This cycle, I designed a small postcard experiment that I outline here. In short, I wrote just over 3,200 postcards to low-turnout Republican voters living in very red areas in Pennsylvania and Florida, arguing that they should change their party affiliation and consider voting for Democrats in November. The four different arguments written on the postcards included highlighting the badness of Republican policies, highlighting the badness of Republican elites, contrasting Democratic and Republican policies, or promoting the general goodness of Democratic policies. Recipients were randomly assigned to receive one of these four different messages, or to a control condition for comparison. Each argument was followed by an ask for these voters to change their voter registration and to consider voting Democratic in November.
Motivation
In Reinventing Political Advertising, Hal Malchow proposed what he called “An Agenda for Better Results” in the field of political communications. This agenda included focusing efforts on
- Lower-turnout voters, who are likely to be more impacted by additional political information because they are “typically ignored by campaigns”
- Lower-information voters, who by definition will learn more from the next marginal political argument they hear than higher-information voters who already have more political information at hand
- Party brands rather than candidates: “Ninety percent of voters are choosing parties. Yet, our approach to advertising has not changed at all. Almost one hundred percent of our advertising dollars are spent on candidate choice… What if we spend that money to promote the Democratic Party and discredit the Republicans? What if our advertising focused on getting votes for all Democrats?”
Malchow also offered some speculation on the value of asking people to change party allegiance directly:
Polling suggests that party-allegiance is more fluid than many would suspect… Surely, if we can convince one voter to become a Democrat, that conversion might benefit candidates in more than one hundred races as opposed to our current strategy of contesting one election.
How much will it cost to change a party allegiance? No one knows. But if we test this proposition we can know. And we can allocate that cost over several races and not just one. Maybe this is not a sound approach but if it works it will influence many more election outcomes over a longer period. We will never know whether advertising for party is a good strategy until we put it to the test.
This cycle, I thought I would take up Hal’s challenge and think about ways to talk to low-information and low-turnout Republicans in ways that focus on party, not candidate, and that take varying approaches to contrasting the party brands. Hal was one of the first people in DC to take me out to lunch and let me ask him dumb questions all afternoon in my early career as a political hack, many years ago during a brief stint of mine at what was then W.V.W.V., and I was excited to see him lay out several new research ideas that struck me as clearly worthy of a test. This experiment is my first attempt at such a test.
This project was a one-man endeavor, covered entirely out of my own pocket, with only the technical resources available to me as an individual. I did not use any of the commercial products - such as propensity scores, professionally designed campaign materials, custom polling, etc. - that a campaign would typically use for a message test of this kind. Neither my employer nor any of my clients had any involvement in this project. I devised the sample and sampling procedure on my own time, on freely available voterfile data. I designed, wrote, and paid for all campaign materials, all the postcards, all the postage, etc., using what knowledge I could bring to bear from my personal experience in my years as a political operative.
Research Design
Target universe: Brand-vulnerable, low-turnout registered Republicans in red areas within Pennsylvania and Florida
The sample universe for this study includes low-turnout Republican voters living in “red” areas of the key battleground states of Pennsylvania and Florida. Pennsylvania and Florida have many attractive features as geographies of study - among them, their free and high-quality voterfiles. These states have partisan voter registration and detailed vote history for those on the file. Their files are updated regularly and reliably. I have a fair amount of experience working with both voterfiles which gave me confidence I would minimize the chance of goofing something up focusing on samples drawn from these states.
I further narrowed the geographic area of study to only include precincts within these states where Donald Trump, the most recent Republican Congressional candidate, the most recent Republican State Senate candidate, and the most recent Republican State House candidate all beat the Democrat they were running against in that precinct, the last time such a contest occurred. This includes the parts of Florida highlighted in red here:
and the parts of Pennsylvania highlighted here:
My reasoning for this limitation is simple: Talking to Republicans is risky, and if the experiment results in backlash, I want to minimize the potential impact on Democratic office-holders.
To pull the samples themselves, I downloaded each state’s voterfile. The Florida sample is based off the state’s February 6, 2024 voterfile vintage and the Pennsylvania sample is based off the state’s February 23, 2024 vintage. To be in the sample universe, I subset the voterfiles to include voters who…
- Reside in a “fully red” precinct, as defined above
- Are registered Republican
- Are marked as “active” on the voterfile
- Have full address information on the voterfile including zip code
- Voted in 2016 but not 2020
- Does not vote in primaries per vote history data available on the voterfile
Those last two bullet points are meant to capture Republicans who may currently be “sour on the brand” (i.e., they voted for Trump once, but not twice), and to include those who are not hyper-involved enough to participate in Republican primaries. This is what I mean by “brand-vulnerable, low-turnout Republicans.”
