Research ArticleSOCIAL SCIENCES

# Religious change preceded economic change in the 20th century

See allHide authors and affiliations

Vol. 4, no. 7, eaar8680

## Abstract

The decline in the everyday importance of religion with economic development is a well-known correlation, but which phenomenon comes first? Using unsupervised factor analysis and a birth cohort approach to create a retrospective time series, we present 100-year time series of secularization in different nations, derived from recent global values surveys, which we compare by decade to historical gross domestic product figures in those nations. We find evidence that a rise in secularization generally has preceded economic growth over the past century. Our multilevel, time-lagged regressions also indicate that tolerance for individual rights predicted 20th century economic growth even better than secularization. These findings hold when we control for education and shared cultural heritage.

## INTRODUCTION

A classic sociological question is whether the decline of religious activity, or secularization, has been caused by economic development (1, 2). A century ago, Durkheim (3) proposed that technological and socioeconomic advances come to displace the functions of religion (4, 5), whereas Weber (6) contended the opposite, that monotheistic religion—the so-called Protestant ethic—made the development of capitalism possible.

Although a correlation between economic development and secularization is evident, in that countries that are highly religious tend to be the poorest (7, 8), it is not obvious which change precedes which through time: whether development causes secularization (9, 10), or vice versa (11), or whether both changes are driven, with different time lags, by a factor such as education or advances in technology (2, 12).

Whereas some studies find a bicausal relationship between income and religion (1315), causality effectively remains unknown as feedbacks may change through time and development. Organized religious charity, for example, might initially encourage certain values that facilitate economic development while restricting individual expression (16), but then the resulting economic development may subsequently reward individualism. Tolerance of individual expression may then feed secularism, partly by undermining religious organizations that provide communities with resources and social capital; other causal arguments also remain feasible (17, 18).

To characterize their temporal relationship, we use 20th century data for both economic development and a measure of secularization extracted from international cultural values surveys. Time-lagged regressions using these 100-year time series for key variables can determine which variable precedes the other. This can be used to rule out certain hypotheses of causality. If, for example, changes in series X precede those in series Y, then one can say that Y does not cause X even while it is not certain that X causes Y.

For indices of cultural values, we use data from the European Values Survey (EVS) and the World Values Survey (WVS) since 1990. To estimate values from all decades of the 20th century, here, we make use of birth cohorts (see Materials and Methods); the utility of this technique is among our key findings. The WVS and EVS featured some of the same survey questions, so we combined these survey data sets, which we refer to as the WEVS. The data we use for economic development are data on historic gross domestic product (GDP) per capita from the Maddison Project (19), which, although more complex indices such as the Human Development Index can be preferable (20), we use because GDP data exist for many nations over the entire 20th century. We also include three control variables, each of which covers the set of nations around the world and extends to the beginning of the 20th century. The first, which recently became available, is an extensive international time series on participation rates in tertiary education, which stretches back to the early 20th century (21). The second is language family, which is used as a proxy for the nested random effect of nonindependent cultural and economic histories between individual countries (22). The third is a measure of tolerance of others, extracted as a factor from the WEVS, as discussed below.

## RESULTS

Having applied exploratory factor analysis (EFA) to the WEVS data set, we selected nine factors to retain, each interpretable using questions that did not overlap between factors (see Materials and Methods). Factor 1, explaining 11% of the variance, was strongly loaded on questions regarding the importance of religion in one’s life (see Materials and Methods). This factor defines our measure of secularization, S, as a composite variable of WEVS responses (table S4 lists all elements of the secularization factor).

Another factor from the EFA, factor 8, was strongly loaded on survey participants’ willingness to tolerate behaviors that are often socially prohibited, such as suicide, homosexuality, or abortion (see table S11). We label this factor as “tolerance,” denoted as V. We explore tolerance, V, as a control variable in our results for two reasons. First, as we will see, the changes in tolerance were closely correlated with the changes in secularism during the 20th century. Second, the tolerance factor was strongly correlated (R = 0.59) with Hofstede’s (23) metric of individualism (we did not extract individualism directly from Hofstede’s data because it is not broken down by birth cohort as we require).

