Michael Jankowski
@michaelj505
Followers
1,110
Following
123
Media
59
Statuses
353
Team Lead Data Science @ Federal Chancellery | PhD in Political Science | Interested in Chess & Badminton
Berlin/Oldenburg
Joined February 2017
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
FEMA
• 1237577 Tweets
Liz Cheney
• 242214 Tweets
#LISAxMoonlitFloor
• 236539 Tweets
MOONLIT FLOOR OUT NOW
• 174165 Tweets
Mets
• 134863 Tweets
Bruce
• 98308 Tweets
Happy Anniversary
• 66384 Tweets
Baker
• 55874 Tweets
Brewers
• 53481 Tweets
Pete Alonso
• 36989 Tweets
天使の日
• 36840 Tweets
Falcons
• 33621 Tweets
Saint Dr MSG
• 31991 Tweets
Mancuso
• 31062 Tweets
Halle
• 27614 Tweets
Reds
• 24008 Tweets
#BakeOffFamosos
• 21300 Tweets
Suárez
• 21119 Tweets
Phillies
• 19181 Tweets
gaston
• 18583 Tweets
Bucs
• 18219 Tweets
Bijan
• 16720 Tweets
SAFE FLIGHT JESBIBLE
• 16651 Tweets
Tito
• 16341 Tweets
Mike Evans
• 15211 Tweets
Mooney
• 13887 Tweets
#ゴンチャのハロウィン準備中
• 12614 Tweets
Milwaukee
• 11938 Tweets
#BakeOffArgentina
• 11897 Tweets
Spiritual Awakening
• 10540 Tweets
Quintana
• 10424 Tweets
Pinned Tweet
New Working Paper with
@JochenRehmert
:
Does gender affect candidate selection and/or list placement in PR systems?
To address this question, we conducted an experiment with party elites in Austria.
Link to paper:
3
15
51
Aus welchen Gründen unterstützen Leute die Petition von Schwarzer und Wagenknecht?
Um das herauszufinden habe ich mir die Kommentare bei der Petition systematisch angeschaut.
118
557
3K
Ich nutze dies, um zu überprüfen, worin sich die Texte von männlichen und weiblichen Unterzeichner unterscheiden. Die Grafik zeigt, welche Begriffe eher von Männern (blau) bzw. eher von Frauen (grau) verwendet werden.
4
23
553
Personen können, wenn sie Petition unterschreiben, auch einen Text schreiben, weshalb sie die Petition unterstützen. Ich habe all diese Kommentare in den ersten drei Tagen der Petition gesammelt. Insgesamt sind es 24.888 Kommentare.
4
6
516
Außerdem geben viele ihren Wohnort an. Unter den 30 am häufigsten genannten erscheinen mir ostdeutsche Städte deutlich überrepräsentiert.
22
10
502
Alle Analysen sind vorläufig und gerade die Befunde bei Topic Models können sich ändern, wenn man an den Modelparametern rumspielt.
16
0
450
Männer scheinen häufiger über die USA/NATO/"den Westen" etc. zu sprechen, während bei Kommentaren von Frauen eher der Wunsch nach Frieden im Vordergrund steht.
2
4
432
Über die Kommentator:innen sind nur wenig Informationen verfügbar. Anhand des Namens lässt sich automatisiert für die meisten das Geschlecht schätzen. Basierend auf dieser Methoden komme ich zu dem Schluss, dass mehr Frauen als Männer unterzeichnet haben.
4
2
425
Die 200 am häufigsten verwendeten Wörter als Wordcloud.
3
4
417
Unter den am häufigsten verwendeten Wörtern sind, wenig überraschend, "krieg" und "frieden". An 27. Stelle findet sich aber auch "USA".
4
3
408
Die meisten Kommentare sind, wenig überraschend, recht kurz und nur wenige Wörter lang (Hinweis: x-Achse ist log10 transformiert). Der Median der Tokenanzahl liegt bei 13.
3
1
389
Anhand von statistischen Verfahren (in diesem Fall: "Keyness-Analyse") lässt sich analysieren, welche Wörter von einer Gruppe im Vergleich zu einer anderen Gruppe besonders häufig verwendet werden.
