1. Defining and measuring high improvement in the quality of electoral democracy

Many analysts agree that democracies exhibit four central characteristics: free and fair elections, universal suffrage, respect for civil liberties and civil rights, and the rule of law (e.g., Mahoney, Reference Mahoney2021). This article focuses on one subcomponent of democracy—electoral democracy—which is essential to any definition of this system of government (Dahl, Reference Dahl2003; Schumpeter, Reference Schumpeter2013). I highlight three central features of electoral democracy—inclusive suffrage, fair districting, and the effective and transparent administration of elections (Grumbach, Reference Grumbach2023). I argue that a state has a stronger electoral democracy to the degree that the state’s policies improve access to the ballot (Pomante et al., Reference Pomante, Schraufnagel and Li2023), increase the fair translation of votes into representation in electoral offices (Warshaw et al., Reference Warshaw, McGhee and Migurski2022), and follow best practices relating to professionalism, efficiency, and transparency in election administration (e.g., MIT Elections and Data Science Lab, 2024).

The outcome of interest in this study is high improvement in the quality of electoral democracy in the American states between 2000 and 2022. Of critical importance, in this study I am not focused on understanding marginal, year-over-year change in democracy from its mean value, as one would when using frequentist statistics; instead, I seek to understand why specific states experienced a fundamental, qualitative shift in the strength of their democracies during this period. This approach to the outcome of interest aligns with the philosophical commitment to the study of necessary and sufficient causation in qualitative states embraced in this study (Goertz and Mahoney Reference Goertz and Mahoney2012; Mahoney, Reference Mahoney2021) and the method of CNA used to analyze the data (Baumgartner and Ambühl, Reference Baumgartner and Ambühl2020, Reference Baumgartner and Ambühl2023). Thus, the outcome measure must establish a qualitative distinction between states that achieved this high standard during this period and those that did not.

Measuring this outcome proceeded in several steps. To assess the overall quality of electoral democracy in the states, I created an index, the Electoral Democracy Score (EDS), based on various indicators of state-level policies. My approach, which sought to gather a large and diverse amount of information about the quality of electoral democracy, closely follows Grumbrach’s (Grumbach, Reference Grumbach2023) exemplary study.Footnote ³ The EDS is based on 53 indicators of ballot access (e.g., felon disenfranchisement laws), partisan gerrymandering (e.g., the partisan efficiency gap), and election administration (e.g., voting wait times), measured in each state in the years 2000, 2004, 2008, 2012, 2016, 2020, and 2022. The ballot access measures were obtained from the Cost of Voting Index (Pomante et al., Reference Pomante, Schraufnagel and Li2023); the partisan gerrymandering measures came from Warshaw (Reference Warshaw2024); and the election administration measures and measures of election outcomes came from the MIT Elections and Data Science Lab (2024). A measure of ineligible individuals convicted of felonies as a percentage of the voting-age population comes from McDonald (Reference McDonald2024). The indicators and their sources are presented in Supplemental Material 1.

Although the 53 indicators gathered for the EDS provide a comprehensive assessment of the quality of state-level electoral democracy, they do not measure every dimension of Americans’ experience with elections. With a few exceptions (e.g., average wait time at the polls), these indicators do not gauge citizens’ experiences with local election administration within their states, which can be consequential for their ability to register and cast a ballot (e.g., White et al., Reference White, Nathan and Faller2015; but see Ferrer et al., Reference Ferrer, Geyn and Thompson2024). However, my focus on state-level policies is consistent with recent studies (e.g., Grumbach, Reference Grumbach2023; Pomante et al., Reference Pomante, Schraufnagel and Li2023) that use measures of state-level institutions and outcomes to characterize the quality of electoral democracy. Future research could advance the study of democracy at the subnational level by incorporating additional indicators of the quality of local democracy.

