Unveiling the Black Box

.title[
# Unveiling the Black Box
]
.subtitle[
## Researching Online Political Microtargeting
]
.author[
### <b>Fabio Votta</b> <i>(University of Amsterdam)</i>
]
.date[
###  <svg viewBox="0 0 512 512" style="height:1em;position:relative;display:inline-block;top:.1em;fill:black;" xmlns="http://www.w3.org/2000/svg"> <path d="M326.612 185.391c59.747 59.809 58.927 155.698.36 214.59-.11.12-.24.25-.36.37l-67.2 67.2c-59.27 59.27-155.699 59.262-214.96 0-59.27-59.26-59.27-155.7 0-214.96l37.106-37.106c9.84-9.84 26.786-3.3 27.294 10.606.648 17.722 3.826 35.527 9.69 52.721 1.986 5.822.567 12.262-3.783 16.612l-13.087 13.087c-28.026 28.026-28.905 73.66-1.155 101.96 28.024 28.579 74.086 28.749 102.325.51l67.2-67.19c28.191-28.191 28.073-73.757 0-101.83-3.701-3.694-7.429-6.564-10.341-8.569a16.037 16.037 0 0 1-6.947-12.606c-.396-10.567 3.348-21.456 11.698-29.806l21.054-21.055c5.521-5.521 14.182-6.199 20.584-1.731a152.482 152.482 0 0 1 20.522 17.197zM467.547 44.449c-59.261-59.262-155.69-59.27-214.96 0l-67.2 67.2c-.12.12-.25.25-.36.37-58.566 58.892-59.387 154.781.36 214.59a152.454 152.454 0 0 0 20.521 17.196c6.402 4.468 15.064 3.789 20.584-1.731l21.054-21.055c8.35-8.35 12.094-19.239 11.698-29.806a16.037 16.037 0 0 0-6.947-12.606c-2.912-2.005-6.64-4.875-10.341-8.569-28.073-28.073-28.191-73.639 0-101.83l67.2-67.19c28.239-28.239 74.3-28.069 102.325.51 27.75 28.3 26.872 73.934-1.155 101.96l-13.087 13.087c-4.35 4.35-5.769 10.79-3.783 16.612 5.864 17.194 9.042 34.999 9.69 52.721.509 13.906 17.454 20.446 27.294 10.606l37.106-37.106c59.271-59.259 59.271-155.699.001-214.959z"></path></svg> favstats.github.io/nefca2023 (Slides) <br>  <svg viewBox="0 0 512 512" style="height:1em;position:relative;display:inline-block;top:.1em;fill:blue;" xmlns="http://www.w3.org/2000/svg"> <path d="M459.37 151.716c.325 4.548.325 9.097.325 13.645 0 138.72-105.583 298.558-298.558 298.558-59.452 0-114.68-17.219-161.137-47.106 8.447.974 16.568 1.299 25.34 1.299 49.055 0 94.213-16.568 130.274-44.832-46.132-.975-84.792-31.188-98.112-72.772 6.498.974 12.995 1.624 19.818 1.624 9.421 0 18.843-1.3 27.614-3.573-48.081-9.747-84.143-51.98-84.143-102.985v-1.299c13.969 7.797 30.214 12.67 47.431 13.319-28.264-18.843-46.781-51.005-46.781-87.391 0-19.492 5.197-37.36 14.294-52.954 51.655 63.675 129.3 105.258 216.365 109.807-1.624-7.797-2.599-15.918-2.599-24.04 0-57.828 46.782-104.934 104.934-104.934 30.213 0 57.502 12.67 76.67 33.137 23.715-4.548 46.456-13.32 66.599-25.34-7.798 24.366-24.366 44.833-46.132 57.827 21.117-2.273 41.584-8.122 60.426-16.243-14.292 20.791-32.161 39.308-52.628 54.253z"></path></svg> favstats <br>  <svg viewBox="0 0 448 512" style="height:1em;position:relative;display:inline-block;top:.1em;fill:#615ff7;" xmlns="http://www.w3.org/2000/svg"> <path d="M433 179.11c0-97.2-63.71-125.7-63.71-125.7-62.52-28.7-228.56-28.4-290.48 0 0 0-63.72 28.5-63.72 125.7 0 115.7-6.6 259.4 105.63 289.1 40.51 10.7 75.32 13 103.33 11.4 50.81-2.8 79.32-18.1 79.32-18.1l-1.7-36.9s-36.31 11.4-77.12 10.1c-40.41-1.4-83-4.4-89.63-54a102.54 102.54 0 0 1-.9-13.9c85.63 20.9 158.65 9.1 178.75 6.7 56.12-6.7 105-41.3 111.23-72.9 9.8-49.8 9-121.5 9-121.5zm-75.12 125.2h-46.63v-114.2c0-49.7-64-51.6-64 6.9v62.5h-46.33V197c0-58.5-64-56.6-64-6.9v114.2H90.19c0-122.1-5.2-147.9 18.41-175 25.9-28.9 79.82-30.8 103.83 6.1l11.6 19.5 11.6-19.5c24.11-37.1 78.12-34.8 103.83-6.1 23.71 27.3 18.4 53 18.4 175z"></path></svg> <a href="mailto:favstats@fosstodon.org" class="email">favstats@fosstodon.org</a> <br>  <svg viewBox="0 0 448 512" style="height:1em;position:relative;display:inline-block;top:.1em;fill:#0085ff;" xmlns="http://www.w3.org/2000/svg"> <path d="M400 32H48C21.5 32 0 53.5 0 80v352c0 26.5 21.5 48 48 48h352c26.5 0 48-21.5 48-48V80c0-26.5-21.5-48-48-48z"></path></svg> favstats                                                                     21st September 2023 - Media Psychology Day NeFCA 2023
]

---

---

![](https://pbs.twimg.com/media/F4yspBJXwAEziim?format=png&name=900x900)

]

![](https://pbs.twimg.com/media/F4y52W0XwAAM4rp?format=jpg&name=large)

]