Altogether, these ended up being some fairly strenuous sample criteria. Many of the voters who would likely qualify for this study are likely registered Independent, but I did not have a propensity model to account for this – all of the population that could qualify for this study had to be registered Republican. Of the roughly 420,000 Floridians who voted in 2016 but not 2020, only 30,679 fulfilled all of the above criteria. Of the roughly 350,000 Pennyslvanians who voted similarly, only 46,439 met all of the above criteria. In other words, the sample universe for this study included just 77,118 voters, 3,200 of whom received postcards.
Prior to finally cutting the sample, I block randomized respondents into treatment or control precincts. The precincts were divided into a control list including 80 percent of the overall geography, and then the remaining 20 percent were randomized into four blocks consisting of 5 percent of precincts apiece, one block of precincts for each of the four message conditions. While individual-level party identification and turnout are measurable from the voterfile, this further level of randomization is necessary to check for aggregate changes in partisan vote choice across treatment and control precincts. The following table summarizes how I divided up the universe into each treatment condition.
In practice, the number of recipients ended up being slightly higher than the design - 401 respondents in Pennsylvania treatment condition 2, 412 respondents in Pennsylvania treatment condition 4, and 401 in each of treatments 2, 3, and 4 in Florida. This happened because I simply mis-counted the number of postcards I’d written on a couple of occasions, and accidentally printed off an extra address label or two from time to time. Rather than throw them away, I included them. Those handful of additional postcards, totaling 16 additional recipients, were drawn from their respective treatment blocks.
The message venue: Handwritten postcards
I chose to use handwritten postcards as the communications means for this study. I did not have much choice in this regard - as a solo project with no budget outside of what I felt like spending, I had no access to phone lists, ad placement services, etc. It is far from clear this is the optimal means of outreach for a project of this kind, but the precision afforded by the postcard approach in terms of reaching who you think you’re reaching is useful, the means of doing so is publicly available, and I knew I could complete the postcards successfully on my own.
I used VistaPrint to buy the postcards, and did so in batch orders placed on November 9, 2023 and on July 15, 2024. The total cost of the postcards used for this study was $245.48, or about $0.07 per postcard including shipping and taxes. On top of that, each postcard was stamped with a $0.56 postcard stamp picked up in rolls of 100 from my local post office, the total cost of which for this project was $1,800.96. Almost all of the stamps had this sailboat design on them but early in the experiment the postcard stamps my post office sold had barns on them, which in my opinion looked much cooler than the sailboats. Such is life. The postcard was a standard 4” by 6” design, with a glossy front and non-glossy back.
I designed the postcard in GIMP, featuring an image I found in the Library Company of Philadelphia’s archive of historic Civil War postage designs. I used the image on this envelope - cleverly titled “Eagle with Union shield envelope” - chiefly because it looks cool, and evokes the patriotic element I felt was important to the narrative around the central ask of the postcard. The typeface used for the glossy front is Caslon Antique Bold, which I felt evoked the kind of old-timey call to political action that might appeal to the recipient. The glossy portion of the postcard read, “Country before party” on top, and then “No more crazies, no more treason! Register Independent today” on the bottom of the glossy front, with the text centered above and below the image.
I printed recipient addresses on my trusty Brother QL-810W label maker on 0.66” * 2.1” DK-1204 labels, which come in rolls of 400 – I bought twelve rolls for the bulk discount for a total cost of $221.00. The labels were printed in the Arial typeface with separate lines for the recipient’s full name, their address, and then their city, state, and ZIP.
While the labels were machine printed, I wrote the actual postcard messages by hand.
The messages: “Democrats are good,” “Democrats are good, Republicans are bad,” “Republican policies are bad,” “Republican elites are bad”
Every postcard followed the format,
Hey, [voter first name]
[Message]
Register Independent at
vote.pa.gov if Pennsylvanian, registertovoteflorida.gov if Floridian
And consider voting Democratic to show no one can take your vote for granted!
On each postcard, the words “Independent” and “Democratic” were written in blue ink and underlined, the URL of the voter registration site for the recipient’s state was written in red ink and underlined, and the remaining text was written in black ink. Several Pilot G-2 pens were sacrificed in the completion of this project, out of a $5.76 pack of pens I bought at my local HEB.