Having extracted these factors measuring secularization, S, and tolerance, V, we aim to extend the information collected in the WEVS during the last quarter-century back to the beginning of the 20th century. To do this, we treat decade of birth, t, recorded for each person surveyed, as a proxy for a historical time period. Although there will be differences from one survey period, p, to the next, the differences apply across all birth cohorts such that the relative differences between birth cohorts were generally maintained through time (5).

Using a likelihood ratio test of the hypotheses presented in Materials and Methods, we confirm that estimates for each birth decade, t, are independent of survey period, p, in that there is no evidence of temporal dependence for St,p or for Vt,p in 91% and 89% of the countries tested, respectively (table S13 shows full results). This confirms that generational trends persist through time. The results make up the elements of an array, Si,t,p, of estimates of secularization in each country, i, during birth decade, t, for survey period, p.

Next, we determined secularization by birth decade in each country, Si,t, by averaging factor 1 across all available survey periods, p, in country, i, corresponding to decade of birth, t. Figure 1 illustrates the temporal trend in secularization across birth decade for several countries with little missing data: Great Britain, the Philippines, Chile, and Nigeria. For countries with missing data, missing values of Si,t were imputed (see Materials and Methods). The same procedure is used to obtain the tolerance score matrix Vi,t.

We compared Si,t versus historical GDP per capita (in 1990 US$) from each country through time. Figure 2 compares Si,t versus the decadal mean GDP, GDPi,t, for the same four countries as Fig. 1. We find evidence that changes in secularization, Si,t, precede changes in GDPi,t, as most clearly seen in reversals of the trend, when a decrease in S occurs shortly before a corresponding decrease in GDP. To test whether changes in Si,t generally precede changes in GDPi,t, or vice versa, we estimate multilevel time-lagged regressions. By including data from all countries in a single test, a multilevel model can maximize the statistical power available in these data. It also allows us to control for non-independence due to shared cultural heritage h, which we do by classifying countries by language family (see Materials and Methods). The multilevel model is(1)(2)where St and GDPt are secularization and economic development in decade t, respectively, and Sty and GDPty are the respective values lagged by y decades. The term hi represents a nested random effect due to the cultural-historic grouping of country i, for which language family is the proxy (22). This term is used as a control for non-independent similarities between individual countries, due to their development and secularization already present at the start of the 20th century. Models 2 and 5 (Table 1) show that changes in St precede those of GDPt and not the other way around. This directionality is independent of time lag, y, as the full results for lags of one decade, two decades, and three decades show (table S13). An increase in St by 1 SD corresponds to$1000 increase in GDPt per capita after 10 years, $2800 after 20 years, and$5000 after 30 years. Our robustness checks show that this result is stable (see Materials and Methods). It is independent of the age we consider a birth cohort economically active (table S15).

Table 1 Selected time-lagged linear regressions (labeled models M2, M5, etc.) between secularization (S), development (GDP), tolerance (V), and education (E).

The time lag is y = 2 decades in all cases (results for y = 1, 2, and 3 decades in table S14). SEs, in parentheses, were determined from the inverse of the negative Hessian matrix (44). N is the number of data points for each autoregression, n is the number of countries included in the data set, i is the percentage of residual variance explained by the random effect (country), and h is the percentage explained by cultural heritage. R2 is the total variance explained. Bonferroni-corrected significance: *P < 0.1, **P < 0.05, ****P < 0.01.

View this table:

### Tertiary education enrollment

We used tertiary education enrollment rates as a proxy measurement for science education. The “Barro-Lee Educational Attainment” data set gives time series for tertiary enrollment, taken mainly from census data and from intergovernmental organizations, and stretches back to 1820 in the most recent edition (21). We took the average rate of enrollment in each decade to correspond to the cultural values data, which is in decadal increments. The coverage is less comprehensive than the WEVS, with only 74 countries covered. Data for most non-Russian former Soviet states are missing because most were not independent states for most of the 20th century; the same is true for Yugoslavia. Some small countries or semiautonomous regions of another country are also missing, such as Northern Ireland. Finally, poorer countries—mainly Islamic or African ones—are missing because tertiary educational enrollment statistics could not be obtained.