---

# Roadmap

1. Introduction to Political Microtargeting

2. Platform-centric methods

+ Toxic Microtargeting
  
      + I. APIs
    
      + II. Scraping
    
3. User-centric methods

+ Collaboration with Who Targets Me

+ III. Tracking users (with browser apps) 
    
  + Algorithm Audit Study with Dutch political parties
  
      + IV. Data donation approach
    
4. Q & A

---

# Slides: favstats.github.io/nefca2023

---

### Who do you think this ad is targeted at?

]

![](img/example1.png)

]

---

### Who do you think this ad is targeted at?

#### Men, 36-65+ year old, Noord-Holland, Zuid-Holland, Noord Brabant, and Gelderland, 11,191 people

*maybe: **excluded** people interested in veganism, green building*, ***included*** *college educated people interested in politics*

]

![](img/example1.png)

]

---

### Who do you think this ad is targeted at?

]

![](img/example2.png)

]

---

### Who do you think this ad is targeted at?

#### Women, 18-34 year old, Noord-Holland, Zuid-Holland, Noord Brabant, 1,975 people

]

![](img/example2.png)

]

---

## What is political microtargeting?

> a set of techniques to leverage ***individual-level data*** for the delivery of political messages to specific target groups that are expected to be more ***susceptible*** to them

What does more "susceptible" mean?

Expectation that it is more *effective*

Based on congruency between recipient and message:

+ **congruency** theory (Aaker 1999)

+ the main idea is:

+ people pay more attention when messages align with self-concept of receiver
  
  + when paying more attention they may have greater impact (Petty and Cacioppo 1986; Wheeler, DeMarree, and Petty 2008)

---

## (Some) Types of Microtargeting

+ *behavioral targeting* (Dobber et al. 2019)
  
  + e.g. people that engage with political content online
  
--
  
+ *psychographic* or *psychological targeting* (Tufekci 2014, Sharp 2018)

+ matching a person's personality (like their degree of extraversion) to the content of an ad can increase an ad's persuasive power and increase the clicks and conversions it generates (Moon 2002, Wheeler 2008)
  
  +  Nai and Maier (2020) show that uncivil attack ads are most effective in lowering perceptions of the attacked politician when the receiver of the message scores high on psychopathy
  
--

+ *issue-based targeting* (Endres 2020)

+ people who care about climate change, economy, abortion rights

---

### Effectiveness

+ In (field) experiments, tentative evidence suggests that microtargeted political ads are indeed effective (Dobber et al., 2020; Dobber
et al., 2023; Endres, 2020; Krotzek, 2019; Tappin et al., 2023; Zarouali et al., 2022).

+ messages about abortion rights were most **effective in increasing vote share of women in competitive congressional districts** (Haenschen, 2022)

+ Krotzek (2019) and Zarouali et al. (2022) find that political ads tailored to match personality traits **enhance positive feelings towards candidates and increase persuasiveness**

![](https://media.tenor.com/UniGtspR-BcAAAAC/damage-thats-a-lot-of-damage.gif)

]

however: see also Coppock et al., 2020; Decker and Krämer, 2023 for more mixed findings

]

---

### Offline vs. Online Political Microtargeting

#### Offline

![](img/offline.jpg)

]

#### Online

![](img/online.png)

]

---

### Online Political Microtargeting

![](img/online.png)

]

- Offline campaigning: limited volunteers & time

- Digital campaigns reach vast audiences quickly

<br>

**2. Accessibility**

- Microtargeting techniques available even for budget-limited campaigns

- Use of platform-provided data and own contact lists

]

---

### Not all platforms are the same for political microtargeting

.pull-left[
**Meta**  allows targeting based on:
  + detailed demographics (age, gender, education, etc.)
  + users "interests" 
  + behavioral targeting 
  + custom and lookalike audiences
  
(***Focus of this presentation***)

**Google** only allows limited targeting 
  + age, gender, location 
  + keywords for political ads
  
**Twitter** disallowed political ads in 2019
  + were brought them back recently (2023)
  + detailed targeting possible (demographics, custom and lookalike audiences)

]

.pull-right[
**Snapchat** offers detailed targeting 
  + demographics, custom and lookalike audiences
  + however is relatively understudied (e.g. Tanusondjaja 2023)

**TikTok** disallows political ads 
  + but researchers have documented its use: https://tiktok-audit.com/blog/2023/tiktok_political_ads/

]

---

## Studying Social Media in the *Post-API Age*

---

![](https://imageio.forbes.com/specials-images/imageserve/6097d7ee81957044af68d9ce/0x0.jpg?format=jpg&width=1200)

---

![](img/gYktqcsiAnhoXz74wM4rKC.jpg)

---

## Studying Social Media in the *Post-API Age*

I. Application Programming Interfaces (APIs)
  
  II. Scraping (Freelon, 2018)

2. user-centric

III. Tracking (via browser apps)
  
  IV. Data donation (asking users to share data)

]

![](img/ohme.png)

]

---

## Microtargeting and Toxic Ads

![](img/toxic.png)

---

### Going Negative

Negative campaigning (*going negative*) is a common strategy used by political campaigns

+ attack opponents 
  + highlight their personal flaws (Geer 2006).
  + point out their voting records

There is a *negativity bias* in human information processing that draws more attention to negative information (Fiske, 1980; Hilbig, 2009; Rozin & Royzman, 2001)

]

![](img/corbyn.png)

]

---

### Going negative may backfire

+ However, negative ads can also **backfire**, i.e. ***backlash effect*** (Garramone 1984; Walter & van der Eijk, 2019)

+ generate sympathy for the target
  
  
  + reduce attitudes towards the attacker

+ So when should campaigns "go negative"?

+ Weighting the rewards and risks (Haselmayer 2019)

]

]

---

## Microtargeting and Toxicity

Microtargeting may lower the risk of backlash effects because:

+ Toxic messages can be targeted towards individuals who are most susceptible to them (Nai & Maier 2020)

+ also means: people likely to dislike toxicity can be excluded

+ A smaller range of individuals are affected if there were to be a backlash

*Hypotheses:*

> H1a: Microtargeted ads from political campaigns are more toxic than ads targeted at more general audiences.

> H1b: Outside groups should be less concerned about reputational hits and therefore show a smaller effect.

