Once a Lounger, Always a Lounger? Tracking Inattention in Longitudinal Panel Data

class: inverse,middle,center
<style type="text/css">
.purpleb {
  font-weight: bold;
  color: #4F2683;
  font-size: 1.25em;
}

.purplebL {
  font-weight: bold;
  color: #4F2683;
  font-size: 1.5em;
}

.footnote {
  position: absolute;
  bottom: 60px;
  padding-right: 4em;
  font-size: .75em;
}

.large {
  font-size:1.5rem;
}

.Small {
  font-size:.9rem;
}

.small {
  font-size:.75rem;
}
.tiny {
  font-size:.25rem;
}
.shift { 
  position:relative; 
  top: -40px;
  }

.plot-callout {
  height: 225px;
  width: 450px;
  bottom: 5%;
  right: 5%;
  position: absolute;
  padding: 0px;
  z-index: 100;
}
.plot-callout img {
  width: 100%;
  border: 4px solid
  #  23373B;
}

.remark-slide table tr:nth-child(even) {
  background: none !important;
}
.remark-slide table thead {
  background: none !important;
}
.remark-slide table thead th {
  border-bottom: 1px solid #666;
}

.pull-leftL {
  float: left;
  width: 57%;
}
.pull-rightS {
  float: right;
  width: 37%;
}

.footer {
    font-size: .75rem;
    position: fixed;
    bottom: -8px; 
    left: 0;
    width: 100%;
    text-align: center;
    padding: 1rem 0;
    color: #4F2683;
}

.footer a {
    margin: 0 3rem;
    text-decoration: none;
    color: #4F2683;
    font-weight: 500;
}

.footer a:visited {
    color: #4F2683;
}

.footer a strong,
.footer a b {
    color: #4F2683;
    font-weight: 700;
}

.footer a:hover {
    text-decoration: underline;
}

</style>

# Once a Lounger, Always a Lounger? 
## Tracking Inattention in Longitudinal Panel Data

### William Poirier

2026-06-02

---

## Plan of presentation

00. Introduction

1. Puzzle

2. Design

3. Results

4. Conclusion

---
layout: true

.footer[[**Introduction**](#intro) [Puzzle](#puzzle) [Design](#design) [Results](#res) [Conclusion ](#conclu) ]

---
name: intro
## Why should I care? 1/2

.small[<a name=cite-huang2015insufficient></a><a name=cite-curran2016methods></a><a name=cite-silber2019impact></a><a name=cite-pyo2021cognitive></a>([Huang, Liu, and Bowling, 2015](#bib-huang2015insufficient); [Curran, 2016](#bib-curran2016methods); [Silber, Danner, and Rammstedt, 2019](#bib-silber2019impact); [Pyo and Maxfield, 2021](#bib-pyo2021cognitive))]
    
]
.pull-right[

<img src="olal_files/figure-html/unnamed-chunk-3-1.png" width="100%" style="display: block; margin: auto;" />
]

---

## Why should I care? 1/2

.pull-left[  
#### Descriptive Inference
  - Conditional IRes `$\rightarrow$` upward/downward bias.
  - Random IRes `$\rightarrow$` increased variance.
  - IR:
    - Bias towards non-attitudes;
    - Scale reliability;
    - Random sampling not quite random.
  - Can think of it as a missing data problem (MCAR, MAR, NMAR).
  
.small[([Huang, Liu, and Bowling, 2015](#bib-huang2015insufficient); [Curran, 2016](#bib-curran2016methods); [Silber, Danner, and Rammstedt, 2019](#bib-silber2019impact); [Pyo and Maxfield, 2021](#bib-pyo2021cognitive))]
]
.pull-right[

]

---

## Why should I care? 1/2

.pull-left[  
#### Descriptive Inference
  - Conditional IRes `$\rightarrow$` upward/downward bias.
  - Random IRes `$\rightarrow$` increased variance.
  - IR:
    - Bias towards non-attitudes;
    - Scale reliability;
    - Random sampling not quite random.
  - Can think of it as a missing data problem (MCAR, MAR, NMAR).
    
