CEO Gender and Beliefs about Firm Performance: an experimental approach

The experiment is an online study ran on Prolific. The main task that participants are asked to perform is to forecast firms’ profits upon the provision of information.
Main Randomization: Participants will be asked to review information about and make predictions for the profits of three real US firm-quarter pairs. For two out of the three pairs, all the information provided will be true. But for one randomly chosen firm-quarter pair out of the three, there will be a fictional component. This fictional component will be the identity of the CEO, which will be female in case the real CEO is male, or vice-versa. For each participant, the three firm-quarter pairs will be randomly selected out of a population of 48 firm-quarter pairs – 24 whose actual CEO in a given time period is a woman and 24 whose CEO is a male.
For each of the three firm-quarter pairs, there will be eight steps that participants will be asked to complete the following steps:
1. Read information about the firm and the CEO.
2. Observe a graph with past profits (past three quarters). Respondents will be asked for a qualitative prediction for the next quarter.
3. Observe the same graph with past profits, but now with an additional feature: two expert forecasts for profits in the next quarter.
4. Make predictions for profits in the next two quarters.
5. Build a confidence interval around the prediction for the second quarter.
6. Learn the actual profit in the first quarter. Make a new prediction for the second quarter.
7. Answer understanding questions.
8. Answer evaluation questions
After completing the eight steps for each of the three firm-quarter pairs, participants will be asked to answer standard demographic and financial literacy questions.

We embed two other randomizations in the experiment:
Optimism vs Pessimism Treatment: For each firm-quarter pair, we randomize whether the pair of experts’ forecasts is optimistic (in the sense that forecasts are higher than the actual realization) or pessimistic (in the sense that forecasts are lower than the actual realization). By providing experts’ forecasts, we aim to benchmark individual’s forecasts. This randomization then allows us to exogenously induce a positive revision (good news)
or negative revision (bad news) in step 6.
"Before and After" vs "Only After" Treatment: For each participant, we randomize whether evaluation questions will be answered "only after" their forecast revision, in step 8, or will be answered both "before and after" making this revision. In the latter treatment, we add a page between steps 1 and 2 where respondents answer evaluation questions. The purpose of this randomization is twofold. It allows us to observe whether "prior" answers are different from "posterior" answers. It also works as a salience treatment: since one of the questions points to the role of the CEO in affecting firm performance,
the relevance of the CEO should be more salient for respondents who answer the evaluation questions before making predictions.

Please read section above "Intervention Public."

2025-03-04

2025-03-31

Forecast errors for profits in the second quarter.

A forecast error is defined as the actual profit in a given quarter minus the participant’s forecast for that quarter. We are interested in the forecast errors based on both participant’s first ("prior forecast error") and second ("revision forecast error") forecasts made for the second quarter in the experiment.

Levels: We are interested in whether the average level of beliefs depends on CEO gender. We plan to test this by regressing forecast errors on a constant and an indicator for whether the observed CEO was female. The coefficient on this indicator indicates whether there is an effect of CEO gender on the level of beliefs. For example, if this coefficient is positive, then participants are more pessimistic about female-led companies relative to their male-led counterparts.

Updating depending on CEO gender: We are also interested in whether the up- dating of beliefs depends on CEO gender. To diagnose this, we plan to run a regression of the "revision forecast error" on a variable called "forecast revision" interacted with the indicator for whether a female CEO was observed. Variable "forecast revision" is defined as the difference between the second and first forecasts made by a participant for the second quarter in the experiment. The rationale for this regression is the following: under full information rational expectations, the coefficient of a regression of forecast errors on forecast revisions should be exactly equal to zero (Coibion and Gorodnichenko, 2015). A positive coefficient suggests under-reaction relative to rational expectations, while a negative coefficient suggests over-reaction. Hence, in our regression, the coeffi- cient on "forecast revision" captures if the updating of beliefs about male-led companies systematically deviates from rational expectations. In turn, the coefficient on "forecast revision x 1(female CEO)" captures if the updating of beliefs about female-led companies systematically deviates from that of their male-led peers.