---

## Methods

**Meta Ad Library API**

+ 912k ads in three months before election day (August 3rd - November 3rd 2020)
  
**Scrape Images and Videos**

+ `metatargetr`
  
  
**Extract Text from Images and Videos**

+ Transcripts of Videos using Mozilla DeepSpeech (Hannun et al., 2014)
  + OCR using Google Cloud Vision API
  
**Toxicity**

+ Scoring via Google's Perspective API

]

---

## Measurement

+ Independent Variable: Toxicity (Perspective API)

+ Dependent Variable: Potential Reach (*Targeting Granularity*)

> `Potential Reach` estimates how many people your ad could potentially reach depending on the targeting and ad placement options you select while creating an ad. ~Facebook Ad Library

.pull-left[
+ 100 - 1.000 audience size
+ 1.001 - 5.000
+ 5.001 - 10.000
+ 10.001 - 50.000
+ 50.001 - 100.000
+ 100.001 - 500.000
+ 500.001 - 1 million
+ +1 million
]

.pull-right[
  <br>
  <br>
  <br>
  The smaller the reach the more microtargeted is an advertisement.
]

Analysis: Multilevel Ordinal Logistic Regression (2nd level: advertiser pages)

---

## Results

![](img/toxres.png)

---

## Results

![](img/toxres1.png)

---

## Results

![](img/toxres2.png)

---

## Results

![](img/toxres3.png)

---

## Results

![](img/toxres4.png)

H1a *confirmed*: official campaigns microtarget toxic ads

H2a *mixed*: outside groups more likely to broadly target toxic ads

---

## Deep Dive into Methods

### Platform-centric approaches

#### I. APIs

---

### I. APIs

`\(\color{green}{\text{Upsides}}\)`

+ Official avenue to retrieve data
  
  + Documentation
  
  + Easier access
  
  + In theory: consistent data formats

+ Reproducibility

]

`\(\color{red}{\text{Downsides}}\)`

+ Reliance on platforms
  
  + May not have the right data
  
  + Rate limiting: don't allow you to retrieve data at scale
  
  + Potential Costs

]

---

### Meta Ad Library (API)

![](img/metaadlibrary.png)

]

---

### Meta Ad Library (API)

+ Meta Ad Library gives access to:

+ Ads about social issues, elections or politics that have run in the past seven years

+ You can get the text, run time (dates), spending, impressions (by age, gender, and location)
  
  + **NEW** since August 2023 *only in EU*: age, gender, and location targeting criteria (thank you DSA!)
  
+ In order to get access you need:

+ A **verified** Meta Developer account (i.e. you need to send in your ID to confirm identity 🤔)
  
  + The steps are outlined here: https://www.facebook.com/ads/library/api/
  
  + R package to access API: [`Radlibrary`](https://github.com/facebookresearch/Radlibrary)

]

![](img/metaadlibrary2.png)

]

---

### Meta Ad Library (API) - It's not enough!

**HOWEVER** Meta Ad Library has been criticized (Dommett & Power, 2023; Edelson et al., 2020; Leerssen et al., 2019):

+ It does not include all political ads
  
+ It includes many ads that are not political
  
+ It does not include the actual targeting criteria used (*only recently in EU and not all of them*)
  
+ Broad spending and impression boundaries make it hard to compare these metrics across accounts

]

![](https://y.yarn.co/09ecffd4-ce51-408d-8f01-428659745570_text.gif)

]

---

![](img/audienceraw.jpg)

---

### II. Scraping

![](https://miro.medium.com/v2/resize:fit:550/0*Bj_O1jRFzZjKxzi4.jpg)

---

### II. Scraping

`\(\color{green}{\text{Upsides}}\)`

+ Get the data you want
  
  + Your access to skillsets is the limit
  
  <br>
  
  <br>
  
  <br>
  
  <br>
  
  <br>
  
  <br>
  
  <br>
  
  <br>
  
  R packages for scraping: [`rvest`](https://rvest.tidyverse.org/), [`httr2`](https://httr2.r-lib.org/), [`Rselenium`](https://docs.ropensci.org/RSelenium/)
  
  Python packages for scraping: [`BeautifulSoup`](https://pypi.org/project/beautifulsoup4/)
  
]

`\(\color{red}{\text{Downsides}}\)`

+ Grey area (probably goes against terms of service)
  
  + Custom solutions require the necessary skillset
  
  + Access unofficial and might be shut down at any point
  
  + Problem for Reproducibility

]

---

### Introducing `metatargetr`

![](img/metatargetr.png)

Link: https://github.com/favstats/metatargetr

---

### `metatargetr`

]

]

The main function is: `get_targeting`:

![](img/metatargetr2.png)

---

### Transparency during Elections

]

]

This has allowed me to create election dashboards

+ 🇸🇪 [2022 Swedish general election](https://favstats.github.io/SwedishElection2022/)
+ 🇺🇸 [2022 United States midterm elections](https://whotargetsme.github.io/midterms2022_dashboard/) 
+ 🇮🇹 [2023 Lazio & Lombardy regional election](https://favstats.github.io/regionali2023/) 
+ 🇪🇪 [2023 Estonian parliamentary election](https://favstats.github.io/EstoniaElection2023/) 
+ 🇳🇱 [2023 Dutch provincial elections](https://favstats.github.io/ProvincialeStatenverkiezingen2023/) 
+ 🇲🇪 [2023 Montenegrin presidential elections](https://refined-github-html-preview.kidonng.workers.dev/favstats/MontenegroPresidentialElection2023/raw/dc4d9baafe3f30b7d79e45206f63c745f51a25b3/index.html) 
+ 🇦🇺 [2023 New South Wales state election](https://favstats.github.io/NSWAustralianElection2023/) 
+ 🇫🇮 [2023 Finnish parliamentary election](https://favstats.github.io/FinlandElections2023/) 
+ 🇹🇷 [2023 Turkish general election](https://refined-github-html-preview.kidonng.workers.dev/favstats/TurkishElection2023/raw/ce6281fe74b8f5a3f99c576c31bd95758cf80dec/index.html) 
+ 🇩🇪 [2023 Bremen State election](https://favstats.github.io/BremenStateElection2023/)
+ 🇬🇷 [2023 Greek Legislative election](https://favstats.github.io/GreeceElection2023/) 
+ 🇹🇷 [2023 Turkish general election](https://favstats.github.io/TurkishElection2023/) 
+ 🇲🇪 [2023 Montenegrin parliamentary elections](https://favstats.github.io/2023MontenegrinParliamentaryElection/) 
+ 🇳🇱 [2023 Dutch parliamentary elections](https://favstats.github.io/TK2023/) 
+ 🇺🇸 [2024 US Presidential Primaries](https://favstats.github.io/USprimaries2024/)

