CHAPTER 4: METHODOLOGY

4.1 Three research phases and three methodologies

The aim of this research is to contribute to the theory and practice of electoral integrity. For contributing towards this, this dissertation seeks to address two main themes: i) when, where and under what conditions are election results accepted? And, derived from this: what can we do to increase credibility in elections? To answer these questions, this research will involve three different research phases. The first phase will develop a multivalue Qualitative Comparative Analysis (mvQCA). Two QCA Models conduct a systematic analysis of key factors (democratic consolidation, political party support for electoral institutions, quality of elections and transparent election results) and allow us to

identify which configurations, or specific combinations of factors, lead to the acceptance of election results. More specifically, it will reveal which conditions are necessary and/or sufficient for accepting election results. Phase two consists of a regression analysis (binary logistic and multilevel regression) analysing the effect of political party support for Electoral Management Bodies –measured through their participation in the appointment of the EMB members - on election credibility. Phase three is a small-N paired focused comparison which addresses the role played by transparency in the election results stage in mitigating post-election protests. The three methods will strengthen and complement one another. At the end, the objective behind this combination of methods is to enhance our insight on the acceptance of election results, on election credibility and on some of the key factors behind this.

4.2 Inspired by mixed methodology

Mixed methods research is a third methodological movement in the social and behavioural sciences that has enabled us to bridge the differences between the two main research paradigms (the positivist paradigm, underlying quantitative methods, and the constructivist paradigm, which underpins qualitative research), (Tashakkori and Teddlie, 2003). The debate between these two approaches, historically focused on philosophical issues such as the nature of reality and establishing causality, has now given way to a more pragmatic approach that combines whatever methods are most appropriate for a study (Tashakkori and Teddlie, 1998). Nowadays, instead of using one single methodology, research can rely on a combination of methods to enhance our knowledge of a particular research problem. The present study is inspired by mixed methodology. Although the methods used do not study exactly the same data, they do complement each other. The idea behind this is to improve our understanding of why elections results are accepted and what makes elections credible by combining different methods. When findings from different methodological

approaches are broadly consistent they enhance the robustness of research claims. The dissertation opens with a Qualitative Comparative Analysis (QCA) which in itself is a quali- quantitative method that provides the initial results. This first analysis is more general and covers five key conditions for the acceptance of results. Then the research integrates quantitative and qualitative methods to improve and strengthen our knowledge of the findings obtained by the QCA. The quantitative study focuses on one key condition across a number of Latin American cases and the qualitative research elaborates a paired comparison of two cases within a single country, Mexico. In short, the study first generates macro level information which is then refined by more specific studies using diverse techniques.

As a result, this research design allows for a degree of triangulation, which is one of the key advantages of mixed methods research. As outlined by Tashakkori and Teddlie (1998) triangulation involves using different methods and data to study the same phenomena within the same study or in complementary studies. The term comes from navigation and implies using various references to pin point more exactly the position of an object (Smith, 1975). Following this metaphor, triangulation can lead to having more robust, complete and valid results. First, combining methods allows complementing their strengths while overcoming their limitations. Second, multiple and independent measures, when their results are congruent, can give us a more robust and complete understanding of the phenomenon (Jick, 1979). Finally, “triangulation in action” (Jick, 1979) means cross checking and getting external validity for our findings.

The study design only allows for a certain degree of triangulation as its component phases do not measure the exact same phenomenon. While the QCA section focuses on the acceptance of election results, the quantitative chapter addresses credibility in elections and the small-N analysis focuses on post-election protests. However, the three methods do

measure related phenomena. As a result, the findings of one chapter become more suggestive when seen in the light of the evidence produced by the other chapters. This combination of methods therefore can produce greater confidence in the results. This is especially true as both the quantitative and small-N chapters support the findings of the more general QCA chapter.

4.3 Three research phases

4.3.1 QCA: acceptance of election results

The first substantive chapter (chapter 5) asks when, where and under which conditions election results are accepted. In particular, it first develops one mvQCA model focusing on 48 presidential elections held in Latin America between 2000 and 2015 and looks at the five following conditions: consolidated democracy, the closeness of an election, the support offered by political parties to electoral institutions, quality in the development of an electoral process, and transparency in election results. Then, in order to provide evidence of how the findings can travel and be generalised to contexts beyond Latin America, it develops a second mvQCA model, which considers 21 close elections held around the world for the same time period of 2000 to 2015.

QCA is a research method that occupies an intermediary position between small-N and large-N analysis (Halperin and Heath, 2012). On the one hand while variable-oriented or quantitative approaches are good for working with a large number of cases and producing generalisations, they do not pay special attention to individual cases or to more concrete questions relevant to the causes of specific phenomena (Ragin, 1987). On the other hand, while case-oriented or qualitative approaches do pay attention to the complexity of specific cases, they are not well suited for generalising findings (Ragin, 1987). QCA combines the strengths of these two approaches. From quantitative research, it allows to analyse a

greater number of cases and to produce generalizable results. QCA requires that each case be reduced to a series of variables which gives it an analytical character (Rihoux and Lobe, 2009: 224). Moreover, as in qualitative research, in QCA each individual case is significant and considered as a complex combination of properties which inform the hypotheses and theories (Berg-Schlosser, et al., 2009: 6). In sum, QCA comes out as a synthetic strategy (Ragin, 1987: 84) that borrows key strengths from both Large and Small- N research methods allowing for a thick, analytical and replicable comparative technique.

By formalising qualitative comparative methods, QCA conducts a systematic analysis of the cases. This means that we can compare across-cases and address their similarities and differences, even if the cases are only more than a handful. This allows to identify patterns and make more general statements about the cases and the relations between them. In order to do this, QCA breaks up cases into independent and dependent variables, much like in quantitative approaches. In particular, QCA transforms its cases into configurations, which represent specific combinations of factors, causal variables or conditions that produce an outcome of interest (Rihoux and Ragin, 2009). As such, it shows how different configurations of conditions can lead to the same outcome. This is refered to as multiple conjunctural causation, where ‘multiple’ refers to the number of possible paths and ‘conjunctural’ expresses that each path is made up of a combination of conditions (Berg- Schlosser, et al, 2009).

In order to do this, this is a set theoretical method based on the language of Boolean algebra, which gives it a more scientific character (Ragin, 1987:x; Rihoux and Ragin, 2009). This language is used to describe logical relations between the cases and analyse formally which combination of factors can lead to the outcome of interest. Cases are evaluated in terms of their membership to sets (supersets and subsets) (Legewie, 2013). Then, Boolean minimisation reduces the complex system of relations between variables and the outcome

to a simpler formula by eliminating redundant terms (Thiem, 2014). This allows identifying which of the conditions are necessary or sufficient for the production of the outcome, and which are superfluous. A necessary condition is always present when the outcome is present, while a sufficient condition alone does not produce the outcome (Berg-Schlosser, et al, 2009).

Conducting a QCA requires a good research design. The research needs to rigorously select the cases under study, followed by a clear identification, definition and measurement of the independent variables and the outcome that needs to be explained. First of all, as will be detailed on the chapter, for the selection of cases the QCA chapter focuses on Latin America. This is as countries in this region share a number of historical, political and even electoral characteristics that facilitate comparison. In particular, these countries have a shared colonial legacy and nowadays have presidential political systems with elections organised by independent and permanent electoral institutions. Second, the research needs to clearly identify the conditions that lead to the outcome. This has to be driven by theory and for the QCA analysis, previous research has highlighted the key role played by five variables in obtaining an accepted election (these are the type of regime, the closeness of the election, the support of political parties for electoral institutions, the quality of the elections and the transparency of results). Third, these conditions have to be clearly conceptualized and measured so that the analysis and its results are reliable.

In relation to this last point, it is important to highlight that conceptualization has been identified as one of the three key challenges in the construction of datasets for the study of democracy (Munck, 2009). Conceptualization refers to the specification of the meaning of the concept and its constituent attributes. Besides properly identifying the concept to be measured, key tasks include avoiding including too many attributes/ attributes that correspond to other concepts or excluding items that are actually part of the meaning of the

concept. For instance, several democracy indices exclude one of the key attributes of democracy as highlighted by Dahl: “participation” (Munck, 2009). Once researchers have identified which attributes will be included as components of the concept to be measured, a second challenge is isolating the ‘leaves’ on the concept by level of abstraction (the concept’s concrete attributes). This is important for having proper measurements of the attributes and for avoiding problems of conflation and redundancy. Conflation occurs when components of attributes are not placed or categorised properly under the attribute they belong to. Then, attribute components at the same level should describe mutually exclusive aspects of the concept, otherwise the logical problem of redundancy arises (Munck, 2009:22)

The conceptualisation challenge and its specific tasks of clearly identifying attributes are aimed at the construction of indices measuring democracy. In particular, these are intended as a framework for having a more solid assessment of a concept (democracy) and its subcomponents (eg. Freedom of association, freedom of the press, right to vote, fair elections, etc). Present research does not involve the definition and measurement of a concept or its constituent components. Rather, as explained above, it is about identifying which conditions can play a role in the acceptance (or rejection) of election results. Nonetheless, this does not mean that the concerns outlined above –and the tasks to overcome them- should not be taken into account when building a solid data set. Therefore, it is important that the conditions/independent variables used in our study are properly identified and isolated. In this research all five conditions are clearly defined –either as contextual variables or as distinct moments and aspects of the electoral cycle- and conceptually and empirically different from each other. Chapter five provides further details of the variables under study, their definitions and measurement.

Once the variables have been properly identified and conceptualised, data must be gathered for their measurement. For this research, data was collected on the acceptance of election results in each country and on the key independent variables under study. This qualitative data then needs to be transformed into categories and numbers for the analysis (Rihoux and Ragin, 2009). Data on cases, variables and the outcome are included in a table. Cases become rows and columns represent the conditions that help explain the outcome and the outcome itself. This raw data table is then synthetized into a more analytical truth table or “table of configurations” (Rihoux and De Meur, 2009). The truth table groups cases into subsets that share the same conditions and outcome. This “minimization” helps us identify which configurations or paths lead to the desired outcome. This is done through specialist QCA software employing Boolean minimization algorithms (Rihoux and De Meur, 2009).

Boolean minimization aims to produce a parsimonious expression that reveals which causal conditions produce the outcome. First, all configurations with a positive outcome are minimized by the software. Then, all configurations linked to a negative outcome are minimised. The results are formulas that give a path or a combination of conditions (which groups several cases) associated with the presence or absence of the outcome. However these formulae are still quite complex. They are labelled “descriptive” formulas as they do not go much beyond the observed cases (Rihoux and De Meur, 2009). For this reason, logical remainders are included in the minimisation process. These are logically potential cases which have not been observed empirically. By assuming that these remainders exist (simplifying assumption) the software is able to produce a more parsimonious formula. This formula tells us which combination of conditions lead to the desired outcome, in this case the acceptance of election results.

There are four different configurational comparative methods. These are crisp-set Qualitative Comparative Analysis (csQCA), fuzzy-set Qualitative Comparative Analysis (fsQCA), multi-value Qualitative Comparative Analysis (mvQCA) and generalized-set Qualitative Comparative Analysis (gsQCA) (Thiem, 2014). ‘Crisp set’ is the first type of QCA to be developed and it is the most widely used. It was developed in the 1980s by Charles Ragin who adapted Boolean algorithms to his own research for simplifying complex data structures and identifying patterns of multiple causation (Rihoux, De Meur, 2009:33). This technique uses conventional Boolean sets which require that all the variables have only two possible values to identify the presence or absence of a given condition. Here, a set is comparable to a binary variable where cases are either non- members or full members of a group, true or false, and therefore coded as ‘1’ or ‘0’ (Halperin and Heath, 2012; Rihoux, De Meur, 2009).

