!"#$%&'$(#)*"*+,*-./$01*'#&2#$34+5$%$674##89%1#$:%&#)$!&%).,*,!;-7+4<,=/$!5.$!4";#>$?%1*'$(@$A+7&,+&>$B.&&$C@$97*-#D+;42#/$A+;4&%)$3+4$-7#$D2*#&-*3*2$D-;'.$+3$(#)*"*+&>$E+)@$FG>$H+@$F$<D#I@>$JKKK=>$II@$LMF8LFN:;O)*,7#'$O./$P)%2QR#))$:;O)*,7*&"$+&$O#7%)3$+3$D+2*#-.$3+4$-7#$D2*#&-*3*2$D-;'.$+3$(#)*"*+&D-%O)#$S(B/$http://www.jstor.org/stable/1387762!22#,,#'/$JFTJUTMUUG$JF/MV

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available athttp://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unlessyou have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and youmay use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained athttp://www.jstor.org/action/showPublisher?publisherCode=black.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printedpage of such transmission.

JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with thescholarly community to preserve their work and the materials they rely upon, and to build a common research platform thatpromotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.

Society for the Scientific Study of Religion and Blackwell Publishing are collaborating with JSTOR to digitize,preserve and extend access to Journal for the Scientific Study of Religion.

http://www.jstor.org

Age and Religiosity: Evidence fromt a Three-'Wave Panel Analysis

AMY ARGUE t DAVID R. JOHNSONt LYNN K. WHITEt

Using pooled time series with random and fixed effects regression models, we examine the effect of age, period, and family life course events on a measure of religious influence on daily life in a panel of 1,339 adults interviewed three times between 1980 and 1992. The results showv a significant, non-linear increase in religiosity with age, with the greatest increase occurring between ages 18 and 30. We also found a significant decline in religiosity between 1980 and 1988, but no evidence of a period effect between 1988 and 1992. Comparison of fixed and random effects solutions found little evidence that a cohort effect accounted for the age findings. The age effect was significantly stronger for Catholics than Protestants and the lower religiosity of males was also significantly stronger for Catholics. Adding children in the range from age two to ten significantly increased religiosity, but family life course events accounted for little if any of the age effect.

Prior research demonstrates substantial ambiguity regarding the extent and degree of an age effect on religiosity, the degree to which this effect reflects aging, family life cycle, or period processes, and the extent to which research findings are affected by measurement and study design. We provide additional empirical evidence on these issues by conducting random effects and fixed effects pooled-time series analyses of multiwave panel data extending over a twelve year period. Although no single technique can resolve the age- period-cohort riddle, this model allows us to estimate age effects controlling for cohort effects and to estimate possible period effects between 1980 and 1992.

BACKGROUND

The observed cross-sectional relationships between-age and religiosity are generally explained by one of three theoretical processes. The "traditional model' (Bahr 1970) focuses on developmental processes related to age per se. Alternatively, a life course model (Chaves 1991) attributes change not to developmental processes but to correlated changes in social roles, particularly in the family. A third interpretation characterizes observed variations in religiosity by age as a statistical artifact associated either with cohort replacement or period effects.

The biggest controversies in studies of contemporary religiosity have been reserved for the study of period effects, specifically secularization (e.g., Chaves 1991; Firebaugh and Harley 1991; Hout and Greeley 1990). Chaves (1989) concludes there is no age effect on church attendance in the 1972-6 period, but most scholars argue that the major processes

t Amy Argue i.s a doctoral candidate in the Department of Sociology, University of North Carolina, Chapel Hill, NC 27599-3210.

David Johnson is a professor of sociology at the University of Nebraska-Lincoln, Lincoln, NE 68588-0324. E-mail: djohnson2@unl.edu.

tLynn White is a professor of sociology at the University of Nebraska-Lincoln, Lincoln, NE 68588-0,324.

?Journal for the Scientific Study of Religion, 1999, 38(3): 423-435 423

JOURNAL FOR THE SCIENTIFIC STUDY OF RELIGION

operating in recent decades are cohort replacement (resulting in declining religiosity) and a relatively stable main effect of age (Firebaugh and Harley 1991; Hout and Greeley 1990). Perhaps the only point of agreement among all of these scholars is that processes differ by religious affiliation and that Catholics have been more affected than Protestants by, if not secularization, disaffection from the church hierarchy (Hout and Greeley 1987). Without exception, however, these studies rely on pooled cross-sectional results and so are limited in their ability to observe change within individuals.