.small[([Huang, Liu, and Bowling, 2015](#bib-huang2015insufficient); [Curran, 2016](#bib-curran2016methods); [Silber, Danner, and Rammstedt, 2019](#bib-silber2019impact); [Pyo and Maxfield, 2021](#bib-pyo2021cognitive))]  
]
.pull-right[

<img src="olal_files/figure-html/unnamed-chunk-5-1.png" width="100%" style="display: block; margin: auto;" />
]

---

## Why should I care? 2/2

.pull-left[  
#### Causal Inference
  - Even random distribution of IRes is problematic: 
    - Control group permeates to the treatment group; 
    - Treatment effect biased toward 0.
    
.small[<a name=cite-kane2025more></a><a name=cite-bruhlmann2020quality></a><a name=cite-desimone2018dirty></a><a name=cite-abbey2017attention></a><a name=cite-berinsky2014separating></a><a name=cite-maniaci2014caring></a>([Kane, 2025](#bib-kane2025more); [Brühlmann, Petralito, Aeschbach, and Opwis, 2020](#bib-bruhlmann2020quality); [DeSimone and Harms, 2018](#bib-desimone2018dirty); [Abbey and Meloy, 2017](#bib-abbey2017attention); [Berinsky, Margolis, and Sances, 2014](#bib-berinsky2014separating); [Maniaci and Rogge, 2014](#bib-maniaci2014caring))]  
]
.pull-right[
<img src="olal_files/figure-html/unnamed-chunk-6-1.png" width="100%" style="display: block; margin: auto;" />
]

---

## Why should I care? 2/2

.pull-left[  
#### Causal Inference
  - Even random distribution of IRes is problematic: 
    - Control group permeates to the treatment group; 
    - Treatment effect biased toward 0.
    
.small[([Kane, 2025](#bib-kane2025more); [Brühlmann, Petralito, Aeschbach et al., 2020](#bib-bruhlmann2020quality); [DeSimone and Harms, 2018](#bib-desimone2018dirty); [Abbey and Meloy, 2017](#bib-abbey2017attention); [Berinsky, Margolis, and Sances, 2014](#bib-berinsky2014separating); [Maniaci and Rogge, 2014](#bib-maniaci2014caring))]  
]
.pull-right[
<img src="olal_files/figure-html/unnamed-chunk-7-1.png" width="100%" style="display: block; margin: auto;" />
]

---

## Why should I care? 2/2

.pull-left[  
#### Causal Inference
  - Even random distribution of IRes is problematic: 
    - Control group permeates to the treatment group; 
    - Treatment effect biased toward 0.
    
.small[([Kane, 2025](#bib-kane2025more); [Brühlmann, Petralito, Aeschbach et al., 2020](#bib-bruhlmann2020quality); [DeSimone and Harms, 2018](#bib-desimone2018dirty); [Abbey and Meloy, 2017](#bib-abbey2017attention); [Berinsky, Margolis, and Sances, 2014](#bib-berinsky2014separating); [Maniaci and Rogge, 2014](#bib-maniaci2014caring))]  
]
.pull-right[
<img src="olal_files/figure-html/unnamed-chunk-8-1.png" width="100%" style="display: block; margin: auto;" />
]

---

## Why should I care? 2/2

.pull-left[  
#### Causal Inference
  - Even random distribution of IRes is problematic: 
    - Control group permeates to the treatment group; 
    - Treatment effect biased toward 0.
    
.small[([Kane, 2025](#bib-kane2025more); [Brühlmann, Petralito, Aeschbach et al., 2020](#bib-bruhlmann2020quality); [DeSimone and Harms, 2018](#bib-desimone2018dirty); [Abbey and Meloy, 2017](#bib-abbey2017attention); [Berinsky, Margolis, and Sances, 2014](#bib-berinsky2014separating); [Maniaci and Rogge, 2014](#bib-maniaci2014caring))]  
]
.pull-right[
<img src="olal_files/figure-html/unnamed-chunk-9-1.png" width="100%" style="display: block; margin: auto;" />
]

---

## Why should I care? 2/2

.pull-left[  
#### Causal Inference
  - Even random distribution of IRes is problematic: 
    - Control group permeates to the treatment group; 
    - Treatment effect biased toward 0.
    