### Language categories (proxy for cultural relatedness)

To avoid Galton’s problem, we have to control for shared culture. Often, this is done using language phylogenies (22), but this requires all societies under study to be from the same language tree with the requisite branch lengths calculated (42). The countries in our global sample speak languages from many different language families, which rules out the use of phylogenetic trees. To control for cultural history, h, in the time-lagged regressions, we discretely categorized the countries based on language families and treat it as a random effect. These data were taken from the Ethnologue database (43), which documents all known extant languages, and the countries in which they are currently the predominant language. The 109 countries were categorized into the following language families (number of countries): Albanian (2), Semitic (17), Italic (23), Greek-Armenian (3), Germanic (23), Turkic (6), Indo-Aryan (4), Balto-Slavic (14), Sino-Tibetan (3), Uralic (3), Kartvelian (1), Austronesian (3), Japonic (1), Niger-Congo (3), Korean (1), Tai (1), and Austroasiatic (1). Table S17 contains the language group assigned to each country.

### Use of birth cohort to extend data set through time

The WVS component of the combined WEVS data set was carried out during five distinct “waves,” carried out at approximately 5-year intervals, between 1990 and 2015. This provides a maximum of five data points per country (not all countries participated in all five waves) in a time series reaching back only 25 years. Given the recorded decade of birth of the survey respondents, however, we can, by assumptions confirmed below, extend these data back to represent all decades of the 20th century. This yields a matrix St,p of values for each country, with decade of birth, t, and survey period, p, as the rows and columns, respectively (for inclusion, a birth cohort must contain at least 100 individuals). To account for birth cohorts that are not represented in all time periods, which could otherwise bias the mean across time periods, we imputed the missing values using the following linear model(4)where t is the birth decade, p is the survey period, and μp and α are the estimated slope and intercept, respectively, for imputation of the missing value(s). Once missing values were imputed, we then defined St for each birth decade, t, as the mean across all survey periods, p. The result is a 10-point time series over the past century (rather than 5 points over 25 years) for the 109 countries in the WEVS, with some countries having only partially complete time series (for example, Nigeria has data from only seven decades). Importantly though, these values should not be interpreted as the true values, which would have been measured had the WEVS existed in earlier decades of the 20th century, except possibly when no period effect is present.

We tested the preservation of generational trends in cultural values (5) by using a likelihood ratio test to determine whether an interaction term between birth decade and survey period provides explanatory value for the data. Specifically(5)(6)where St,p is secularization, but could also be tolerance of homosexuality and abortion Vt,p. Each country was subjected to this test. We reported the likelihood ratio and the proportion of the variance explained by H1, not explained by H0. Further, using the χ2 distribution, we calculated asymptotic significance values to quantify the evidence that H1 was a better explanation for the WEVS data than H0, that is, whether estimates for each birth decade t were independent of survey period p. This test was carried out for 79 countries because we were limited to those who appeared in two or more waves of the WEVS.

### Multilevel time-lagged linear regressions

We chose a time-lagged model to express secularization (St) as a function of historical development (GDPty) while controlling for historical secularization data (Sty), where y is the lag in decades. Unlike a standard time-lag test, however, which normally requires two long individual time series, we have many time series (103 countries) that have 10 points or fewer (limited to number of decades in the 20th century). To control for cultural non-independence between countries, which is a nested random effect, we categorized countries by language family—as the best available proxy for cultural similarity—designated by variable hi for country i. This amounted to two nested random effects for each designated cultural heritage h, within each country i. To avoid multiple testing and low statistical power, we formulated a multilevel model to incorporate data from all countries into a single test(7)where (1|hi) is the nested random effect for a country i from language category h, ϵ is the error, and we let y = 1, 2, or 3 decades. Using the control variable, hi, present in all of the time-lagged equations (Eqs. 1, 2, 7, and 12), we found that this nested random effect did not substantially change our results (Table 1 and tables S14 to S17). This indicated that religious change predicted economic change while controlling for language as a proxy for shared history.

To deal with missing data in the GDPt and/or St time series for certain countries, we chose to omit the missing values rather than attempt to impute them without an obvious universal model to describe how secularization or GDP changes. However, despite omitting variables, we still obtained sizable contributions from the major cultural groups (except for the ex-Soviet states that lack credible GDP data before 1990). We also reported the number of countries represented in the data and the number of total data points; both depended on the time lag used.