Using the 53 indicators gathered for this study, I modeled the quality of state electoral democracy for each state-year as a latent variable (Treier and Jackman, Reference Treier and Jackman2008). This approach estimates an “ideal point” on a latent dimension of state electoral democracy for each state-year that best predicts the observed values of the indicators in the dataset.Footnote ⁴ Following Grumbach (Reference Grumbach2023), I used Bayesian factor analysis for mixed data (Quinn, Reference Quinn2004) because this approach can handle the combination of binary, interval, and continuous data in my dataset.Footnote ⁵

The Bayesian factor analysis estimates the quality of electoral democracy for each state-year. Positive values of the index indicate higher quality electoral democracy, while negative values suggest lower quality. I present the discrimination parameter for each policy in the EDS—the slope of the relationship between an indicator and a state’s latent EDS—in Supplemental Material 2. A positive discrimination parameter means that the presence (or higher level) of an indicator increases a state’s latent EDS, and a negative discrimination parameter suggests that the presence (or higher level) of the indicator decreases the score.

Supplemental Material 2 shows that, while a small number of indicators do not load well on the latent electoral democracy dimension (and thus have discrimination parameters close to zero) and two (% Absentee Ballots Rejected and % Military Ballots Not Returned) load in the opposite direction than expected, the discrimination parameters for the overwhelming majority of indicators conform to theoretical expectations. This provides initial confidence that the Bayesian factor analysis is working as expected and producing a valid measure of the quality of state electoral democracy.Footnote ⁶

To further increase confidence that the EDS provides a valid measure of the quality of electoral democracy in the states, I conducted a series of validation tests in which I compare states’ EDS scores to their scores on other well-known measures of democratic quality (the State Democracy Index [Grumbach, Reference Grumbach2023] and the Cost of Voting Index [Pomante et al., Reference Pomante, Schraufnagel and Li2023], respectively) and examine the relationship between state EDS scores and voter turnout. These validation tests, discussed in detail in Supplemental Materials 3–5, respectively, provide strong empirical evidence of the validity of the EDS as a measure of the quality of state electoral democracy. In evaluating the EDS, a final factor to consider is whether and how state EDS scores early in the period under examination affect the prospects for observing dramatic improvement in the quality of democracy in some states. In Supplemental Material 6, I show that our ability to observe significant improvement in the quality of electoral democracy is not compromised by “ceiling effects” at the beginning of the panel.

I used the EDS to construct my final measure, High Improvement in the Quality of Electoral Democracy (HDEM), which establishes a fundamental qualitative distinction between states that experienced dramatic gains in democratic quality and those that did not. I first subtracted each state’s 2000 EDS value from its 2022 EDS value. This provided a measure of change in the quality of electoral democracy between 2000 and 2022, with higher scores indicating more improvement. As Figure 1 shows, states had very different experiences with change in electoral democracy between 2000 and 2022, with some experiencing declines in quality, and others experiencing improvement. In fact, according to the EDS, a majority of states enjoyed improvement in the quality of their democracies during this period. Among states experiencing improvement, some had small gains while others had much larger positive changes.

Figure 1.

Change in electoral democracy score by state, 2000–2022.

The finding that a majority of states experienced an improvement in the quality of democracy between 2000 and 2022 might surprise scholars attuned to serious declines in democracy in some states (Bentele and O’Brien, Reference Bentele and O’Brien2013; Anderson, Reference Anderson2018; Grumbach, Reference Grumbach2023). To ensure that these findings were not idiosyncratic to the measure used in this study, in Supplemental Material 7, I examined the correlation between change in the quality of democracy as measured by the EDS and change measured by Grumbach’s SDI. The two measures of change are highly positively correlated (0.76), which suggests that my findings are not idiosyncratic.

As noted above, given the objectives of this research and my methodological approach, it is essential to establish a qualitative distinction between High Improvement in the Quality of Democracy and Not High Improvement. In the configurational causal modeling tradition, reconfiguring raw quantitative data to reflect the qualitative states of interest (e.g., “democracy” vs. “nondemocracy,” and “capitalist state” vs. “noncapitalist state”) is known as “calibration” and is a conventional practice (Ragin, Reference Ragin2000; Mello, Reference Mello2021). Because the EDS is roughly continuous with no obvious qualitative breakpoint, I used the empirical distribution to set the threshold for high improvement. I established the threshold for High Improvement at 0.5 standard deviations above the mean improvement during this period (threshold = 1.054). I set this threshold because 0.5 standard deviations above the grand mean is generally considered “high” relative to the mean. I coded this condition in a binary fashion, so that all states with improvement in the EDS above 1.054 were scored as achieving high improvement, and all those that did not reach this threshold were scored as not having high improvement. In Supplemental Material 11, I provide a detailed discussion of my choice of binary conditions in this study; and in Supplemental Material 15, I consider how the choice of alternative thresholds for inclusion in the condition High Improvement in the Quality of Electoral Democracy affects my inferences about the contributors to this phenomenon.

