Jon Mellon
@jon_mellon
Followers
4,352
Following
1,389
Media
363
Statuses
5,908
Co-director @besresearch . Associate Professor West Point DSE. polisci methods/political behavior. Tweets do not represent employer.
Joined October 2010
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
SEVENTEEN
• 1097411 Tweets
Taylor
• 223357 Tweets
FELIP TAKES ON BBPHSTAGE
• 100582 Tweets
The Next Prince Q9
• 90992 Tweets
#TimnasDay
• 82056 Tweets
Carat
• 76435 Tweets
オーストラリア
• 63348 Tweets
#サッカー日本代表
• 49171 Tweets
Tattvadarshi Sant Rampal Ji
• 48430 Tweets
BBPH ANN1V WITH BINI
• 42509 Tweets
#あのクズ
• 41130 Tweets
Asnawi
• 37624 Tweets
オウンゴール
• 30728 Tweets
最終予選
• 28076 Tweets
ミンギュ
• 27182 Tweets
誹謗中傷
• 26601 Tweets
Zaplana
• 26040 Tweets
ジョシュア
• 20624 Tweets
ドギョム
• 20075 Tweets
スングァン
• 19643 Tweets
نعيم قاسم
• 19274 Tweets
Witan
• 17482 Tweets
كوريا
• 16619 Tweets
Arhan
• 14969 Tweets
中村敬斗
• 12469 Tweets
Tuchel
• 11549 Tweets
報道ステーション
• 10833 Tweets
Pinned Tweet
Out now open access at
@AJPS_Editor
. 194 potential exclusion-restriction violations for studies using weather as an instrumental variable
14
192
654
Everyone's focusing on the gender breakdown but I really admire the chutzpah of the 2% of pensioners who think they could win a point off Serena
31
248
2K
Update to my weather IV paper. IV assumes effect of instrument on Y works only through instrumented variable X. But scientists have linked weather to 192 variables, which makes exclusion restriction assumption implausible. This plot is just for rainfall:
17
188
908
Just updated my weather IV working paper:
tldr: all IV weather papers represent potential exclusion restriction violations for all other IV weather papers. I find 1̵3̵7̵ 176 possible exclusion restriction violations
8
151
569
Just put up a research note on using weather as an instrumental variable.
tldr: all IV weather papers represent exclusion restriction violations for all other IV weather papers. I find 137 possible exclusion restriction violations
23
80
389
New working paper! What drives electoral change?
Do election results change because of party switching or turnout?
Party switching is the most important factor in 96% of elections (across 103 individual-level panel surveys in 18 countries)
/thread
10
123
283
Situations when another study using the same instrument (W) as you *doesn't* imply an exclusion restriction violation
7
37
270
Final seat counts against exit poll (in brackets)
Con 318 (316)
Lab 262 (265)
LD 12 (13)
SNP 35 (34)
UKIP 0 (0)
PC 4 (3)
Grn 1 (1)
19
212
255
New paper on using AI to code survey responses!
@PoliSciJack
@ralphascott
@Marta_Miori
@jamesbreckwoldt
and I show chatGPT/GPT-3 can achieve near-human performance on coding “most important issue”
@besresearch
survey responses with zero training
7
48
265
It's too early to quantify the size of any polling misses in the US, but a lot of research has consistently found no evidence for shy voters and lots of evidence for unrepresentative samples and weighting problems
review article with
@caprosser
here
12
49
207
Alternative: Excel is a great piece of software that's *so* good that people keep using it for things it's woefully unsuited for.
6
14
193
I'm delighted to report that I have been promoted to senior lecturer at Manchester
17
2
188
@stochastician
Alternate theory: Satoshi Nakamoto is a time-travelling singularity AI from 2038 who created bitcoin to increase demand for GPUs and create a large decentralized GPU network that could be highjacked by a breakout AI
1
37
169
New draft of my weather IV paper with exciting new exclusion restriction violations including UFO sightings
3
47
175
Vote switching from 2015 to the end of 2016 in
@BESResearch
among remain and leave voters with
@caprosser
13
176
162
Just found out that I’m being promoted to Associate Professor with tenure! Huge thanks to everyone in the Department of Systems Engineering at West Point and to all my mentors throughout my career for helping make it happen!