Four recent studies that use panel data on single cohorts of high school or college students followed for 10 to 15 years provide a complement to these pooled cross-sections. Despite a diverse set of dependent variables, the single cohort panels of young adults demonstrate substantial similarities in outcomes. Becoming married and a parent are associated with less likelihood of apostasy (Sandomirsky and Wilson 1990), more likelihood of returning after a period of apostasy (Wilson and Sherkat 1994), and greater increases in religion as a source of life satisfaction (Hoge and Hoge 1984). In the most comprehensive of these studies, Stolzenberg, Blair-Loy, and Waite (1995) demonstrate that parenthood and marriage are associated with increased likelihood of belong to a church or synagogue. They show that the maximum effect of children on church membership occurs when children reach age 10 and declines thereafter. Similarly to Sandomirsky and Wilson (1990), they show that men's religiosity is more affected than women's by changing family statuses. Restriction to a relatively narrow age range and to a single cohort, however, makes it. virtually impossible for these studies to distinguish between age and period effects.

As far as we are aware, no prior study has applied pooled time series techniques to multiwave panel data of religious change. The data that we use are from a twelve year, four-wave national study of 2,033 persons, all of whom were married and under age fifty- five at the time of the first interview in 1980 (Booth, Amato, Johnson, and Edwards 1993). The age range of this sample (ages eighteen to fifty-five in wave one) is substantially larger than in previous panel studies, and the twelve year period examined in the study is long enough to permit us to observe both period and life course effects. These advantages come with two costs. First, because all respondents were married at the first wave, we cannot pxamine the effect of transition to marriage as a mediator of the age effect. We can, how- ever, examine birth of the first child, the presence of school-aged children, the empty nest, divorce, and widowhood. Second, the indicator of religiosity available in this study is a measure of religious influence on daily life rather than the usual church attendance measure. Considering criticisms of church attendance measures (for example, Hadaway, Marler, and Chaves 1998; Hout and Greeley 1998), this analysis may provide useful evidence on the robustness of age relationships across alternative measures of religiosity.

DATA AND METHODS

Sample

The Marital Instability Over the Life Course study interviewed a national sample in 1980, 1983, 1988 and 1992. A clustered random-digit dialing design generated sample households, and a random technique determined whether the husband or wife was inter- viewed. A comparison of the original sample of 2,033 with U.S. Census data revealed that the respondents were representative of the U.S. population of married persons with respect to age, race, household size, presence of children, home ownership, and region. By the fourth wave, 42% of the respondents were no longer in the sample, and the resulting panel is slightly less representative with respect to race, college education, and renter status (Booth et al. 1993). The first step of the analysis examines how selective attrition is related to variables used in this study.

424

AGE AND RELIGIOSITY

The present analysis includes 1,187 respondents with complete data on the selected variables who participated in waves one, three and four. The 1983 wave was excluded because it did not contain the measure of religiosity used here. In addition, 152 respondents interviewed in waves one and three but not wave four are included. When these data are pooled, the analysis includes 3,865 records. Standard errors are adjusted for the variance between and within individuals, so that the change in apparent sample size does not underestimate actual sample variability.

Measures

Dependent Variable. The measure of religiosity used is an indicator of the consequences of religious beliefs on the individual's daily life: "In general, would you say your religious beliefs influence your daily life (5) very much, (4) quite a bit, (3) some, (2) a little, or (1) not at all?" This measure was chosen because it was included in three of the four panels and spanned the full twelve year period (1980-1992). Other measures of religiosity, including the more commonly used measure of church attendance, were available only on the 1988 and 1992 panels. Table 1 presents the correlations between the measure used here and five other measures included on the 1988 wave and also shows how each measure is related to age.

TABLE 1

CORRELATIONS AMONG MEASURES OF RELIGIOSITY AND AGE FROM THE 1988 WAVE (n = 1,321)

Influence on Church Religiosity Indicator Daily Life Attendance Age

Religious Influence on Daily Life 1.0 .639 .090 Frequency of Church Attendance .639 1.0 .036 Participation in Church Activities .567** .652** .014 Frequency of Prayer .669 .559 .065 Frequency Watch Religious Broadcasts .366** .273* .109** Frequency Read Bible or Other

Religious Materials .646 .619 .074

p < .05; p < .01.

