Phil Swatton
@philswatton.bsky.social
Work as a senior data scientist at the Alan Turing Institute, background in political science. Views my own and not necessarily shared by my employer.
https://philswatton.github.io/
https://philswatton.github.io/
I think my overall interpretation is close to yours, from the limitations/robustness section and the conclusion, respectively:
November 23, 2025 at 11:18 AM
I think my overall interpretation is close to yours, from the limitations/robustness section and the conclusion, respectively:
One of the things I got out of using vote choice (beyond capturing knowing how someone votes given some characteristic) is a straightforward relationship between the predictability of ideology/demographic models + largest 2/1 party vote share 3/3
November 21, 2025 at 5:38 PM
One of the things I got out of using vote choice (beyond capturing knowing how someone votes given some characteristic) is a straightforward relationship between the predictability of ideology/demographic models + largest 2/1 party vote share 3/3
I do explore including constituency variables in the paper, at least for the UK portion, to capture the issue of tactical voting etc.
TLDR is even these models helped w/ LD, PC, SNP, but weren't sufficient for smaller third parties. Partly because main party voters denser in the same space 2/3
TLDR is even these models helped w/ LD, PC, SNP, but weren't sufficient for smaller third parties. Partly because main party voters denser in the same space 2/3
November 21, 2025 at 5:37 PM
I do explore including constituency variables in the paper, at least for the UK portion, to capture the issue of tactical voting etc.
TLDR is even these models helped w/ LD, PC, SNP, but weren't sufficient for smaller third parties. Partly because main party voters denser in the same space 2/3
TLDR is even these models helped w/ LD, PC, SNP, but weren't sufficient for smaller third parties. Partly because main party voters denser in the same space 2/3
If, like me, you are an enjoyer of Ben's patented bubble plots, you'll have likely noticed that voters cluster by ideology and not demographics.
My new preprint is on exactly that: osf.io/preprints/so...
My new preprint is on exactly that: osf.io/preprints/so...
November 21, 2025 at 4:47 PM
If, like me, you are an enjoyer of Ben's patented bubble plots, you'll have likely noticed that voters cluster by ideology and not demographics.
My new preprint is on exactly that: osf.io/preprints/so...
My new preprint is on exactly that: osf.io/preprints/so...
Headline empirical findings:
Models using demographics are not much better at prediction than just assuming all voters vote the same way.
Models using ideology however are better, and are close to models using past vote choice w/ constituency vars
This result generalises from BES to EES
Models using demographics are not much better at prediction than just assuming all voters vote the same way.
Models using ideology however are better, and are close to models using past vote choice w/ constituency vars
This result generalises from BES to EES
November 21, 2025 at 7:17 AM
Headline empirical findings:
Models using demographics are not much better at prediction than just assuming all voters vote the same way.
Models using ideology however are better, and are close to models using past vote choice w/ constituency vars
This result generalises from BES to EES
Models using demographics are not much better at prediction than just assuming all voters vote the same way.
Models using ideology however are better, and are close to models using past vote choice w/ constituency vars
This result generalises from BES to EES
I conclude by summarising my findings and interpreting them in light of the three kinds of contribution predictive modelling can make to the abductive reasoning process. (22/25)
November 20, 2025 at 8:09 AM
I conclude by summarising my findings and interpreting them in light of the three kinds of contribution predictive modelling can make to the abductive reasoning process. (22/25)
After discussing some limitations, I provide a discussion of why I believe demographics are so prominent in our political discourse despite their limited ability to predict vote choice (21/25).
November 20, 2025 at 8:09 AM
After discussing some limitations, I provide a discussion of why I believe demographics are so prominent in our political discourse despite their limited ability to predict vote choice (21/25).