16
0
111
Assuming St Ives goes Conservative, the final result versus our exit poll prediction is (predicted/real)
Con 368/365
Lab 191/203
LD 13/11
SNP 55/48
PC 3/4
GRN 1/1
OTH 19/18
@StephenDFisher
@robfordmancs
@PME_Politics
@whatukthinks
Jouni Kuha
10
30
86
Replication data/code for our paper on using AI to code open-ended survey responses is available here
Includes LLM prompts, API calls, and code to run local supervised language models
2
20
86
@JohnHolbein1
I do a google sheet like this. (current coauthor responsibility for each project redacted). Target date is when I aim to have a substantive next step taken on a project (i.e. a new draft or round of analysis).
5
4
82
Heading to ASA to present a paper on what determines social status in Britain using a new conjoint experiment that had people 30,720 status judgements (screenshot below)
6
21
83
Just published a paper in
@Res_Pol
showing that LLMs can code “most important issue” survey responses nearly as well as humans and better than SOTA supervised methods.
1
31
81
Class voting has shifted quite dramatically in 2019, with the working class more supportive of the Conservatives than the middle class (possibly for the first time ever)
6
67
70
Just published a blogpost questioning the uncritical coverage of Google Trends analysis in the news
tldr: google trends does not validate against reliable sources most of the time, so you should be sceptical about most claims made using it
2
35
64
Farage says he has spent 10 years trying to make immigration synonymous with EU membership. Our evidence suggests it worked
9
74
58
Life update: I’m joining West Point’s Department of Systems Engineering as an Assistant Professor at the end of June! I’ll be continuing my political behavior research and work as BES co-Director as well as starting some new projects in statistics and policy.
12
0
57
Update: in 2019, it’s neck and neck but plural data is still just leading singular data.
12
21
56
@bryan_caplan
The great shutdown is an indirect effect of coronavirus but the coronavirus is not an indirect effect of the great shutdown. So clearly coronavirus has to be worse even if it only gave 1 person the sniffles.
5
0
49
Summary:
1) lots of potential exclusion restriction violations for weather
2) those violations are strong enough to overturn many results
3) sensitivity analysis should be used systematically in IV analysis to examine how believable IV assumptions are for your models
0
4
45
there are possibilities other than exclusion-restriction violations as I show here:
2
4
48
Just made a YouTube video asking: Do Politicians Keep Their Promises?
5
14
43
Early analysis of social status out in UKICE report with
@PoliSciJack
& Geoff Evans. Other measures of social status rely on proxies such as friendship patterns. We use a conjoint experiment to directly ask people to make social status judgments
5
12
44
unfortunately, I find that the existing IV literature suggests that it looks closer to this:
1
3
43
I tried out the new bing chatbot and it’s just as weird as everyone said. I tried asking it about papers I wrote. It made up some papers and then accused me of impersonating myself after I corrected it
5
9
43
People pointed out that it wasn’t always fair to assume every study was relevant to every other study. I’ve therefore also split the DAGs up by income level, weather type and whether the study is looking at short or long-term weather.
1
4
42
Heading into exit poll silence with Curtice,
@StephenDFisher
@robfordmancs
Jouni Kuha,
@PME_Politics
and
@robfordmancs
. See everyone at 10pm
4
9
40
We find very strong correlations between mentions of immigration in tabloids (particularly the Daily Mail r=0.97) and immigration concern. The lines are essentially identical.
2
34
38
People like to complain about peer review but answering a reviewer comment was what led to this piece of the analysis that adds quite a bit to the paper's contribution on understanding how different studies' claims affect each other
there are possibilities other than exclusion-restriction violations as I show here:
2
4
48
2
2
40
Just put together this summary of vote recall accuracy (one election later) across various studies (it's pretty bad).
4
10
39
Con manifesto launching tomorrow. Our research shows they fulfilled less than half of the major promises in their last manifesto
📣 Prime Minister
@BorisJohnson
will launch our 2019 manifesto tomorrow 📣
It’s time to get Brexit done and move forward with our ambitious and exciting domestic agenda to make life better for everyone across the UK 🇬🇧
🗳
#VoteConservative
✔️
61
49
171
6
35
37
@sam_atis
Mostly yes, although the stronger the first stage and the more specific the instrument the more I tend to trust the result. Leah Stokes has a paper using wind speed as an instrument for building wind farms in Ontario. I think that paper is still pretty convincing
2
2
37
PSA: Make your stata plots nice forever by running
ssc install blindschemes, replace all
set scheme plotplainblind, permanently
@DanBischof
4
7
37
What’s the best 1-2 papers to read to get me up to speed with what’s best practice for diff-in-diff (and the whole TWFE bun fight that’s going on)
8
3
36
@aecoppock
Generalizing from this: every study using rainfall as IV for a different independent variable represents an enormous excludability violation for all the other ones
2
10
34
My article on immigration, Europe and UKIP with Geoff Evans is out in Party Politics /thread
2
10
31
New working with
@caprosser
disputing recent APSR claim that mass polarization is driven by declining survey cooperation rates.