The influence measure is highly correlated (r = .639) with the attendance measure, and the two items have similar relations with measures such as participation in church activities and frequency of prayer, watching religious broadcasts, and reading the Bible. Overall, the high correlations of the influence of religion with the other indicators suggest it is a valid and reliable measure of religiosity. As previous studies have suggested (Campbell and Curtis 1994), all measures of religiosity have a positive correlation with age (column three). These bivariate, cross-sectional correlations assessed when respondents were twenty-six to sixty-three, however, show that the linear association between age and influence is greater than that between age and attendance. This may reflect the fact that disability associated with age is more likely to retard church attendance than general religiosity (Williams 1994). Everything else equal then, the measure of religiosity that we use makes it more likely that will find age effects.

Family Life Course Measures. The indicators of life course transitions focus on family and marital events. These include divorce, widowhood, and parental status. For each wave of the study, two dummy variables were created indicating whether the respondent's marital status was divorced (0 = no; 1 = yes) or widowed. Currently married was the omitted (reference) group. The parental status variables consisted of the number of children present

425

JOURNAL FOR THE SCIENTIFIC STUDY OF RELIGION

in the household in the age ranges of under age two, two through five, six through ten, and eleven and over. Another parenting indicator (empty nest) is a dummy variable coded one if the parents live in a household in which all the children have left home.

Control Variables. The control variables include respondent's sex, education, race, and religious preference. Respondent's sex is a dummy variable with females coded zero and males one. Education is the respondent's self reported number of years of schooling. Race is a dummy variable with Anglos coded as one and all other race/ethnic groups coded zero. We had insufficient cases to identify other race/ethnic groups separately. Religious preference is coded into three categories represented by the dummy variables Catholic and other. Protestants are the omitted (reference group). The other category includes those with no religion as well as those who identified affiliations other than Catholic or Protestant. Obviously, such a mixed category cannot be interpreted substantively. In the absence of sufficient cases to analyze these groups separately, the present procedure permits a clean comparison between Catholics and Protestants without muddying either category with the inclusion of those with no or other religious affiliation. Table 2 presents the mean, standard deviation, and range for each of the variables used in these analyses.

TABLE 2

DESCRIPTIVE STATISTICS FOR VARIABLES USED IN THE REGRESSION ANALYSIS

Variable Mean Std. Dev. Minimum Maximum

Importance of Religion 3.6994 1.2284 1.00 5.00 Age of Respondent 41.8561 10.2240 17.00 68.00 Ln (Age-5) 3.5660 .2934 2.48 4.14 Wave 1 .3462 .4758 .00 1.00 Wave 3 .3464 .4759 .00 1.00 Wave 4 .3074 .4615 .00 1.00 Divorced (1 = yes) .0347 .1830 .00 1.00 Widowed (1= yes) .0083 .0906 .00 1.00 Emptied Nest (1 = yes) .0996 .2995 .00 1.00 No. Children age < 2 .1415 .3785 .00 3.00 No. of Children 2-4 .1607 .4124 .00 3.00 No. of Children 5-10 .4197 .7004 .00 4.00 No. of Children 11+ .4895 .7818 .00 4.00 Gender (1 = male; 0 = female) .3837 .4863 .00 1.00 Years of Schooling 13.9413 2.6529 .00 30.00 Ethnicity (Anglo = 1) .9182 .2740 .00 1.00 Catholic .2611 .4393 .00 1.00 Protestant .5946 .4910 .00 1.00 Other religion .1444 .3515 .00 1.00

NOTE: n = 3865.

Method of Analysis

The primary method of analysis for this study is the random effects pooled-time series model (Johnson 1995). Pooled time series approaches are superior to the more conventional approach of regression models with lagged dependent variables in that we can model the relationship between age and religiosity explicitly, include variables assessing period effects, and model the effects of life course events on the dependent variable (Allison 1994).

The fixed effects approach to estimation is concerned only with variation within individuals over time. Because variations between individuals do not enter into the compu- tation, this technique automatically controls for characteristics that remain the same in individuals over time (specifically, birth cohort) as well as all other unmeasured time-

426

AGE AND RELIGIOSITY

invariant differences among individuals (Allison 1994). Thus, no time-invariant variables need to be, or can be, included in the analysis. If for example, we find a significant rela- tionship of religiosity to age in a fixed effects solution, we can be quite confident that it does not reflect a cohort effect because cohort membership is a time-invariant characteristic and is controlled in the analysis.

