Spanish Flu in Britain: The Diff-in-Diff That Never Was
When it became clear that COVID-19 was going to be a serious disruption to life in the UK, comparisons with the 1918 Spanish Flu pandemic were everywhere. Whether it was to emphasise that COVID-19 was not as physiologically dangerous, or older articles highlighting the long-lasting institutional changes that a pandemic can bring about, discussion of the present through the lens of the past was abundant.
This got me thinking. At the time of the Spanish Flu British politics was in a state of flux. By the time the world war ended and the pandemic began, the British state had swelled, spending more money than ever in history while simultaneously exerting unprecedented control over the economy. Bellicist accounts of state-making, spearheaded by Charles Tilly, argue that increased state spending in wars fuels the need for a deeper administrative apparatus, leading to the development of stronger states in the long run. Britain in the 1910s seems to be a great example of this mechanism.
Notice what I just said; “by the time the world war ended and the pandemic began”. The impact of the first world war is what is most written about in the history books, but we should not gloss over the transformative potential of the global pandemic that occurred right afterwards. It seems plausible that a pandemic might lead people to expect more healthcare provision and a more generous welfare system. But a world war might also lead people to want such things. In other words, two enormous events took place almost simultaneously, and both might be expected to have similar macro-level shifts in what people expect from the state. We have a problem of confounding; British state-society relations did transform, the state did get and stay bigger, the welfare state expanded, but we do not know whether this was due to the war, the pandemic, or both.
This was the puzzle I had in my mind when I sceptically read comparisons with COVID-19. To be sure, the Spanish Flu occurred just before a transformation of the British state. But its effects cannot be easily disentangled from those of the war. Hence saying that COVID will lead to outcome X, because the Spanish Flu occurred just before outcome X, should ring alarm bells. This is before one considers the physiological differences too. COVID, for instance, seems to be more deadly to older people, whilst the Spanish Flu was deadly for both young adults and the old, with middle-aged people being the least at risk. A pandemic in which young adults have a lot to fear would surely create very different generational coalitions than those produced by COVID.
So I had the puzzle, and, being in lockdown, suddenly had a lot of time. After a bit of reading, it then struck me that the Spanish Flu pandemic, occurring between 1918 and 1919, took place on either side of the UK 1918 general election. At this point I was also working (if not that productively) on a piece of Causal Inference coursework, which put my brain into overdrive looking for natural experiments to exploit.
My idea was as follows. If people’s political preferences transform because of the pandemic or the war, we might expect to see this reflected in voting patterns. Notably, the rising Labour Party campaigned on a public spending oriented platform.
A crucial difference between the war and the pandemic is timing. Whilst the war lasted years, everybody knew about it and had their life impacted by it, the pandemic was slightly different. Although the precise origins of Spanish Flu in Britain are uncertain, the disease spread in waves from city to city, most likely by returning soldiers travelling home by train. The wave-like nature means that different parts of the country were affected by the pandemic at different times, unlike the war which, by 1918, was more homogenous. There were three waves of the pandemic, but only two of these saw a substantial death rate. Areas which were not as badly hit in the first wave of the disease were typically worse hit by the second. In an era of minimal social distancing, perhaps this is the product of some form of herd immunity.
If we factor in the fact that fewer people were literate, and that national broadcasting did not yet exist, it is plausible to suggest that areas which saw fewer pandemic-related deaths would not perceive it as being so urgent a problem. Whilst the modern 24 hour news cycle means that everyone in Britain knew about COVID-19 quite quickly, this would probably not have been the case a hundred years ago.
So, if we assume that every electoral constituency is equally “treated” by the war, but that they are differentially “treated” by the pandemic, we have roughly set the scene for a difference-in-difference research design. Under such a design, we allocate constituencies that only receive the war treatment as the “control” group, the baseline against which constituencies affected by the pandemic, the “treatment” group, can be compared.
The difficulty, though, is that the previous election in 1910 took place under different suffrage laws and with different constituencies. Making direct comparisons to 1918 seems tricky. This was the first major problem.
To get around it, I planned to focus on how Labour vote share changed between the 1918 and 1922 elections. My thinking was that, if being “hit” by the pandemic permanently changes one’s preferences, then people who voted for the Labour Party in 1918 because of the pandemic would almost certainly do so again in 1922. Constituencies that were only hit by the pandemic after the 1918 election should see a corresponding, disproportionate boost in Labour vote share in 1922, since the 1918 Labour gains from the pandemic would not be priced in yet. But this is quite a strong, untestable assumption, introduced before delving into the tricky question of how to actually allocate groups. Do we select an arbitrary cut-off in the first peak death rate which we use to demarcate treatment and control? Do we use death rate as a continuous measure? What if all the treated constituencies are the ones that would be most likely to vote Labour for other reasons, namely by being urban and industrial?
Already, then, it was clear that the design had holes, but I started searching for data nonetheless. It was here that I encountered the second major problem. Although the Spanish Flu did arrive in waves, by the December election it had hit almost every part of the country quite severely. The below plots show death rate data aggregated into regions, showing two distinct waves of deaths. The vertical red lines indicate the election date. I had been severely hoping that, for some regions, the red line would fall just before, or in the very early stages of, the first peak. But this is not the case - by December 1918 everywhere seems to have gone through a wave of the disease, so allocating parts of the country to “treatment” and “control” seems silly. I used digitised death rate data (available here), which aggregates at the level of boroughs and counties. Even at this more granular level, there are only a handful of places (N = about 4) in which the first peak occurred after the election. These were rural constituencies that would not be expected to have many Labour voters anyway.
If history had gone a little bit differently, then diff-in-diff could have been the perfect way to tease out the effects of the war from the pandemic. Say there had somehow been an election in early 1918, before the flu arrived at all, and it resulted in a hung parliament and further election in the early autumn, when many parts of the country had still not received the first peak. Just a few years later, political squabbling led to elections in 1922, 1923 and 1924. In 1974 a hung parliament in February led to a fresh election in October. Such rapid elections are unlikely, but not impossible in British politics. We’d then have been able to compare constituencies between February and October 1918, with real variation in exposure to the pandemic but equal exposure to the war.
Alternatively, if the disease had somehow spread through Britain a bit more slowly, such that some areas were not hit by December 1918, then there would be real variation to exploit. But this would still be unideal, because of the tenuous assumptions about existing Labour voters’ behaviour in future elections.
In sum, this post is a track record of my unsuccessful attempt to do some COVID-inspired political science. The papers we see in journals are polished drafts number N, which have gone through copious discussion and debate with academics of far greater capability than me. But this masks the fact that so many papers get started and later binned, either due to uninteresting results, or because the required conditions for the econometrics do not actually hold. Getting a paper to work entails luck at multiple levels, and a lot of the time this luck isn’t there. I’m sure I’m not the only person who has spent too much time getting excited about a research design, only to find that it doesn’t work in real life. But it was still quite fun, and something to help get me through lockdown. So I suppose the moral of my story, besides the fact that I should have binned the project a bit sooner, is that research is hard and doesn’t always work, but that this is no reason to stop trying.