]

![](img/ps2023.png)

[Blog Post](https://www.favstats.eu/post/provincial_elections/)

]

---

### Transparency during Elections

]

]

![](img/ps2023_1.png)

]

---

![](https://media4.giphy.com/media/9V1F9o1pBjsxFzHzBr/giphy.gif)

---

### Downloading Images and Videos

]

]

With `metatargetr` you can also download the images and videos of any ad!

Just use the function `get_ad_snapshot` on any ad id:

![](img/imgvid.jpeg)

---

#### Now that we have image and videos, what can we do with that?

+ Extract texts from images and videos

+ Detect emotions in music, speech or color schemes (Mendoza 2023, work in progress)

---

## Optical Character Recognition (OCR)

![](img/xxaas.png)

---

## Optical Character Recognition (OCR)

+ Tesseract Engine ([tesseract R package](https://cran.r-project.org/web/packages/tesseract/vignettes/intro.html))

+ Huggingface - Keyword: OCR ([e.g. TrOCR](https://huggingface.co/microsoft/trocr-base-printed))

+ access from R using [huggr](https://github.com/benjaminguinaudeau/huggr)

]

.pull-right[
  Google Vision API (try [here](https://cloud.google.com/vision/docs/drag-and-drop))
  
  ![](img/vision.png)
  
  + [googleCloudVisionR](https://github.com/emartech/googleCloudVisionR)
  + [Python Client for Google Cloud Vision](https://cloud.google.com/python/docs/reference/vision/latest)
]

---

## Video to Audio to Text

The best approach to turn video to audio is the library [`FFmpeg`](https://ffmpeg.org)

+ A complete, cross-platform solution to record, convert and stream audio and video
  
Once you obtained audio:

+ I utilized Mozilla DeepSpeech (Hannun et al., 2014)
  
**HOWEVER**:

+ I would now recommend [`whisper`](https://github.com/openai/whisper) from OpenAI

]

.pull-right[
![](https://uploads-ssl.webflow.com/621de55357719363b658d18c/64cd057290d5360c1e382d9d_1_y4FrBZQtxGPjwgbVFbcWzg.png)
]

---

# TOXICITY

![](https://y.yarn.co/e4bfcab1-4916-45e7-a570-fe722eededc6_text.gif)

---

## Perspective API - Models

> What is Perspective? Perspective is a free API that uses machine learning to identify toxic comments, making it easier to host better conversations online.

---

## Perspective API - Models

| Model name                 | Type  | Description                                                                                                                                                                                                                                                                             | Available Languages                                                                               |
|--------------------------------|-------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------|
| `SEVERE_TOXICITY`              | prod. | A very hateful, aggressive, disrespectful comment or otherwise very likely to make a user leave a discussion or give up on sharing their perspective. This attribute is much less sensitive to more mild forms of toxicity, such as comments that include positive uses of curse words. | en, fr, es, de, it, pt, ru                                                                        |
| `SEVERE_TOXICITY_EXPERIMENTAL` | exp.  | ^[] | ar                                                                                                |
| `IDENTITY_ATTACK`              | prod. | Negative or hateful comments targeting someone because of their identity.                                                                                                                                                                                                               | de, it, pt, ru, en                                                                                |
| `IDENTITY_ATTACK_EXPERIMENTAL` | exp.  | ^[]                                                                                                                                                                                                              | fr, es, ar                                                                                        |
| `INSULT`                       | prod. | Insulting, inflammatory, or negative comment towards a person or a group of people.                                                                                                                                                                                                     | de, it, pt, ru, en                                                                                |
| `INSULT_EXPERIMENTAL`          | exp.  | ^[]                                                                                                                                                                                                   | fr, es, ar                                                                                        |

---

## Perspective API - Models

| Model name                 | Type  | Description                                                                                                                                                                                                                                                                             | Available Languages                                                                               |
|--------------------------------|-------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------|                                                                                     |
| `PROFANITY`                    | prod. | Swear words, curse words, or other obscene or profane language.                                                                                                                                                                                                                         | de, it, pt, ru, en                                                                                |
| `PROFANITY_EXPERIMENTAL`       | exp.  | ^[]                                                                                                                                                                                                                       | fr, es, ar                                                                                        |
| `THREAT`                       | prod. | Describes an intention to inflict pain, injury, or violence against an individual or group.                                                                                                                                                                                             | de, it, pt, ru, en                                                                                |
| `THREAT_EXPERIMENTAL`          | exp.  | ^[]                                                                                                                                                                                             | fr, es, ar                                                                                        |
| `SEXUALLY_EXPLICIT`            | exp.  | Contains references to sexual acts, body parts, or other lewd content.                                                                                                                                                                                                                  | en                                                                                                |
| `FLIRTATION`                   | exp.  | Pickup lines, complimenting appearance, subtle sexual innuendos, etc.                                                                                                                                                                                                                   | en

---

### peRspective R package

![](img/perspective.png)

Link: https://github.com/favstats/peRspective

---

### peRspective R package

]

]

`my_text <- "You wrote this? Wow. This is dumb and childish, please go f**** yourself."`

prsp_score(`
           `text = my_text,` 
           `languages = "en",`
           `score_model = peRspective::prsp_models`
           `)`

> Don't forget to validate outputs of this measurement!

See: van Atteveldt et al. (2021); Chan et al. (2021)

]

![](https://github.com/favstats/peRspective/raw/master/man/figures/README-unnamed-chunk-9-1.png)

]

---

## Takeaways

+ Official APIs are not dead..