The fixed effects model can provide strong inferences about the effects of changes in age on changes in religiosity, but it has several disadvantages. Because it only focuses on variation within individuals, the method sacrifices statistical power and the standard errors of the estimates may be substantially higher than those found in the random effects approach. Another disadvantage of the fixed effects approach is the problem of separating period from aging effects. Year of the survey and age of the respondent both cannot be included as predictors in a linear form in the equation because change in age is perfectly correlated with change in years between the waves. A partial solution to this collinearity problem is to represent age or the year of the survey in a nonlinear functional form in the equation (Jasso 1985). A third disadvantage of the fixed effects approach is that time- invariant factors cannot be included in the model except as interaction terms with time varying factors. The random effects solution overcomes a number of these disadvantages.

The random effects solution is a generalized least squares solution that uses covari- ation both within and between individuals. The parameters are estimated from a covariance matrix in which the covariances within and between individuals are weighted to reflect their respective variance components. The resulting solution is more efficient than fixed effects, because a larger part of the total variance in the variables is included in the estima- tion. Since both within and between individual variance are included in the solution, time- invariant variables can be included in the equation and the dependency between change in age and period found in the fixed effects solution is relaxed. The primary disadvantage is that estimates are no longer unbiased by both observed and unobserved time-invariant differences between individuals. For example, in a random effects solution, we cannot be as certain that an effect of age on religiosity found in the analysis reflects change in religiosity with age. It might reflect cohort or other unobserved differences between the individuals of different ages.

We have chosen to analyze the data primarily with the more flexible and robust random effects approach. However, because some prior research has suggested a significant cohort effect (Chaves 1989; Miller and Nakamura 1996), we repeat some of the analyses employing the fixed effect solution.

FINDINGS

Assessing the Impact of Attrition

Over the twelve year span of this survey, 42% of the sample was lost through refusals or non-contact. To ascertain how this attrition biases our findings, a logistic regression analysis was computed in which attrition from the panel was regressed on perceived influence of religion on daily life in the first (1980) wave. Respondent's gender, education, and age in 1980 were included as control variables. Attrition between 1980 and 1988 was related to religiosity (p < .05), with higher religiosity related to lower attrition. For this period, moving from one religious influence category to another was associated with a 9% change in probability of attrition. For example, if the overall attrition rate was 30%, then the attrition rate for similar individuals with one point lower religiosity would be 33%. The effect over twelve years was slightly stronger, and a decrease of religiosity of one point was associated with a 10% increase in attrition. The effects of religiosity on attrition are much smaller than the effects of gender, age, and educational attainment. The major consequence

427

JOURNAL FOR THE SCIENTIFIC STUDY OF RELIGION

for this study is that the persons we analyze display a slightly higher level of religious influence over their lives than would a sample that had no attrition. We found no interaction between age and religiosity on attrition, however, so it does not appear that the relationships between age and religiosity reported below are biased by attrition.

Random Effects Models

Previous studies differ on whether the relationship between age and religiosity is linear or curvilinear (Ploch and Hastings 1994; Sasaki and Suzuki 1984 ), so the first step in the analysis is to establish the form of the relationship in this data set. Because it imposes no assumptions about the shape of the curve representing the relationship, our initial approach expresses age in nine categories and treats this as a set of dummy variables in the regression model. The categorical approach allows us to model the shape of the relationship better than a single measure, but this comes at the cost of decreased statistical power. Thus, we test a number of more parsimonious functional forms that provide a good fit to the data but make use of fewer degrees of freedom and allow for more efficient tests of some of the interaction effects.

