On the Internet,No One Knows You’re an Activist: Patterns of Participation and Response in an Online,Opt-in Survey Panel

Abstract

Recent years have seen a dramatic increase in surveys of online respondents who participate absent random selection. Using an uncompensated, opt-in panel of 11,000 Pennsylvanians (2020–2022), we benchmark self-reports against vote histories, campaign contributions, election returns, and Census data. Ours is among the first such benchmarking exercises in more than a decade, a period when online survey research has transformed. It is also among the first analyzing a longitudinal panel. Our panelists are approximately balanced on partisanship; attrition is limited. Notably, respondents are very politically engaged, with high rates of turnout, campaign contributions, and rally attendance. Ethnic/racial minorities are under-represented. Political attitudes show high over-time stability, and an experiment uncovers few partisan differences in response rates. Such survey samples are cost-effective ways to track the views of highly engaged citizens over time, but less engaged citizens prove hard to recruit, even with supplemental efforts. Survey participation is itself a political behavior.

Keywords

online panels opt-in surveys survey experiments non-response activists political engagement

In recent decades, the use of online, opt-in surveys to study Americans' political attitudes has exploded. To date, survey experiments with opt-in samples have shown noteworthy external validity, whether respondents are recruited via online markets or empaneled directly to take surveys (Berinsky et al. 2012; Coppock et al. 2018; Mullinix et al. 2015).¹

Still, there are outstanding questions about non-probability samples (Cornesse et al. 2020; Eyal et al. 2021; Krupnikov et al. 2021). In contrast to earlier practice, researchers using these samples are typically not involved in sample recruitment, and so aren’t as familiar with the procedures that lead respondents to be included in their data (Enns and Rothschild 2023). The fraction of the target population available for surveys is thus unclear but is likely to be small (Stewart et al. 2015). In particular, online, opt-in surveys risk over-representing the highly politically engaged (Cavari and Freedman 2022; Jackman and Spahn 2021).

We address these questions by detailing a multi-year panel survey which we directed with the survey firm Civiqs serving as recruiter and interviewer.² Certainly, previous studies have sought to understand online, opt-in survey respondents by benchmarking them relative to other populations, principally making comparisons to high-quality governmental surveys (e.g., Ansolabehere and Schaffner 2014; Dutwin and Buskirk 2017; Keeter 2018; Kennedy et al. 2016; MacInnis et al. 2018; Malhotra and Krosnick 2007; Rivers and Bailey 2009; Vavreck and Rivers 2008; Yeager et al. 2011). Yet among those benchmarking studies, the most recent data in a peer-reviewed article comes from 2014, which was a decade ago and prior to the significant polling errors in the U.S. and U.K. (Clinton et al. 2022; Prosser and Mellon 2018). The population of respondents to online, non-probability samples can change significantly in a few months (Aronow et al. 2020), to say nothing of a decade. Many of the main providers of online, opt-in samples for contemporary political science did not exist a decade ago (Coppock and McClellan 2019).

Even beyond its recent vintage, the panel we report has several features which make it an especially valuable tool through which to assess the opportunities and limits of samples recruited online to take uncompensated surveys. As Appendix Table 8 illustrates, Pennsylvania is broadly representative of the U.S. overall. Recruiting in this large American state enabled us to compare our survey-based estimates with external benchmarks, including public records of vote history and campaign contributions alongside election returns, Census data, and newspaper reports. In fact, this study is among the only to use official, individual-level voter records as an externally validated benchmark when evaluating the bias in online, opt-in survey data—and the first to our knowledge in more than a decade (see also Rivers and Bailey 2009; Vavreck and Rivers 2008).

We also contribute by evaluating the properties of not just a one-time cross-section but a multi-wave panel (see also Vavreck and Rivers 2008). Doing so is especially important given the increased use of panels to measure the outcomes of field experiments (Broockman et al. 2017). By assessing who participates in such panels, we will be better positioned to understand the populations to whom such field experiments generalize.

Via online advertisements on Facebook, Google, and Instagram, we recruited sizable numbers of Democrats and Republicans to take brief surveys, with high re-interview rates that facilitate over-time study. These respondents were recruited via Civiqs’ usual recruitment channels and were combined with pre-existing members of Civiqs’ panel who were Pennsylvania residents—48% of our full panel was newly recruited for this project, with the remaining 52% coming from Civiqs’ existing panelists. However, our sample is very politically engaged: of those we were able to match to voter files, 87% voted in the 2018 midterm election and 72% in the 2020 primary. The actual figures in Pennsylvania were significantly lower, at 58% and 34%, respectively. We also find levels of campaign contributions and Trump rally attendance far higher than among Pennsylvanian adults overall. Our respondents' vote choice preferences are very stable over a campaign, even relative to a national, population-based panel. Deliberate efforts to reduce such biases by targeting lower-engagement Pennsylvanians yielded few additional respondents.

These results reinforce an expressive account of survey response, in which those who are most motivated to participate are highly engaged partisans (see also Graham and Huber 2021). However, in contrast to Gelman et al. (2016), such motivations vary little over time among this sample. An experiment before and after the November 2021 election uncovers no evidence that individuals are motivated to respond to our survey by events that are positive or negative for their party.

Our results also corroborate prior findings (Hillygus et al. 2014; Kennedy et al. 2016; MacInnis et al. 2018): racial/ethnic minorities are under-represented in our panel relative to their share of the population, as are younger voters. Comparisons with benchmarks from the Current Population Survey (CPS) illustrate that lower income Black registered voters and Black or Latino registered voters without college degrees are particularly under-represented.

Researcher involvement in the design and recruiting of panelists provides procedural transparency and opportunities for experimentation.³ Unlike some similar online, non-probability surveys, our panel does not offer respondents material incentives (see MacInnis et al. 2018). Recruiting respondents to take free, short surveys avoids the need to solicit personal information for payment, and so has the potential to include some respondents who would not otherwise participate while reducing the incentive for disruptive bots (Aronow et al. 2020).

While many contemporary online, opt-in surveys provide compensation (e.g., Ansolabehere et al. 2023), it is nonetheless key to understand samples motivated by non-monetary incentives, in part because such expressive incentives are likely to play some role even in surveys in which respondents are modestly compensated. That said, recent research has commonly employed uncompensated surveys, at times because certain research designs make compensating otherwise anonymous survey respondents difficult and/or because of the need to maintain anonymity (e.g., Carpenter 2012; Hopkins et al. 2024; Hopkins and Gorton 2024; Medeiros and Gravelle 2023; Weiss 2021; Williams, Gravelle and Klar 2022; Zhang et al. 2018). (Surveys administered through other modes, such as telephone surveys and exit polls, are also commonly uncompensated—e.g., Croco et al. 2019; Tyler and Iyengar 2023; Sambanis, Simonson, and Yaylacı 2023—so insights from this study may also apply to those survey modes.)

We define activists as those who engage in political activities beyond voting (Hopkins and Noel 2022). In recent years, such activists have arguably been understudied relative to both political elites and the mass public (Blum 2020; Han 2014; Parker and Barreto 2014). These results clarify that online, opt-in samples can effectively survey such highly engaged partisans repeatedly at relatively low cost.⁴ When coupled with prior results on treatment effect stability (Berinsky et al. 2012; Coppock et al. 2018; Mullinix et al. 2015), they indicate a meaningful role for free, online, opt-in samples in contemporary research (Krupnikov et al. 2021). This approach thus opens up many opportunities for studying activists from both parties. However, these results also clarify that our ability to sample lower-engagement voters is quite limited—and that in cases where researchers are interested in that side of the engagement divide (Krupnikov and Ryan 2022), or in Black or Latino citizens without college degrees, they will need to employ alternatives. Moreover, free surveys may disproportionately attract those with expressive motivations—such as strong, highly engaged partisans and activists—and may also have lower response rates (Yan et al. 2018).

Prior Research

The popularity of surveys administered online has soared recently (Tourangeau et al. 2013). The increasing use of online surveys has transformed two different elements of survey research: the dominant interview mode and the typical sampling procedure (Gooch and Vavreck 2016). Our emphasis here is on samples of respondents available online for opt-in surveys.⁵

The popular “Total Survey Error” framework encourages survey researchers to move beyond a focus on sampling error (Weisberg 2010). It distinguishes between errors of nonobservation (including coverage error) and errors of observation (such as social desirability bias; Groves and Lyberg 2010, 855), and we focus squarely on the former. Specifically, we emphasize coverage errors resulting from the differential accessibility of some in the target population when surveying online.

