How does color influence our political conversations?

Damon C. Roberts

Background

Key findings from other chapters

Question guiding this chapter

Does color influence whether we talk about politics?

Building a theory of snap-judgment based on politically-relevant visual information

A common language

Valance: a positive or negative response to an object.
Affect (or affective state): feeling positive or negative in response to an object.
- Pre-conscious response
Emotion: Much more complex labels assigned to affective responses.
- Often requires some amount of conscious processing to assign an appropriate label.
- Slower
Attitudes are an expression of valanced responses to some object (Fazio 2007).

A cognitive architecture

Memories are stored prior experiences (Cunningham, Haas, and Jahn 2012).

often encode affective responses

Memories are multi-dimensional (Kahana, Diamond, and Aka 2022).
Sensory and affective information is often encoded in memories (Kensinger and Fields 2022).

Memories are a complex network (Kahana, Diamond, and Aka 2022).

Interconnected by different bits of information

Social groups and hierarchies are reflected in this network (Santavirta et al. 2023).

A snap-judgment model of politically-relevant visual information

%%{init: {'theme':'base', 'themeVariables':{'primaryColor':'#ffffff', 'primaryBorderColor': '#000000'}}}%%
graph TD
    A[Detection of politically-relevant visual information] --> B(Activation of retrieval mechanism)
    B --> C[Memory and contiguous paths hot]
    C --> D[Appraisal of hot autonomous nervous system]
    D --> E1[Negative Valence]
    D --> E2[Positive Valence]
    E1 --> F1[Disengagement motivation]
    E2 --> F2[Attention activated]
    F1 --> |Neg. new information| G1[Encoded as negative]
    F1 --> |Pos. new information| G2[Encoded as positive]
    F2 --> |Neg. new information| G3[Encoded as negative]
    F2 --> |Pos. new information| G4[Encoded as positive]
    G1 --> H1[Strengthens negative path]
    G2 --> H2[Weakens negative path]
    G3 --> H3[Weakens positive path]
    G4 --> H4[Strengthens positive path]
    G2 --> |forgotten| H1
    H2 --> |reinforced| G2
    H3 --> |reinforced| G3
    G3 --> |forgotten| H4

Figure 4: Snap-judgment model of politically-relevant visual information

Hypotheses

People notice the color of clothing a potential conversation partner is wearing.
When primed to think about politics, people are less likely to want to have a conversation with someone wearing a blue shirt if they are a Republican and a red shirt if they are a Democrat.

When learning more information about someone that fits with our initial impression, our motivation of engagement or disengagement is stronger than if the information were incongruent with our initial impressions.
When learning more information about someone that does not fit with our initial impression, the new information can shift our motivations to have the conversation or to disengage. However, this difference is smaller than if both forms of information were congruent.

Proposed research design

Study 1

Table 1: \(2 \times 2\) factorial design
🚫 Prime, Red shirt	Prime, Red shirt
🚫 Prime, Blue shirt	Prime, Blue shirt

Prime condition: political attitudes questionnaire before treatment vs. after
Person wearing shirt might be White Female (but would like some feedback on this)

Study 2

Table 2: \(2 \times 2\) factorial design
Red shirt, 🚫 congruent	Red shirt, congruent
Blue shirt, 🚫 congruent	Blue shirt, congruent

Congruency is whether or not policy attitudes after image matches with presumed partisanship from color of shirt
- Policy positions will be on: Support for Gun Control, whether pro-choice or pro-life, support for Universal Basic Income

Outcomes

\(H_1\): What color shirt was your political discussion partner wearing?
\(H_2\) & \(H_3\): To what extent are you willing to have a conversation with this person? (5-item response)

References

Cunningham, William A., Ingrid Johnsen Haas, and Andrew Jahn. 2012. “Attitudes.” In The Oxford Handbook of Social Neuroscience, edited by Jean Decety and John T. Cacioppo. New York: Oxford University Press. https://doi.org/10.1093/oxfordhb/9780195342161.013.0013.

Fazio, Russell H. 2007. “Attitudes as Object-Evaluation Associations of Varying Strength.” Social Cogition 25 (5): 603–37. https://doi.org/10.1521/soco.2007.25.5.603.

Kahana, Michael J., Nicholas B. Diamond, and Ada Aka. 2022. “Laws of Human Memory.” In Oxford Handbook of Human Memory, edited by Henry L. Roediger III and Okyu Uner. New York: Oxford University Press. https://memory.psych.upenn.edu/Oxford\_Handbook\_of\_Human\_Memory.