.small[([Kane, 2025](#bib-kane2025more); [Brühlmann, Petralito, Aeschbach et al., 2020](#bib-bruhlmann2020quality); [DeSimone and Harms, 2018](#bib-desimone2018dirty); [Abbey and Meloy, 2017](#bib-abbey2017attention); [Berinsky, Margolis, and Sances, 2014](#bib-berinsky2014separating); [Maniaci and Rogge, 2014](#bib-maniaci2014caring))]  
]
.pull-right[
<img src="olal_files/figure-html/unnamed-chunk-10-1.png" width="100%" style="display: block; margin: auto;" />
]

---

## Why should I care? 2/2

.pull-left[  
#### Causal Inference
  - Even random distribution of IRes is problematic: 
    - Control group permeates to the treatment group; 
    - Treatment effect biased toward 0.
    
.small[([Kane, 2025](#bib-kane2025more); [Brühlmann, Petralito, Aeschbach et al., 2020](#bib-bruhlmann2020quality); [DeSimone and Harms, 2018](#bib-desimone2018dirty); [Abbey and Meloy, 2017](#bib-abbey2017attention); [Berinsky, Margolis, and Sances, 2014](#bib-berinsky2014separating); [Maniaci and Rogge, 2014](#bib-maniaci2014caring))]  
]
.pull-right[
<img src="olal_files/figure-html/unnamed-chunk-11-1.png" width="100%" style="display: block; margin: auto;" />
]

---

## Why should I care? 2/2

.pull-left[  
#### Causal Inference
  - Even random distribution of IRes is problematic: 
    - Control group permeates to the treatment group; 
    - Treatment effect biased toward 0.
    
.small[([Kane, 2025](#bib-kane2025more); [Brühlmann, Petralito, Aeschbach et al., 2020](#bib-bruhlmann2020quality); [DeSimone and Harms, 2018](#bib-desimone2018dirty); [Abbey and Meloy, 2017](#bib-abbey2017attention); [Berinsky, Margolis, and Sances, 2014](#bib-berinsky2014separating); [Maniaci and Rogge, 2014](#bib-maniaci2014caring))]  
]
.pull-right[
<img src="olal_files/figure-html/unnamed-chunk-12-1.png" width="100%" style="display: block; margin: auto;" />
]

---

## Inattention as a concept

Theory of optimal answering <a name=cite-tourangeau1988cognitive></a>([Tourangeau and Rasinski, 1988](#bib-tourangeau1988cognitive)):

---

## Inattention as a concept

Theory of optimal answering ([Tourangeau and Rasinski, 1988](#bib-tourangeau1988cognitive)):

---

## Inattention as a concept

Theory of optimal answering ([Tourangeau and Rasinski, 1988](#bib-tourangeau1988cognitive)):

---

## Inattention as a concept

Theory of optimal answering ([Tourangeau and Rasinski, 1988](#bib-tourangeau1988cognitive)):

---

## Inattention as a concept

Theory of optimal answering ([Tourangeau and Rasinski, 1988](#bib-tourangeau1988cognitive)):

---

## Inattention as a concept

Theory of optimal answering ([Tourangeau and Rasinski, 1988](#bib-tourangeau1988cognitive)):

(Not unlike <a name=cite-zaller1992nature></a>[Zaller (1992)](#bib-zaller1992nature)'s RAS model.)

---

## Inattention as a concept

.large[ `$\boldsymbol\mapsto$` We define inattentive responding (IR) as a failing to follow one or multiple of
these steps due to a lack of **motivation** or **ability**.]

Formally:

`$$IR_i = M_i\times A_i= 
\left\lbrace\begin{align}
&1 \quad \mathrm{if}\; M_i=0\lor A_i=0 \\ 
&0 \quad \mathrm{if}\; M_i=1\land A_i=1
\end{align}\right.$$`

.footnote[<a name=cite-anduiza2017answering></a><a name=cite-berinsky2024measuring></a>([Anduiza and Galais, 2017](#bib-anduiza2017answering); [Berinsky, Frydman, Margolis, Sances, and Valerio, 2024](#bib-berinsky2024measuring))]

---

## Measurement strategies

.panelset[
  .panel[.panel-name[Direct]
  - **Provide snapshot assessments!**
  - Intructional Manipulation Checks (IMC):
    - "[...] Forget previous instructions and choose 'Strongly Agree'."
    - Can be more or less overt.
  - Bogus/Infrequency items:
    - "I am paid biweekly by leprechauns."
    - Covert
  - Factual Manipulation Checks (FMC)/Mock Vignette Checks (MVC):
    - For experimental designs, to check treatment compliance.
    