+ but often have limited data
  
  + can be taken away at a whim

+ Scraping remains important

+ However: high hurdle to implement
  
  + Hopefully tools like `metatargetr` can reduce  that burden

+ `peRspective` can help with analyzing text data

+ However: important to validate for your use-case!

---

### Targeting during 2021 German Bundestag elections

#### Collaboration with Who Targets Me

##### User-centric approaches

##### III. User Tracking

---

## Who Targets Me - Tracking users to see how they are tracked

]

.pull-right[
![](https://whotargets.me/wp-content/uploads/2021/04/Screenshot-2021-04-30-at-09.49.54.png)
]

---

## Why Am I Seeing This

+  WTM scrapes the text from the ‘Why am I seeing this?’ label

![](https://www.impactplus.com/hubfs/Screen%20Shot%202019-07-15%20at%202.11.21%20PM.png)

---

### III. User Tracking

`\(\color{green}{\text{Upsides}}\)`

+ Study effects on users
  
      + over time
  
  + Users just have to install tracking app
  
      + removes the burden from users

]

`\(\color{red}{\text{Downsides}}\)`

+ Biased samples 
  
      + who is more likely to install an app?
  
  + Creation of tracker software
  
      + who has the necessary skillset for that?
      
      + substantial expertise and effort

]

---

### Who Targets Me

+ ZDF Magazin Royale collaborates with Who Targets Me (April 2021)

+ German equivalent of *De Avondshow met Arjen Lubach*

+ 17k users in Germany sign up with WTM in the next months

+ only ~5k users (~30%) see political ads

+ 150k political ad impressions

+ I join the project end of July 2021

]

]

---

### ZDF Magazine Royale Show 24. September 2021

https://targetleaks.de/

---

## Political Advertisers in Germany

![](img/ad_impressions.png)

---

## Bias in data - Gender

![](img/gender_bias.png)

---

## Bias in data - Age

![](img/age_bias.png)

---

## Bias in data - Bundesland

![](img/region_bias.png)

---

## Solution - Weighting

+ Weighting is a common procedure when working with survey data to enhance *representability* after data was collected

+ Detailed [blog post](https://www.pewresearch.org/decoded/2020/03/26/weighting-survey-data-with-the-pewmethods-r-package/) about how to use `pewmethods` for weighing

![](img/pewresearch.png)

---

### Weighting the data

#### according to German population metrics (using `pewmethods`)

![](img/gender_weighted.png)

]

<br>

Link to Shiny app to explore results: [favstats.shinyapps.io/btw21_wtm/](https://favstats.shinyapps.io/btw21_wtm/)

]

---

### Now what can we do with this (weighted) data?

We could study

+ **Voter turnout**: Ads might motivate or demotivate certain demographics from participating in the election.

+ **Shift in voting intent**: Exposure to certain ads might make voters reconsider their choice.

+ **Misinformation**: Microtargeted ads might spread misleading or false information tailored to specific demographics, influencing perceptions.

+ **Polarization**: Specific ads might increase division and animosity between groups, especially when they focus on contentious issues.

---

## Takeaways

+ Tracking studies may require weighting because of self-selection bias

+ R package `pewmethods` can help create weights
  
  
<br>

+ Who Targets Me helps create transparency around elections

+ they are very open to research
  
  + they can be approached if you would like to conduct studies with them

---

# The Role of Algorithms

# in Political Microtargeting

![](https://media3.giphy.com/media/3o6Yg4GUVgIUg3bf7W/giphy.gif)

(*IV. Data Donation + Algorithm Audit Study*)

---

### IV. Data Donation

---

### IV. Data Donation

`\(\color{green}{\text{Upsides}}\)`

+ non-public data (e.g., private messages, web browsing history)

+ Study (historical) records of users
  
      + over time
  
  + Completeness of data (e.g. when using Google Takeout)
  
  
]

`\(\color{red}{\text{Downsides}}\)`

+ Biased samples 
  
      + who is more likely to give up data?
  
  + Very privacy-sensitive data might need to be collected
  
      + how to ensure privacy?
      
      + reproducibility?
      
  + data is often less structured or documented

]

---

### Online Political Microtargeting of Political Ads - the "bad actors"-story
 
 
.pull-left[
<img src="img/cambridge_analytica.png" width="100%">

]

---

### Online Political Microtargeting of Political Ads - the "bad actors"-story
 
 
.pull-left[
<img src="img/cambridge_analytica.png" width="100%">

*The explicit assumption here that advertisers typically have strong control over who sees which ad*

] 
  
  
--

**But there is more than *just* targeting criteria that decides who sees political ads:**

+ advertisers can set targeting *boundaries*

+ *ad delivery algorithms* "decide" which individual users get ads from which advertiser

]

---

<img src="img/plantuml00.png" width="80%">

---

<img src="img/plantuml01.png" width="80%">

---

---

---

<img src="img/plantuml05.png" width="80%">

---

### Who decides who sees which ad on Meta?

+ **Ad auctions** = an auction takes place that determines which ad by whom is shown

---

### Who decides who sees which ad on Meta?

+ **Relevance** = how relevant is the ad to the user

[(Meta Business Help Center, 2022)](https://www.facebook.com/business/help/430291176997542)

---

### Who decides who sees which ad on Meta?

+ **Ad auctions** = an auction takes place that determines which ad by whom is shown: based on *budget*

+ **Relevance** = how relevant is the ad to the user

##### *Ad delivery algorithms* finding *relevant* audiences for ads: we term this **algorithmic microtargeting**

---

### A (silly) example

.pull-left[
    <img src="https://images-na.ssl-images-amazon.com/images/S/compressed.photo.goodreads.com/books/1465341854i/12111823.jpg" width = "60%">
]

---

### Pricing differences in the US 2020 election

+ Biden campaign paid more than **6x more** 
  + compared to Trump campaign when targeting older voters. [(The Markup 2020)](https://themarkup.org/election-2020/2020/10/29/facebook-political-ad-targeting-algorithm-prices-trump-biden)
  