1
0
377
Die Ergebnisse legen nahe, dass sich insgesamt vier Hauptthemen unterscheiden lassen. (Hinweis: die Interpretation von Topic Models ist immer etwas subjektiv).
Dies sind:
1) Eskalation/Weltkrieg
2) USA etc.
3) Verzweiflung über das Leid der Menschen
4) Glaube an Diplomatie
2
5
372
Um einen Eindruck davon zu bekommen wie viele Kommentare einen Bezug zur USA o.Ä. herstellen, habe ich alle Kommentare identifiziert, die entweder "usa", "nato" oder ähnliche Begriffe enthalten. Die Wordcloud zeigt alle identifizierten Begriffe.
2
3
368
In 1648 Kommentaren (=6.6%) kommt mind. 1 dieser Wörter vor. Unterscheidet man hier nach Geschlecht, bestätigt sich der Befund der Keyness-Analyse. Bei Frauen liegt der Anteil nur bei 3.7%, bei Männern über 10%.
1
2
349
Das Topic "Leid" macht dabei den größten Teil der Texte aus, gefolgt von Diplomatie, USA/NATO und dann der Eskalationsangst.
2
2
338
Auch hier kann man schauen, welche Themen häufiger bzw. seltener von Männern/Frauen angesprochen werden. Erneut bestätigt sich der Befund, dass Männer häufiger über USA/NATO reden, Frauen dafür mehr über die Themen "Leid" bzw. "Diplomatie". Bei Eskalation keine Unterschiede.
3
3
334
Um die gesamten Inhalte der Kommentare besser analysieren zu können, habe ich noch ein Topic Model (konkret: keyATM) geschätzt. Dieses identifiziert automatisiert Themen in den Texten.
1
0
328
Is a study robust when its findings depend on the inclusion of a single observation?
In a new report,
@Ben_Guinaudeau
and I discuss this question based on the replication of an
@apsrjournal
article. A thread on our main findings.
We have a new reproduction and response by the authors! The original article is entitled "When Do Männerparteien Elect Women? Radical Right Populist Parties and Strategic Descriptive Representation" (by Weeks et al.) The reproducers are Guinaudeau and Jankowski.
1
10
60
2
54
198
Career update: Starting in July, I will be working in the data lab of the German Chancellery.
6
5
105
Yesterday,
@Ben_Guinaudeau
provided some very interesting analyses of
#EPSA2023
's geographic composition. Here are some additional analyses regarding the gender composition.
3
20
91
@Gorki45930058
Eindeutig feststellen lässt sich das nicht. Aber ich werde in Zukunft versuchen zu schauen, ob sich Hinweise auf Bots o.Ä. finden lassen.
6
2
71
The APSR desk rejected the letter as “it primarily serves to correct the record by identifying issues that call into question the reliability of the findings […and…] does not significantly advance scholarly debate about the substantive or methodological issues at stake”.
5
9
76
Wenn man im text-as-data Seminar nur so halb zugehört hat...
Im Koalitionsvertrag der Ampel steht häufiger „Klima“ als „Deutschland“
465
109
574
1
4
75
In a new working paper (),
@Robert_A_Huber
and I discuss a recently proposed method for deriving populism scores from texts. We argue that a more comprehensive validation raises concerns about the validity of the suggested approach.
4
18
73
Candidates from the CDU/CSU changed their position toward same-sex marriage quite strongly in '21. Analysis based on GLES Candidate Surveys from '13/'17/'21. Plot shows predicted probabilities of disagreement with the statement "Same-sex marriages should be prohibited by law".
2
9
56
Das Ergebnis ist recht eindeutig. Die Abbildung zeigt die 30 Städte mit den meisten Kommentare pro 100.000 Einwohnern. Berlin habe ich mal zum Westen gezählt, damit wenigstens etwas farbliche Variation entsteht. Auf Platz 33 käme Stuttgart.
9
7
55
@iriswasapupil
Dachte eigentlich, dass durch die transparente Benennung von bspw. 3% vs. 10% oder auch der Ausgabe der Topic Proportions klar wird, wie die Verteilung der Themen ist. Die Keyness-Analyse gibt in der Tat nur relative Häufigkeitsunterschiede wieder.