Updating depending on CEO gender and sign of the news: We are further interested in understanding if the effect of CEO gender on beliefs is intermediated by whether participants observed good news or bad news. Since a forecast revision captures the participant’s interpretation of the underlying information received, we define good news as a positive forecast revision and bad news as a negative forecast revision. We then interact the independent variables of interest with an indicator of the sign of the news (sign of the forecast revision).

Additional dimensions of heterogeneity: Another dimension we are interested in investigating is whether potential biases according to CEO gender depend on participants own gender. To analyze this, we further interact the independent variables of interest with an indicator for the gender of the participant. We might also explore heterogeneity across other participant characteristics (such as political affiliation, age, financial literacy, etc.).

Note: In our main analysis, we aim to isolate variation that is due to CEO gender and not to a participant’s characteristics or to the specific firm-quarter pairs that a participant observed. Hence, we plan to control our regressions for participant fixed effects and by firm-quarter fixed effects.

Qualitative prediction for profits in the next quarter; Confidence intervals on point forecast; Evaluation questions.

Qualitative prediction: Participants observe a graph with profits in the last three quarters, and are asked whether profits in the next quarter will increase (strongly or moderately), decrease (strongly or moderately), or stay the same. We call the answer to this question "qualitative prediction" and code it into a 5-point likert scale ranging from -2 to +2. We plan to conduct an analysis of whether these answers differ systematically based on the gender of the observed CEO: regressing the "qualitative prediction" on an indicator for whether the observed CEO is female.

Evaluation questions: We ask three evaluation questions: (i) ‘In your opinion, how challenging is the economic environment navigated by this company?’ (not challenging at all - very challenging); (ii) ‘In your opinion, how qualified is the CEO to lead the company through this economic environment?’ (not qualified at all - very qualified); and (iii) ‘In your opinion, how effectively does the company’s headquarters support the op- erations of this company?’ (not effectively at all - very effectively). The answer to each question is coded into a 7-point likert scale ranging from -3 to +3. We are interested in participants’ answers for the first two questions, while the third question (about the headquarters) is aimed at making participants feel that the pictures of both the CEO and the headquarters, which are shown to them when they review the background of the company, are equally important. These evaluation questions are presented to some (ran- domly selected) participants after they review the scenario for a company – we call these answers "prior evaluation responses" –, and are presented to all participants after they revise their numeric prediction for the second quarter – we call these answers "posterior evaluation responses". We plan to investigate the answers to these questions in three ways. First, we investigate whether the average response (both "prior response" and "posterior response") is different based on CEO gender. Second, we explore whether the average "posterior response" is different based on CEO gender and sign of news. Third, we investigate whether the change from the "prior response" to the "posterior response" is different by CEO gender and sign of news.

Confidence intervals on forecasts: Participants are asked to build a confidence inter- val around their point forecast for the second quarter in the experiment by spreading 100 points across five bins (with values centered at their point forecast). We plan to compute a measure of uncertainty based on these answers using three methodologies. The first methodology is to assume that these answers define a discrete probability distribution over the mid-point of each bin and calculate the standard deviation of this implied dis- tribution. The second and third methodologies are based on measures resembling an interquantile range. To this matter we order bins 1 through 5 (with 1 associated with lowest values for profits, and 5 with the highest) and cumulate the probability across the bins. The two measures of interest are: (i) the difference between the cumulative proba- bilities for the outer bins, 1 and 5; (ii) the difference between the cumulative probabilities for inner bins, 2 and 4. We then plan to study whether the measure of uncertainty differs by CEO gender.

Participants will be asked to review information about and make predictions for the profits of three real US companies. For two out of the three companies, all the information provided will be true. But for one randomly chosen company out of the three, there’ll be a fictional component. This fictional component will be the identity of the CEO, which will be female in case the real CEO is male, or vice-versa. For each participant, the three companies will be randomly selected out of a population of 48 company-quarter pairs – 24 whose actual CEO in a given time period is a woman and 24 whose CEO is a male.