Online, opt-in survey respondents often select into taking surveys by proactively seeking them out or by being among the few people who respond to online advertisements, raising concerns about representativeness (Baker et al. 2010; Callegaro et al. 2014; Chang and Krosnick 2009; Mercer et al. 2018; Yeager et al. 2011). Among this self-selected group are professional respondents who take part for the incentives or based on their interest in the survey topic (Hillygus et al. 2014).

For researchers to make inferences about a population from a survey sample, respondents must be drawn from the broader population with known, non-zero probabilities. Non-response bias is a particular concern when it is associated with key measures of interest and when it is non-ignorable, meaning not fully explained by covariates captured in the survey. For example, if contemporary survey samples are disproportionately highly politically engaged, we risk over-estimating the fraction of those with strong partisan identities (Huddy 2018; Jackman and Spahn, 2021; Cavari and Freedman 2022; but see Mellon and Prosser 2021).

Non-response biases may be getting worse. Telephone response rates have been declining for decades (Leeper 2019; Luiten, Hox, and de Leeuw 2020), and researchers increasingly use online, probability-based or opt-in samples where response rates may be lower still (MacInnis et al. 2018; Mutz 2011). Given researchers’ growing reliance on online, opt-in samples—and the substantively meaningful differences between Americans who are or are not politically engaged (Krupnikov and Ryan 2022)—the question of who participates in such surveys is crucial (Porter and Whitcomb 2005).

Prior research has already begun to address this question. In an analysis of Pew telephone and face-to-face interviews, Keeter (2018) finds that surveys with low response rates tend to over-represent the politically and civically engaged, though not people with any particular political leaning. To date, research validating online, opt-in recruitment has typically compared such samples to Census data or other survey benchmarks (see esp. Malhotra and Krosnick 2007; Vavreck and Rivers 2008; Rivers and Bailey 2009; Yeager et al. 2011; Ansolabehere and Schaffner 2014; Kennedy et al. 2016; Dutwin and Buskirk 2017; Keeter 2018; MacInnis et al. 2018). Of those studies, the only two to benchmark opt-in samples using voter file data were published more than a decade ago (e.g., Rivers and Bailey 2009; Vavreck and Rivers 2008).

Partisan Motivations

One possibility is that opt-in respondents—and especially those who aren’t paid or quasi-professional (Hillygus et al. 2014)—will respond (or not) in ways that express their partisan identity. For example, Gelman et al. (2016) show that in the 2012 presidential elections, the shift in polls against Democrat Barack Obama in the wake of his poor debate performance was driven primarily by a pro-Republican shift in survey participation. Relatedly, Clinton et al. (2022) show that in 2020 Democrats were more likely to respond to telephone polls than Republicans or independents. Given the strength of contemporary partisan identities and their integration with many Americans’ self-concepts (Huddy et al. 2015; Iyengar et al. 2019; Malka and Adelman 2022; Osmundsen et al. 2021), it seems quite plausible that split-second decisions about whether to participate in a survey or how to respond to a question may be shaped by a desire to associate one’s self with copartisans (see also Schuessler 1999).

Consistent with an expressive account of survey participation, Graham and Huber (2021) find that when offered the opportunity to answer additional survey questions, between 50 and 80% of respondents accept, suggesting that respondents receive expressive benefits from answering survey questions (see also Schaffner and Luks 2018). Acceptance was especially high for questions previewed as explicitly partisan, such as presidential approval. Respondents’ desire to express their partisanship seems to be a potent motivation for taking surveys, one likely to influence who participates in uncompensated surveys.

Recruitment Procedure

We focus on a single state so as to collect benchmark data from various sources. (The COVID-19 pandemic forced us to cancel a planned 2020 statewide exit poll which was one benchmark.) As Appendix Table 8 shows, in key respects, Pennsylvania’s demographics closely mirror those of the nation overall; it is also a swing state which has voted for the winner of every presidential election since 2004.⁶

We first worked with Civiqs to increase its base of Pennsylvania panelists. Civiqs scaled up its usual online recruitment, placing various advertisements via multiple web platforms—of the 5189 panelists for whom source data is known, 67% came from Facebook, 31% from Google, and the remainder from Instagram. The recruitment advertisements varied but typically focused on current political issues; Appendix Figures 3 and 4 illustrate examples of the recruitment done for this study.⁷ (It is plausible that these ads' political content encouraged politically engaged respondents to reply; below we report the results of a test with a less explicitly political advertisement. See also Hopkins et al. 2024, which employ Facebook ads with images of respondents' neighborhoods to recruit respondents.) In all, 48% of our panelists were drawn from Civiqs' augmented recruitment in Pennsylvania while the remaining 52% had been empaneled previously. Below, we report attempts to recruit respondents via less political materials. On the potential to recruit diverse samples via Facebook, see Zhang et al. (2018) and Sances (2021).

Sampling these respondents, Civiqs then sent email invitations to take brief surveys (between 10 and 20 questions) for free. For the initial interviews, Civiqs drew panelists from this augmented sample; for some subsequent surveys, the researchers identified panelists to be re-interviewed (to measure change over time) or to be excluded (to avoid interviewing people twice in close proximity). Key elements of the sampling among recruited panelists were dictated by the researchers and the requirements of the research design, so this does not constitute an evaluation of Civiqs’ other surveys’ vote choice estimates. Table 1 shows the dates, sample sizes, and response rates for select waves in which we targeted respondents statewide. Figure 1 provides a flowchart illustrating the timelines for recruitment, survey, and voter file matching. Given that our interest is precisely in the characteristics of our sample, we primarily report unweighted results below, but do in places employ demographic weights to see whether our estimates are sensitive to reweighting.

Table 1.

Survey Dates, Sample Sizes, and Response Rates for Select Waves. *Wave was Entirely Re-Interviews. Waves Employed for Other Projects Omitted.

Wave Number	Dates	N Respondents	N Invited	Response Rate (%)
Wave 1	2/27/20–3/3/20	2462	9889	25
Wave 2	3/14/20–3/18/20	1598	6556	24
Wave 3	4/4/20–4/8/20	1912	8567	22
Wave 5	5/30/20–6/2/20	2045	5729	36
Wave 6	6/6/20–6/11/20	1221	3698	33
Wave 8*	10/17/20–10/21/20	1557	3029	51
Wave 11*	10/31/21–11/5/21	1622	3230	50
Wave 13*	5/13/22–5/16/22	1740	4521	38

Figure 1.

Flowchart illustrating timing for recruitment, surveys, and voter file matching.

Compliance with Ethical Standards

This research adheres to the American Political Science Association’s Principles and Guidance on Human Subject Research. It was reviewed by the University of Pennsylvania Institutional Review Board twice, with the survey and intervention determined exempt. In response to Civiqs advertisements on Facebook, Google, and Instagram,⁸ subjects volunteered to take occasional, brief online surveys for free. Survey links were provided by email, and subjects could decline to participate.^9,10

Comparisons with Benchmarks

How do our panelists compare to registered Pennsylvania voters? To answer, we acquired validated vote history information. Partnering with two outside vendors—Civiqs and TargetSmart—we were able to obtain such information for 65% of our respondents. The remaining 35% of respondents could be unmatched for multiple reasons: either they did not provide Civiqs with the correct personally identifying information (PII) necessary or they did not appear in TargetSmart’s database of registered voters.¹¹ The 65% match rate is quite similar to the 69% of 2022 Cooperative Election Study respondents who matched to TargetSmart (Ansolabehere et al. 2023).

Civiqs’ panelist recruitment and matching to TargetSmart data was ongoing during the ad-ministration of this panel. Thus, while we acquired TargetSmart data three separate times, these analyses focus on 610 panelists whose vote histories were obtained from TargetSmart on June 20, 2021. We opt to highlight this subset because all of these respondents were recruited as part of this panel, meaning that it excludes respondents recruited by Civiqs for surveys prior to the inception of this project and so reduces heterogeneity by recruitment date.

One straightforward concern is that those who do match to the voter file might be more politically engaged than other respondents. However, when analyzing self-reported political giving, we find that respondents who did match in June 2021 reported lower levels of donations (41% versus 49%; t test p-value = 0.01) compared to those who did not. The two groups do not differ meaningfully in their self-reported attendance at Trump rallies (10% versus 9%; t test p-value = 0.64). This suggests that by analyzing the turnout patterns of respondents who did match to the voter file, we are not systematically excluding less-engaged individuals.