Using the threshold described above, fourteen states achieved High Improvement in the Quality of Electoral Democracy: California, Colorado, Connecticut, Hawaii, Illinois, Michigan, Nevada, New Jersey, New York, Oregon, Pennsylvania, Vermont, Virginia, and Washington.

2. Period of study

This study examines changes in the quality of electoral democracy in the states between 2000 and 2022. A detailed discussion of the rationale for this period of study can be found in Supplemental Material 18. As I note in the conclusion to this article, future research could extend the analysis backward in time to further investigate trends in state-level electoral democracy.

3. Hypotheses

I assess the potential contribution of five state-level factors to high improvement in the quality of electoral democracy: strong unions, high Democratic Party control of state government, a liberal state Democratic Party, a large population of people of color, and a liberal public mood.

Recall that, in this study, I seek to explain the qualitative state of high improvement in democratic quality and measure this concept accordingly as a binary factor. In this section, I conceptualize each explanatory factor in the same way, as a qualitatively distinctive state rather than a continuous (or interval or ordinal) variable (Beach and Pedersen, Reference Beach and Pedersen2016; Goertz, Reference Goertz2017). In Supplemental Material 12, I calibrate (Ragin, Reference Ragin2000; Mello, Reference Mello2021) a raw variable to construct my final measure of each explanatory factor to conform to this conceptualization.

As discussed in greater detail below, I adopt a configurational causal modeling approach in this study, in which individual factors exercise influence as insufficient but nonredundant parts of combinations of multiple conditions (known as conjunctions) that are unnecessary but sufficient for the outcome of interest (INUS; Mackie, Reference Mackie1974; Baumgartner and Ambühl, Reference Baumgartner and Ambühl2020). How the factors relate to a configurational understanding of high improvement in the quality of electoral democracy is discussed in detail in Supplemental Material 17. My approach to understanding the factors aligns with the regularity theory of causality (Mahoney and Acosta, Reference Mahoney and Acosta2021) and the method of analysis—CNA (Baumgartner and Ambühl, Reference Baumgartner and Ambühl2020)—adopted in this study.

4. Configurational modeling and coincidence analysis

I use CNA (Baumgartner, Reference Baumgartner2009, Reference Baumgartner2013; Baumgartner and Ambühl, Reference Baumgartner and Ambühl2020, Reference Baumgartner and Ambühl2023) to study the causes of high improvement in the quality of electoral democracy in the American states. CNA is one member of a family of CCMs that includes Quantitative Comparative Analysis (QCA; Ragin Reference Ragin1987, Reference Ragin2000), Logic Regression (Kooperberg and Ruczinski, Reference Kooperberg and Ruczinski2023), and Combinatorial Regularity Analysis (CORA; Thiem, Mkrtchyan and Sebechlebská, Reference Thiem, Mkrtchyan and Sebechlebská2024), which are built for detecting relations of sufficiency and necessity in quantitative data and are used widely in the social and behavioral sciences (Baumgartner and Ambühl, Reference Baumgartner and Ambühl2020; Swiatczak and Baumgartner, Reference Swiatczak and Baumgartner2024).Footnote ¹⁰ Important differences between CNA and QCA, a CCM method used widely in political science (particularly in comparative politics and international relations) since the 1980s, are discussed in Supplemental Material 9; and key differences between CNA and regression analysis methods (RAMs) are discussed in Supplemental Material 10.

CNA models the INUS theory of causation, in which X is a cause of Y if, and only if, X is an insufficient but non-redundant part of a condition that is unnecessary but sufficient to produce the outcome (Mackie, Reference Mackie1974). The INUS theory of causation also emphasizes equifinality, in which different combinations of factors can produce the same outcome. The INUS theory of causation belongs to the family of regularity theories of causation, which have their origins in the work of Hume (1748/Reference Hume2011) and assert that “causation is ultimately a deterministic form of dependence, such that whenever the same complete cause occurs the same effect follows” (Baumgartner and Ambühl, Reference Baumgartner and Ambühl2020:4, emphasis added). In practice, however, departures from strict determinism are accepted given “insufficient control over background influences generating noise” in most real-world datasets (Baumgartner and Ambühl, Reference Baumgartner and Ambühl2020:4).

The CNA algorithm takes a rectangular m x k dataset in which m is the number of cases and k is the number of factors (the outcome and explanatory conditions), and searches for explanatory conditions that are sufficient and/or necessary for the outcome in particular cases. CNA searches “from the bottom up,” “first check[ing] single factor values for their sufficiency for the outcome and conjunctively add[ing] additional factors only if no less complex conjunction is sufficient” (Swiatczak and Baumgartner, Reference Swiatczak and Baumgartner2024:1916). The CNA algorithm, thus, for example, starts with assessing whether condition X is sufficient for the outcome Y. If not, it evaluates whether the combination of X and another factor, Z, is sufficient for Y, and continues with increasingly complex combinations of conditions (Haesebrouck, Reference Haesebrouck2023:358). This procedure produces a set of sufficient combinations (conjunctions, or conjuncts) of factors for the outcome Y. Next, CNA assesses whether the sufficient combinations can be combined into alternative combinations, known as disjunctions, necessary for the outcome (that is, always present when the outcome is present). Ultimately, “[t]he result of this procedure are formulas that represent minimally necessary disjunctions…of minimally sufficient conjuncts” for the outcome (Haesebrouck, Reference Haesebrouck2023:358). Importantly, CNA is designed to generate redundancy-free solutions—that is, solutions that contain only factors that are indispensable to account for (at least one instance of) the outcome—and are thus potentially causally interpretable within the INUS theory (on this interpretation, see especially Baumgartner and Ambühl, Reference Baumgartner and Ambühl2023; Baumgartner and Falk Reference Baumgartner and Falk2024).

CNA employs user-designated standards of model predictive power (consistency) and recall (coverage) as the thresholds for returning models. The consistency measure, which can vary between 0 and 1, provides information about how well the CNA solution predicts observed instances of the outcome in cases. Thus, for example, a consistency score of 0.80 indicates that 80 percent of cases that present the solution found by the CNA also instantiate the outcome. The coverage measure, which also can vary between 0 and 1, indicates the proportion of cases of the outcome accounted for by the solution. A coverage score of 0.80 would suggest that the CNA solution accounts for 80 percent of the cases of the outcome (Ragin, Reference Ragin2000).Footnote ¹¹

Because real-world data rarely exhibit perfectly sufficient or necessary conditions due to incomplete causal models, the absence of some empirically possible configurations of factors from the dataset, measurement error, and so forth (Baumgartner and Ambühl, Reference Baumgartner and Ambühl2023:9), users can permit CNA to recognize sufficient or necessary conditions even if consistency and/or coverage are less than 1.Footnote ¹² If no model meets the user’s explicit consistency and coverage standards CNA will not return any solutions. To guard against omitted variable bias, it is conventional to only accept CNA solutions with consistency and coverage scores ≥ 0.75 (Baumgartner and Ambühl, Reference Baumgartner and Ambühl2020, Reference Baumgartner and Ambühl2023).

As with any observational data method, CNA faces obstacles to valid inference. Most importantly, unobserved factors that are associated with both the explanatory factors and the outcome of interest may be “confounders” that cast doubt on the causal interpretability of the CNA (Baumgartner and Ambühl, Reference Baumgartner and Ambühl2023:9–10). Concerns about unobserved confounders are most acute when the model’s coverage of the outcome is low. In contrast, “the higher the coverage, the less likely it is that we are facing data confounding, [and] the more reliable a causal interpretation of issued models becomes” (Baumgartner and Ambühl, Reference Baumgartner and Ambühl2020:531). Although I seek to limit confounding by only accepting models with coverage ≥ 0.75, confounding could be present in any model where coverage < 1.Footnote ¹³

5. Data and measures

I included all 50 states in my study. Because I conceptualize each condition as a qualitative state, as discussed above, I calibrate (Ragin, Reference Ragin2000; Mello, Reference Mello2021) the raw scores for each variable to a binary factor representing the condition of interest. For each binary factor, I calibrate the factor to emphasize the distinction between a high degree of the factor and a not high degree. A detailed discussion of my decision to use binary factors is provided in Supplemental Material 11.