1
9
31
Was a blast doing the ITV election analysis with
@ProfJaneGreen
tonight. Probably the weirdest election night I've analyzed.
3
1
30
Not all violations will necessarily undermine a specific claim either. So I use
@analisereal
's sensitivity analysis approach to assess how strongly related an exclusion variable and DV would have to be to cause problems
1
4
30
Can AI code open-text survey responses? We find the answer is: it depends which one. Here’s the comparison of large language models, supervised approaches (trained on 1,000 cases), and a human for two of our tests (link: )
2
12
29
1: New paper on how turnout measurement is very broken. Let’s start with registered voter turnout in the UK, US and Sweden (left) compared with corrected versions (right). w/
@caprosser
@ProfJaneGreen
@ed_fieldhouse
Geoff Evans (thread)
3
16
25
One of the most common statistical errors I see is not including random slopes when modeling cross-level interactions.
I explain this problem and show it doesn’t apply to cluster robust standard errors in a new blog
1
6
29
In some cases, the Z~Y relationship would have to be huge to overturn a finding. But in many cases, plausible Z variables could explain as little as 0.001% of variance in Y and still overturn the result.
1
0
28
Bourdieu: elites use arbitrary cultural signifiers to gatekeep social spaces and access to opportunities, reproducing privilege and status
Sociologists: we will only take you seriously in the field if you can correctly perform an understanding of Bourdieu
0
0
28
It’s really neat to see an out of sample test of the analysis Geoff Evans and I did. Really interesting that the tabloid → public concern link remained strong even as the immigration → tabloid coverage link broke
There is some cause for tentative optimism, though.
Here’s that same chart updated for 2022: tabloid coverage of immigration has ticked up again, but as of yet public concern is not following its lead.
7
78
505
1
4
29
@GaribaldoValdez
The whole point of researchers teaching (as opposed to just having dedicated teaching faculty) is that they are passing on information they get from working at the frontier of knowledge in that area. I think teaching your own research is an important part of that
0
0
27
This was also my first time doing computational reproducibility at a journal. The paper reproduced the first time, but in prepping the documentation I realized a small missing data issue that I would otherwise have missed, so definitely see how these checks will improve things
Out now open access at
@AJPS_Editor
. 194 potential exclusion-restriction violations for studies using weather as an instrumental variable
14
192
654
2
2
27
Really excited to be working on this tonight for ITV with
@ProfJaneGreen
!
‘Trump vs Biden: The Results’ this Tuesday w/
@tombradby
. I really can’t overstate the importance of this election. I’ll be in the London studio analysing the results with
@jon_mellon
in the US. Big week for us all. Stay safe.
3
6
60
0
2
26
New blogpost up summarizing the evidence around the Brexit party, UKIP and whether they are a threat to Lab or Con
2
26
26
Interested in what's went wrong with UK/US polls?
@caprosser
and I review (nearly) everything written about it in a new
@govandopp
article.
TLDR: weighting and representativeness matter more than lying voters
Link:
Open access:
4
9
25
@BrendanNyhan
It’s almost like this whole identification strategy can generate arbitrarily large effects for any independent variable
3
2
25
Overall evidence is that polls haven't got worse over time
We also looked at whether falling response rates have made polls worse. The evidence doesn't seem to be there for that either
3
11
26
Very early draft on authoritarianism up w/
@caprosser
.Interested in thoughts(presenting at
#APSA2017
#EPOP2017
) 1/15
5
8
23
Congratulations to my amazing wife on her dissertation defense!
Passed my dissertation defense! Since it’s official, I’m excited to also share that I’ll be starting a postdoc
@PSTC_Brown
this Fall! Thanks so much again to my committee and to everyone who supported me through the PhD journey!!
10
3
64
1
1
25
New article with
@caprosser
and
@PoliSciJack
using
@BESResearch
. We find COVID politics fit into traditional economic left-right competition rather than liberal-authoritarianism which structured fights over Brexit and immigration
OA link:
1
12
24
Final
#GE2015
forecast:
Con 274
Lab 278
LD: 23
UKIP: 0
SNP: 52
PC: 2
Oth: 1
with
@ed_fieldhouse
18
64
24
Looking forward to talking about whether electoral change is driven by party switching or turnout change at APSA on Thursday morning at
#APSA2024
2
2
25
I'll be giving a webinar for the UK data service tomorrow (3pm Wednesday April 1st) about
@BESResearch
data resources.