Appendix Table 9 provides evidence that these respondents are broadly similar to panelists outside this sample on other dimensions, although those who matched are (unexpectedly) more Republican-leaning. Appendix Table 9 suggests that the pro-Republican lean may originate partly from the matching process—which respondents are willing to give valid names and addresses—although the partisan skew is less pronounced when looking at respondents who matched in December 2020.¹² Accordingly, the partisan skew may also be partly an artifact of the timing by which Republicans were willing to provide personal information.

Separately, we obtained voter history information for a random sample of 20,000 Pennsylvania registered voters which serves as a baseline.¹³ For both the random sample of voters and our matched panelists, we calculated the fraction who had voted in 0, 1, 2, or 3 of the most recent three federal elections (2016, 2018, and 2020); the results are displayed in Figure 2. Among the baseline sample of registered voters (light grey bars), 11% voted in none of those three elections, 18% voted in one, 19% in two, and the remaining 53% in all three. However, among the matched panelists (dark grey bars), turnout is much higher: 80% voted in all three elections and 14% voted in two. As compared to a random sample of Pennsylvania’s registered voters, our matched panelists are much more likely to consistently vote.

Figure 2.

Fraction of a random sample of 20,000 Pennsylvania registered voters and our panelists (n = 610) by number of validated votes cast in general elections, 2016–2020.

Table 2 extends this analysis using other measures available through TargetSmart. While 34% of Pennsylvania registered voters participated in the 2020 primary, the figure for our panelists is 72%. In addition, our panelists are less likely to be registered Democrats (46% statewide versus 40% in our panel), less likely to not be registered with a major party (15% statewide versus 8% in our panel), and more likely to be registered Republicans (39% statewide versus 52% in our panel).¹⁴

Table 2.

Means for a Random Sample of 20,000 Pennsylvania Registered Voters Compared to 610 of Our Panelists With Voter File data. Racial Demographics and Marital Status Estimated by TargetSmart.

	Voter File Sample	Validated PA Panel
Proportion female	0.524	0.479
Mean birth year	1970	1968
Proportion 18–29	0.157	0.099
Proportion 65+	0.262	0.292
Proportion White	0.850	0.938
Proportion Black	0.080	0.034
Proportion Hispanic	0.026	0.008
Proportion Asian American	0.010	0.002
Proportion native	0.000	0
Proportion married	0.439	0.511
Proportion rural	0.302	0.375
Proportion suburban	0.413	0.400
Proportion urban	0.284	0.220
Proportion registered Democrat	0.464	0.402
Proportion registered Republican	0.388	0.521
Proportion registered other	0.147	0.077
Proportion voted in 2020 primary	0.344	0.720
N	20,000	610

TargetSmart also provides model-based estimates of various demographics; these show that our panelists are more likely to be White (88% of statewide registered voters via TargetSmart or 90% via the CPS versus 94% in our panel) and less likely to be Black (8% statewide via TargetSmart or 7% via the CPS versus 3% in our panel),¹⁵ Hispanic (3% statewide via TargetSmart or the CPS versus 1% in our panel), or Asian American (1.0% statewide via TargetSmart or the CPS versus 0.2%).¹⁶ The online panel also under-represents urban residents (28% versus 22%). Our panelists are born 2 years earlier on average, making them slightly older than the state population. Stereotypes that older voters are harder to reach online appear to be incorrect.

When asked in Wave 1 (February/March of 2020) if they had “donated money to any political candidates in the last 2 years,” 47.8% of unweighted respondents said “yes,” and that number changes only slightly—to 47.5%—with weights.¹⁷ As a baseline comparison, we found that only 0.47% of the Pennsylvanian voting age population made an itemized contribution to a federal candidate during the same period.¹⁸ That figure is almost certainly an under-estimate, as not all donations to federal candidates are itemized. Hughes (2017) reports that in the 2016 cycle, 15% of national survey respondents reported making a political donation. Our panel’s estimate (48%) clearly reflects a difference from the general population—the question is of what order of magnitude.

The panel’s first wave also asked respondents, “Have you ever attended a political rally for Donald Trump?” 9.3% of all respondents—and 17.9% of all Republican respondents—said “yes.”¹⁹ As a baseline comparison, we used real-world crowd-size reports to estimate the proportion of Pennsylvania voters who have attended Trump rallies at between 0.7 and 1.4%.²⁰ Trump rallygoers, many of whom are strong partisans, are vastly over-represented in our data set. Also, unlike our validated measures of turnout, it is possible that respondents may over-report donations or rally attendance in surveys, perhaps due to social desirability biases. Sampling may not be the only reason for our panel’s inflated estimates.

Race, Ethnicity, and Socioeconomic Status

Given that a majority of Americans identify as White, accuracy assessments in surveys targeting the population overall will be driven more by that group than others. As a consequence, it is particularly important to investigate who within different racial/ethnic groups participates in surveys.

We accordingly used the November 2020 CPS—accessed via IPUMS—to benchmark family income and education measures by race/ethnicity. We identified 1714 Pennsylvania respondents to that CPS, enabling us to isolate registered voters and employ that wave’s weights. Given the high levels of political engagement evident in our sample, we view registered voters as a valuable benchmark, even as we acknowledge that our panel’s sampling frame is technically adult residents of Pennsylvania rather than registered voters. Prior research demonstrates that registered voters tend to be older, more highly educated, and are more likely to be White than adults (Fraga 2018; Leighley and Nagler 2013; Verba et al. 1995), so if anything, our use of registered voters as a benchmark understates the extent of the biases.

As Table 3 shows, the mean family income estimate for weighted Black CPS respondents is $56K, a figure that is 62% of the $91K estimate for White respondents. However, possibly due to the coarseness of our panel’s measure of income, the differences are more muted among our panelists. The mean income estimate among Black respondents is $72K, while for White respondents it is $82K, a smaller gap than estimated via the CPS. Among Latinos, however, the estimates are almost identical: $73K in the CPS versus $76K in the panel. In general, the panel-based average income estimates are not far off. But they do under-state the Black-White income gap considerably.

Table 3.

Income Comparisons From 2020 November CPS (Registered Voters; n = 1714) and PA Panel Respondents (n = 4521). Income Estimates in Thousands of Dollars.

	Asian Am	Black	Latino	White
CPS mean	137	56	73	91
PA panel mean	88	72	76	82
CPS SD	59	54	67	59
PA panel SD	59	56	55	63
CPS N	18	128	58	1,550
PA panel N	15	240	78	3,174

In Table 4, we report the same comparisons with respect to two levels of educational attainment—having completed high school and college. As Table 4 illustrates, high school completion is nearly universal in our panel, with rates of 99% for White respondents, 95% for Asian American respondents, and 100% for Black and Latino respondents. In the CPS, 89% of weighted Black respondents have a high school degree, while the numbers for Latino and White respondents are 91% and 96%, respectively. The vast majority of those identifying with each group completed high school, but our panel almost entirely excludes the 4% to 11% who have not.

Table 4.

Proportion of Respondents Completing College by Racial/Ethnic Group; Data From 2020 November CPS (Registered Voters; n = 1714) and PA Panel Respondents (n = 4521).

	Asian Am	Black	Latino	White
CPS HS mean	1.00	0.89	0.91	0.96
PA panel HS mean	0.95	1.00	1.00	0.99
CPS BA mean	0.75	0.23	0.38	0.41
PA panel BA mean	0.76	0.54	0.61	0.48

With respect to college degrees, the biases are even more pronounced. According to the CPS, 23% of Black registered voters in Pennsylvania have college degrees, but in our sample, the estimate is 54%. Among Latinos, the estimates are 38% (CPS) versus 61% (PA panel); for White respondents the estimates are meaningfully closer, at 41% and 48% (PA panel). In short, our sample significantly under-represents Black and Latino respondents without college degrees. While those without college degrees are under-represented, that under-representation varies by race/ethnicity. Given the association between educational attainment and political engagement, one likely effect is to overstate political engagement (Verba et al. 1995; see also Krupnikov and Ryan 2022).

Response Stability and Vote Choice

Our panel also allows us to estimate stability in vote intentions over the 2020 campaign, a valuable feature not available in prior, cross-sectional studies. One concern is that a highly engaged sample may also be more stable in partisan preferences than the population overall. In Table 5, we report intentions to vote for Biden, Trump, or neither in the late February/early March 2020 wave and the mid-October 2020 wave. Those who are undecided or opt for neither are a small fraction—just 2% in October. Partisan stability is also high, with 95% of all respondents reporting the same vote intention in both waves. Such stability is especially noteworthy given that the first wave took place before the COVID-19 pandemic.

