This text is a condensation of the main findings of my PhD thesis “On the Record: Understanding A Century of Congressional Lawmaking Through Speech and Vote Behavior.” Published via Linköping University Press, doi:10.3384/9789181182828.
Most modern democratic parliaments publish their proceedings and vote results in publicly accessible databases, and they are often broadcasted live. U.S. Congressional debates are available on C-SPAN, while German federal debates are accessible through public broadcasting services such as Phoenix. Despite this openness, we know surprisingly little about how these parliaments actually function. All this transparency for little understanding of what causes policymakers to vote in a certain way, and without a real explanation of the unexpected votes which prevent laws that were expected to pass.
In my research, I argue that there are fundamentally two reasons for this: current research typically focuses on parties as the relevant actors and does not take social mechanisms into account; and until recently, a large-scale analysis of parliamentary proceedings was simply not feasible. In this article, I want to reflect on these, and outline the opportunities political sociology can gain from a deeper understanding of the profoundly social mechanisms at work in the “machine rooms” of democracy.
First, parliamentary research typically does not take social mechanisms into account. Most research on the various democratic parliaments worldwide shares a central assumption: the relevant actors acting in parliaments are parties, not individual lawmakers. This assumption makes sense, given that research has over and over demonstrated that the best predictor of vote results is party affiliation. But at the same time, all the data that parliaments produce are on the individual-level.
In the U.S., political scientists have long realized that their models, which operate on the individual level, can’t really say much about the parties these individuals are part of — even though the polarized clusters recovered from them are oddly party-shaped. When representatives switch party affiliation (most famously Ronald Reagan, who was for most of his pre-presidential career a member of the Democratic Party) this looks like an error in most models, rather than a deliberate decision. What is lacking for a better and deeper understanding of parliamentary politics is a sociological lens on parliament, one that attempts to explain all the various actions representatives take as results from social mechanisms — homophily, for example, or how socio-economic markers such as educational attainment can have a stronger impact on vote agreement than party affiliation.
There is a second reason for a lack of individual explanations for representative behavior. Despite the amounts of data being made public day after day in many parliamentary systems, most of the data fall into one of two categories: it’s either vote data, or speech data. Vote data is a staple of parliamentary research, with many approaches being very successful in studying polarization using vote results. This is the data that Keith Poole and Howard Rosenthal have used over four decades ago to demonstrate the immense amounts of polarization observable in U.S. Congress. But while this data is easy to analyze, it also comes with an Achilles’ Heel: it only tells us how representatives have voted, but not why. That, I argue, is the primary reason that we still understand little about social mechanisms in parliaments.
Underlying this propensity to analyze vote results was a fundamental inability to make sense of the deluges of textual data, transcribed speech from the individual representatives. And this has only started to change recently. While quantitative approaches to text analysis date back to the 1930s, it was until recently not possible to quantitatively analyze parliamentary records guided by theoretically meaningful categories. Only when Google researchers developed the transformer model — the underlying technology of both decoder models such as ChatGPT, and classifiers like BERT — did research gain a meaningful new method that allows coding large amounts of text. This was the missing puzzle piece researchers needed in order to be able to analyze not just how parties interact in parliaments, or how polarized representatives were, but what actually causes the coalitions we see, the vote results we observe, and the general mechanisms at play during parliamentary debates. Analyzing the speeches of representatives in conjunction with their votes, one crucial finding emerges: despite the power of parties to steer their members, social mechanisms are sometimes stronger.
This is work that I performed in my PhD thesis. Across three chapters I can show that it was not Ronald Reagan alone who instigated the large-scale rework of the tax system in the United States during the 1980s. Instead, I can show how the economic crises of the 1970s have significant influence in this. Second, I can show that polarization is multidimensional in showing that economic discourse has indeed become polarized, but orthogonal to party lines. Lastly, I show that it is possible to precisely measure the influence of parties on the voting process if one adopts an individual-level framework driven by mechanism-based social explanation, rather than focusing only on parties.
Now that LLMs and other text-analysis tools have unlocked parliamentary debates, it is time for political sociologists to develop a deeper and better understanding for the social mechanisms at work in parliaments. The applications are manifold. For example, we can explain the decline of U.S. Congressional parties since the 1980s. Or, and much more pressing as European multi-party systems face a threat from populism, we can use this to understand the inability of non-populist parties in reigning populism in. Many political phenomena can be explained via social mechanisms, and by tapping into the preferences representatives utter in their speeches we can unlock an entirely new area that demands our attention.
