Does inequality undermine democracy? The evidence is weak

Intuitively, on a personal level, it makes a lot of sense to me that high inequality (or over-time increases in inequality) should lead to weaker democratic institutions in practice and on paper.

More importantly, there’s a clear theoretical case to be made in this regard. When elites become disproportionally economically powerful, it should be more likely for them to use their disproportionate resources to dismantle democracy, thus disempowering the masses, and tilting the political and economic system even more in their own favor. There are at least two separate arguments here, one having to do with motivation and the other with means.

First, when income/wealth gaps grow or get really big, democracy starts to feel more threatening to the rich. If most voters live comparatively meager while a small group controls most of the income/wealth, those voters might think to elect politicians who promise higher taxes on the wealthy or tougher business regulations. Fearing this, the elites would rationally want to rejigger democratic institutions such that people can’t actually touch their money. They might press to enact constitutional changes that weaken legislatures or try and expand presidential powers to bypass normal checks and balances. That is, they might be motivated to dabble in a bit of democratic backsliding.

Second, disproportionally rising income/wealth at the same time enables the elite to do what it wants. When a small group controls most of a country’s income and assets, they can fund political parties, bribe politicians, finance legal challenges, acquire media outlets, and do whatever else you need to ensure backsliding. Moreover, because any potential opposition controls a comparatively small amount of resources, it’s harder for them to fight back. The playing field isn’t levelled, and so it’s even easier for the elite to get what it wants and do so at a faster pace than anyone else.

Marxists and other radical theorists are the ones who typically make these points, but I don’t think you have to be either to see the logic at work here. Just use basic rational choice and public choice theories.

What the existing empirical evidence says

But what systematic empirical evidence do we have that this is the case? To my big surprise, the evidence is pretty thin, or at least uncomfortably mixed. I couldn’t find good meta-analyses (or even any meta-analysis) on the topic. But there are individual studies.

A 2009 paper published in World Politics shows that inequality harms democratic consolidation but has “no net effect on democratization.” If anything, inequality “weakly promotes democratization, though not significantly.” The method is nothing impressive, but control variables are fairly sensible.

However, a 2015 paper that does much more to address the endogeneity issue, employing internal instruments with sGMM models, argues that there’s actually no effect either way. As the author puts it

Despite being correlated, there is very little evidence of any effect of income inequality on level of democracy once employing appropriate model specifications. Furthermore, there is no robust evidence that inequality systematically affects either democratization prospects or democratic stability.

A more recent 2024 paper published in PNAS pushes back and tries reestablishing the original result, namely, that unequal income distribution causes democratic erosion. Unfortunately, the authors don’t even mention the word “endogeneity” in the study. Their main specification is severely confounded, as there’s almost no control variables. They remedy this somewhat later on by adding a few extra controls, such as polarization, age of democracy, and state capacity. The negative association between the Gini coefficient and democracy they find is fairly robust, but it’s purely correlational.

Finally, there’s a paper from just a few months ago, which also finds a negative correlation in a sample of developed societies between 2006 and 2017. Unfortunately, the study is almost completely worthless as it presents pooled OLS regressions, which is not how you model panel data. Typically, there’s unobserved time-invariant confounding that has to be partialed out by using fixed effects or correlated random effects. The author gestures in this direction by saying that a supplementary fixed-effects regressions shows the same results as the main pooled OLS regression. But the fixed-effects model he has in mind isn’t what you’d expect, because in it he only includes one general dummy for Anglo-Saxon countries. This doesn’t solve unobserved time-invariant confounding at all. You need individual country dummies, not one single regional dummy. (The author also thinks he doesn’t need to cluster standard errors, even though it’s very likely there’s autocorrelation in such panel data, which requires clustering.)

And that’s mostly it as far as published research goes. In 2023, a Master’s thesis examined the same topic and found “inconclusive” results, arguing that “no definitive relation could be claimed between inequality and democracy,” after applying different estimation techniques (meaning pooled OLS and fixed-effects regressions). By the way, the author here did a better job of controlling for confounders than the PNAS study.

There’s a also a 2025 working paper that should be mentioned. It affirms the impression I’ve been laying out so far, which is that “empirical analyses of this interplay [i.e., the relationship between economic development, inequality, and democratization] have been limited and have yielded overall mixed and inconsistent results.” The authors then present their own panel regression analyses and find “no discernible net effect of inequality on democracy,” adding that their “results are robust across different identification strategies, outcome variable choices, and subsample analyses.”

Interesting, right? If you’re like me, you were pretty sure that there’s a lot of quantitative research on the topic of inequality—>democracy, and that the results mostly point in the negative direction. I guess not.

Let’s take a closer look at the data

So, I decided to take a look myself. I picked three standard measures of inequality: shares of total disposable income held by the top 1% and the top 0.1%, and the share of net wealth held by the top 1%. I added the standard control variables (GDP per capita, urbanization, education, age structure, and so on), and ran cross-sectional and panel regressions with different measures of democracy (V-Dem, Polity, Freedom House).

With income inequality, there’s a pretty clear and strong bivariate relationship in the expected direction under a cross-sectional design (see below). The relationship even survives when controlling for economic development. Wealth inequality shows nothing statistically significant no matter what I do.

But okay, this is virtually useless because omitted-variable bias and reverse causality are pervasive here.

Simple bivariate scatterplots.

When I turn to more informative panel regressions, though, it’s the mixed picture from before (see the coefficient plot below). Here, country-fixed effects are included, so we’re looking at how inequality and democracy are related over time within each country. Time-invariant confounding is controlled for, as is a number of important observed time-varying confounders.

Even when a statistically significant relationship emerges, it’s absolutely tiny! I lagged inequality by 5 years to try and control for reverse causation a bit, though nothing much changes if I drop the lag.

I also played around with sGMM models to try and do something more serious about endogeneity. There, almost every result was statistically non-significant.

Fixed-effects panel regressions with control variables.

I then ran a bunch of robustness tests, and the results stay almost completely the same, that is, highly mixed. No matter whether I drop all controls, add region-year interactions or time dummies, control for possible time trends – it’s just as mixed, if not more, as before.

Trying different specifications.

As always, what we really need is good, convincing causal inference, not just cross-sectional OLS, fixed-effects panel regressions, or even sGMM models. But in lieu of a proper causal design, and going off of these different statistical tests (and the existing literature), the claim that inequality is destroying democracy seems a bit iffy to me. It’s not well-established at all.

(A reader also asked about critical thresholds. Perhaps inequality only starts becoming bad for democracy after reaching a high enough level. That’s plausible. I tested a few thresholds and ran a bunch of different specifications. I find some evidence for the link when I set the threshold quite high; for instance, after the top 1% income share surpasses 15% or 20% (the latter being quite a bit higher than even in South Africa today). However, when I then start employing lags to somewhat guard against reverse causality, the results turn out to be statistically non-significant. Cross-sectional evidence is just as mixed.)