Table 5.

Transitions in Vote Intention for 952 Pennsylvania Respondents to the 2/27–3/3/20 and 10/17–10/21/2020 Waves.

	Biden Oct	Neither Oct	Trump Oct
Biden Feb/Mar	0.511	0.003	0.002
Neither Feb/Mar	0.024	0.012	0.006
Trump Feb/Mar	0.005	0.007	0.430

There are no other publicly available panels of Pennsylvania residents. However, because Pennsylvania demographics are broadly similar to those of the U.S. (see Appendix Table 8), there is some benefit to comparing these stability estimates with those from a contemporaneous national panel. In Table 6, we provide a benchmark using the Institute for the Study of Citizens and Politics panel (see also Hopkins 2021; Mutz 2018), an online, population-based panel administered by Ipsos. In 2020, this panel only included American respondents 30 and older, a design which may lead it to overstate stability relative to the overall population (see also Hobbs 2019). Yet even in this older population, those recruited via population-based methods such as random digit dialing and address-based sampling are more likely to be undecided and less stable in their responses than our Pennsylvania panelists. 89% of national respondents reported the same vote intention between January and October 2020, a number which is high in absolute terms but meaningfully lower than the 95% stability uncovered in our online Pennsylvania panel. The activists interested in taking free, online surveys are unusually stable in their vote preferences, even relative to other online survey takers.

Table 6.

This Table Reports Transitions in Vote Intention for 970 Respondents to a Separate Population-Based Panel Administered to a National Sample of Respondents 30 and Over.

	Biden Oct	Neither Oct	Trump Oct
Biden Jan	0.457	0.007	0.013
Neither Jan	0.042	0.056	0.019
Trump Jan	0.013	0.015	0.377

The January 2020 wave was administered from the 24th to the 31st; the October 2020 wave was administered from October 7th to 22nd.

Our panel’s sampling for the October 2020 wave focused on February/March respondents, and so was not designed to estimate Pennsylvania voters’ preferences. However, the weighted breakdown was 51.6% for Biden versus 46.4% for Trump. In reality, Biden won Pennsylvania by a similar but smaller margin of 50.0 to 48.8.

The May 2022 Pennsylvania primaries provide another opportunity to test whether our panelists under-represent support for specific types of candidates. Voting in primaries is not anchored by partisanship to nearly the same extent as in general elections, meaning that primaries represent a harder test for survey research. As Appendix Table 10 illustrates, after removing respondents who said they would not vote in the GOP primary, 29.6% of our weighted panelists expressed support for Senate candidate Kathy Barnette, 21.2% for Mehmet Oz, and 17.8% for David McCormick.²¹ In reality, Barnette received 24.7% of the vote, Oz got 31.2%, and McCormick secured 31.1%. Very conservative, Barnette lacked the campaign budget of her chief rivals and relied more on media appearances to garner support; her support appeared to grow late in the campaign, and she may have done especially well among voters who followed the campaign closely. The over-representation of Barnette support is thus consistent with a characterization of our panelists as very politically active. Still, why our sample under-represented support for Oz and (to a greater degree) McCormick is not clear, as both (Trump-backed) Oz and McCormick had well-funded campaigns and ran millions of dollars in advertisements.

In the GOP’s 2022 gubernatorial primary, 39.5% of respondents backed Doug Mastriano while 19.3% backed Lou Barletta and another 11.8% backed Bill McSwain. These results were reasonably close to the election returns, which put the Trump-backed populist Mastriano in the lead with 43.9% of the vote, ahead of the Trump-style former Congressperson Barletta with 20.3% and former prosecutor McSwain (whom Trump pointedly did not endorse) with 15.8%. In the Democratic Senate primary, 57.9% of panelists backed John Fetterman versus 17.2% for Conor Lamb and 12.1% for Malcolm Kenyatta. Those numbers are not far from the reality of Fetterman winning with 58.7% of votes, followed by Lamb with 26.3%, and Kenyatta with 10.8%, although the survey did under-state support for the more moderate Conor Lamb.

Overall, these results indicate that our panelists are reasonably representative of primary voters, and that more populist, Trump-style candidates are not consistently disadvantaged. If anything, our poll under-represented support for more establishment-oriented candidates like McSwain and Lamb while overstating support for Barnette. In sum, our poll does not appear to have some of the biases noted in other surveys, such as a consistent bias against Trump-style candidates or candidates from minority ethnic/racial or religious backgrounds (Hopkins 2009).

Recruiting Lower-Turnout Voters

Given the high levels of political engagement detailed above, we attempted a separate effort to survey low-turnout Pennsylvania voters, which was conducted outside of our partnership with Civiqs. We first drew a random sample of 9937 registered Pennsylvania voters who had voted no more than one time in the 2012, 2014, 2016, and 2018 cycles from a February 2019 Pennsylvania statewide voter file. Via TargetSmart, we then successfully matched 9332 of these registered voters to its database, and were able to purchase 9627 phone numbers for these individuals as well as 2849 email addresses. We next uploaded the TargetSmart data to Facebook using its Custom Audience feature, and served 1321 separate targeted individuals with advertisements inviting them to take a survey (see Appendix Figure 5 for the advertisement). Note that this advertisement was less explicitly political than some of those employed in the Civiqs recruitment detailed above. In total, we paid $235.53. Facebook records indicate 66 individuals clicked on the advertisement. Of these, 7 began the survey, 6 completed it, and 1 left an email address for possible follow-up surveys.

Even if we only consider those who saw the Facebook advertisement, the response rate was 0.4%. While assessing low-engagement registered voters via these methods is much more costly, such costs are broadly in line with those for recruiting population-based samples. Thus, such methods may be useful in some cases, especially when complementing the lower-cost recruitment of more engaged respondents. Future research should consider survey invitations that are entirely apolitical, such as those inviting respondents to participate in a survey about entertainment or sports.

Differential Partisan Response: An Experiment

Our panelists are more partisan and politically engaged than registered Pennsylvania voters. At least for these panelists, panel enrollment and survey response have the appearance of being opportunities to express their political identities (see also Malka and Adelman 2022). The evidence to date is consistent with a model of respondents as motivated partisans who participate in surveys to represent their partisan side (see also Clinton et al. 2022; Graham and Huber 2021). These observations lead to a follow-up question: is there a temporal dimension to their survey participation, with panelists more likely to take surveys when their partisan side is advantaged? Gelman et al. (2016) report that in the 2012 election, sample composition explained more of the change in polling than did actual vote intentions. It’s possible such threats are especially pronounced with highly engaged samples.

To test that prospect, we conducted an experiment straddling the November 2021 elections. Specifically, in late October 2021, we blocked 3230 panelists from prior survey waves into pairs using demographics such as racial self-identification and political partisanship (Moore 2012).²² Within each matched pair, we then randomly assigned respondents to be invited to take a survey either on Sunday, October 31st or Wednesday, November 3rd. In the intervening period, Virginia Republican gubernatorial candidate Glenn Youngkin won an unexpected victory over his Democratic rival on Tuesday evening, November 2nd. While this was a governor’s race outside Pennsylvania, it was closely followed by political activists, and made national headlines; a Republican won a statewide race in Virginia for the first time since 2009. Our hypothesis was straightforward: Republicans invited to participate just after Youngkin’s victory would be energized and would be more likely to complete the survey, while Democrats would be less likely. Of the 1615 panelists randomly assigned to be invited before the election, 806 completed the survey, for a response rate of 49.9%. After the election, of the other 1615 panelists, 816 responded (50.5%). A test of equal proportions indicates no meaningful difference between the samples (p = 0.75). Overall, the response rate was virtually identical before and after the election.

Table 7 illustrates the same tests when conducted with only the respondents who took the survey. For example, we see that among those who were assigned to a pre-election invitation and completed the survey, 45.4% were Democrats while 38.8% were Republicans. The numbers among those assigned to the post-election survey were almost identical, at 44.0% for Democrats and 39.0% for Republicans. Those differences are statistically insignificant, with p-values of 0.567 and 0.955. The actual samples were virtually identical before and after the election, indicating that there was not differential survey response induced by the GOP’s unexpected Virginia win. Certainly, it’s possible that a more politically salient event—including a successful presidential debate—might induce different responses. But we do not observe that the Republicans' successful off-year election induced differential partisan responsiveness.²³ The panelists are not differentially responsive to surveys, perhaps because they are so highly and consistently engaged with politics.

Table 7.