While this project and my day job had zero overlap, it was impossible for me to compose the messages for this experiment without keeping in mind what I’ve learned from polling this cycle. As a pollster, I worked in a few general things I think are true about the narrow subset of voters who are Republicans that are fed up with the Republican brand. This includes:
- These Republicans value Social Security very highly, to the point where many simply refuse to believe the GOP has any intentions to defund Social Security – aided by the kinds of idiotic “fact-checking” that leads to absurd statements like “The Heritage Foundation, but not Project 2025, proposes Social Security cuts” – I think it is obvious the GOP plans to cut Social Security in practice, and I incorporated this into the messaging
- These Republicans are sincere when they say they don’t like crime. This includes not just the crime they believe is rampant in large cities, but also the type of crime for which their party leader has received multiple felony convictions
- These Republicans are much more likely than “MAGA diehards” to support NATO and Ukraine, and are also alarmed by Republican elites’ calls to defund public education
- These Republicans overwhelmingly support Democratic policies that capped the price of insulin and prescription drugs, and overwhelmingly agree with Democrats that the cost of housing is way too high - even though they don’t currently give Democrats credit for any of these things
I incorporated this knowledge into the design of each message. I prefaced each ask with a gentle caveat that I thought might be both inviting and eye-catching while also establishing that what the recipient was about to read was true (“We may not agree on everything, but we all know…”). The messages used in this study included:
Or, as they would look rendered in my at best semi-legible script:
Note that even in the random sample of postcards I drew from the full pile of 3,200, there are minor typos and inconsistencies – for example, sometimes I wrote the longer version of a URL and sometimes the shorter one (with both leading to the same page), and sometimes I corrected around typos and misspellings. In about 200,000 handwritten words there will be an error term! I like the handwritten approach but this experiment has made me skeptical it is worth the tradeoffs - a subject for future research.
Project logistics
But in the meantime, that is the approach I took. Over the course of 2024 - aiming for about 10-20 postcards per day but often more like fits and starts, with more knocked out over a weekend than during the week, no progress during busy travel weeks followed by large bursts while catching up on TV, etc - I wrote the postcards. I ended up filling two entire USPS bins:
These postcards were sent out from my local post office in the Austin, Texas area on September 30, 2024. Here they are at the post office, saying goodbye to their author:
Many thanks to my local post office, who did not bat an eyelash even once during my repeated visits for large quantities of stamps, for letting me borrow the USPS boxes you see above to contain the postcards, and for getting them out the door. Confederates living within the sample geographies confirmed receiving example postcards on October 5th.
Outcomes of interest
The central ask of each postcard is explicit: It asks the recipient to change their party registration. Everyone who receives this postcard is currently registered Republican. First and foremost, once the voterfiles update after the election, I will be looking for the share of voters who change their party identification in each condition compared to control precincts and to the state overall.
At the same time, I kind of suspect that if these messages have an impact, they may also impact the decision to vote at all. I am worried about backlash - I’m reminding Republicans of the existence of Democrats, and asking them to think about the badness of their own party! - but I’m also hopeful these arguments may give these voters permission to stay home if they are not successfully persuaded to alter their general voting preference. Democrats want good things to happen and Republicans do not, so perhaps there is some peace of mind to be had in reminding some Republican voters that their decision not to vote if they were going to vote Republican might be just fine. However, if this type of outreach induces backlash, a small-scale experiment of this kind feels like an optimal place to observe it.
I used block randomization on precinct to see if I can observe any changes in Democratic vote choice at the precinct group level but I’m aware of the power limitations on a study of this kind. I will conduct an analysis of aggregate data with this outcome in mind, but my prior as of time of pre-registration is that it is unlikely I will observe major differences in Democratic vote share across conditions.
My overall hunch, however, is that to the degree that I find any differences across conditions, the Democratic/Republican contrast message, followed by the Democratic positive message, will emerge as the most effective. Although I believe decent shares of Republicans are fed up with their party’s elites as well as their party’s policies, I’m not sure that only attacking along only one of these dimensions is enough to jolt a voter into a major change like a change in party affiliation.
That said, I also think the difference between getting the ask and not getting the ask is larger than the difference between the different kind of asks. That is, I think the differences in impact of the postcards will be smaller across treatments than they will be when it comes to treatment versus control. I don’t have strong priors in this regard.
Ultimately, this is a small-scale experiment, totaling less than $2,300 in cost. How much anything moves in any direction is somewhat dubious. But I believe Malchow’s challenge is worth taking seriously. And I am onboard with the intuition that working on the brands is worthwhile compared to working on individual candidates, in many cases. (Frankly, I think the Democratic brand ain’t great right now, and that is a big problem for us) I designed the experiment with these key learning objectives in mind, and I look forward to updating with a full report when voterfiles are refreshed in 2025.