1
0
43
Writing replications which contradict the original findings is no fun task, but it is nonetheless an essential part of academic publishing, especially when the topic matters that much.
1
2
40
Here is our full replication report:
The response of the authors:
4
1
34
In the thread below, one of the authors summarizes their arguments against our replication. A few important comments on their points from me and
@Ben_Guinaudeau
.
For anyone following the debate about our article “When Do Männerparteien Elect Women?” here is a brief thread summarizing our response to the comment by Guinaudeau and Jankowski. I encourage those interested to read it in full here:
2
43
138
3
10
32
Arguing that small-n studies are known to be sensitive to outliers does not convince me. Neither the authors nor the papers citing the study discuss the limitations arising from small-n studies. The severe sensitivity of the results is not publicly known and deserve awareness.
1
3
31
I must admit that APSR’s arguments were not very convincing. To provide code that accurately reproduces the article’s results is a good signal, but it is far from representing a reasonable benchmark when assessing the robustness of a paper.
1
0
30
I turned this analysis into a very short paper ("Gender Update") which is now published in the
@EJPGjournal
. You can find the article here:
Candidates from the CDU/CSU changed their position toward same-sex marriage quite strongly in '21. Analysis based on GLES Candidate Surveys from '13/'17/'21. Plot shows predicted probabilities of disagreement with the statement "Same-sex marriages should be prohibited by law".
2
9
56
1
8
30
However, the APSR started an investigation which lasted until December ’22 and concluded that no corrigendum is necessary. This decision was based on an undisclosed response from the authors and a single reviewer report, summarized by the APSR as described in the picture below.
2
1
30
It was a pleasure to collaborate with
@Ben_Guinaudeau
on the project, and we are very grateful to the Institute for Replication (
@I4Replication
) for providing a platform structuring these difficult and nonetheless important exchanges between authors and replicators.
1
0
30
Great guest lecture by
@Robert_A_Huber
on "populism, climate change, and covid-19" in my seminar on "current challenges to liberal democracies".
1
3
24
Die Aussage bezog sich sich natürlich auf die Einwohnerzahl. Es befinden sich recht viele ostdeutsche Städte unter den 30 häufigsten "kommentierenden Städten", die aber nicht zu den 30 einwohnerreichsten Städten gehören.
1
1
23
The paper is now forthcoming at the EJPR
New Working Paper with
@JochenRehmert
:
Does gender affect candidate selection and/or list placement in PR systems?
To address this question, we conducted an experiment with party elites in Austria.
Link to paper:
3
15
51
1
4
22
Um das systematischer zu zeigen, habe ich nun die Einwohnerzahlen zu den Städten hinzugefügt und dann die Anzahl der Kommentare pro 100.000 Einwohner ausgerechnet. Ich betrachte nur Städte aus denen mind. 20 Kommentare kamen.
1
0
22
Overall, this has been a very interesting but also challenging experience for me, and I would like to thank all colleagues and friends who supported me.
3
0
21
Ergänzung zu den Städten: Es gab dort die Nachfrage, warum ich ostdeutsche Städte als "überrepräsentiert" bezeichne, wenn doch unter den 30 häufigsten Städten mehr westdeutsche als ostdeutsche Städten zu finden sind.
1
0
21
Last day to submit your paper/panel to
@christinajuen
,
@Robert_A_Huber
, and my section on "Contemporary debates on populism and polarised political behavior" at the
@ECPR
General Conference. We are looking forward to your submissions.
1
6
20
We then evaluated how the results are affected by excluding this observation. We ran the regression models without it and display how the interaction effect changes compared to the published result.
1
2
17
Although the Front National may play a significant role in analyzing populism, relying on the outcome of a single candidate’s election hardly constitutes a robust conclusion. I submitted a research note describing these issues first to the authors and later to the APSR in May '22
1
0
17
The effect is gone. In the most favorable model specification, the coefficient of the interaction effect changes from -0.87 (p < 0.01) in the published paper to -0.15 (p = 0.71) in the replication. In other specifications, the interaction equals zero with a p-value > 0.99.