  
<center>
<img src="img/older.png" width="60%" />
</center>

---

### Prior Research (Ali et al., 2020,2021)

---

### Prior Research (Ali et al., 2020,2021)

When targeting the same audience, at the same time, with the same budget:

+ Ad delivery is heavily skewed along gendered and racial stereotypes
  + even without the intent of the advertiser [(Ali et al. 2020)](https://dl.acm.org/doi/10.1145/3359301)
  
--

Images invisible to humans but still detectable by algorithm:

+ yield **similar skews** in delivery

+ highlights importance of algorithm

+ less based on differences in user behavior/preferences
]

---

### Prior Research (Ali et al., 2020,2021)

When targeting the same audience, at the same time, with the same budget:

Regarding political ads [(Ali et al., 2021)](https://dl.acm.org/doi/pdf/10.1145/3437963.3441801):

+ Political ads more often delivered to ideologically congruent audience 
      + Bernie ads → higher % D; 
      + Trump ads → higher % R

+ **Increased cost**

+ Liberal ad to a liberal audience: *21 Dollar per 1000 users*; 
  + Conservative ad delivered to liberal audience: *40 Dollar per 1000 users*.
]

+ when tricking Facebook into classifying non-partisan ads as partisan

]

---

---

## Research Question

### How does the Meta ad delivery algorithm<br>influence the pricing & distribution of political ads<br>in the Netherlands?

---

# Research Design

---

### Research Design

+ Algorithm audit study

+ Place the same ads targeting the same audiences (9 different ones)

+ Collaborate with Dutch parties to place political ads

+ Final collaboration with 3:

1. GroenLinks (Green party)
  2. VVD (centre-right party of PM Rutte)
  3. PvdA (social democrats)

+ Place ads before nationwide local elections on March 16th 2022
  + 1st to 7th February 2022

+ Spend 2 Euros a day on 45 ad copies

+ in total: 630 Euro per party
  
  
--
  
+ Pre-registered research design and hypotheses

---

### Dependent Variables

+ Price per 1k users reached

+ this measure is an industry standard

+ Unique users reached

+ this measure is co-linear with the price
  
---

### Ad Relevance

We theorize two different levels of (predicted) relevance:

1. Relevant audience for party (i.e. source of ad)

+ Ads from an environmentalist party more likely to be relevant for audience interested in environmentalism.

<br>

<ol start="2">
  <li>Relevant audience for ad content (e.g. political message)</li>
</ol>

+ Political message likely to be relevant for people interested in politics

---

### Hypotheses

![](img/relevant_quote.png)

[(Meta Business Help Center, 2022)](https://www.facebook.com/business/help/430291176997542)

> **H1:** **The more relevant** an audience is for an ad, **the cheaper is the cost** for reaching 1000 users in that audience.

> **H2:** **The more relevant** an audience is for an ad, **the more are ads delivered** to that audience.

We expect that ads by party with a greater share of supporters are less expensive (H3a) and reach more people (H3b)

> **H3a:** Parties with a greater share of supporters pay less for reaching 1000 users.

> **H3b:** Parties with a greater share of supporters reach more people than smaller parties.

---

### Targeting criteria (Sub-hypotheses for H1 & H2)

We used 9 different (paired) targeting criteria for our advertisements

1. Political interests
2. Excluding political interest

<ol start="3">
  <li>Higher educated audience</li>
  <li>Lower educated audience</li>
</ol>

]

**Relevant audiences for ad content**

> Targeting political ads to **politically interested** and **higher-educated** audiences

> *is less expensive*

> *deliver more*

> than targeting politically uninterested and lower-educated audiences.

]

---

### Targeting criteria (Sub-hypotheses for H1 & H2)

We used 9 different (paired) targeting criteria for our advertisements

1. Political interests
2. Excluding political interest
  
<ol start="3">
  <li>Higher educated audience</li>
  <li>Lower educated audience</li>
</ol>

<ol start="5">
  <li>Environmental interests</li>
  <li>Excluding environmental interests</li>
</ol>

<ol start="7">
  <li>Economic interests</li>
  <li>Excluding Economic interests</li>
</ol>

]

**Relevant audience for party**

> Targeting political ads to issues that party has issue ownership over

>  *is less expensive*

>  *delivers more*

compared to other parties

]

---

### Targeting criteria (Sub-hypotheses for H1 & H2)

We used 9 different (paired) targeting criteria for our advertisements

1. Political interests
2. Excluding political interest

<ol start="3">
  <li>Higher educated audience</li>
  <li>Lower educated audience</li>
</ol>

<ol start="5">
  <li>Environmental interests</li>
  <li>Excluding environmental interests</li>
</ol>

<ol start="7">
  <li>Economic interests</li>
  <li>Excluding Economic interests</li>
</ol>
  
<ol start="9">
  <li>No Targeting</li>
</ol>

]

---

# Ad Creative and Setup

---

## How the ad looked like on Desktop

---

## How the ad looked like on Desktop

---

## Results

---

### Between-party differences

`\(\rightarrow\)` we consistently find one party that pays less and reaches more people

---

#### Between-party differences (per individual ad)

.font80[PvdA pays the least (**10-12 cents less** or: 8-10%) & reaches more people (~**1.1 - 1.3k more** per ad)]

```
## # A tibble: 15 × 5
##    party      reach share targeting        relevance
##    <chr>      <dbl> <dbl> <chr>                <dbl>
##  1 PvdA       13138  52.1 Higher Education         1
##  2 PvdA       12917  51.7 Higher Education         1
##  3 GroenLinks 11938  51.7 Higher Education         2
##  4 VVD        11528  51.6 Higher Education         1
##  5 VVD        11845  51.6 Higher Education         1
##  6 GroenLinks 11622  51.6 Higher Education         2
##  7 PvdA       12860  51.6 Higher Education         1
##  8 GroenLinks 11727  51.4 Higher Education         2
##  9 PvdA       12729  51.1 Higher Education         1
## 10 GroenLinks 11486  51.1 Higher Education         2
## 11 VVD        11388  51.0 Higher Education         1
## 12 PvdA       12632  50.9 Higher Education         1
## 13 GroenLinks 11509  50.9 Higher Education         2
## 14 VVD        11344  50.8 Higher Education         1
## 15 VVD        11260  50.6 Higher Education         1
```

]

]

---

#### Between-party differences (per target audience)

---

### Within-party differences

---

### Within-party differences - Price per 1k

Ads **cost less for**:

+ *higher-educated* vs. *lower-educated audience*

Ad price **does not statistically differ for**:

+ Audience *interested in the economy* vs. *not interested*

+ Audience *interested in politics* vs. *not interested*

Ads **cost more for**:

+ Audience *interested in the environment* vs. *not interested*

]

![](img/diffs1.png)

]

---

**18-24 year olds and women are reached less (and cost more to reach)**

![](img/priceshare.png)

---

## Summary

---

### Summary

Our findings do not always align with expectations.

H1: More "relevant" audiences were not always cheaper

H2: More "relevant" audiences were not always reached more

H3: Party with greatest audience did not reach more or get cheaper prices

**However:**

> We **still** find that Meta ad delivery algorithm prioritizes certain parties and audiences for political advertising

1. PvdA pays least and reach most
2. Lower-educated, people interested in environment, women and younger people more expensive to reach

---

### Limitations

+ Only three political parties

+ Study first-of-its-kind

+ needs more research!
  
+ Relevance might need to be measured differently?

+ We do not vary content.. although studies suggest this is important (Ali et al. 2021, 2022)

---

### Implications

+ Unequal playing field

+ Meta (dis-)advantages certain parties

+  the findings presented in this paper show that political parties were not charged the same price for the same service
  
--

+ Potential for deepening political, social and geographical inequalities

+ Some groups of people and regions are **systematically** less likely to receive political advertisements and more expensive to reach

+ isolating these groups from receiving election-related information

+ Little to no transparency by Meta about these systematic biases

+ difficult to research and make visible instances of unequal treatment and price discrimination
  + highlighting importance of access to data
  
--
  
+ Simply "banning" microtargeting would be inadequate

+ more power to the black box algorithm
  
---

# Zooming out

Hopefully you found the methods, studies, and results in this talk interesting!

---

## Thank you for your attention! Questions?

Link to presentation: *favstats.github.io/nefca2023*

![](https://c.tenor.com/Q9qk5zN5EesAAAAM/space-kitten.gif)
![](https://c.tenor.com/Q9qk5zN5EesAAAAM/space-kitten.gif)

]

![](https://c.tenor.com/Q9qk5zN5EesAAAAM/space-kitten.gif)
![](https://c.tenor.com/Q9qk5zN5EesAAAAM/space-kitten.gif)

]

---

## Literature

Aaker, J. L. (1999). The Malleable Self: The Role of Self-Expression in Persuasion. Journal of Marketing Research, 36(1), 45–57. https://doi.org/10.1177/002224379903600104

Ali, M., Sapiezynski, P., Bogen, M., Korolova, A., Mislove, A., & Rieke, A. (2019). Discrimination through Optimization: How Facebook’s Ad Delivery Can Lead to Biased Outcomes. Proceedings of the ACM on Human-Computer Interaction, 3(CSCW), 1–30. https://doi.org/10.1145/3359301

Ali, M., Sapiezynski, P., Korolova, A., Mislove, A., & Rieke, A. (2021). Ad Delivery Algorithms: The Hidden Arbiters of Political Messaging. Proceedings of the 14th ACM International Conference on Web Search and Data Mining, 13–21. https://doi.org/10.1145/3437963.3441801

Dobber, T., Trilling, D., Helberger, N., & de Vreese, C. (2019). Spiraling downward: The reciprocal relation between attitude toward political behavioral targeting and privacy concerns. New Media & Society, 21(6), 1212–1231. https://doi.org/10.1177/1461444818813372

Dobber, T., Metoui, N., Trilling, D., Helberger, N., & de Vreese, C. (2020). Do (Microtargeted) Deepfakes Have Real Effects on Political Attitudes? The International Journal of Press/Politics, 26(1), 69–91. https://doi.org/10.1177/1940161220944364

Dobber, T., Trilling, D., Helberger, N., & de Vreese, C. (2023). Effects of an issue-based microtargeting campaign: A small-scale field experiment in a multi-party setting. The Information Society, 39(1), 35–44. https://doi.org/10.1080/01972243.2022.2134240

---

## Literature

Chan, C. H., Bajjalieh, J., Auvil, L., Wessler, H., Althaus, S., Welbers, K., ... & Jungblut, M. (2021). Four best practices for measuring news sentiment using ‘off-the-shelf’dictionaries: A large-scale p-hacking experiment. Computational Communication Research, 3(1), 1-27.

Coppock, A., Hill, S. J., & Vavreck, L. (2020). The small effects of political advertising are small regardless of context, message, sender, or receiver: Evidence from 59 real-time randomized experiments. Science Advances, 6(36), eabc4046. https://doi.org/10.1126/sciadv.abc4046

Decker, H., & Krämer, N. (2023). Is Personality Key? Persuasive Effects of Prior Attitudes and Personality in Political Microtargeting. Media and Communication, 11(3), 250–261. https://doi.org/10.17645/
mac.v11i3.6627

Endres, K. (2020). Targeted Issue Messages and Voting Behavior. American Politics Research, 48(2), 317–328. https://doi.org/10.1177/1532673X19875694

Fiske, S. T. (1980). Attention and weight in person perception: The impact of negative and extreme behavior. Journal of Personality and Social Psychology, 38(6), 889–906. https://doi.org/10.1037/0022-3514.38.6.889

Freelon, D. (2018). Computational research in the post-API age. Political Communication, 35(4), 665-668.

Garramone, G. M. (1984). Voter Responses to Negative Political Ads. Journalism Quarterly, 61(2), 250–259. https://doi.org/10.1177/107769908406100202

---

## Literature

Geer, J. G. (2006). In defense of negativity: Attack ads in presidential campaigns. University of Chicago Press.