TABLES

RANDOM EFFECTS REGRF.FSION MODELS RELATING AGE AND LIFE COURSE VARIABLES TO RELIGIOUS IMPORTANCE

Independent Variables Model 1 Model 2 Model 3 Model 4 Model 5 Model 6

Gender (male = 1; female = 0) -0.3971** -0.4137** -0.4010** -0.4223** -0.3472** -0.3472** Years of Schooling 0.0030 0.0057 0.0040 0.0067 0.0076 0.0062 Ethnicity (Anglo = 1) -0.2521* -0.2600 -0.2537 -0.2624* -0.2820** -0.2740 Protestant (reference group) Catholic -0.1991 -0.1899 -0.1967 -0.1852 -1.2944 -1.2467 Other Religious Affiliation -0.8337** -0.8206** -0.8297 -0.8210** -0.8256** -0.8253 Age 17-24 (reference group) Age 25-29 0.0887 0.1622 Age 30-34 0.2253** 0.3891** Age 35-39 0.2312 0.4612 Age 40-44 0.2352 0.5063** Age 45-49 0.2756** 0.5825** Age 50-54 0.2833** 0.7312** Age 55-59 0.2401* 0.6605** Age 60+ 0.2658 0.7117 * Age (In(age-5)) 0.2307 0.6892 0.5952 0.6232 1980 Wave (reference group) 1988 Wave -0.2130 -0.2345 -0.2339 -0.2216 1992 Wave -0.1944 -0.2389 -0.2376 -0.2103 Divorced ( = yes) -0.0927 Widowed (1 = yes) -0.1843 Emptied Nest (1 = yes) -0.1150 No. of Children age < 2 -0.0110 No. of Children age 2-4 0.0884** No. of Children age 5-10 0.0817* No. of Children age 11+ 0.0475* Age X Catholic 0.3415 0.3233* Gender X Catholic -0.2863 -0.2810

Constant 3.9849 3.8330 3.3771 1.8684 2.1818 2.0222 R-squared 0.093 0.102 0.095 0.103 0.108 0.118

NOTES: p < .05; p < .01. n = 3,865. Unstandardized regression coefficients.

428

AGE AND RELIGIOSITY

In Model 1, the only variables in the equation are the control variables of gender, years of schooling, ethnicity, religious affiliation, and the eight dummy variables repre- senting age. The relationship between age and religiosity is statistically significant, although much of the observed effect is the increase in religiosity between ages eighteen and thirty. After age thirty, religiosity increases only slightly to around age fifty, then declines a small amount. This suggests only a limited relationship between aging and religiosity, with much of the effect occurring before age thirty. This analysis is likely to be biased if a period effect is present.

Model 2 adds a control for period, treating the three waves of the study as a set of two dummy variables with the 1980 wave as the reference (omitted) group. This model found a statistically significant period effect between the 1980 and the 1988 waves. Religiosity declined by around .2 points between 1980 and 1988 with virtually no change occurring between the 1988 and the 1992 waves. Controlling for the period effect dramatically affected the estimates of an age effect, and religiosity now is shown to increase steadily with age. As in Model 1, the largest rate of increase occurs before age thirty, but increases are steady between ages thirty and fifty. After age fifty, religiosity changes little. The magnitude of the age effect more than doubles when the control for period is included (the largest difference between age categories is .71 after controlling for period compared to .27 before controlling for period). The declining religiosity documented in the period effect between the first two waves appears to suppress the observation of an age effect.

Several of the control variables were also statistically significant in this and subse- quent models. Males were found to have lower levels of religiosity than females (around .4 less), and Anglos, compared to minority racial and ethnic groups, also had lower religiosity. Among the three religious affiliation groups, Protestants, the reference group, had the highest religiosity scores, Catholics scored an average of .2 less than Protestants, and Others, as would be expected, had the lowest religiosity levels.

Our next step was to estimate a functional relationship between age and religiosity that treated age as a continuous variable. Since the curve observed in the categorical analysis generally follows a smooth non-linear trend, we tested several functional forms, including linear, quadratic, and several log functions. Each of the functions was fit to the data and the explained variance from the equations and the shapes of the resulting curves were compared. The best fitting function represented age by the natural log of age minus 5 years (ln(age - 5)). Models 3 and 4 repeat the first two models measuring age by the natural log rather than as a set of categories. The models with the log functions explain slightly more variance in the uncorrected R-square statistics in both models, but the substantive findings are identical: age is significant before and after period is controlled, but the effect is substantially stronger after period is controlled. The estimates for the period effect are very similar to those in the categorical model. Figure 1 presents the curves of the effect of age on religiosity controlling for period from models 2 and 4. The continuous curve produced by the logged model is very similar to the categorical curve from model 2, except that the curve based on the logged model does not show the decline in religiosity after age fifty observed in the categorical model. Because the logistic model has a marginally better fit to the data with considerably fewer degrees of freedom used and because the decline in religiosity after age fifty does not appear to be statistically significant, we have chosen to estimate the remainder of the models using the logged form of age.