This Table Reports Two-Sided t-tests Evaluating Respondents Who Completed the Survey Either Before or After the GOP Won the Virginia Governor’s Race on November 2nd, 2021.

	Mean	Mean	P-	Total
	Before	After	Value	N
Democrat	0.454	0.440	0.567	1,622
Republican	0.388	0.390	0.955	1,622
Trump (1) vs. Biden (−1)	−0.063	−0.029	0.489	1,622
Age	58.605	58.610	0.995	1,622
Black	0.051	0.048	0.775	1,622
Hispanic/Latino	0.014	0.012	0.804	1,622
White	0.924	0.929	0.725	1,622
Female	0.460	0.480	0.418	1,622
Education (years)	15.535	15.250	0.013	1,620

Conclusion

Contemporary political science hinges on researchers' capacity to survey populations of interest. In recent years, declining response rates via traditional telephone polls—alongside new opportunities afforded by online survey administration—have led to an explosion of online polls, many with non-probability samples (e.g., Ansolabehere and Schaffner 2014; Rivers and Bailey 2009). For researchers, such polls have provided newfound opportunities to conduct cost-effective survey experiments with broader, more diverse populations. They also afford the opportunity to resample respondents at relatively low cost (see also Broockman et al. 2017).

Nonetheless, as survey response rates decline, understanding the correlates of survey participation—and the effects of various recruitment strategies—becomes increasingly vital. Survey non-response can produce a distorted image of the electorate which overstates political awareness and polarization (Ansolabehere and Hersh 2012; Cavari and Freedman 2022; Krupnikov and Ryan 2022). In 2020, polls under-estimated support for incumbent president Donald Trump, in part because of the under-representation of less engaged voters (see also Clinton et al. 2021; Clinton et al. 2022; Krupnikov and Ryan 2022).

Already, researchers have undertaken several efforts to validate online, opt-in samples (e.g., Ansolabehere and Schaffner 2014; Dutwin and Buskirk 2017; Keeter 2018; Kennedy et al. 2016; MacInnis et al. 2018; Malhotra and Krosnick 2007; Yeager et al. 2011) as well as online recruitment methods (Sances 2021; Zhang et al. 2018). This paper contributes by focusing on online respondents from one large American state who opt to take short, free surveys. This methodological approach has multiple advantages. For one, it enables us to benchmark various estimates against public records or other external estimates. It also permits researchers to be involved in the recruitment of online, opt-in respondents directly. Without compensation, bots may be less of a concern. Even today, telephone polling is typically uncompensated. Moreover, with rare exceptions (Shineman 2018), political participation itself is an uncompensated activity, so uncompensated surveys may help researchers understand political engagement and motivations. However, there are certainly downsides to uncompensated surveys. In some contexts, researchers may want or need to compensate subjects for their time, whether for ethical reasons or to enable varying experimental incentives.

We find that at an affordable cost, researchers can recruit sizeable numbers of highly engaged Republicans as well as Democrats to take uncompensated surveys—and that such respondents are frequently available for re-interviews, facilitating panels and experiments (Broockman et al. 2017) and opening new areas of inquiry. Activists are a vital, under-studied actor in American politics (but see Skocpol and Williamson 2012; Parker and Barreto 2014; Blum 2020), and research designs like those detailed here can access them repeatedly. In contrast to other research (Clinton et al. 2022; Gelman et al. 2016), we do not find evidence of differential partisan responsiveness. Activists from both sides are thus consistently available for surveys via this method. While we study participation in uncompensated surveys here, the prominent role of expressive partisanship in motivating survey responses may extend to decisions to participate in compensated surveys, too. Survey participation itself is a political behavior and needs to be understood as such.

In an era of low response rates and declining trust in researchers (Leeper 2019), many strategies for drawing samples will have limitations, and the strategy employed here is no exception. The free, online panel we have built under-represents registered voters who are Black, Latino, or Asian American, especially those with low educational attainment. That means that online, opt-in samples of respondents from these groups may be particularly skewed, a worrisome prospect given that those groups are often under-studied and under-represented (Segura and Rodrigues 2006). The danger that some emerging findings may not generalize to racial/ethnic minority populations is acute and normatively troubling. Given the stability of our respondents' attitudes, it is also possible that experiments with similar populations will under-state the effects of interventions.

It is possible that our results are influenced by the period of time in which we undertook this study. The COVID-19 pandemic disrupted many lives and may have influenced the population interested in taking online surveys. Another limitation of the sampling strategy detailed here is that activists may be particularly stable in their attitudes and hard to influence. As a consequence, experiments run with such samples may be more likely to generate null results than those using more broadly representative samples with more heterogeneity on political engagement.

The advantages and limitations uncovered here do not apply equally to all opt-in surveys. In- stead, they apply chiefly to those surveys using similar designs in which individuals are recruited online to periodically take short surveys for free. That, in turn, underscores that even among the surveys termed “opt-in,” there are important differences. The strengths and weaknesses of one opt-in sample may differ from those of another. It’s thus key to catalog the specific features of samples collected using varying strategies.

Still, opt-in designs like those studied here are commonly employed, especially in evaluating the impact of experimental interventions or in settings where maintaining anonymity is paramount (e.g., Carpenter 2012; Ceelen et al. 2019; Hopkins and Gorton, 2024; Williams, Gravelle and Klar 2022). Moreover, this research identifies straightforward ways to improve current practice. Researchers using online, opt-in samples would do well to learn and report key facts about the recruitment methods employed in sample construction. They also benefit from collecting and reporting external benchmarks not just on demographic measures but also political engagement, especially verifiable behaviors such as vote history. Experiments with heterogeneous effects across levels of political engagement may not generalize (Egami and Hartman 2022), making it important to test key results for heterogeneity on political engagement. Employing alternative survey modes such as mail or in-person surveys can be key supplements for hard-to-reach populations, too (Hopkins et al. 2024; Yan et al. 2018) (Supplemental material).

Supplemental Material

Supplemental Material - On the Internet, No One Knows You’re an Activist: Patterns of Participation and Response in an Online, Opt-in Survey Panel

Supplemental Material for On the Internet, No One Knows You’re an Activist: Patterns of Participation and Response in an Online, Opt-in Survey Panel by Daniel J. Hopkins and Tori Gorton in Political Research Quarterly.

Footnotes

Acknowledgments

This research was reviewed by the University of Pennsylvania IRB (834884 and 834094). The authors express deep thanks to partners who made this possible: Drew Linzer and Rachel Sinderbrand of Civiqs. David Azizi, Camilla Alarcon, Will Halm, Oliver Kleinman, Sophia Leung, Matthew Rabinowitz, and Gall Sigler provided stellar research assistance, and Nejla Asimovic, Matt Graham, Greg Huber, Josh Kalla, Silvia Kim, Matt Levendusky, Efrén Pérez, Nicholas Sambanis, and Chris Warshaw provided astute comments.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Daniel J. Hopkins

Supplemental Material

Supplemental material for this article is available online.

Notes

Appendix

Figure 3.

Sample Civiqs advertisement placed on Facebook to recruit survey respondents for this study.

Figure 4.

Sample Civiqs advertisements placed on Facebook to recruit survey respondents for this study.

Figure 5.

Advertisement placed on Facebook to recruit survey respondents.

Table 8.

Comparing PA Census Demographics, Election Results, and Campaign Donations to the U.S.

Demographic	PA Estimate	U.S. Estimate
Percent non-Hispanic white	75.7	60.1
Percent Black	12.0	13.4
Percent Hispanic	7.8	18.5
Percent Asian American	3.8	5.9
Percent American Indian	0.4	1.3
Percent in poverty	10.9	11.4
Percent of over 25s with a college degree	32.3	32.9
Median household income	$63,627	$64,994
Percent voted for Biden in 2020	50.0	51.3
Percent voted for Trump in 2020	48.8	46.9
Percent donated more than $200 to a federal	0.47	0.49
election campaign between 2018 and 2020^a

Sources: U.S. Census Bureau (2021a, 2021b), Wasserman et al. (2024), Federal Election Commission (2018, 2020).

^aUsing public contribution-level data obtained from the Federal Election Commission (FEC) on individual contributions to federal election campaigns, we were able to identify the number of unique individuals in Pennsylvania, and the United States more broadly, who had donated at least $200 total within the election cycle, between January 2018 and March 2020.

Table 9.

Comparing Validated Sample to June 2020 Respondents, Full Panel. June 2020 Chosen as Final Unrestricted Survey Sample. Full Panel Includes all Respondents Who Participated in at Least One Statewide Survey.