Haenschen, K. (2022). The Conditional Effects of Microtargeted Facebook Advertisements on Voter Turnout. Political Behavior, 1–21. https://doi.org/10.1007/s11109-022-09781-7

Hannun, A., Case, C., Casper, J., Catanzaro, B., Diamos, G., Elsen, E., Prenger, R., Satheesh, S., Sengupta, S., Coates, A., & Ng, A. Y. (2014). Deep Speech: Scaling up end-to-end speech recognition. arXiv:1412.5567 [cs]. Retrieved January 30, 2021, from http://arxiv.org/abs/1412.5567

Haselmayer, M. (2019). Negative campaigning and its consequences: A review and a look ahead. French Politics, 17(3), 355–372. https://doi.org/10.1057/s41253-019-00084-8

Hilbig, B. E. (2009). Sad, thus true: Negativity bias in judgments of truth. Journal of Experimental Social Psychology, 45(4), 983–986. https://doi.org/10.1016/j.jesp.2009.04.012

Krotzek, L. J. (2019). Inside the Voter’s Mind: The Effect of Psychometric Microtargeting on Feelings Toward and Propensity to Vote for a Candidate. International Journal of Communication; Vol 13 (2019). https://ijoc.org/index.php/ijoc/article/view/9605

Nai, A., & Maier, J. (2020). Is Negative Campaigning a Matter of Taste? Political Attacks, Incivility, and the Moderating Role of Individual Differences. American Politics Research, 49(3), 269–281. https://doi.org/10.1177/1532673X20965548

---

## Literature

Moon, Y. (2002). Personalization and Personality: Some Effects of Customizing Message Style Based on Consumer Personality. Journal of Consumer Psychology, 12(4), 313–325. https://doi.org/10.1016/S1057-7408(16)30083-3

Mutz, D. C., & Reeves, B. (2005). The New Videomalaise: Effects of Televised Incivility on Political Trust. American Political Science Review, 99(1), 1–15. https://doi.org/10.1017/S0003055405051452

Ohme, J., Araujo, T., Boeschoten, L., Freelon, D., Ram, N., Reeves, B. B., & Robinson, T. N. (2023). Digital Trace Data Collection for Social Media Effects Research: APIs, Data Donation, and (Screen) Tracking. Communication Methods and Measures, 1-18.

Petty, R. E., and J. T. Cacioppo. 1986. The elaboration likelihood model of persuasion. Advances in Experimental Social Psychology 19:123–205. doi: 10.1016/S0065-2601(08)60214-2.

Rozin, P., & Royzman, E. B. (2001). Negativity Bias, Negativity Dominance, and Contagion. Personality and Social Psychology Review, 5(4), 296–320. https://doi.org/10.1207/S15327957PSPR0504_2

Sharp, B., Danenberg, N., & Bellman, S. (2018). Psychological targeting. Proceedings of the National Academy of Sciences, 115(34), E7890–E7890. https://doi.org/10.1073/pnas.1810436115

Tanusondjaja, A., Michelon, A., Hartnett, N., & Stocchi, L. (2023). Reaching Voters on Social Media: Planning Political Advertising on Snapchat. International Journal of Market Research, 65(5), 566–580. https://doi.org/10.1177/14707853231175085

---

## Literature

Tappin, B. M., Wittenberg, C., Hewitt, L. B., Berinsky, A. J., & Rand, D. G. (2023). Quantifying the potential persuasive returns to political microtargeting. Proceedings of the National Academy of Sciences, 120(25), e2216261120. https://doi.org/10.1073/pnas.2216261120

Tufekci, Z. (2014). Engineering the public: Big data, surveillance and computational politics. First Monday. https://doi.org/10.5210/fm.v19i7.4901

Van Atteveldt, W., Van der Velden, M. A., & Boukes, M. (2021). The validity of sentiment analysis: Comparing manual annotation, crowd-coding, dictionary approaches, and machine learning algorithms. Communication Methods and Measures, 15(2), 121-140.

Walter, A. S., & van der Eijk, C. (2019). Unintended consequences of negative campaigning: Backlash and second-preference boost effects in a multi-party context. The British Journal of Politics and International Relations, 21(3), 612–629. https://doi.org/10.1177/1369148119842038

Wheeler, S. C., DeMarree, K. G., & Petty, R. E. (2008). A match made in the laboratory: Persuasion and matches to primed traits and stereotypes. Journal of Experimental Social Psychology, 44(4), 1035–1047. https://doi.org/10.1016/j.jesp.2008.03.007

Zarouali, B., Dobber, T., De Pauw, G., & de Vreese, C. (2022). Using a Personality-Profiling Algorithm to Investigate Political Microtargeting: Assessing the Persuasion Effects of Personality-Tailored Ads on Social Media. Communication Research, 1066–1091. https://doi.org/10.1177/0093650220961965

---

## Appendix

---

### Within-party differences

Reach and cost **over time**

Potential *market shock* on February 4th?

---

### Within-party differences per day - Reach and Cost

---

### Within-party differences

Reach and cost **over time** and **per party**

`\(\rightarrow\)` party differences remain constant despite *"shock"*

---

## Price differences per day

We observe:

*Consistent results*

+ "Market shock" **hits all parties equally**

+ Environment audience consistently *more expensive*

+ Higher educated audience consistently *less expensive*

*Inconsistent results*

+ Audiences interested in economy & politics are typically cheaper except on the day of the spike

]

---

### Price differences per day

---

### Bulk Discount?

]

]

---

## Skewed delivery

in terms of gender, age and region

---

## Differences in delivery by gender

.pull-left[
+ *Line at zero* shows empirical equilibrium of target audiences (i.e. the observed share of men and women in target audience)

+ *Deviation from zero* are algorithmic biases

+ above zero: prioritization
  
  + below zero: de-prioritization

+ Ads *deliver to more men* for every party

+ However: bias towards men seems smaller for GroenLinks
]

---

## Differences in delivery by age group

+ Ads *deliver less to young people*

+ aged 18-24

+ Consistent for each party

]

---

## Region differences

+ Ads deliver more to some regions

+ for example: Limburg, Friesland, Drenthe

+ Ads deliver less to other regions

+ Utrecht, North Holland, North Brabant

+ Consistent for each party

]

---

.font80[If we exclude economic interests/target environmental interests: VVD reaches less people and cheaper than GL]

]