Because previous research suggests that changes in religiosity with age may differ for Catholics and Protestants and for men and women, we tested two- and three-way inter- actions of age, gender, and religious affiliation. Two of three two-way interactions were statistically significant when evaluated individually. When added as a set, the interactions between age and Catholic and between gender and Catholic remained statistically signi-

429

JOURNAL FOR THE SCIENTIFIC STUDY OF RELIGION

ficant atp < .05. Adding the three-way interaction to a model containing all three two-way interactions did not add significantly to explained variance.

FIGURE 1

RELIGIOSITY BY AGE CONTROLLING FOR PERIOD

5-

4.5 -

41

, 3.5-

. - 0e 3-

--Categodcal 2.5 -

-Continuous (log)

2 -

1 .5 I i I I I I l I li 18 23 28 33 38 43 48 53 58 63

Age Categories

Model 5 in Table 3 presents the results of the regression including these interactions. The interaction between age and Catholicism is positive, indicating that Catholic religiosity increased more with age than did that of other religious groups. The additive effect of age (now read as the effect of age among non-Catholics), however, remains strong. The second interaction shows that the negative effect of being male on religiosity is intensified (nearly doubled) for Catholic males compared to other males. When these factors are taken together (see Figure 2), the broad picture is one of increasing religiosity with age for all groups, but with some variations. Male Catholics are the least religious at all ages, but, because Catholic religiosity increased more with age, the gap between Protestant and Catholic males decreased with age. The most dynamic picture is for Catholic females, who start out lower even than Protestant males but, by age fifty, surpass Protestant females and become the most religious of all groups. These findings are remarkably similar to those Ploch and Hastings (1994) report in their three-dimensional analysis of GSS trend data on church attendance, which suggests that these age patterns by gender and Catholicism are robust and not artifacts of particular measurement or study design.

The next step in the analysis introduced the family life course variables into the equation. Because all of our sample was married in the first wave, our analysis misses one of the important family life course events that is postulated to increase religious involvement. The coefficients from the model including these variables are found in model 6 of Table 3. The results do not support the hypothesis that the observed relationship between age and religiosity is accounted for by family life cycle events. The effect of age changed very little when the life course measures were introduced, and all variables statistically signifi- cant in model 5 were also significant in model 6. The presence of children age two and over did have a small significant positive effect on religiosity. The largest effects were for children age two to four and age five to ten. Adding a child in these groups increased religiosity by around .08. Children under age two had no effect. Divorcing, becoming

430

AGE AND RELIGIOSITY

widowed, and emptying the nest decreased religiosity, but none of these effects was statistically significant.

FIGURE 2

RELIGIOSITY AND AGE FOR CATHOLIC AND PROTESTANT MALES AND FEMALES

5

*I

4.5

2.5 -Catholic Males ..... Protestant Males -Catholic Females

2 -- - - Protestant Females

1.5 -

18 23 28 33 38 43 48 53 58 63 68

Age

Because earlier research suggested that the effects of family life course variables will depend on the ages at which they occur (Stolzenberg, Blair-Loy, and Waite 1995), we tested to see whether the life course variables interacted with age in their effects on religiosity. Three multiplicative terms (age-by-divorced, age-by-widowed, and age-by-a dummy variable for children in the household) were added to model 6 in Table 3. The increment to explained variance was not statistically significant. We conclude that the age effects demonstrated here are not modified significantly by family life course measures included in this analysis.

As a final step in the analysis, we replicated the analysis in model 6 using a binary probit model, in which religiosity scores of 4 and 5 were contrasted with scores of 3, 2, 1, and 0, to test the sensitivity of the analysis to the modest skewness in the dependent variable. A comparison of results using a continuous and dichotomous form of the dependent variable (not shown) showed that the substantive outcomes, including interaction tests, are robust across forms of measurement in the dependent variable.

Fixed Effects Models

The random effects models made the assumption of no significant cohort effects. If cohort effects are present, they could account for the increase in religiosity with age. The final step in our analysis is to repeat the basic models employing a fixed effects statistical model. By examining only within-individual variation, the fixed effects model controls for all time-invariant characteristics and effectively controls for cohort effects. If the results are similar to those in the random effects solution, we can assume that a cohort effect is not responsible for results reported in Table 3.