@nspmartin
will be discussing Understanding Society. Sign up here:
2
16
22
Without running it, about what value will the following R code produce?
mean(abs(rnorm(2000)))
0.8
298
1
320
1.2
121
1.65
77
3
3
24
@waitbutwhy
Fairly sure this is wrong. If you assume 30 yr generations 50 generations takes us back 1500 years. There have been populations that haven’t had a common ancestor with Europeans for much longer than that and there hasn’t been enough time for mixing the whole populations since
1
0
24
All of them! But here's some examples from my work and others
2
11
23
The goal of this causal platform is to collaboratively build a single shared causal graph for social scientists based on claims in the literature. Get in touch if you're interested in helping to test it out!
Working paper on how radical right parties affect military spending w/ Jordan Becker (short answer: they don’t). It was a great opportunity to test out a causal network platform I’m building.
2
2
16
1
3
23
@BrendanNyhan
I do appreciate being able to add UFO sightings to the next version of this paper though
0
1
21
My research with
@caprosser
showed overengaged samples caused 2015 polling miss, I am sceptical about polls where that's clearly not fixed
4
16
23
@MarkFabian_PAIS
Most statistical routines have a bunch of defaults (some of which are editable by the user and some not). Saying what software you used is specifying what those choices likely are
1
0
23
@analisereal
Unfortunately, the answer is often not very strongly related, so a lot of weather IV studies are likely to have serious exclusion-restriction issues with their results
2
1
23
Important finding in our youthquake paper. It's become increasingly important how we deal with non-eligible residents in weighting targets. The eligible voting population is less educated and older than the UK population
2
12
23
Sure glad we got to the bottom of this before there were any political consequences...
0
15
21
New paper on timing of 2015-2017 vote switching. Pre-campaign switching structured around Brexit but campaign saw general rise in Lab support w/ Geoff Evans,
@ed_fieldhouse
,
@ProfJaneGreen
@caprosser
@BESResearch
1
17
22
Just published Twitter and Facebook are not representative of the general population with
@caprosser
open access
1
26
22
I’ve been asked for advice on writing effective prompts for this type of task a few times.
1: run your prompt on a few examples to see where it fails. Edit the prompt to eliminate these errors in the most general way possible (ie always use lower case for the labels)
Just published a paper in
@Res_Pol
showing that LLMs can code “most important issue” survey responses nearly as well as humans and better than SOTA supervised methods.
1
31
81
3
4
21
Among many other goodies in here, we already have coded most important issue data thanks to integrating the LLM method we validated here into our pipeline
🚨NEW DATA🚨
The British Election Study team are pleased to announce the release of the 2024 Election Provisional Panel Dataset.
This abridged beta version includes data from May 2024 and the new constituency boundaries.
You can find out more here:
3
27
49
1
5
21
Can the next trend in polisci be including data with reviews by default? I've had the situation multiple times where I find a possible problem with a paper, and without the data it's hard to know if it's a devastating problem or a minor variation.
2
4
21
Testing out the next round of our social status experiment. Person 2 is fascinating.
2
0
21
Not only is party switching the biggest factor in electoral change overall, it’s the most important factor on average in every single one of the 18 countries:
1
5
20
Ideally, the causal structure for an IV estimate should look like this:
1
1
21
What's a short intuitive example of conditioning on a collider? Some reviewers seem confused when I mention it and it would be nice to have an example that is really obvious.
9
2
21
@SolomonKurz
@ajordannafa
social science overly focuses on existence of effects rather than quantifying their magnitude. But you’d probably explain more about the world by better understanding these causal effects and their variability than proving a dozen tiny effects are non-zero
2
7
20
Come read our analysis of vote switching between 2015 and 2017. Was it driven by Brexit or Corbyn? Free to read!
Was the 2017
#election
a ‘Brexit election’ or a ‘Corbyn election,’ or something else entirely?
@jon_mellon
@ed_fieldhouse
@ProfJaneGreen
@caprosser
@HansardSociety
0
13
14
0
10
19
One indication Brexit is a truly cross cutting issue is that Lab and Con voters were more likely to switch to the other major party than at previous elections
3
21
18
The behavior of don’t knows is a legitimate concern but weighting after excluding them is definitely a bad idea. Weighting is about approximating a representative sample, conditioning messily on your DV is making that a lot harder
Could Conservative "Don't Knows" be causing the next big polling error?
Looking at
@StackStrat
data, the decision to weight a survey before vs after excluding "Don't Knows" is the difference between a 16% and 21% Labour lead:
12
27
74
5
4
18