I used quantitative indicators commonly employed in research in American politics to operationalize each of the explanatory factors—Strong Unions, High Democratic Control, Liberal Democratic Party, High Population of People of Color, and Liberal Public Mood—in this study. A full discussion of the indicators and calibration decisions is provided in Supplemental Material 12. Consistent with best practices in the CCM literature, I present a table with the final scores on each of the factors for all 50 states in Supplemental Material 13. I use Parkkinen and Baumgarnter’s (Parkkinen and Baumgartner, Reference Parkkinen and Baumgartner2023) fit-robust approach to select the best possible CNA model from the data at hand. This approach is thoroughly discussed in Supplemental Material 14.

Because the cut-point for establishing membership in each binary factor is somewhat arbitrary, I conducted robustness tests that applied alternative thresholds. Robustness testing of calibration decisions is standard practice in the tradition of configurational causal modeling (e.g., Schneider and Wagemann Reference Schneider and Wagemann2012; Mello, Reference Mello2021). The robustness tests are discussed fully in Supplemental Material 15.

7. Robustness tests

I conducted two different sets of robustness tests of the main CNA results presented in this study. The first set of robustness tests follows CCM best practices by systematically varying the thresholds for the presence or absence of each factor and checking how this affected the CNA solutions (Mello, Reference Mello2021). The results of these robustness tests are presented in Supplemental Material 15. Overall, they provide strong support for the main results presented here.

In the second robustness test presented in Supplemental Material 16, I re-estimated my CNA model using an alternative measure of the outcome, High Improvement in the Quality of Electoral Democracy, derived from Grumbach’s (Reference Grumbach2023) updated SDI. Overall, the results in Supplemental Material 13 reaffirm several of my findings, while also providing distinctive insights.

8. Conclusion

Over the past decade, scholars and pundits have highlighted democratic deterioration in some American states. While this attention is warranted, it has obscured another important—and more promising—phenomenon: dramatic improvement in the quality of electoral democracy in other states. Indeed, as I show using the electoral democracy score (EDS), an index of 53 policies relating to ballot access, fairness in districting, and election administration, more than a quarter of states experienced large gains (of more than 0.5 standard deviations above the mean) in the quality of their electoral democracies between 2000 and 2022. Previous research leaves us unprepared to explain the dramatic improvement in democratic quality in these states.

Why did some states experience large improvements in the quality of their electoral democracies? In this article, I investigate five hypotheses about factors that may contribute to major democratic gains. I suggest strong unions, long periods of unified Democratic control of state government, liberal Democratic Parties, large populations of people of color, and a liberal public mood may each contribute, most likely as INUS causes, to a substantial improvement in democratic quality.

Using CNA, a method custom-built for detecting relations of sufficiency and necessity in quantitative data, I scrutinize whether these factors contribute to a dramatic improvement in the quality of electoral democracy. The results of my CNA suggest that there are three minimally sufficient paths to this outcome. States had a High Improvement in the Quality of Electoral Democracy if, and only if, they had High Democratic Control AND Not Liberal Public Mood; OR had Strong Unions AND High Populations of People of Color; OR had Liberal Democratic Parties AND Liberal Public Mood AND Strong Unions. The results of the CNA had high coverage and consistency and showed robustness to sensitivity tests.

This analysis points most clearly to the importance of the Democratic Party and strong unions as INUS causes of substantial gains in democratic quality, while also revealing roles for large populations of people of color and a liberal public mood. These findings provide a valuable extension of previous research on trends in democracy in the American states (Bentele and O’Brien, Reference Bentele and O’Brien2013; Rocha and Matsubayashi, Reference Rocha and Matsubayashi2014; Hicks et al., Reference Hicks, McKee, Sellers and Smith2015; Anderson, Reference Anderson2018; Levitsky and Ziblatt, Reference Levitsky and Ziblatt2019; Grumbach, Reference Grumbach2022, Reference Grumbach2023; Pomante et al., Reference Pomante, Schraufnagel and Li2023). While previous research (e.g., Grumbach, Reference Grumbach2023) has indicated that unified Republican control of state government contributes to democratic deterioration, this study shows that both a long period of Democratic unified control of state government and a particularly liberal Democratic Party are INUS contributors to dramatic improvement in democracy. This is a distinctive finding: while previous research has shown a correlation between party control and the quality of electoral democracy, my work extends these insights by revealing that Democrats can contribute to high improvement in the quality of electoral democracy via two alternative paths (one featuring a high degree of Democratic control of state government, and the other involving an especially liberal Democratic Party without a high degree of Democratic control). Moreover, the results of this study indicate that high Democratic control of state government is neither sufficient nor necessary for high improvement.

