Experimental Design
Primary experiment:
Subjects will be split into two groups: Senders and Receivers.
There will be several factual questions of the form "the answer is greater/less than [number]." Senders will be told the correct answer, and receivers will give their perceived prior probability P(answer is greater). Each sender will, via the strategy method for each possible receiver prior, choose whether to send a message that either says "the answer is greater than [number]" or "the answer is less than [number]." One of these messages is truthful, and one is false.
The binary choice, "true news" or "false news," is the main dependent variable for senders and main hypotheses (a)-(c). Since senders answer via the strategy method, there will be several (eleven) choices per question for each sender.
There will also be several quotes whose content is either accurate or inaccurate. On these questions, senders will only choose one message instead of one message for each possible receiver prior; otherwise, the definition of true/false is the same.
There is within-subject randomization and between-subject randomization.
Senders:
Between subjects:
- Each sender will either be given reputation incentives, be unincentivized, or be given competition incentives.
- Each sender will either be matched with a receiver who learns the sender's party or a receiver who does not. Senders in the competition incentives condition will not have their party revealed to receivers.
Within subjects:
- For each question, each sender will be randomly assigned to either learn that the receiver is a Democrat, a Republican, or not learn this information.
- Subjects randomly see 7 of 11 "greater/less than" questions, and 4 "quote accuracy" questions. The "greater/less than questions" appear first, but the order within each block is randomized.
Receivers:
Between subjects:
- Receivers will be told that when matched, their sender will either be given reputation incentives, be unincentivized, or be given competition incentives. In the reputation and unincentivized conditions, the receiver will be asked to predict, via strategy method, P(sender's message is truthful). In the competition condition, the receiver will be asked to predict, via strategy method, which of two senders sent more truthful messages over the course of the experiment.
- Receivers will either be matched with a sender who knows their party or a receiver who does not.
Within subjects:
- For each question, receivers in the reputation and unincentivized conditions will be randomly assigned to either learn that the sender is a Democrat, a Republican, or not learn this information. Receivers in the competition condition will not learn the two senders' party.
- Subjects randomly see 7 of 11 "greater/less than" questions, and 4 "quote accuracy" questions. The "greater/less than questions" appear first, but the order within each block is randomized.
Secondary experiment:
Before this experiment, a separate set of subjects ("receivers") were asked to complete a task intended to identify motivated reasoning (Thaler, WP). They were first asked to state their median belief about a factual question on various political and neutral topics. Then, they were given two binary messages that say "The answer is greater than [your median belief]." and "The answer is less than [your median belief]." One of these messages comes from "true news" and the other from "fake news," and receivers are asked to predict the likelihood of each. A Bayesian would assess the news sources as equally likely to be true, while a motivated reasoner would think the party-aligned news would be more likely to be true.
Hypothesis (d): Subjects in this experiment are asked to predict the share of receivers' news judgments that assess Democratic news as being more likely to be true, Republican news as being more likely to be true, or both being equally likely. They will give one prediction for Democratic receivers and one for Republican receivers. I average subjects' predicted motivated reasoning, measured as 1 * party-aligned news more True + 0.5 * both news equally likely + 0 * party-misaligned news more True (such that a Bayesian would have a value of 0.5), for each party.
Hypothesis (e): Subjects are asked to "choose" a message after learning information about the receiver. On each question, subjects are matched with either a Democrat or a Republican who have the same median belief. Subjects learn whether this belief is greater than or less than the true answer, but may choose either message. Messages are not sent to receivers, but some subjects are incentivized to choose messages that receivers believe is true. This binary choice, "true news" or "fake news," is the dependent variable.
Hypothesis (f): Subjects are asked whether they prefer to receive a higher expected payment and learn the receivers' party with probability 1/2, or receive a lower expected payment and learn the receivers' party with probability 1. This binary choice, "demand," is the dependent variable.