To test the alternative hypothesis that economic development precedes secularization, we ran a similar test to see whether St in a birth cohort predicted GDP y decades later(8)We also tested the effect of tolerance of behaviors such as homosexuality and abortion, V, on either S or GDP. We added V as a control in the time-lagged regressions(9)(10)We also wanted to test the effect of advanced education E, so we similarly added a variable representing the tertiary education enrollment rate. Once again, testing a lag of y = 1, 2, or 3 decades(11)(12)We normalized St and Vt so that the SD of each is equal to 1. This allowed us to state the dollar improvement in GDP resulting from 1 SD change in both St and Vt.

### Robustness checks

When comparing GDP data versus our estimates of secularization (St) for given birth decade, t, we make no assumption about the age at which a birth cohort begins to affect the economy; economic development can affect cultural values during formative years, whereas people will not normally influence the economy until they are older. To ensure that our results are robust to this uncertainty, we ran the S-GDP regressions considering coincidence points between development and secularization in birth cohorts: childhood (+0), teenage years (+10), and twenties (+20). The results in table S15 show that, under all of these scenarios, secularization precedes economic development and not the other way around.

We also tested the robustness of our multilevel, time-lagged regressions to ensure that random noise in the EFA factors did not affect the regression results. To do this, instead of defining secularization with EFA factors, we defined secularization as the simple mean of six relevant WEVS variables (see table S16), each normalized to mean zero and unit variance. We found that the predictive structure that emerges (table S16) is the same as when we used the factors derived through EFA.

## SUPPLEMENTARY MATERIALS

Section S1. WVS and EVS

Section S2. Exploratory factor analysis

Table S1. Participation in the different waves of the WVS and EVS.

Table S2. Participating countries in the WEVS.

Table S3. Questions common to all eight waves of the WVS and EVS.

Table S4. Secularization.

Table S5. Institutional confidence.

Table S6. Openness to intrinsic differences.

Table S7. Prosociality.

Table S8. Interest of politics.

Table S9. Wellbeing.

Table S10. Political engagement.

Table S11. Tolerance of prohibited behaviors.

Table S12. Openness to extrinsic differences.

Table S13. Independence of birth decade, t, versus WEVS phase, p, for secularization, St,p, and tolerance, Vt,p.

Table S14. Multilevel time-lagged linear models (see Materials and Methods) demonstrating that secularization predicts GDP and not vice versa (models 1 to 6); tolerance predicts GDP better than secularization (models 7 to 12) and education predicts future GDP, but not secularization (models 13 to 18).

Table S15. Time-lagged models, models 1 to 6 (see Materials and Methods), of S versus GDP for cohorts in their first decade or childhood (y = 0 decades, top row), teenage years (y = 1 decade, middle row), and twenties (y = 2 decades, bottom row).

Table S16. Multilevel time-lagged models, but with secularization (Salt) measured using the average of six indicators, which are subjectively associated with religiosity.

Table S17. Language categories assigned to WEVS countries, using Ethnologue data.

Fig. S1. The ordered factor loadings on WEVS survey questions, following EFA analysis with oblique rotation.

This is an open-access article distributed under the terms of the Creative Commons Attribution license, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

## REFERENCES AND NOTES

Acknowledgments: Funding: D.J.R. is supported by an Engineering and Physical Sciences Research Council grant to the Bristol Centre for Complexity Sciences (EP/I013717/1). D.J.L. was supported by Wellcome Trust grant number WT104125AIA. R.A.B. and D.J.R. were further supported by a grant from the Hobby Center for Public Affairs, University of Houston. High-performance computing facilities were provided by Blue Crystal at the Advanced Computational Research Centre, University of Bristol, UK. Author contributions: D.J.R., R.A.B., and D.J.L. designed research; D.J.R., D.J.L., and R.A.B. performed research; D.J.L. contributed new analytic tools; D.J.R. analyzed data; and D.J.R., R.A.B., and D.J.L. wrote the paper. Competing interests: D.J.L. is a director of GENSCI Ltd. All other authors declare that they have no competing interests. Data and materials availability: WVS data: www.worldvaluessurvey.org/WVSDocumentationWVL.jsp. EVS data: www.europeanvaluesstudy.eu/page/longitudinal-file-1981-2008.html. Historic GDP data: www.ggdc.net/maddison/maddison-project/home.htm. All other data and author-written code for this study: https://github.com/dr2g08/Religious-change-preceded-economic-change-in-the-20th-century. All data needed to evaluate the conclusions in the paper are present in the paper, Supplementary Materials, and/or the listed repositories. Additional data related to this paper may be requested from the authors.
View Abstract