This study also builds on recent research (e.g., Frymer and Grumbach, Reference Frymer and Grumbach2021; Frymer, Grumbach, and Hill, Reference Frymer, Grumbach and Hill2025) by highlighting the critical role of labor unions in contributing to the expansion of democracy. As I show, strong labor unions are INUS contributors on two minimally sufficient paths to substantial improvement in the quality of democracy. Whereas most previous studies of struggles over electoral democracy in the states pay relatively little attention to unions (Bentele and O’Brien, Reference Bentele and O’Brien2013; Grumbach, Reference Grumbach2023; Pomante et al., Reference Pomante, Schraufnagel and Li2023; but see Frymer, Grumbach, and Hill, Reference Frymer, Grumbach and Hill2025), my research puts unions at the center of these conflicts and highlights how powerful unions contribute to high-quality electoral democracies. Future research should make unions more central to their analyses of why state democracies flourish or decline.

Finally, this study reveals that large populations of people of color, when combined with strong unions, can contribute to significant democratic gains. This implies that communities of color may advance democratic renewal through voting, social movement activity, and lobbying, possibly through labor unions and independently. While some previous studies (e.g., Bentele and O’Brien, Reference Bentele and O’Brien2013; Pomante et al., Reference Pomante, Schraufnagel and Li2023) have found that large populations of people of color are correlated with democratic decline (attributable to Republican-led voter suppression efforts against communities of color), my study suggests that people of color can exercise political agency, working with labor unions to enhance democratic quality. That said, if support among communities of color for the Democratic Party continues to erode (e.g., Sommer and Franco, Reference Sommer and Franco2024a), the relationship between the size of communities of color and improvement in the quality of electoral democracy could disappear.

Of course, much research on the dynamics of electoral democracy in the states remains to be done. In this study, I examined patterns in the quality of electoral democracy during a relatively brief interval (2000-2022) in American history. Consequently, we cannot know how well the conclusions of this research apply to earlier periods. It would be valuable to extend the EDS backward in time, to trace (and explain) patterns in democratic quality in the American states over a much longer period. Doing this is challenging, because contemporary interstate differences in democratic quality are relatively modest compared to earlier periods when, for example, states varied in whether (and the degree to which) they allowed African Americans, Latinx Americans, and women to vote (e.g., Grumbach, Reference Grumbach2023:15). Still, expanding the temporal scope of analysis of the dynamics of state democracy would help contextualize the fundings of this study, and perhaps provide insights on likely future trends.

Additionally, while this research provided strong cross-case evidence for some of my hypotheses, it does not provide direct insights into the causal mechanisms by which factors like strong unions, Democratic Party control of state government, or a large population of people of color contributed to a high improvement in the quality of democracy. Future research should build on the cross-case patterns identified in this study with qualitative case studies that trace the causal mechanisms linking the factors to substantial democratic improvement. Such process-tracing research would provide a more complete explanation of cases of high improvement in democratic quality.

Future qualitative research should also study the deviant cases—both those like Massachusetts, which possessed the causal factors but failed to instantiate substantial improvement; and those like Pennsylvania, Virginia, and Vermont, which achieved substantial improvement despite lacking the relevant causal conditions—to understand the unexpected outcomes in these cases. Such research could contribute to further theory development and testing, thus enabling greater coverage of cases of high improvement. This could help address lingering concerns about omitted factor bias in the present study. Finally, while this study has focused on explaining large gains in electoral democracy, future research should examine why states achieve substantial improvement in the quality of other substantive dimensions of democracy, such as social welfare, civil rights, and the rule of law.

Ultimately, this research underscores that the advance of state-level democracy is profoundly influenced by political struggles within states. I hope this work encourages further research into how and why ordinary citizens, labor movements, and party organizations can advance the cause of democracy in the United States and abroad.