Research Summaries

Question: Does joining a union make someone more liberal and active in politics?

Why is it important?: People are interested in how powerful labor unions are in politics. One way that labor unions have long been considered powerful by politicians, scholars, and activists is that labor unions can influence their members' political behavior. The hope was that if they have this power then they could ensure that there was reliable bloc of voters who wanted more economically egalitarian policies and that union members would participate (e.g., volunteer, donate, etc.) on behalf of candidates who supported those policies.

What did we know?: There was a lot of research on this question, but there were two key shortcomings. First, previous research tended to use cross-sectional regressions that were vulnerable to omitted variable bias, however, they consistently found a strong correlation between being a union member and liberal political attitudes and being more active in politics. Second, research that used causal inference research designs tended to either not actually target the estimand of interest, whether joining a union changed one's politics, still were vulnerable to omitted variable bias, or had data issues. Nevertheless, all of these results also suggested that joining a union would lead to more liberal political attitudes and political activity. Therefore, before starting the project, the evidence was strongly on one side; joining a union should make someone more liberal and politically active.

Research design: I collected 14 panel studies with close to 50,000 respondents from the 1950s to the present. These surveys asked the same survey respondents the same questions about their political behavior and union membership at least twice. Before getting into technical details, I want to lay out the intuition about what this allows me to do. First, by looking at the same person at two points in time, I can investigate how their attitudes changed over time. Second, I can then compare how two people's attitudes changed across time when one of those individuals join a union and the other person does not. In other words, we have two differences, the difference over time and the difference between people. In the literature, that is called a difference-in-differences design. 

This improves upon previous work in two key ways. First, because we are comparing the same people to themselves across time, characteristics about individuals that generally do not change over time (e.g., race) are automatically controlled for regardless of whether they are observed or not. Second, because I collected so much data on so many time periods with 89 outcomes, I can get a more comprehensive look at the effects. If there are no effects with this much data and with this many tests, we can be more confident that there is little evidence of an effect.

The details are a bit more complicated. Some panels had more than two time periods and some respondents joined their labor union at different points in time. To deal with this staggered membership, I use two-way fixed effects estimators that are robust to treatment cohort heterogeneous treatment effects (see this for a nice review). In my case, I use the Callaway and Sant'Anna estimator where I find the weighted average of the effect across different cohorts. But the basic intuition is the same. I am comparing people across time and then comparing those who joined a union against those who did not join a union. 

The key assumption behind this approach is the parallel trends assumption. That is, the people who joined a labor union would change similarly to those the folks who did not join a labor union. There is no way to prove that this assumption holds, but I conduct a few tests to show that there is limited evidence of pre-trends biasing the estimates.

Results: With all of these data and the estimator, I conducted 347 regressions to show that there is little evidence that joining a union changes one's political attitudes or activity. Below, I show the key figure from the paper. I look at 12 key political attitudes for which we have strong priors that they should be affected by union membership. I precision weight, or take the weighted average weighting by the inverse of the standard errors, all of the estimates related to these 12 outcomes. I show that the precision-weighted estimates are all near zero and are much smaller than previous estimates and that they are actually more likely to be negative than positive, or in other words, union membership is more likely to make someone conservative than liberal. And, these estimates suggest that there is little evidence for an effect even without multiple testing corrections. I show how robust this finding is by looking at voter turnout, voter registration, heterogeneous effects by time period, heterogeneous treatment effects by race, controlling for other characteristics that might change due to union membership, and a whole host of other checks. I consistently find the same thing: there is very little evidence that joining a union changes someone's politics which suggests that one of the key ways that we had believed labor unions exerted power in the United States for over close to a century may not hold up.

Unfortunately, this paper is not open access, but I'm happy to share it if you email me.

Question: How do American workers want their workplaces to be structured and are they willing to take on more managerial responsibilities if it means more work?

Why is it important?: We want to know how to make workers' lives better. Given that the vast majority of people spend much of their waking hours at work, we wanted to understand if they would prefer more of a say over their company's corporate governance, or the people in charge at work.

What did we know?: When my co-author and I started this project, there was little empirical work written about the topic. There were some older empirical works on labor unions and there was a large political theoretical tradition on the topic, but no one had directly asked people. This also meant that there was a lot of basic research that needed to be done.

Research design: We fielded two surveys, one a conjoint experimental survey on YouGov and the other a survey experiment on Lucid. Starting with the conjoint survey, we wanted to know which company an individual would rather work for, one where they would have a say over who their manager was or a more traditional workplace, so we randomly assigned two company profiles in two job offers and asked the survey respondent which job offer they would accept. However, there are a lot of features about job offers besides the management structure. People are typically interested in how much they would be paid, the quality of health insurance, retirement benefits, and other features about the company. We randomized all of those features as well, creating a fully randomized factorial experiment where respondents were asked to choose between two randomly generated job offers. We asked 1,002 respondents to compare two firms four times, giving us 8,016 respondent-firm choices. 