	Validated Sample	June ’20 Sample	Full PA Panel
Proportion female	0.47	0.45	0.48
Mean age	52.11	55.90	54.17
Mean years of education	15.02	15.34	15.11
Mean income	73.19	86.02	81.14
Proportion Democrat	0.32	0.37	0.39
Proportion Republican	0.46	0.40	0.37
N	610	1221	4521

Table 10.

Comparing PA Panel Estimates for 2022 Pennsylvanian Primaries Against Actual Results.

Primary Race	Candidate	PA Panel Est.	Result
Democratic Senate	John Fetterman	57.9	58.7
	Conor Lamb	17.2	26.3
	Malcolm Kenyatta	12.1	10.8
	Someone else	4.3	4.2
	Unsure	8.4
Republican Senate	Kathy Barnette	29.6	24.7
	Mehmet Oz	21.2	31.2
	David McCormick	17.8	31.1
	Carla Sands	6.2	5.4
	Jeff Bartos	4.3	4.9
	Sean Gale	1.8	1.5
	Someone else	5.2	1.2
	Unsure	13.9
Republican Governor	Doug Mastriano	39.5	43.9
	Lou Barletta	19.3	20.3
	Bill McSwain	11.8	15.8
	Dave White	7.6	9.6
	Charlie Gerow	1.7	1.3
	Joe Gale	2.2	2.1
	Melissa Hart	1.2	4.0
	Jake Corman	1.0	1.9
	Nche Zama	0.7	1.2
	Unsure	15.0

Table 11.

This Table Reports two-Sided t-tests Evaluating Respondents Randomly Assigned to be Surveyed Before or After the GOP Won the Virginia Governor’s Race on November 2nd, 2021.

	Mean	Mean	P-	Total
	Before	After	Value	N
Democrat	0.420	0.417	0.831	3,230
Republican	0.372	0.371	0.971	3,230
Trump (1) vs. Biden (−1)	−0.035	−0.028	0.857	3,230
Age	54.913	54.854	0.917	3,230
Black	0.069	0.069	1.000	3,230
Hispanic/Latino	0.020	0.021	0.902	3,230
White	0.896	0.891	0.648	3,230
Female	0.483	0.474	0.623	3,228
Education (years)	15.191	15.083	0.180	3,211

References

Ansolabehere

Stephen

Schaffner

Brian F.

. 2014. “Does Survey Mode Still Matter? Findings from a 2010 Multi-Mode Comparison.” Political Analysis 22 (3): 285–303. doi:10.1093/pan/mpt025.

Ansolabehere

Stephen

Schaffner

Brian

Shih

Marissa

. 2023. Cooperative Election Study Common Content, 2022. CCES Dataverse. https://dataverse.harvard.edu/dataverse/cces

Ansolabehere

Stephen

Hersh

Eitan

. 2012. “Validation: What Big Data Reveal about Survey Misreporting and the Real Electorate.” Political Analysis 20 (4): 437–459. doi:10.1093/pan/mps023.

Aronow

Peter Michael

Kalla

Joshua

Orr

Lilla

Ternovski

John

. Evidence of Rising Rates of Inattentiveness on Lucid in 2020, 2020, pp. 59–63.

Baker

Blumberg

S. J.

Brick

J. M.

Couper

M. P.

Courtright

Dennis

J. M.

Dillman

Frankel

Martin R.

Garland

Philip

Groves

Robert M.

Kennedy

Courtney

Krosnick

Jon

Lavrakas

Paul J.

Lee

Sunghee

Link

Michael

Piekarski

Linda

Rao

Kumar

Thomas

Randall K.

Zahs

Dan

. 2010. “Research Synthesis: AAPOR Report on Online Panels.” Public Opinion Quarterly 74 (4): 711–781. doi:10.1093/poq/nfq048.

Berinsky

Adam J.

Huber

Gregory A.

Lenz

Gabriel S.

. 2012. “Evaluating Online Labor Markets for Experimental Research: Amazon.Com’s Mechanical Turk.” Political Analysis 20 (3): 351–368. doi:10.1093/pan/mpr057.

Blum

Rachel M

. 2020. How the Tea Party Captured the GOP: Insurgent Factions in American Politics. Chicago, IL: The University of Chicago Press.

Broockman

David E.

Kalla

Joshua L.

Sekhon

Jasjeet S.

. 2017. “The Design of Field Experiments with Survey Outcomes: A Framework for Selecting More Efficient, Robust, and Ethical Designs.” Political Analysis 25 (4): 435–464. doi:10.1017/pan.2017.27.

Callegaro

Mario

Baker

Reginald P

Bethlehem

Jelke

Göritz

Anja S

Krosnick

Jon A

Lavrakas

Paul J

. 2014. Online Panel Research: A Data Quality Perspective. Chichester, West Sussex: John Wiley & Sons.

10.

Carpenter

Christopher J.

2012. “Narcissism on Facebook: Self-Promotional and Anti-social Behavior.” Personality and Individual Differences 52 (4): 482–486. doi:10.1016/j.paid.2011.11.011.

11.

Cavari

Amnon

Freedman

Guy

. 2022. “Survey Nonresponse and Mass Polarization: The Consequences of Declining Contact and Cooperation Rates.” American Political Science Review 117 (1): 332–339. doi:10.1017/s0003055422000399.

12.

Ceelen

Manon

Dorn

Tina

van Huis

Flora S.

Reijnders

Udo J.

. 2019. “Characteristics and Post-Decision Attitudes of Non-reporting Sexual Violence Victims.” Journal of Interpersonal Violence 34 (9): 1961–1977. doi:10.1177/0886260516658756.

13.

Chang

Linchiat

Krosnick

Jon A.

. 2009. “National Surveys via RDD Telephone Interviewing versus the Internet.” Public Opinion Quarterly 73 (4): 641–678. doi:10.1093/poq/nfp075.

14.

Clinton

Josh

Agiesta

Jennifer

Brenan

Megan

Burge

Camille

Connelly

Marjorie

Edwards-Levy

Ariel

Fraga

Bernard

Guskin

Emily

Hillygus

D. Sunshine

Jackson

Chris

Jones

Jeff

Keeter

Scott

Khanna

Kabir

Lapinski

John

Saad

Lydia

Shaw

Daron

Smith

Andrew

Wilson

David

Wlezien

Christopher

. Task Force on 2020 Pre-election Polling: An Evaluation of the 2020 General Election Polls. New York: American Association for Public Opinion Research, 2021.

15.

Clinton

Joshua D

Lapinski

John S

Trussler

Marc J

. 2022. “Reluctant Republicans, Eager Democrats? Partisan Nonresponse and the Accuracy of 2020 Presidential Pre-election Telephone Polls.” Public Opinion Quarterly 86 (2): 247–269. doi:10.1093/poq/nfac011.

16.

Coppock

Alexander

McClellan

Oliver A.

. 2019. “Validating the Demographic, Political, Psychological, and Experimental Results Obtained from a New Source of Online Survey Respondents.” Research & Politics 6 (1): 205316801882217. doi:10.1177/2053168018822174.

17.

Coppock

Alexander

Leeper

Thomas J.

Mullinix

Kevin J.

. 2018. “Generalizability of Heterogeneous Treatment Effect Estimates across Samples.” Proceedings of the National Academy of Sciences 115 (49): 12441–12446. doi:10.1073/pnas.1808083115.

18.

Cornesse

Carina

Blom

Annelies G

Dutwin

David

Krosnick

Jon A

De Leeuw

Edith D

Legleye

Stéphane

Pasek

Josh

Pennay

Darren

Phillips

Benjamin

Sakshaug

Joseph W

Struminskaya

Bella

Wenz

Alexander

. 2020. “A Review of Conceptual Approaches and Empirical Evidence on Probability and Nonprobability Sample Survey Research.” Journal of Survey Statistics and Methodology 8 (1): 4–36. doi:10.1093/jssam/smz041.

19.

Croco

Sarah E.

Suhay

Elizabeth

Blum

Rachel

Mason

Lilliana

Noel

Hans

Ladd

Jonathan

Bailey

Michael A.

. 2019. “Student-Run Exit Polls 101.” PS: Political Science & Politics 52 (2): 361–366. doi:10.1017/s1049096518002330.

20.

Devaux

Martin

Egami

Naoki

. 2022. “Quantifying Robustness to External Validity Bias.” SSRN Electronic Journal. doi:10.2139/ssrn.4213753.

21.

Dutwin

David

Buskirk

Trent D.