2
1
15
Based on the country of the presenters' home institution, we find that Israel and Swedish universities had the highest share of women (>50%) presenting at
#EPSA2023
.
1
3
15
📢 Call for applications: W1 professorship with tenure track (to W2) for Digital Social Science at the University of Oldenburg: 👉
0
15
15
@ryancbriggs
A top journal recently argued that a paper (which i criticized for a lack of robustness) is robust because the replication code runs without errors...
1
0
15
An in-depth analysis of the case, thus, reveals that the observation of Front National in 2012 is not particularly suitable to support the theoretical mechanism: Essentially, the main result of the paper is contingent on the election Marechal-Le Pen.
1
0
13
If Marechal-Le Pen was not elected in 2012, which would have been likely if the PS candidate withdrew from the second round, the value of the dependent variable would drop from 50% to 0% which radically changes the results.
1
0
13
While technically correct, it is important to note that Front National in 2012 only elected two legislators, of which one was female (Marion Maréchal-Le Pen, the niece of Marine Le Pen). Front National did not elect more women; they elected a single woman.
2
0
13
Based on the institutions, Frankfurt/Main had the highest share of female presenters.
1
1
13
However, the plot above is based on an interaction effect estimated from a sample of 58 cases. Small sample sizes can be challenging for empirical analysis as they have low power and results are more likely to be unduly influenced by individual outliers.
1
0
12
This is quite a drastic change in the regression estimates and it strongly affects the conclusions drawn from the statistical analysis. The inclusion of a single observation has a massive effect on whether we reject the null hypothesis of no effect.
1
0
11
@mlewandowsky
"von anderen Parteien enttäuscht" und "von der AfD überzeugt" schließt sich auch nicht gegenseitig aus, wie die Frage/Abbildung suggeriert.
1
0
11
Die Daten des OES21 können hier abgerufen werden:
Neuer Beitrag in der PVS: Jankowski et al. berichten Ergebnisse des „Open Expert Surveys“ zur Bundestagswahl 2021. Wo verorten über 300 Politikwissenschaftler:innen die Parteien und welche Veränderungen gab es zu 2017? Den Beitrag gibt es
#openaccess
hier:
0
7
32
0
2
11
Finally, let's have a look at the number of all-male ("manels") and all-female ("fanels") panels. There were 9.2% manels and 2.4% fanels. (We only analyzed panels with at least 3 papers, N = 251; gender of chair & discussant excluded!).
1
0
11
This is a very fundamental logic of data analysis and using increasingly complex models to counter this point does not really help.
0
1
10
In 2013 and especially in 2017, left-right self-placement of CDU/CSU candidates explained their positions toward same-sex marriage. In 2021, there is no effect.
1
0
10
In the case of WMKC’s analysis, it is even more extreme as the finding does not only depend on the one observation, but on the election of a single person.
1
1
10
Next, we can look at the share of women in each section (we estimate the share of women among all submissions). In the Gender & Diversity section, the share of women is >50%. IR/Conflict is close to 50%. The share of women is very low in the methods and formal modeling sections.
1
0
10
This raises the question: Why does the Front National in 2012 have such a strong effect? This question is less trivial than it seems.
1
0
10
Overall, our paper suggests that deriving populism scores from texts remains an important challenge for political science research and that supervised machine-learning models often need to be broader validated.
0
1
9
We can further analyze the share of women among the paper presenters. In the "Gender & Diversity" section, the share of women is 70% among the presenters. In the Formal Models section, only a single paper was presented by a woman.
1
0
9
In our replication, we applied standard regression diagnostics to the regression models, including estimating Cook’s D, which measures how strongly each observation influences the regression coefficients.
1
0
9
@mlewandowsky
Trotzdem erstaunlich wie schlecht Laschet dabei abschneidet (auch im Vgl. zu Scholz). Die Frage, ob Rauchen ungesund sei ist nun wirklich nicht "ultrahart". Da mit "Nicht auf Lunge" zu antworten anstatt zu sagen "Ja, ich rauche und das ist doof. Fang bloß nicht damit an." 🤷♂️
1
0
8
One might argue that Front National in 2012 needs to be included because it is such an informative case for the analysis. There is no doubt that Front National is a very prominent example of a PRRP in Europe. But what makes this observation so impactful?