Afterwards, we analyzed the data by regressing a binary variable representing whether a respondent wanted to work at the firm or not on dummy variables representing the specific combination of firm features that the respondent was randomly assigned to see about the firm using ordinary least squares and clustering the standard errors at the respondent-level to account for correlated errors within respondents. This regression provides the average marginal component effect (AMCE), or the effect of a specific feature about the firm conditional on all of the firm features. We were also interested in who might want more of a say over who their boss was because several Democratic senators proposed legislation that would provide workers seats on corporate boards. We used generalized random forests to predict the individual-level treatment effects and then looked at the distribution of effects across key covariates, like partisanship.

We also conducted a traditional survey experiment with 2,105 respondents because we were concerned that respondents did not truly internalize the potentially large costs from having more managerial duties. It's easy to say that you want more of a say in the workplace if you don't consider the extra time that you have to spend on work. To that end, we randomly assigned respondents to receive information about the benefits, costs, or both benefits of costs of having more say over corporate governance, or no additional information. Afterwards, we simply plotted respondents' preferences for different corporate governance structures across the experimental conditions because we could leverage randomization.

Results: Across both experiments, we found 1) that workers wanted more of a say over how their workplaces operated, 2) they wanted more of a say even when we told them about the personal costs that they might pay (e.g., more managerial responsibilities), and 3) there was mixed evidence of partisan polarization on the topic. I will focus on the last part because it will give you a glimpse of the overall preference while also showcasing the distribution of preferences across an important covariate. 

The figure below shows the predicted individual-level effects of a type of corporate governance structure compared to private ownership by non-workers. I will walk through the top left facet first to build intuition. Every dot represents an individual survey respondent's predicted relative preference for having workers on the corporate board compared private ownership by non-worker shareholders on their preference to work at the firm. The bars around each dot is the predicted 95% confidence interval around that predicted individual-level effect. Blue dots represent Democrats, red dots reprsent Republicans, and white dots are Independents. The gray line and bar represents the average treatment effect and the accompanying 95% confidence interval. That top left facet shows that Americans on average prefer to work at firms that have workers on the corporate board, or codetermination, and Democrats seem to like this corporate governance structure more than Republicans. The next two facets to the right show that workers on average prefer working at firms where workers are shareholders and even prefer firms where workers can elect their managers. We found little evidence of partisan polarization on firms where workers are shareholders but we saw strong evidence of partisan polarization when workers elect managers. 

We also asked respondents whether they thought workers have more power at the firm as a manipulation check. We should expect that respondents understand that workers would have more power under these different corporate governance structures, and that's what we found. Across all three corporate governance structures, respondents understood that workers would have more power and we found little evidence of partisan polarization. 

This is just one key result, but I encourage you to check out the rest of the article (it's open access!).

Question: Do women have more negative interactions with voters than men do?

Why is it important?: I was asked by NextGen America, a political organization, to investigate whether women volunteers systematically received more negative responses when texting voters about political issues. They had received complaints from their women volunteers that they felt they would receive sexist and offensive responses, while men volunteers rarely reported receiving such messages. To that end, I proposed running two field experiments with them to investigate how prevalent and severe the problem was so that the organization could adapt their texting strategy to deal with these issues.

What did we know?: There was a great deal of work showing that sexism is a problem in politics and the workplace, however, there was little work investigating the causal impact of interpersonal interactions when one individual was a woman. Furthermore, much work focused on political violence, an obviously important topic, but we wanted to understand whether discrimination leaked into these more mundane political interactions.

Research design: We fielded two field experiments with over 130,000 voters in total. The randomization was straightforward. Voters were randomly assigned to receive a text from a volunteer using a female name, male name, no name, or a gender-neutral name. We then recorded how often voters responded with an offensive, silencing, and withdrawal text. Offensiveness meant the text included language that was meant to insult the volunteer (e.g., "No. Your fat."). Silencing meant the text impolitely ended conversations with the volunteer (e.g., "STFU"). Finally, withdrawal texts were firm but polite messages that ended the conversation. We had a team of research assistants code the text responses for those three elements. We analyzed the data by simply regressing the outcome on dummy variables representing the experimental conditions with the gender-neutral condition serving as the baseline using ordinary least squares with robust standard errors.

Results: Across all three measures, we found that voters were more likely to respond with offensive, silencing, and withdrawal messages when they perceived the volunteer to be a woman. Furthermore, female names were consistently more likely to receive these responses than male names. Here is the result for offensiveness. Female names received more offensive replies from voters when compared against the gender-neutral name condition and the male name condition. Interestingly, female names and no names received similarly offensive messages. One potential explanation is that people were more likely to believe that the no name condition was a bot and thus they felt more willing to reply offensively. 

I encourage you to read the rest of the study (it's open access!).