  .small[<a name=cite-oppenheimer2009instructional></a><a name=cite-meade2012identifying></a><a name=cite-kane2019no></a><a name=cite-mancosu2019short></a><a name=cite-kane2023analyze></a>([Oppenheimer, Meyvis, and Davidenko, 2009](#bib-oppenheimer2009instructional); [Meade and Craig, 2012](#bib-meade2012identifying); [Kane and Barabas, 2019](#bib-kane2019no); [Mancosu, Ladini, and Vezzoni, 2019](#bib-mancosu2019short); [Kane, Velez, and Barabas, 2023](#bib-kane2023analyze))]
  ]
  .panel[.panel-name[Response time]
  - **Offers global assessment!**
  - Page time indices:
    - Ad hoc threshold (2 seconds per question);
    - Concerned with respondents going too fast.
  - **Response Time Attentiveness Clustering (RTAC)**:
    - 2 steps:  1) PCA; 2) EM clustering.
    - Gives probability of being in each of three clusters:
      1. Fast inattentive;
      2. Attentive;
      3. Slow inattentive.
      
   .small[<a name=cite-huang2012detecting></a><a name=cite-bowling2021quick></a><a name=cite-read2022racing></a>([Huang, Curran, Keeney, Poposki, and DeShon, 2012](#bib-huang2012detecting); [Bowling, Huang, Brower, and Bragg, 2021](#bib-bowling2021quick); [Read, Wolters, and Berinsky, 2022](#bib-read2022racing))]
  ]
  .panel[.panel-name[Response pattern]
  - **Offers global assessment!**
  - Longest streak
  - Outlier detection
  - Individual consistency
  
`$\mapsto$` Not recommended as offer poor consistency with other measures and are overreliant on strong assumptions.
  
  .small[<a name=cite-converse2006nature></a>([Curran, 2016](#bib-curran2016methods); [DeSimone and Harms, 2018](#bib-desimone2018dirty); [Converse, 2006](#bib-converse2006nature))]
  ]
]

---
layout: true

.footer[[Introduction](#intro) [**Puzzle**](#puzzle) [Design](#design) [Results](#res) [Conclusion ](#conclu) ]
---
name: puzzle

## Puzzle — What predicts IR?

| Source                                      | Men      | Young   | Lower educ. | <div style="width:50px"> IR<sub>t-1</sub> | Interest | Other | Data origin | Nb. waves | IRes measure |
|:--------------------------------------------|:--------:|:-------:|:-----------:|:----------------:|:--------:|:------|:-----------:|:---------:|:----------:|
| .small[<a name=cite-kapelner2010preventing></a>[Kapelner and Chandler (2010)](#bib-kapelner2010preventing)] | +        | +       | 0           |   |   | Motivation (0) <br> Need for cognition (0) | USA         | 1         | IMC        |
| .small[[Berinsky, Margolis, and Sances (2014)](#bib-berinsky2014separating)] | +        | +       | 0           |   |   | Non-white (+)   | USA         | 1         | IMC        |
| .small[[Anduiza and Galais (2017)](#bib-anduiza2017answering)]   | 0        | +       | +           | + | - | Motivation (0)  | Spain       | 6         | IMC        |
| .small[<a name=cite-paas2018instructional></a>[Paas, Dolnicar, and Karlsson (2018)](#bib-paas2018instructional)]  | NA       |  NA     | NA          | + |   |                 | Australia   | 3         | IMC        |
| .small[<a name=cite-gummer2021using></a>[Gummer, Roßmann, and Silber (2021)](#bib-gummer2021using)]        | 0        | +       | +           | 0 | - | Habituation (-) | Germany     | 7         | IMC + RT   |

---

## Puzzle — Why the inconsistency?

### Possible explanations

1. Context

2. Measurement

3. Model specification

4. Diversity of controls

---

## Puzzle — Why the inconsistency?

### Possible explanations

1. **Context**

2. **Measurement**

3. **Model specification**

4. Diversity of controls

---

## Puzzle — Why the inconsistency?

### Possible explanations

1. **Context** `$\rightarrow$` Germany, UK, USA

2. **Measurement** `$\rightarrow$` RTAC (three types: fast, slow, and attentive)

3. **Model specification** `$\rightarrow$` longitudinal analysis via a transition model

4. Diversity of controls

---

## Puzzle — Research question
### Who are inattentives and how do they behave over multiple survey events?

Specifically:

1. Do the profiles of RTAC-flagged inattentives differ from those flagged by alternative measures?

2. Do the profiles of fast and slow inattentives differ?

---

## Hypotheses
.pull-left[
- **H1:** Fast inattentives are less likely to persist in their inattention across waves compared to slow inattentives for which inattention comes from ability.