Next, I use PCA to explore the separability of third party voters. I show that, typically, wherever third party voters are in the latent space, there is greater density for voters of a main party. This explains why they are difficult to predict without constituency information. (20/25)
November 20, 2025 at 8:09 AM
Next, I use PCA to explore the separability of third party voters. I show that, typically, wherever third party voters are in the latent space, there is greater density for voters of a main party. This explains why they are difficult to predict without constituency information. (20/25)
I perform some post-hoc analysis to explore my findings. First, I use mutual information to explore the pairwise reductions in uncertainty in the BES train set. I find ideological variables most reduce uncertainty in vote choice. This gives confidence I have not underfit my models (18/25)
November 20, 2025 at 8:09 AM
I perform some post-hoc analysis to explore my findings. First, I use mutual information to explore the pairwise reductions in uncertainty in the BES train set. I find ideological variables most reduce uncertainty in vote choice. This gives confidence I have not underfit my models (18/25)
Exploring the point estimates, I show that the vote share of the two largest parties acts as a ceiling on the performance of ideology models, while demographics models broadly trace the vote share of the single largest party in their performance (17/25)
November 20, 2025 at 8:09 AM
Exploring the point estimates, I show that the vote share of the two largest parties acts as a ceiling on the performance of ideology models, while demographics models broadly trace the vote share of the single largest party in their performance (17/25)
Switching to the EES, I find that in none of the 15 countries is there sufficient evidence to rule either modelling approach better, as the 95\% confidence intervals consistently overlap one another.
(15/25)
(15/25)
November 20, 2025 at 8:09 AM
Switching to the EES, I find that in none of the 15 countries is there sufficient evidence to rule either modelling approach better, as the 95\% confidence intervals consistently overlap one another.
(15/25)
(15/25)
Breaking down the results by party, it becomes clear that without constituency variables, no models other than one using past vote choice are particularly good at predicting third party voters. (14/25)
November 20, 2025 at 8:09 AM
Breaking down the results by party, it becomes clear that without constituency variables, no models other than one using past vote choice are particularly good at predicting third party voters. (14/25)
In the results for the BES, I find that models using demographics are not much better than assuming all voters are just voting for the largest party. Models based on ideology, by contrast, do somewhat better.
(13/25)
(13/25)
November 20, 2025 at 8:09 AM
In the results for the BES, I find that models using demographics are not much better than assuming all voters are just voting for the largest party. Models based on ideology, by contrast, do somewhat better.
(13/25)
(13/25)
I therefore select for each election study a set of demographic indicators and some ideological scales. These are redistribution and EU attitudes for the BES, and 7 issue dimensions for the EES. I avoid using left-right placements because they are known to contain a partisan component. (10/25)
November 20, 2025 at 8:09 AM
I therefore select for each election study a set of demographic indicators and some ideological scales. These are redistribution and EU attitudes for the BES, and 7 issue dimensions for the EES. I avoid using left-right placements because they are known to contain a partisan component. (10/25)
I argue that in this approach, it is important to exclude both moderators and variables which are essentially tautological.
An example of the latter might be party 'like' ratings - if we say people who like party X vote for party X, we leave ourselves with the problem of explaining X.
(9/25)
An example of the latter might be party 'like' ratings - if we say people who like party X vote for party X, we leave ourselves with the problem of explaining X.
(9/25)
November 20, 2025 at 8:09 AM
I argue that in this approach, it is important to exclude both moderators and variables which are essentially tautological.
An example of the latter might be party 'like' ratings - if we say people who like party X vote for party X, we leave ourselves with the problem of explaining X.
(9/25)
An example of the latter might be party 'like' ratings - if we say people who like party X vote for party X, we leave ourselves with the problem of explaining X.
