unsurprisingly striking: It is the relationship between party unity and the type of government as the reader can tell from the following table. [insert table 5 here] The parliamentary type of government plausibly correlates with higher levels of party unity. However, it is neither a sufficient nor a necessary condition. That would not even be the case if the measurement of party unity would be done by simply using two-levels (high unity, not high unity). On the one hand, the examined PPGs from the Czech Chamber of Deputies and the Italian DC and PSI act in a parliamentary type of government but they do possess a medium level of party unity. On the other hand, variance within the presidential type of government is quite extraordinary and especially the Brazilian PT and the Mexican PRI are two PPGs in my data set with a high level of party unity that one can hardly explain with the type of government. The first finding, therefore, is that we cannot simply make the type of government dichotomy (parliamentary vs. presidential) expecting a perfect relationship with the levels of party unity (however, many textbooks do so; for this finding cf. Kailitz 2008). 9 It becomes clear that variance within one and the same legislature is noticeable, especially in presidential systems. For example, Weldon 2002 shows that in the Mexican Cámara PRI is significantly more united than PAN and PRD (as the reader can tell from the table 1 above, as well). Consider the Brazilian PT and compare it with the other PPGs in the Brazilian Câmara: While the PT is clearly similarly united as many parties in most of the Western European parliaments, the other PPGS within the Câmara are only slightly more united than the parties in the US Congress. Bivariate coefficient scores put it in a nutshell: Tau-c is.748 and gamma is.943, showing the relevance of the parliamentary type of government for explanation. 3.2.2 QCA, step I: Few Variables This pattern asks for some multivariate analysis hoping for some additional explaining value of other conditions. As with statistical analysis the researcher may construct different causal models for the QCA analysis and as with statistical analysis, too technically this is done with the selection of variables out of the data set. I do the same, here, and start with the type of government and the variable which belongs to the electoral connection : the electoral system. 3.2.2.1 Bivariate Patterns Before analyzing these two variables with QCA, I present the cross tabulation of party centeredness of the electoral system and party unity layered by i.e. controlled for the type of government. [insert table 6 here] As we can note from the cross tabulation above, party centered electoral systems in presidential systems seem to show a certain trend towards more united voting patterns. The Brazilian, Russian, Argentine and Mexican PPGs confirm the personal vote argument. In Brazil the electoral system is candidate centered and correspondingly party unity is low; in 9 Recently, large N research has underlined the finding presented here: The type of government is a good predictor but it is neither a necessary nor a sufficient condition for the levels of party unity (see Kailitz 2008). 10

Russia, Argentina and Mexico the party centered electoral system may account at least in part to the middle level of party unity in presidential systems; the Mexican PRI is even a highly disciplined PPG under the examined constellation. The Brazilian PT is an extraordinary outlier and the relatively poorly united caucuses of the US House operate under conditions of an intermediate system. The argument can be underlined by a tau-c-score of.57 and a gamma-score of.75. 10 What about the parliamentary systems? The Italian DC and PSI show that their relatively low level of party unity corresponds to candidate centered electoral systems. However, the Danish PPGs and the Italian PCI disconfirm the personal vote argument regarding party unity. They show a high level of party unity and have a candidate centered electoral system. Possibly, the absence of intermediate or party centered systems induces some degree of freedom to the variance of the phenomenon in question and the causes for high or medium level of party unity depend on other factors. The French and British PPGs as well as the PPGs of Belgium and Germany do not seem to speak against the personal vote hypothesis, or they confirm it, respectively. The whole relationship can be expressed parsimoniously with a Tau-c score of.48. 11 3.2.2.2 Inspection of the Truth table and QCA-Solutions Now, let me show the observed configurations of the two analyzed conditions: type of government and party centeredness of the electoral system. However, due to technical reasons I need to dichotomize my independent variable party unity, i.e. I need to make a decision which outcome I want to explain. Here, I start with the high level of party unity ( 1, else: 0 ). The following configurations have been observed as presented in table 7: [insert table 7 here] In this truth table, there are configurations of conditions (lines in the truth table) whose corresponding cases have different output values (different levels of party unity), so called contradictories (indicated by a C in the PU-column). Consider the following three contradicting expressions: There are PPGs in presidential systems with a candidate-centered electoral system but different levels of party unity (the Brazilian PPGs vs. Brazilian PT). There are PPGs in presidential systems with party-centered electoral systems but different levels of party unity (Russian, Argentine PPGs, Mexican PAN and PRD vs. Mexican PRI). There are PPGs in parliamentary systems with a candidate-centered electoral system but different levels of party unity (Danish PPGs, Italian PCI vs. Czech PPGs and Italian DC and PSI). The researcher has two ways to tackle the problem of contradictory lines when he uses the minimization algorithm (for other strategies before minimization takes places see below pp. 13f). Both way have some interesting implications with regard to the consistency and 10 For calculating the scores I used a special dataset with each line representing a PPG and not, as outlined above on page 4, a dataset where PPGs with the same combination of investigated characteristics were grouped together (cf. table 1 where the case column shows the grouped together cases and the PPG column shows the investigated PPGs as they are represented with a line in the dataset for the correlations). By doing this, I avoid a bias toward those lines that imply few cases. 11 The gamma-score is 1.0. However, this value is induced technically because of the mathematical sensitivity of gamma towards L-distributions in cross tables. 11

coverage of the theory the researcher may obtain by the QCA minimization. First, one may use the conservative approach : That is, the researcher excludes the corresponding line of the truth table that implies contradictory cases. The theory the researcher gains after minimization is fully consistent, however, it does not include all the cases he actually intended to explain, i.e. the coverage of cases is not the best and, therefore, the range of the obtained theory diminished. The second way is a liberal approach : One may accept the inconsistency and use the contradictory combination of conditions for the minimization as well as for the theory development. Consequently, the resulting theory is not fully consistent but its coverage is more extensive, i.e. its range has increased. 12 Note that with statistical analysis the researcher is forced to accept that the obtained theories, i.e. the statistical coefficients, apply to the analysis of the whole data and that they are always imperfect, i.e. inconsistent, if the estimated relationship is not perfect or the explained variance is far from 100 per cent. However, in contrast to statistical analysis, the here done QCA is essentially case-oriented and uncovers to the researcher which cases exactly cause the inconsistency. Therefore, it is much easier to make theoretical progress by examining the deviant cases in detail. That the researcher can clearly recognize where the inconsistency in his theory remains and which cases are responsible is a nice advantage of the liberal approach. The following result is obtainedbyignoring the contradictory lines, i.e. being conservative. PZENT{2} * REGSYS{1} + PZENT{1} * REGSYS{1} PU(high). (DB, BAK) (UKC, FANRPR, FANUDF, FANPS, FANPCF) Of course, no minimization occurred. However, let me interpret this theoretical statement, thereby showing how the coverage of the theory was limited by excluding contradictory lines. These terms of the Boolean solution (although no minimization of conditions occurred) show that either party centered or intermediate electoral systems in conjunction with parliamentary government cause high levels of party unity. However, this theory does not include a statement with regard to the causal processes leading to high levels of party unity in the Danish PPGs and the Italian PCI, because they helped to constitute a contradictory line that has been ignored taking the conservative approach of minimizing. Therefore, the statement of the expression above simply ignores their empirical occurrence, and consequently its coverage is limited. The same is true for the Mexican PRI and the Brazilian PT whose outcomes (high level of party unity) lead to two out of the three above depicted contradictory configurations. Their terms are not parts of the Boolean solution again, showing that they have been ignored, thereby limiting the coverage of the theory again. Let me now show what happens when the parliamentary contradictory is treated different. Instead of ignoring the Danish PPGs and the Italian PCI I will explicitly use their information to make causal inferences now. In consequence, I conclude: There are three configurations with parliamentary system and a high level of party unity but each with a different electoral 12 This shows the dilemma of comparative research. If and only if contradictories do not exist then both consistent theories and theories with a huge coverage may be developed. The more contradictories the more problematic everything will be. However, it is hard to avoid contradictions. The likeliness of getting contradictories increases when the number of cases increases and/or the number of variables diminishes (and/or the categories per variable, respectively). So the dilemma arises: If comparativists use small N in order to avoid contradictions they will not be able to make causal inferences any more. If comparativists use more variables (and/or finer grained concepts) it will be equally problematic to make causal analysis. Therefore, both ways of trying to avoid contradictions lead to the n/v-problem. If comparativists handle the n/v-problem this may lead back to inconsistency or limited theories. Therefore, one has to try to go the middle route of optimizing n/v-ratio and causal conclusions at the expense of inconsistency and the shortcomings of limited theories. 12

system so the electoral system can be considered irrelevant with respect to the outcome high level of party unity. The Boolean expression above would simplify to: PU(high) REGSYS (1). Of course, one should keep in mind that the Czech PPGs and the Italian DC and PSI differ from that pattern. This theory explicitly covers the Danish PPGs and the Italian PCI; however, it is not as consistent as the theory above. Some cases in my data set are not in line with this theoretical statement: the Czech PPGs and the Italian DC and PSI those cases which, of course, lead to the contradictory configuration (together with the Danish PPGs and the Italian PCI). 3.2.2.3 Results of step I The bivariate patterns did not allow to state empirical arguments on perfect relationships between the party centeredness of the electoral system and party unity controlled for either type of government. This is, however, usual business in the social sciences. As the analytical arguments outlined here show QCA does not as sometimes claimed seek to find perfect explanations. Contradictory cases force the analyst to formulate imperfect theories having the choice: either by limiting its coverage or its consistency. Therefore, it can be summed up that the bivariate coefficient perspective and the QCA perspective on my data set come to the same conclusion: In tendency, the parliamentary government goes hand in hand with higher levels of party unity, and the more party centered electoral systems are the more likely it seems that higher levels of party unity can be observed. 