. 2017. “Apples to Oranges or Gala versus Golden Delicious?” Public Opinion Quarterly 81 (S1): 213–239. doi:10.1093/poq/nfw061.

22.

Egami

Naoki

Hartman

Erin

. 2022. “Elements of External Validity: Framework, Design, and Analysis.” American Political Science Review 117 (3): 1070–1088. doi:10.1017/s0003055422000880.

23.

Enns

Peter K.

Rothschild

Jake

. 2023. “Do You Know where Your Survey Data Come from?” Medium. https://medium.com/3streams/surveys-3ec95995dde2

24.

Federal Election Commission . 2018. “Campaign Finance Data: Contributions by Individuals 2017-2018 [Dataset].” Federal Election Commission. https://www.fec.gov/files/bulk-downloads/2018/indiv18.zip

25.

Federal Election Commission . 2020. “Campaign Finance Data: Contributions by Individuals 2019-2020 [Dataset].” Federal Election Commission. https://www.fec.gov/files/bulk-downloads/2020/indiv20.zip

26.

Fraga

Bernard L.

2018. The Turnout Gap. Race, Ethnicity, and Political Inequality in a Diversifying America. Cambridge: Cambridge University Press.

27.

Gelman

Andrew

Goel

Sharad

Rivers

Douglas

Rothschild

David

. 2016. “The Mythical Swing Voter.” Quarterly Journal of Political Science 11 (1): 103–130. doi:10.1561/100.00015031.

28.

Gooch

Andrew

Vavreck

Lynn

. 2016. “How Face-To-Face Interviews and Cognitive Skill Affect Item Non-response: A Randomized Experiment Assigning Mode of Interview.” Political Science Research and Methods 7 (1): 143–162. doi:10.1017/psrm.2016.20.

29.

Graham

Matthew H

Huber

Gregory A

. 2021. ““The Expressive Value of Answering Survey Questions.” Essay.” In The Politics of Truth in Polarized America, edited by Barker

David C

Suhay

Elizabeth

, Vol. 83. Oxford University Press.

30.

Groves

R. M.

Lyberg

. 2010. “Total Survey Error: Past, Present, and Future.” Public Opinion Quarterly 74 (5): 849–879. doi:10.1093/poq/nfq065.

31.

Han

Hahrie

. 2014. How Organizations Develop Activists: Civic Associations and Leadership in the 21st Century. Oxford: Oxford University Press.

32.

Hillygus

D. Sunshine

Jackson

Natalie

Young

McKenzie

. 2014. “Professional Respondents in Nonprobability Online Panels.” In Online Panel Research: A data quality perspective, edited by Callegaro

Mario

Baker

Reg

Bethlehem

Jelke

Göritz

Anja S.

Krosnick

Jon A.

Lavrakas

Paul J.

, 219–237. Hoboken, NJ: John Wiley & Sons. doi:10.1002/9781118763520.ch10.

33.

Hobbs

William R.

2019. “Major Life Events and the Age-Partisan Stability Association.” Political Behavior 41 (3): 791–814. doi:10.1007/s11109-018-9472-6.

34.

Hopkins

Daniel J

Halm

William

Huerta

Melissa

Torres

Josearmando

. Getting the Race Wrong: A Case Study of Sampling Bias and Black Voters in Online. Opt-In Polls 2024. https://ssrn.com/abstract=4591422.

35.

Hopkins

Daniel J.

2009. “No More Wilder Effect, Never a Whitman Effect: When and Why Polls Mislead about Black and Female Candidates.” The Journal of Politics 71 (3): 769–781. doi:10.1017/s0022381609090707.

36.

Hopkins

Daniel J.

2021. “The Activation of Prejudice and Presidential Voting: Panel Evidence from the 2016 U.S. Election.” Political Behavior 43 (2): 663–686. doi:10.1007/s11109-019-09567-4.

37.

Hopkins

Daniel J.

Noel

Hans

. 2022. “Trump and the Shifting Meaning of ‘Conservative’: Using Activists’ Pairwise Comparisons to Measure Politicians’ Perceived Ideologies.” American Political Science Review 116 (3): 1–8. doi:10.1017/s0003055421001416.

38.

Hopkins

Daniel J.

Gorton

Tori

. Unsubscribed and Undemanding: Partisanship and the Minimal Effects of a Field Experiment Encouraging Local News Consumption. American Journal of Political Science 2024; 1–17. doi:10.1111/ajps.12845

39.

Huddy

Leonie

Mason

Lilliana

Aarøe

Lene

. 2015. “Expressive Partisanship: Campaign Involvement, Political Emotion, and Partisan Identity.” American Political Science Review 109 (1): 1–17. doi:10.1017/s0003055414000604.

40.

Huddy

Leonie.

2018. “The Group Foundations of Democratic Political Behavior.” Critical Review 30 (1-2): 71–86. doi:10.1080/08913811.2018.1466857.

41.

Hughes

Adam.

2017. 5 Facts about U.S. Political Donations. Pew Research Center. https://www.pewresearch.org/short-reads/2017/05/17/5-facts-about-u-s-political-donations/

42.

Iyengar

Shanto

Lelkes

Yphtach

Levendusky

Matthew

Malhotra

Neil

Westwood

Sean J.

. 2019. “The Origins and Consequences of Affective Polarization in the United States.” Annual Review of Political Science 22 (1): 129–146. doi:10.1146/annurev-polisci-051117-073034.

43.

Jackman

Simon

Spahn

Bradley

. 2021. “Politically Invisible in America.” PS: Political Science & Politics 54 (4): 623–629. doi:10.1017/s1049096521000639.

44.

Keeter

Scott.

2018. ““The Impact of Survey Non-response on Survey Accuracy.” Essay.” In The Palgrave Handbook of Survey Research, edited by Vannette

David L

Krosnick

Jon A

, 373–381. Cham: Springer International Publishing.

45.

Kennedy

Courtney

Mercer

Andrew

Keeter

Scott

Hatley

Nick

McGeeney

Kyley

Giminez

Alejandra

. 2016. Evaluating Online Nonprobability Surveys. Pew Research Center. https://www.pewresearch.org/methods/2016/05/02/evaluating-online-nonprobability-surveys/

46.

Kreuter

Frauke

Presser

Stanley

Tourangeau

Roger

. 2008. “Social Desirability Bias in CATI, IVR, and Web Surveys: The Effects of Mode and Question Sensitivity.” Public Opinion Quarterly 72 (5): 847–865. doi:10.1093/poq/nfn063.

47.

Krupnikov

Yanna

Ryan

John Barry

. 2022. The Other Divide: Polarization and Disengagement in American Politics. Cambridge: Cambridge University Press.

48.

Krupnikov

Yanna

Nam

H. Hannah

Style

Hillary

. Convenience Samples in Political Science Experiments. Advances in Experimental Political Science 2021; 165–183. doi:10.1017/9781108777919.012

49.

Leeper

Thomas J.

2019. “Where Have the Respondents Gone? Perhaps We Ate Them All.” Public Opinion Quarterly 83 (S1): 280–288. doi:10.1093/poq/nfz010.

50.

Leighley

Jan E.

Nagler

Jonathan

. 2013. Who Votes Now? Demographics, Issues, Inequality and Turnout in the United States. Princeton, NJ: Princeton University Press.

51.

Luiten

Annemieke

Hox

Joop

de Leeuw

Edith

. 2020. “Survey Nonresponse Trends and Fieldwork Effort in the 21st Century: Results of an International Study across Countries and Surveys.” Journal of Official Statistics 36 (3): 469–487. doi:10.2478/jos-2020-0025.

52.

MacInnis

Krosnick

Jon A

Annabell S

Cho

Mu-Jung

. 2018. “The Accuracy of Measurements with Probability and Nonprobability Survey Samples: Replication and Extension.” Public Opinion Quarterly 82 (4): 707–744. doi:10.1093/poq/nfy038.

53.

Malhotra

Neil

Krosnick

Jon A.

. 2007. “The Effect of Survey Mode and Sampling on Inferences about Political Attitudes and Behavior: Comparing the 2000 and 2004 Anes to Internet Surveys with Nonprobability Samples.” Political Analysis 15 (3): 286–323. doi:10.1093/pan/mpm003.

54.

Malka

Ariel

Adelman

Mark

. 2022. “Expressive Survey Responding: A Closer Look at the Evidence and its Implications for American Democracy.” Perspectives on Politics 21 (4): 1198–1209. doi:10.1017/s1537592721004096.

55.

Medeiros

Mike

Gravelle

Timothy B.