(9/25)
I highlight three ways in which predictive inference can contribute to the abductive reasoning process:
1) Pragmatic inference
2) Testing a theoretical implication
3) Prediction as discovery
(7/25)
1) Pragmatic inference
2) Testing a theoretical implication
3) Prediction as discovery
(7/25)
November 20, 2025 at 8:09 AM
I highlight three ways in which predictive inference can contribute to the abductive reasoning process:
1) Pragmatic inference
2) Testing a theoretical implication
3) Prediction as discovery
(7/25)
1) Pragmatic inference
2) Testing a theoretical implication
3) Prediction as discovery
(7/25)
As Spirling & Stewart argue (doi.org/10.1086/734280), poli sci proceeds through inference to the best explanation. This means there are three kinds of contribution:
1) Proposing new facts
2) Proposing new explanations of observed facts
3) Selecting between existing explanations
(3/25)
1) Proposing new facts
2) Proposing new explanations of observed facts
3) Selecting between existing explanations
(3/25)
November 20, 2025 at 8:09 AM
As Spirling & Stewart argue (doi.org/10.1086/734280), poli sci proceeds through inference to the best explanation. This means there are three kinds of contribution:
1) Proposing new facts
2) Proposing new explanations of observed facts
3) Selecting between existing explanations
(3/25)
1) Proposing new facts
2) Proposing new explanations of observed facts
3) Selecting between existing explanations
(3/25)
The sociological tradition models vote choice as downstream of demographics. The spatial tradition models vote choice as downstream of ideology. (2/25)
November 20, 2025 at 8:09 AM
The sociological tradition models vote choice as downstream of demographics. The spatial tradition models vote choice as downstream of ideology. (2/25)
Political scientists are interested in explaining vote choice. This has been a concern from the beginning of the discipline. In this paper I consider two influential traditions: the sociological tradition and the spatial tradition (1/25)
November 20, 2025 at 8:09 AM
Political scientists are interested in explaining vote choice. This has been a concern from the beginning of the discipline. In this paper I consider two influential traditions: the sociological tradition and the spatial tradition (1/25)
I have a new preprint: 'Predictive Modelling Shows Demographics Do Not Predict Vote Choice at the Individual Level'
DOI: doi.org/10.31235/osf...
Thread below
DOI: doi.org/10.31235/osf...
Thread below
November 20, 2025 at 8:09 AM
I have a new preprint: 'Predictive Modelling Shows Demographics Do Not Predict Vote Choice at the Individual Level'
DOI: doi.org/10.31235/osf...
Thread below
DOI: doi.org/10.31235/osf...
Thread below
I've written a blog post about using metric unfolding to extract latent dimensions from data, using the ANES feeling thermometers & BES 'like' scales as examples.
It's a high-level intro to unfolding and using the smacof R package. These plots are the results.
philswatton.github.io/2025/11/05/e...
It's a high-level intro to unfolding and using the smacof R package. These plots are the results.
philswatton.github.io/2025/11/05/e...
November 7, 2025 at 11:55 AM
I've written a blog post about using metric unfolding to extract latent dimensions from data, using the ANES feeling thermometers & BES 'like' scales as examples.
It's a high-level intro to unfolding and using the smacof R package. These plots are the results.
philswatton.github.io/2025/11/05/e...
It's a high-level intro to unfolding and using the smacof R package. These plots are the results.
philswatton.github.io/2025/11/05/e...
Here's the equivalent distributions for ANES 2024, they look much spikier (but possibly you still get more out of it, if e.g. people use values ending 0 or 5, that's still 21 meaningful values vs 11 in 0-10 or 7 in 1-7)
November 4, 2025 at 9:49 PM
Here's the equivalent distributions for ANES 2024, they look much spikier (but possibly you still get more out of it, if e.g. people use values ending 0 or 5, that's still 21 meaningful values vs 11 in 0-10 or 7 in 1-7)
Made this example w/ the 2024 data, will upload a blog post or gist over the next couple of days:
November 4, 2025 at 9:42 PM
Made this example w/ the 2024 data, will upload a blog post or gist over the next couple of days:
1-7 scales are on issues
0-100 are warmth, but you can recover a liberal-conservative dimension from them (I think I have a better plot somewhere):
0-100 are warmth, but you can recover a liberal-conservative dimension from them (I think I have a better plot somewhere):
November 4, 2025 at 5:33 PM
1-7 scales are on issues
0-100 are warmth, but you can recover a liberal-conservative dimension from them (I think I have a better plot somewhere):
0-100 are warmth, but you can recover a liberal-conservative dimension from them (I think I have a better plot somewhere):
I'm not sure actually - I wouldn't use 0-100 for issue scales - when I wish for them it's more about having warmth ratings towards lots of different stimuli. I helped make a presentation to the RSS on the 2020 presidential w/ them, they were really interesting for nonmetric unfolding:
November 4, 2025 at 5:31 PM
I'm not sure actually - I wouldn't use 0-100 for issue scales - when I wish for them it's more about having warmth ratings towards lots of different stimuli. I helped make a presentation to the RSS on the 2020 presidential w/ them, they were really interesting for nonmetric unfolding: