How Many People Attend Religious Services? Part 1

I knew there were serious problems with at least one of the statistics as soon as I skimmed the newspaper article. Paragraph 2 of Magruder (2024) states:

But since 2000, church attendance has dropped in the U.S., from 42% of Americans worshipping every week or so back then, to 30% today, according to a 2024 Gallup poll. Only 11% of Americans report being in the pews monthly, and 56% seldom or never are. 

But paragraphs 15 and 16, on the following page, state:

Nationally, the COVID-19 pandemic did not help church attendance. 

Beforehand, 75% of Americans said they attended religious services at least monthly, a number that fell to 68% by mid-2022, according to report by American Enterprise Institute and the University of Chicago.

The difference between the two statistics on how many Americans attend religious services at least once a month (41% from the statistic attributed to the Gallup poll, 68% from the statistic attributed to the American Enterprise Institute, AEI) is much too large to be explained by sampling error. Curious about the discrepancy, and thinking that a 68% number for monthly attendance seemed way too high, I decided to track down the data sources and investigate.

It turns out that the 68% statistic is much too high. Magruder (2024) had a typo: The AEI report says that 68% of US adults attend religious services at least once a year, not once a month (Benz et al., 2023, pp. 3-4).

But the statistics on religious service attendance from the two polls are still farther apart than can be explained by sampling error (the margin of sampling error is less than 1.5 percentage points for each survey). Let’s look at the statistics and explore some other factors that might affect them. This type of investigation could be done for any pair of data sources giving different results, and might form the basis for a class discussion on evaluating statistics.

Gallup and American Enterprise Institute Survey Estimates

Table 1 gives the statistics on religious service attendance from the Gallup and AEI surveys. Because the two surveys have different question wording and ask about different frequency categories for religious service attendance, I lined up categories that are approximately comparable (according to the descriptors from each survey) in Table 1. For the approximate lineups in the last two rows, the AEI attendance statistics are somewhat lower than those from Gallup. AEI has 24% of adults attending “regularly” while the Gallup polls from 2021 to 2023 have 30% attending every or almost every week.

Table 1. Estimated percentages of U.S. adults reporting various frequencies of attending religious services. Source: Jones (2024) and Benz et al. (2023).

Let’s look at some possible explanations for these differences. To do this, we need to look at how each sample was selected, what questions were asked, and how estimates were calculated — information that should be given in the methodology report for a survey.

How Were the Samples Selected?

Gallup asks about religious affiliation in most of its surveys, and the statistics quoted in the newspaper article were calculated by aggregating data from telephone surveys conducted between 2021 and 2023 (about 32,000 respondents altogether; see Jones, 2024). Data from the 9,425 persons used to calculate the AEI statistics were collected for the 2022 American Religious Benchmark Survey by NORC at the University of Chicago (see Benz et al., 2023). These sample sizes are much larger than those for a typical poll (which usually has between 500 and 2000 respondents) in order to allow separate attendance frequency statistics to be calculated for specific religions.

The methodology report for the Gallup Social Survey says that landline and cell phone numbers for an individual poll were sampled using random digit dialing. The AEI sample was drawn from AmeriSpeak®, NORC’s online panel that serves as a source for many of its surveys. Both sample selection methods yield representative samples of U.S. adults if every eligible person who is selected to be in the sample participates in the poll.*

Many people, however, do not answer their telephones or agree to participate in online surveys. If people who do not participate in the survey have different opinions or behavior from those who take the survey, estimates may be systematically off target. The Gallup Social Survey methodology report does not give the response rate or other details about the quality of the data, but Marken (2018) reported a response rate of 7% in 2017. The overall response rate for a typical AmeriSpeak® survey, which includes participating in the panel and then agreeing to be in a particular survey, is between 10% and 15%.** We’ll come back to potential effects of nonresponse after looking at the questions and how people might answer them.

Question Wording and Ordering

How a question is worded can affect the answers that people give. Gallup asked: “How often do you usually attend church, synagogue, mosque or temple — every week, almost every week, about once a month, seldom or never?” The AEI survey asked: “How often do you attend religious services?” with response categories never, less than once per year, about once or twice a year, several times a year, about once a month, two or three times a month, nearly every week, every week, or several times a week.

The questions are similar, but not identical: the Gallup question asks about attending services at four specific types of religious centers, while the AEI question asks about religious services in general. The AEI question on attendance follows a question about what the respondent’s present religion is (if any), with 12 response options. It’s possible that some types of religious service attendance (for example, a service held in someone’s home or a school gymnasium) might be reported on the AEI survey but not the Gallup survey. On the other hand, the specificity of the Gallup question may prompt people who attend at those locations to report their service attendance.

The order of response categories can also affect responses. People have a tendency to pick earlier categories in a list, so those presented with the weekly-to-never ordering (Gallup) might report a higher frequency of attending than those presented with a never-to-weekly ordering (AEI).

Social Desirability

Survey researchers have long known that respondents tend to overreport behavior and opinions that society views as “good” and they tend to underreport behavior viewed as “bad.” This is called social desirability bias. Thus, some people exaggerate their charitable donations or their physical activity levels to a survey interviewer, while others say they have never shoplifted or cheated on their taxes (when in fact they have).

There is a lot of research indicating that self-reports of religious service attendance are subject to social desirability bias. An early study in one Ohio county by Hadaway et al. (1993) found that levels of church attendance estimated from survey responses were twice as high as levels of attendance estimated from church records. Chaves and Stephens (2003), Brenner (2011), and Brenner et al. (2024) review some of the studies that have been done to investigate social desirability bias, and suggest methods for reducing it.

One method that has been shown to reduce social desirability bias is using a self-administered survey, as opposed to having an interviewer ask questions (Presser and Stinson, 1998; Kreuter et al., 2008). This may partially explain why the AEI poll estimates lower levels of religious service attendance than the Gallup poll. In the Gallup poll, questions are asked by an interviewer over the telephone. For the AEI poll, respondents answer the questions directly on a screen, with no worries that an interviewer might think poorly of them if they say they don’t attend services.

Nonresponse

Nonresponse is another potential sources of bias in the survey estimates. The response rates for both polls are low — fewer than 15% of the people selected for the sample end up participating. Survey organizations use weighting to attempt to compensate for nonresponse but bias will remain after weighting if, for each weighting subgroup, nonrespondents tend to have higher (or lower) levels of religious service attendance than respondents.***

When answers to the question of interest are correlated with the weighting variables, weighting can substantially reduce nonresponse bias. Estimates about volunteer activity or civic involvement, however, often exhibit substantial bias even after weighting. People who volunteer or participate in other civic-minded activities are also more likely to participate in surveys when asked (Abraham et al., 2009). This tendency for volunteers to be more likely to participate in surveys occurs across demographic groups, so weight adjustments based on demographic information alone do not remove the bias.

It is unclear whether religious service attendance has the same type of nonresponse bias as volunteerism, though. Many researchers investigate effects of nonresponse by comparing the people who readily respond to the survey when contacted (initial respondents) with people who require a lot more effort to secure their participation (followed-up respondents). If initial and followed-up respondents have different levels of religious service attendance, then nonrespondents may have different levels as well. Nonrespondents are often assumed to be more similar to the difficult-to-get respondents than the easy-to-get respondents, but we cannot tell for sure unless we get data from the nonrespondents.

The studies to date have mixed evidence on the direction of nonresponse bias for religious service attendance. Brenner (2019) found slightly higher rates of church attendance among the initial telephone respondents than among those who did not respond by telephone but were followed up in person.

On the other hand, the AmeriSpeak® methodology report reported that churchgoers are less likely to be initial respondents when recruited for the panel: “For example, compared to the panelists recruited during the initial stage, NRFU panelists tend to be more conservative politically, more likely to attend church, less interested in current events or topics in the news report, less knowledgeable about science, less likely to be in favor of gun control policies, less likely to read a print newspaper (more likely to read the news online and use social media), more likely to eat at fast-food restaurants, and so on.”

Looking at poll responses alone, without information on the actual behavior of the respondents and the answers and behavior of nonrespondents, cannot tell us whether there is social desirability or nonresponse bias for measuring religious service attendance. We need to look at alternative data sources to do that, which is the subject of the next post.

Copyright (c) 2025 Sharon L. Lohr

Footnotes and References

*Gallup’s method gives a randomly selected sample of telephone numbers and, if everyone answered the calls, would result in a representative sample of U.S. adults with telephone access, which is almost everyone (Pew Research estimates that 98% of people now have a cell phone). Gallup purchased samples of cell and landline numbers from a data vendor, and designed the survey so that at least 80% of the telephone numbers in the sample were from cell phones. In the 1960s, more than 95% of US households had landline telephones and most people answered the telephone when called. As people began using cell phones and ditching their landlines, however, a telephone survey that called only landline numbers missed large parts of the population. Survey organizations started taking two samples of telephone numbers — one of landline numbers and one of cell numbers — to be able to include cell-phone-only households in the survey. Calculating estimates gets a little tricky, since some people have both landline and cell phone and sometimes multiple adults share a telephone, and survey organizations weight the data so that people with multiple chances of being selected for the survey aren’t overcounted in the results (see Lohr, 2021, 2025, for descriptions of how weighting is used).

What is random digit dialing (RDD)? Early telephone surveys were conducted by selecting numbers at random out of telephone directories, but this missed people with unlisted numbers. To give all residential telephone numbers — including unlisted ones — a chance to be in the sample, telephone survey organizations often used RDD. A U.S. telephone number has a three-digit area code followed by a three-digit prefix and an additional four digits. For example, the telephone number for the switchboard of the U.S. House of Representatives is (202) 224-3121, where the area code is 202 and the prefix is 224. Survey organizations would take a random sample of area codes and prefixes in use. If, say, that sample contained area code 202 and prefix 224, they would then generate a random number between 0000 and 9999 (let’s say the random number generated was wxyz) and call the number (202) 224-wxyz. Some numbers, of course, were unassigned or belonged to businesses or organizations (and hence did not lead to an eligible respondent), and numerous innovations were introduced to make RDD more efficient. Today, however, many people do not answer their phones and it is often unknown whether an unanswered number is ineligible for the survey or it belongs to a nonrespondent. There are thus multiple ways of calculating a response rate for a telephone survey, depending on how one treats unanswered calls.

According to the methodology report for AmeriSpeak®, NORC’s online panel is recruited from a variety of sampling frames using probability methods. Initial recruitment is by USPS or telephone contact. This ensures that people without internet access are eligible to be selected for the panel. Recruited persons are given the option of participating by web or by telephone. The set of persons selected for the panel is a probability sample from the U.S. adult household population; if everyone who was selected agreed to participate in the panel and in subsequent surveys, the participants would be representative of the U.S. household population.

**These response rates are actually higher than those for a typical public opinion poll today. Election polls often have response rates of 1% or lower. Many polls are taken via opt-in samples, which have a response rate close to zero since the sample is composed of volunteers.

***See Chapter 5 of Lohr (2019) for a nontechnical illustration of how weighting works. For example, if the sample ends up with 250 men and 750 women but the population has half men and half women, each man in the sample would be assigned a weight three times as large as the weight for a woman in the sample. Estimates would be calculated (conceptually) by making two extra copies of each man’s answers and then averaging the 1500 answers in the pseudo data set. But this completely removes nonresponse bias only if all of the differences between respondents and nonrespondents can be explained by gender. If churchgoing men and women are both more likely to respond to the survey, weighting by gender may reduce the bias but will not take care of all of it.

Abraham, K.G., Helms, S. and Presser, S. (2009). How social processes distort measurement: The impact of survey nonresponse on estimates of volunteer work in the United States. American Journal of Sociology, 114(4), 1129-1165.

Benz, J., Witt-Swanson, L., and Cox, D.A. (2023). Faith After the Pandemic: How COVID-19 Changed American Religion. Washington, DC: American Enterprise Instititute.

Brenner, P. S. (2011). Exceptional behavior or exceptional identity? Public Opinion Quarterly 75(1), 19–41.

Brenner, P. S. (2019). Does survey nonresponse bias estimates of religious service attendance? Evidence from an address-based sample from the Boston aArea. Sociology of Religion, 80(2), 247-262.

Brenner, P. S., Laplante, J., and Reed, T.L. (2023). Sources of inconsistency in the measurement of religious affiliation: Evidence from a survey experiment and cognitive interviews. Sociology of Religion, 85, 404-428.

Chaves, M. and Stephens, L. (2003). Church attendance in the United States. In Handbook of the Sociology of Religion, ed. M. Dillon. New York: Cambridge University Press, 85-95.

Hadaway, C. K., Marler, P. L., and Chaves, M. (1993). What the polls don’t show: A closer look at US church attendance. American Sociological Review 58, 741–52.

Jones, J.M. (2024). Church attendance has declined in most U.S. religious groups. https://news.gallup.com/poll/642548/church-attendance-declined-religious-groups.aspx.

Kreuter, F., Presser, S. and Tourangeau, R., 2008. Social desirability bias in CATI, IVR, and web surveys: The effects of mode and question sensitivity. Public Opinion Quarterly 72(5), 847-865.

Lohr, S.L. (2019). Measuring Crime: Behind the Statistics. Boca Raton, FL: CRC Press.

Lohr, S.L. (2021). Multiple-frame surveys for a multiple-data-source world. Survey Methodology, 47, 229-263.

Lohr, S.L. (2025). Multiple frame methods for combining data sources. The Survey Statistician, in press.

Magruder, J. (2024). Tempe church launches new way to worship. Tempe Tribune (December 1), pp. 10-11.

Marken, S. (2018). Still listening: The state of telephone surveys. https://news.gallup.com/opinion/methodology/225143/listening-state-telephone-surveys.aspx

Presser, S. and Stinson, L. (1998). Data collection mode and social desirability bias in self-reported religious attendance. American Sociological Review, 63(1), 137-145.