For each of the three companies, there will be six steps that participants will be asked to complete:
1. Read information about the company and the CEO.
2. Observe a graph with past profits and two expert forecasts for profits in the next quarter.
3. Make predictions for profits in the next two quarters.
4. Build a confidence interval around the prediction for the first quarter.
5. Learn the actual profit in the first quarter. Make new prediction for the second quarter.
6. Answer some questions regarding one’s personal opinion.
For each company there is a “good news” versus “bad news” treatment which will be randomly assigned. Under the “good news” treatment, the participant will observe two expert forecasts in step 2 that are both below the actual profit in that period. Hence, in step 5, this participant will receive “good news” because the actual profits will be above what experts expected. Under the “bad news” treatment, the two expert forecasts will be both above the actual profit in that period so that there will be “bad news” in step 5. This allows me to understand if the gender-treatment effects depend on whether good or bad news is observed. Note that under both treatments – good news or bad news – participants are observing true information that has been curated to that specific case. In the database of analysts’ forecasts, the companies in this experiment are among those that have at least four analyst forecasts: at least two forecasts above and at least two forecasts below the realization for a given quarter. After completing the 6 steps for each of the three companies, participants will be asked to answer standard demographic and financial literacy questions.

Please refer to the "Intervention" section above for a thorough explanation of the experiment.

Attention checks: We have two attention checks in the survey where respondents are asked to select a specific answer provided in the question. We exclude from the experiment respondents who fail both attention checks (i.e., these participants are not paid and are removed from the survey completely). Moreover, we plan to include in our analysis only those participants who answered both attention checks correctly (i.e., we plan to exclude from the analysis those participants who fail one attention check).

Understanding questions: For each firm-quarter pair a participant observes, we ask two questions that capture the extent to which respondents understand the tasks given to them. In particular, in step 7, participants are asked about the result for the first quarter (whether it was better, worse or in line with their expectation) and about their forecast revision for the second quarter (whether they revised it up or down, or performed no revision). We plan to include in the analysis only those participants who answered all of the understanding questions correctly (six questions in total – two for each of the three firm-quarter pairs observed).

Incentives: In this survey, we incentivize all nine numeric predictions made by par- ticipants (two predictions made in step 4 and one prediction made in step 6 for three firm-quarter pairs). To that matter, we compute a bonus ranging from $0 to $2 in the following manner (participants are informed of this prior to starting the survey): one out of the nine numeric forecasts made by a participant during the study will be randomly selected and compared to the actual profits in that period, according to the following rule:
• If the selected prediction is within $50 from the actual profit, then the bonus will be determined by the following formula: (1-|Actual Profit-Predicted Profit|/50)×$2.
• No bonus for predictions more than $50 from the actual profit.

Importantly, the bonus is contingent on answering both attention checks and all undestanding questions correctly.

Finally, participants are paid a base compensation of $3 (the average length of the survey is 15 minutes). We expect the average bonus to be around $1.5. Hence, the average total compensation should be around $4.5.

No Deception: Participants are not deceived since we inform them that there will be a fictional component for one of the firm-quarter pairs they review. In particular, participants are told that: "all the information provided is real for two out of the three companies to be reviewed. For one randomly selected company out of the three, there will be a fictional component."

Randomization coded in the software (i.e., done by a computer).

The randomization of the CEO gender is performed within individuals (i.e., at the firm- quarter level). Every individual will have the CEO gender of one of the three firm-quarter pairs altered.

No

We plan to survey 1,500 respondents in total, excluding respondents who fail both attention checks (these respondents are indeed "dropped" directly in Prolific and will not count towards the target of 1,500). However, the target size of 1,500 participants includes those respondents that will be excluded from the analysis, either because they failed one or more of the understanding questions or because they failed one of the attention checks.

Since each respondent will perform the forecasting exercise for three firm-quarter pairs, we plan to have 4,500 firm-quarter-respondent observations in total. Note that some of these observations will be removed from the analysis as explained above.

One third of the firm-quarter-respondent observations are treated (i.e., they see a manip- ulated CEO gender) and two thirds firm-quarter-respondent observations are not treated (i.e., they observe the true CEO gender). This will result in half of the sample observing a male CEO and the other half observing a female CEO.

Half of the respondents will answer the evaluation questions both before and after the forecasting exercise and half of the sample will answer these questions only after the forecasting exercise.

Half of the firm-quarter-respondent observations will be randomly assigned a pair of opti- mistic experts’ forecasts and the other half will be randomly assigned a pair of pessimistic experts’ forecasts.