3.2.2.4 Implications for the Research Design What can be learned by the researcher from the analysis for his further research? Contradictories in QCA, and correlations that are far from being perfect alike, let the researcher draw the conclusion that the causal model consisting of the type of government and the type of electoral system is insufficient to separate the cases into two groups: those who have high levels of party unity and those who do not have high unity (i.e. it prevents the researcher observing perfect relationships). There is a general epistemological problem at work here: The more parsimonious the modeling is the more it is endangered by being inconsistent, i.e. leading to contradictory expressions. Obviously, QCA shows in a simple way what is quite normal for every research effort, not least for statistical analysis. With regression analysis coefficient of determination R², or PRE-measures alike, the researcher gets an indicator of the quality of his theory vis-à-vis data. The smaller R² is the worse the model fits to the analyzed data which can easily be interpreted as variance of the dependent variable explained by the independent variables. Bivariate coefficients smaller than one refer to the same fact. For case-oriented researchers the quality of their theoretical argument can be assessed by the number of cases which do not fit the propositions. Each contradictory in the QCA sense is necessarily induced by such a case, because: If the proposition is that a certain configuration of conditions leads to a certain outcome, say 1, then the theoretical argument is lacking the more plausibility the more cases exist that have the configuration in question showing a different outcome, say 0. Exactly this has happened above when the model consisted of the type of government and the electoral system, only. Therefore, the model is not suitable for finding sufficient explanation; rather, it seems to be too simple for the complex patterns of the social world investigated in this research effort. As a result of this the researcher has some possible strategies to cope with the situation before going to minimization in order to solute contradictories all of these solutions imply further research efforts. 13

- Adding variables/conditions: Either the researcher has ideas about new variables to collect data on or he needs to analyze more variables he already has in his data set. To put it in statistical terms: More variables need to be integrated into the model in order to find a possible explanation of more variance. - Adjusting variables/conditions: This would ask for a refinement of the categories in which the researcher thinks about the phenomenon he examines. It is typical for qualitative research that the researcher does not consider the definitions and measurement as given. Instead, knowledge is hardly adequate to measure fine grained or at least open for revision within one and the same research effort. This may contain the pure re-organization of data by adding properties that are to record with this variable. (It is also possible to adjust the simple model by redefining such properties or by thereby making it even more simple and losing information merging two or more properties.) It may also mean the total revision of concepts and their theoretical claims, or it may mean the substitution of new concepts and measurement scales for old ones. 13 It is important to note that all these strategies of clearing the analysis by adjusting the categories while the researcher has to think about his empirical material is a dialogue between theory and data and never theory alone. - Data correction: Another possible route to improve the analysis is to go back to data in order to improve measurement. This should be done by the researcher if he has the suspicion that some of the very problematic cases observations (i.e. observations regarding those cases that form contradictory configurations or that prevent bivariate coefficient scores close to one) are not correct. The correction of these observations may lead to new patterns; it may also lead to the questioning of those categories the researcher uses to think about his phenomenon (see the point above). - Case selection and the construction of populations: It may also be helpful to think about the cases that have been selected. One of the possible options is to think about the exclusion of cases that lead to contradictory lines. However, the researcher needs good arguments for that and should not simply exclude cases in order to erase contradictories in the data set. Another possible solution could be to think about the split of the cases into two different populations. This would be equal to an attempt to find distinct theories for each of the two populations that have different characteristics, yet trying to explain the same phenomenon (cf. Ragin 2004 and point five of my conclusion on page 24). All these strategies underline how hermeneutic a qualitatively comparative research effort really is or might be. I go on in the next section with adding some promising variables in order to make my analysis really multivariate. There, I will obtain some insightful conclusions regarding the adjustment of my research design. 3.2.3 QCA, step II: Adding Candidate Selection Methods 3.2.3.1 Bivariate Patterns: Candidate Selection Methods and Party Unity In this step, I integrate the candidate selection methods into my model. The model consisting of the type of government, the electoral system, and the candidate selection methods should be the most promising in explaining the vast part of the variance of party unity. Before doing 13 I have already done exactly this when I tried to capture the incentives of electoral systems to cultivate personal votes. I first started with the concepts and scales proposed by Carey/Shugart 1995; later on, I used Mitchell 2000 whose classification was rather more parsimonious. However, it is possible to show that the observations of the electoral systems I analyzed in this research do correspond to each other. I cannot do this here due to limited space. 14