1
0
8
A blog post summarizing
@Robert_A_Huber
,
@christinajuen
and my recent
@EJPRjournal
paper on the relationship betweem thick ideologies, salience and populism.
🔮
#FutureOfPopulism
No.9
You might assume
#populism
to be highest among Eurosceptic, anti-immigration parties. But
@Robert_A_Huber
@michaelj505
&
@christinajuen
argue that the crucial factors are collectivist positions on parties’ salient issue dimension
0
7
8
0
3
7
At first glance, the observation is a perfect example for the theoretical mechanism. Front National in 2012 had a large lagged male-to-female voter ratio, which is used for measuring a predominance of male voters in the previous election.
1
0
8
Therefore, the values of Front National in 2012 in the independent and dependent variables perfectly align with WMKC’s hypothesis. The problem is, however, that the value of 50% in the dependent variable is somewhat misleading.
1
0
8
Using automated methods, we have estimated the gender of all authors based on their first names. Of course, this is a simplification and only considers two genders. These limitations should be kept in mind when interpreting the results.
1
0
8
We can also estimate the share of women conditional on the number of authors of a submission. Quite interestingly, the share of women is highest among solo-authored papers.
1
0
8
That's it. Let us know if you have any additional questions.
3
0
7
The circumstances under which Maréchal-Le Pen was elected also question the relevance of the case. Maréchal-Le Pen was only elected by a relative majority, because the PS candidate refused to withdraw from the race in the second round, against the advice of the party leadership.
1
0
6
This analysis reveals that one observation, the Front National (now Rassemblement National) in the French Parliamentary Election of 2012, has a very strong effect on the regression coefficients.
1
0
6
Moreover, Front National also had a very large value in the dependent variable (share of elected female candidates). In fact, with 50% of the elected candidates being female, it has the highest value of all 58 cases.
1
0
6
An inspection of the most important features identified by the supervised machine-learning models suggests that the models strongly rely on party names to predict populism.
1
0
5
@Raffzahn
Ja, das stimmt. Wenn man nur Städte ab 100.000 Einwohnern betrachtet sieht es so aus... Also auch recht eindeutig.
1
0
5
Front National in 2012 also lost votes in the previous election. Thus, the party was in the exact situation for which WMKC’s argument expects an increase in the number of elected female politicians.
1
0
5
These claims are misleading. Identifying influential individual observations is not wrong in multilevel models. If a single observation influences the results, the model is not robust. This holds true for both individual observations and clusters in multilevel models.
1
0
5
We estimate the share of women among (1) all submissions, (2) all unique authors, and (3) all presenters. In all cases, the share of women is between 35% and 40%.
1
0
5
We also find that the effect is conditioned by a respondent's gender. Female party elites are particularly like to select female candidates.
1
0
4
@DebusMarc
@pluggedchris
@JocMuel
Interesting! There seems to be no difference between CDU and CSU candidates.
2
0
4
WMKC show a significant and substantial interaction effect which supports their hypothesis. The figure below depicts the main finding from the APSR paper.
1
0
4
Thus, both models remain dominated by this observation. Robust regression models can be useful for addressing influential observations, but here they fail. Even with robust regression, the 2012 Front National still dominates the results.
1
0
4
Lastly, a general point: I started my tweet about our replication with a question: can a study be considered robust when results depend on a single observation? It is obviously rhetorical. When an analysis depends fully on a single observation, it is not robust.
1
0
4
Yet, even if we accept this model confusion, WMKC only shows the re-estimated results for Model 2. The results do not hold once the model-5-influential cases are excluded from any other model but model 2.
1
0
4
Under OLS, the Front National in 2012 has a hat-value of 0.45 (vs an average of 0.06), showing undue influence. In the robust regression, this value drops slightly to 0.40, still much higher than average.
1
0
4
We show that if you exclude all potentially influential observations from the respective model, none of the models provide evidence for their hypothesis.
1
0
4