. 2023. “Pandemic Populism: Explaining Support for the People’s Party of Canada in the 2021 Federal Election.” Canadian Journal of Political Science/Revue Canadienne de Science Politique 56 (2): 413–434. doi:10.1017/s000842392300015x.

56.

Mellon

Jonathan

Prosser

Christopher

. 2021. “Correlation with Time Explains the Relationship between Survey Nonresponse and Mass Polarization.” The Journal of Politics 83 (1): 390–395. doi:10.1086/709433.

57.

Mercer

Andrew

Lau

Arnold

Kennedy

Courtney

. 2018. For Weighting Online Opt-In Samples, what Matters Most? Pew Research Center. https://www.pewresearch.org/methods/2018/01/26/for-weighting-online-opt-in-samples-what-matters-most/

58.

Moore

Ryan T.

2012. “Multivariate Continuous Blocking to Improve Political Science Experiments.” Political Analysis 20 (4): 460–479. doi:10.1093/pan/mps025.

59.

Mullinix

Kevin J.

Leeper

Thomas J.

Druckman

James N.

Jeremy

Freese.

. 2015. “The Generalizability of Survey Experiments.” Journal of Experimental Political Science 2 (2): 109–138. doi:10.1017/xps.2015.19.

60.

Mutz

Diana C.

2018. “Status Threat, Not Economic Hardship, Explains the 2016 Presidential Vote.” Proceedings of the National Academy of Sciences 115 (19): E4330–E4339. doi:10.1073/pnas.1718155115.

61.

Mutz

Diana Carole

. 2011. Population-based Survey Experiments. Princeton, NJ: Princeton University Press.

62.

Osmundsen

Mathias

Bor

Alexander

Bjerregaard Vahlstrup

Peter

Bechmann

Anja

Petersen

Michael Bang

. 2021. “Partisan Polarization Is the Primary Psychological Motivation behind Political Fake News Sharing on Twitter.” American Political Science Review 115 (3): 999–1015. doi:10.1017/s0003055421000290.

63.

Parker

Christopher S

Barreto

Matt A

. 2014. Change They Can’t Believe in: The Tea Party and Reactionary Politics in America-Updated Edition. Princeton, NJ: Princeton University Press.

64.

Peer

Eyal

Rothschild

David

Gordon

Andrew

Evernden

Zak

Damer

Ekaterina

. 2021. “Data Quality of Platforms and Panels for Online Behavioral Research.” Behavior Research Methods 54 (4): 1–20. doi:10.3758/s13428-021-01694-3.

65.

Porter

Stephen R.

Whitcomb

Michael E.

. 2005. “Non-Response in Student Surveys: The Role of Demographics, Engagement and Personality.” Research in Higher Education 46 (2): 127–152. doi:10.1007/s11162-004-1597-2.

66.

Prosser

Christopher

Mellon

Jonathan

. 2018. “The Twilight of the Polls? A Review of Trends in Polling Accuracy and the Causes of Polling Misses.” Government and Opposition 53 (4): 757–790. doi:10.1017/gov.2018.7.

67.

Rivers

Douglas

Bailey

Delia

. 2009. “Inference from Matched Samples in the 2008 US National Elections.” In Proceedings of the Joint Statistical Meetings, Vol. 1, 627–639. Palo Alto, CA: YouGov/Polimetrix.

68.

Sambanis

Nicholas

Simonson

Matthew

Yaylacı

Şule

. 2023. “Once We Too Were Strangers: Can a Heritage of Displacement Be Leveraged to Build Support for Present-Day Refugees?” SSRN Electronic Journal. doi:10.2139/ssrn.4509421.

69.

Sances

Michael W.

2021. “Missing the Target? Using Surveys to Validate Social Media Ad Targeting.” Political Science Research and Methods 9 (1): 215–222. doi:10.1017/psrm.2018.68.

70.

Schaffner

Brian F

Luks

Samantha

. 2018. “Misinformation or Expressive Responding? What an Inauguration Crowd Can Tell Us about the Source of Political Misinformation in Surveys.” Public Opinion Quarterly 82 (1): 135–147.

71.

Schuessler

Alexander A.

1999. “Ecological Inference.” Proceedings of the National Academy of Sciences 96 (19): 10578–10581.

72.

Segura

Gary M.

Rodrigues

Helena Alves

. 2006. “Comparative Ethnic Politics in the United States: Beyond Black and White.” Annual Review of Political Science 9 (1): 375–395. doi:10.1146/annurev.polisci.9.072204.175806.

73.

Shineman

Victoria Anne.

2018. “If You Mobilize Them, They Will Become Informed: Experimental Evidence that Information Acquisition Is Endogenous to Costs and Incentives to Participate.” British Journal of Political Science 48 (1): 189–211. doi:10.1017/s0007123416000168.

74.

Stewart

Neil

Ungemach

Christoph

Harris

Adam J.

Bartels

Daniel M.

Newell

Ben R.

Paolacci

Gabriele

Chandler

Jesse

. 2015. “The Average Laboratory Samples a Population of 7,300 Amazon Mechanical Turk Workers.” Judgment and Decision Making 10 (5): 479–491. doi:10.1017/s1930297500005611.

75.

Tourangeau

Roger

Conrad

Frederick G.

Couper

Mick

. 2013. The Science of Web Surveys. Oxford: Oxford University Press.

76.

Tyler

Matthew

Iyengar

Shanto

. 2023. “Learning to Dislike Your Opponents: Political Socialization in the Era of Polarization.” American Political Science Review 117 (1): 347–354. doi:10.1017/s000305542200048x.

77.

U.S. Census Bureau QuickFacts: Pennsylvania. U.S. Census Bureau, 2021. https://www.census.gov/quickfacts/PA

78.

U.S. Census Bureau QuickFacts: United States. U.S. Census Bureau, 2021. https://www.census.gov/quickfacts/fact/table/US/PST045221

79.

Vavreck

Lynn

Rivers

Douglas

. 2008. “The 2006 Cooperative Congressional Election Study.” Journal of Elections, Public Opinion and Parties 18 (4): 355–366. doi:10.1080/17457280802305177.

80.

Verba

Sidney

Schlozman

Kay Lehman

Brady

Henry E.

. 1995. Voice and Equality: Civic Voluntarism in American Politics. Cambridge, MA: Harvard University Press.

81.

Wasserman

David

Andrews

Sophie

Saenger

Leo

Cohen

Lev

Flinn

Ally

Tatarsky

Griff

. 2024. “2020 Popular Vote Tracker.” In Cook Political Report. https://www.cookpolitical.com/2020-national-popular-vote-tracker. Accessed June 12, 2024.

82.

Weisberg

Herbert F.

2010. The Total Survey Error Approach: A Guide to the New Science of Survey Research. Chicago, IL: University of Chicago Press.

83.

Weiss

Chagai M.

2021. “Diversity in Health Care Institutions Reduces Israeli Patients’ Prejudice toward Arabs.” Proceedings of the National Academy of Sciences 118 (14): e2022634118. doi:10.1073/pnas.2022634118.

84.

Williams

Isabel

Gravelle

Timothy B.

Samara

Klar.

. 2022. “The Competing Influence of Policy Content and Political Cues: Cross-Border Evidence from the United States and Canada.” American Political Science Review 116 (4): 1375–1388. doi:10.1017/s0003055422000351.

85.

Williamson

Vanessa

Skocpol

Theda

. 2012. Tea Party and the Remaking of Republican Conservatism. Oxford: Oxford University Press.

86.

Yan

Alan

Kalla

Joshua

Broockman

David E.

. 2018. “Increasing Response Rates and Representativeness of Online Panels Recruited by Mail: Evidence from Experiments in 12 Original Surveys.” SSRN Electronic Journal. doi:10.2139/ssrn.3136245.

87.

Yeager

David S.

Krosnick

Jon A.

Chang

LinChiat

Javitz

Harold S.

Levendusky

Matthew S.

Simpser

Alberto

Wang

Rui

. 2011. “Comparing the Accuracy of RDD Telephone Surveys and Internet Surveys Conducted with Probability and Non-probability Samples.” Public Opinion Quarterly 75 (4): 709–747. doi:10.1093/poq/nfr020.

88.

Zhang

Baobao

Mildenberger

Matto

Howe

Peter D.

Marlon

Jennifer

Rosenthal

Seth A.

Leiserowitz

Anthony

. 2018. “Quota Sampling Using Facebook Advertisements.” Political Science Research and Methods 8 (3): 558–564. doi:10.1017/psrm.2018.49.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.42 MB

0.00 MB