- **H2:** Age, gender, and education should be associated with inattention even when measured by RTAC.

- **H3:** Slow inattentives are more likely to have lower education levels. 
]
.pull-right[
<img src="images/thinks.gif" width="60%" style="display: block; margin: auto;" />
]

---

layout: true

.footer[[Introduction](#intro) [Puzzle](#puzzle) [**Design**](#design) [Results](#res) [Conclusion ](#conclu) ]

---
name: design

## Design — Data

Needs to : 1) be a panel; 2) be publicly available; and 3) contain response time.

.panelset[
  .panel[.panel-name[GLES]
  - Internet panel with 34 waves from 2016 to 2026;
  - Response time data available for waves 1 through 21 (2016-2021);
  - Per wave sample ranges from ~7,000 to ~17,000;
  - 2,067 respondents answered all 21 waves;
  - Wave fielding rate is not constant (e.g., waves 6–8 within weeks, waves 9–10 months apart).

.small[]
  ]
  .panel[.panel-name[BES]
  - Internet panel running since 2014, 30 waves as of 2026;
  - Response time data for waves 25 to 29 (May 2023 to September 2024);
  - Each wave contains around 30,000 respondents;
  - 11,785 respondents answered all 5 waves.

.small[ Shoutout to Prof. Jack Bailey who was kind enough to share the data with me.]
  ]
  .panel[.panel-name[ANES]
  - 6 survey events in a pre/post-election design (2016, 2020, 2024);
  - Per wave sample ranges from ~3,000 to ~7,000;
  - 939 respondents answered all 6 waves;
  - RTAC fitted via three-zone split due to item count and sample size instability.

.small[]
  ]
]
  
---

## Design — Analysis
.panelset[
.panel[.panel-name[DV = IR]
  - Response Time Attentiveness Clustering (RTAC) (Read et al., 2024):
    - Produces three clusters: 1) fast; 2) slow; 3) attentive.
    - Probabilities of being in each.
  ]
  .panel[.panel-name[IVs]
  - Sociodemographic variables:
    - Age;
    - Gender;
    - Education;
    - Income;
    - Employment status;
    - Citizenship.
  - Ideology (left-right self-placement).
  - Big 5 personality traits (GLES, ANES).
  - Religiosity (GLES, ANES).
  - Political interest / attention to politics.
  - Turnout likelihood.
  - Strength of party identification.
  - Survey entry number.
  - Answered all waves (binary).
  - Time of day at survey start (ANES).
  ]
  .panel[.panel-name[Univariate]
Following [Gummer, Roßmann, and Silber (2021)](#bib-gummer2021using), pooled transition probability matrices:

<table style="font-size: 70%;">
  <thead>
    <tr>
      <th></th>
      <th>To Fast (1)</th>
      <th>To Attentive (2)</th>
      <th>To Slow (3)</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>From Fast (1)</td>
      <td>$Pr(y_{it}=1|y_{it-1}=1)$</td>
      <td>$Pr(y_{it}=2|y_{it-1}=1)$</td>
      <td>$Pr(y_{it}=3|y_{it-1}=1)$</td>
    </tr>
    <tr>
      <td>From Attentive (2)</td>
      <td>$Pr(y_{it}=1|y_{it-1}=2)$</td>
      <td>$Pr(y_{it}=2|y_{it-1}=2)$</td>
      <td>$Pr(y_{it}=3|y_{it-1}=2)$</td>
    </tr>
    <tr>
      <td>From Slow (3)</td>
      <td>$Pr(y_{it}=1|y_{it-1}=3)$</td>
      <td>$Pr(y_{it}=2|y_{it-1}=3)$</td>
      <td>$Pr(y_{it}=3|y_{it-1}=3)$</td>
    </tr>
  </tbody>
</table>