Multi-value QCA (mvQCA) originates from csQCA and shares its basic principles and aims to explain an outcome by simplifying complex data. This method differs from csQCA as it moves beyond the use of dichotomies to have a more flexible structure for classifying or coding cases. Therefore, instead of taking two possible values (“0” or “1), variables can take multiple values (“0”, “1”, “2”, “3”, etc.). These multiple values can be useful, for example, for a more detailed coding of the data or to incorporate scale categories into the research.

Qualitative Comparative Analysis can also be conducted by using fuzzy sets. Fuzzy set QCA (fsQCA) also seeks to unravel causality through set theoretic relationships but does so in a different way. FsQCA is based on a “diversity-oriented approach” where “diversity exists not only in the different configurations of set memberships that social phenomena exhibit but also in the degree to which they belong to such sets and configurations” (Ragin, 2000: 149). This means that while crisp set and multi-value QCA allow cases to take one of

two values, or one out of three or four categories (dichotomies and multichotomies, respectively), in fsQCA cases can take an unlimited number of values and have partial membership in the different sets. This results in a possible number of infinite configurations, which, in contrast to other QCA variants, does not allow constructing truth tables (Tiem, 2014). Membership scores are part of a scale that ranges from 0.0 to 1.0 (allowing to have values such as 0.25, 0.5, 0.9, etc.) and are useful for more fine grained data, such as the literacy rate or the GDP per capita in a country.

Finally, there is a fourth type of QCA, denominated generalized-set Qualitative Comparative Analysis (gsQCA). Thiem (2014) argues that crisp, multi-value and fuzzy set QCA are in fact not so different and can be included as part of a broader type of QCA, gsQCA. For this, he advances the concept of multivalent fuzzy set variable, a variable which includes features of all existing QCA variable types. As a result, this type of variable helps solving the problem of inaccurately classifying cases that, for example, have partial membership in different sets. In addition, this more general type of variable allows the construction of truth tables that was not possible under standard fuzzy-sets. However, although GsQCA includes all three previously existing QCA methods, it must be noted that it can also be identified as a variant on its own, with its own technique and notation.

4.3.1.2 QCA and the acceptance of election results

QCA is selected as an appropriate methodology for understanding when, where and under what conditions are election results accepted. First, this technique is appropriate for methodological reasons. QCA is especially well suited for small and intermediate-N designs (Ragin, 1987; Rihoux and Ragin, 2009) with medium sized samples. The cases selected for this research are 48 elections in Latin America and then, for the second QCA model, 21 elections around the world. This number is limited for using quantitative research methods,

which require a number of cases sufficiently large to enable robust inferences. At the same time, the number of cases in this research is quite large for conducting small-N analyses, which involve the comparison of few cases, typically two or four and not exceeding more than a dozen or so (Halperin and Heath, 2012). Working with 48 or 21 cases would imply doing many paired comparisons which complicate making a systematic analysis.

Small-N and large-N analysis offer some advantages but also present some shortcomings for answering our research question. On the one hand, small N studies allow for an in depth analysis of case studies and allow to understand the context of the cases selected, leaving out neither the particular nor the general (Halperin and Heath, 2012). Moreover, they are useful as a source of theory and for linking theories to evidence (Keman, 2011). However, this method has some weaknesses that limit its usefulness for our research. Most importantly, with small N studies the problem of selection bias is always a risk. As is well known in the political science discipline, the cases we choose affect the answers we get (Geddes, 1990) and this is more likely to occur when we analyse just a few cases. When cases are not selected carefully, results can be biased and misleading. In addition, this risk is always present and it is more difficult to arrive to more generalizable results from only a small number of cases. This is connected to the issue of weak external validity. While small N gives us a good understanding of the mechanisms at work in the cases under study (internal validity) we also lose the power to extend our findings to other cases not included