Kensinger, Elizabeth A., and Eric Fields. 2022. “Affective Memory.” In Oxford Handbook of Human Memory, edited by Henry L. Roediger III and Okyu Uner. New York: Oxford University Press. https://memory.psych.upenn.edu/Oxford\_Handbook\_of\_Human\_Memory.

Santavirta, Severi, Tomi Karjalainen, Sanaz Nazari-Farsani, Matthew Hudson, Vesa Putkinen, Kerttu Seppälä, Lihua Sun, et al. 2023. “Functional Organization of Social Perception Networks in the Human Brain.” NeuroImage 272: 1–16. https://doi.org/10.1016/j.neuroimage.2023.120025.

Appendix

Sensitivity of conclusions to sample size, research design, and modeling strategies

the results of a bayesian model are kind of like an average of your prior beliefs and your data that you have on hand
- what this means is that I start off with a belief about how things are, then add data to see if that distribution changes
for hypothesis testing, I’ll start off by saying that the most likely outcome is that my hypotheses are wrong and that the effect is likely 0.
- though, I’ll allow for some uncertainty to be expressed by saying that there is some amount of probability that the effect could be positive or negative, but most effects are going to be pretty small substantively speaking.
This is reflected with a normally distributed prior and a standard deviation of 1
Now my task is to see whether, given a DGP that I know how it looks in the population, my particular research design and modeling strategies can produce a posterior destribution that help me conclude that the effects are not zero when my DGP says that they aren’t zero. That is, can my research design and modeling strategies bring me to the correct conclusions and provide evidence in support of my hypotheses as opposed to the null hypothesis?

\[ \begin{align} \text{Age} = uniform(18, 85) \\ \text{Gender} = bernoulli(0.5) \\ \text{Education} = normal(-0.1 * Age + 0.1 * gender + \\ 14 + normal(0, 1), 0.5) \\ \end{align} \]

\[ \begin{align} \text{Income} = normal(2 * age + 0.7 * gender + \\ 40000 + normal(0, 1), 200000) \\ \text{Conflict Avoidant (latent)} = normal(0.3 * gender + \\ 2.4 + normal(0, 1)) \\ \end{align} \]

\[ \begin{align} \text{Party identification (latent)} = normal(0.4 * age - \\ 0.6 * gender + 0.5 * income + normal(0, 1)) \\ \text{attention (latent)} = normal(0.5 * age - \\ 0.3 * gender + 0.1 * income + normal(0, 1)) \end{align} \]

\[ \begin{align} \text{treatment}_{prime} = bernoulli(0.5) \\ \text{treatment}_{blue} = bernoulli(0.5) \\ \end{align} \] \[ \begin{align} \text{notice} = bernoulli(1 * prime) \\ \text{willing (latent)} = normal(0.1 * treatment_{prime} + \\ 0.1 * treatment_{blue} + 0.1 * \text{party identification} + \\ 0.5 * treatment_{prime} * treatment_{blue} * \text{party identification} + \\ normal(0, 1)) \end{align} \]

Fitting the models on the simulated data

Model 1

\[ \begin{align} Pr(notice_i = 1) \sim logistic(\phi_i + \alpha) \\ \phi_i = \beta_1 * treatment_{prime_i} \\ \beta_1 \sim Normal(0, 1) \\ \alpha_i \sim student(0,3,2.5) \end{align} \]

Model 2

\[ \begin{align} Willing_i \sim CDF(logistic(\phi_i + \alpha)) \\ \phi_i = \beta_1 * treatment_{blue_i} + \beta_2 * treatment_{prime_i} + \\ \beta_3 * treatment_{blue_i} * treatment_{prime_i} + \\ \beta_4 * gender_i + \beta_5 * education_i + \\ \beta_6 * income_i + \beta_7 * conflict_i + \beta_8 * attention_i \\ \beta_j \sim Normal(0, 1) \\ \alpha_j \sim student(0, 3, 2.5) \end{align} \]

Model 2 (alternate)

\[ \begin{align} Willing_i \sim CDF(logistic(\phi_i + \alpha)) \\ \phi_i = \beta_1 * treatment_{blue_i} + \\ \beta_2 * treatment_{prime_i} + \beta_3 * \text{Party ID}_i + \\ \beta_4 * treatment_{blue_i} * treatment_{prime_i} * \text{Party ID}_i + \\ \beta_4 * gender_i + \beta_5 * education_i + \\ \beta_6 * income_i + \beta_7 * conflict_i + \beta_8 * attention_i \\ \beta_j \sim Normal(0, 1) \\ \alpha_j \sim student(0, 3, 2.5) \end{align} \]

Performance

Model 1

Model 2