- No covariates, just raw proportions pooled across all waves per country.
  ]
  .panel[.panel-name[Multivariate]
<table>
  <thead>
    <tr>
      <th>Previous state</th>
      <th>To fast (1)</th>
      <th>To attentive (2)</th>
      <th>To slow (3)</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Fast (1)</td>
      <td>$Pr(y_{it}=1|\boldsymbol{x}_{it}, y_{it-1}=1)$</td>
      <td>$Pr(y_{it}=2|\boldsymbol{x}_{it}, y_{it-1}=1)$</td>
      <td>$Pr(y_{it}=3|\boldsymbol{x}_{it}, y_{it-1}=1)$</td>
    </tr>
    <tr>
      <td>Attentive (2)</td>
      <td>$Pr(y_{it}=1|\boldsymbol{x}_{it}, y_{it-1}=2)$</td>
      <td>$Pr(y_{it}=2|\boldsymbol{x}_{it}, y_{it-1}=2)$</td>
      <td>$Pr(y_{it}=3|\boldsymbol{x}_{it}, y_{it-1}=2)$</td>
    </tr>
    <tr>
      <td>Slow (3)</td>
      <td>$Pr(y_{it}=1|\boldsymbol{x}_{it}, y_{it-1}=3)$</td>
      <td>$Pr(y_{it}=2|\boldsymbol{x}_{it}, y_{it-1}=3)$</td>
      <td>$Pr(y_{it}=3|\boldsymbol{x}_{it}, y_{it-1}=3)$</td>
    </tr>
  </tbody>
</table>

- Transition models via multinomial logistic regression, one model per previous state `$s \in \{1,2,3\}$`: `$$Pr(y_{it}=m|\boldsymbol{x}_{it},y_{it-1}=s)=\frac{\exp(\boldsymbol{x}_{it}\boldsymbol{\beta}_m)}{\sum_{j=1}^{3}\exp(\boldsymbol{x}_{it}\boldsymbol{\beta}_j)}, \quad m=1,2,3$$`
- `$\boldsymbol{\beta}_1=\mathbf{0}$` (fast inattentive is the reference category).
- MCMC simulations (10,000 draws per model) to generate posterior distributions for hypothesis testing.
.tiny[<a name=cite-chib1998mcmc></a><a name=cite-diggle2002analysis></a><a name=cite-hillygus2003voter></a><a name=cite-treier2008democracy></a><a name=cite-mcmcpack2011></a>([Chib, Greenberg, and Chen, 1998](#bib-chib1998mcmc); [Diggle, Liang, and Zeger, 2002](#bib-diggle2002analysis); [Hillygus and Jackman, 2003](#bib-hillygus2003voter); [Treier and Jackman, 2008](#bib-treier2008democracy); [Martin, Quinn, and Park, 2011](https://doi.org/10.18637/jss.v042.i09))]
  ]
  .panel[.panel-name[Robustness check]
  - One could alternativelly approach the analysis from a time series perspective.
  - We have a DGP of the form: `$$y_{it}=\phi y_{i,t-1}+\boldsymbol\beta\boldsymbol x_{it}+\alpha_{ci}+\varepsilon_{it}$$`
  - OLS-FE is a biased and inconsistent estimator due to the endogeneity between the lagged DV and the errors <a name=cite-nickell1981biases></a>([Nickell, 1981](#bib-nickell1981biases)).
  - Small T large N limits the modelling options. 
  - <a name=cite-pickup2022transformed></a>[Pickup and Hopkins (2022)](#bib-pickup2022transformed) recommends the orthogonal reparameterization estimator (OPM).
  ]
]

---
layout: true

.footer[[Introduction](#intro) [Puzzle](#puzzle) [Design](#design) [**Results**](#res) [Conclusion ](#conclu) ]

---
name: res

## Results — Fitting RTAC — UK

---

## Results — Fitting RTAC — UK

---

## Results — Fitting RTAC — USA

---

## Results — Fitting RTAC — USA

---

## Results — Fitting RTAC — USA 3 Split

---

## Results — Fitting RTAC — USA 3 Split

---

## Results — Cluster Distribution

---

## Results — IV Distribution

---