Model 1 in Table 4 examines the relationship between logged age and religiosity with no other variables in the equation (similar to model 3 in Table 3), and model 2 adds the period effect measures. For model 3, we added the life course measures and re-examined all

431

JOURNAL FOR THE SCIENTIFIC STUDY OF RELIGION

two-way interactions between gender, age, and religion. The age-by-Catholic and gender-by- Catholic interactions are insignificant, but the age-by-gender effect is statistically significant in the fixed-effects solution. In this model, religiosity increased significantly less with age for males than females. The R-square terms in these models are substantially larger because the fixed effect equation includes a dummy variable for each individual represented in the pooled data set. The R-square may be read as showing that over 70% of the variation in religiosity in the three wave data set is between individuals and less than 30% is within individuals.

TABLE 4

FIXED EFFECTS SOLUTION OF THE EFFECT OF AGE (LOGAN (AGE-5) AND PERIOD ON RELIGIOSITY

Independent Variables Model 1 Model 2 Model 3

Age (in(age-5)) 0.0235 1.6074** 1.3180** 1980 Wave (reference group) 1988 Wave -.4433 -0.3510** 1992 Wave -.4585** -0.4065** Divorced (1 = yes) -0.0232 Widowed ( 1 = yes) -0.0988 Emptied Nest ( 1 = yes) -0.0182 No of Children age < 2 -0.0530 No of Children age 2-4 0.0374 No of Children age 5-10 0.0446 No of Children age 11+ 0.0104 Gender X Age -0.2714* R-squared 0.773 0.776 0.778

NOTE: *p < .05; **p < .01. n= 3,865. Unstandardized regression coefficients.

The results are similar, but not identical, to those in the random effects model. Age was not related to religiosity significantly before the period measures were included (model 1), but the addition of period effects in model 2 results in strong and significant age as well as period effects. Because the age effect in the fixed effects solution cannot be biased by cohort, we conclude that the effects of age and period on religiosity estimated in the random effects model are not due to birth cohort. Perhaps because of the lower power of the fixed- effects solution, none of the life course variables is statistically significant although all are in the expected direction. A difference between the fixed- and random-effects solutions is that the two-way interactions with Catholic drop out and the age-by-gender interaction becomes significant. Although this finding could capitalize on chance variations in the sample, the difference in designs suggests a plausible substantive hypothesis. The fixed effects solution controls all time-invariant characteristics, whereas the random effects solution controls only a few basic socio-deomographics such as education, race, and ethnicity. If Catholics differ from other religious groups on invariant dimensions that are plausibly related to religiosity but are not controlled in the random-effects solution, e.g., foreign born, ethnic identity or ethnic enclave, or religious background, this would explain why the effects fall out in the fixed-effects solution. It is possible that the greater increase in religiosity for females of all religions that is revealed in Table 4 is a more general process than the more complicated interactions with religion. However, more detailed analysis of Catholicism would be needed to examine this hypothesis.

432

AGE AND RELIGIOSITY

SUMMARY AND DISCUSSION

Using a random effects pooled time-series panel model, we test the relationship between age and religious influence on daily life, the influence of family life course change on that relationship, and differences among male and female Protestants and Catholics. The use of panel data is an improvement over prior research because they allow us to track religious and life cycle change in the same individuals over a span of up to twelve years. The sample does have some limitations. All respondents were married at the first wave of the study, which may have biased the sample toward more conventional respondents. Further, the oldest respondent in the sample was aged 68 in 1992, preventing us from studying the relationship between old age and religious influence on daily life.

The relationship between age and religious influence on daily life appears to be a smooth curvilinear function with the greatest change in the early adult years. We found a strong period effect operating from 1980 to 1988 in both the categorical and natural log function models. The period effect was one of declining influence of religion on daily life and counteracted the age effect of increased religious influence on daily life. Once this period effect was controlled, the age effect nearly doubled. Consistent with prior research (Chaves 1989; Hout and Greeley 1987), we found Catholics increased religiosity more with age than other groups but that male Catholics showed less increase than female Catholics.

The strong effect of period on the age effect raises the question whether period is a simple additive effect, with the same age effects at each period or whether age effects depend on period. If they do, of course, this would suggest that cohort effects were operating. To explore this possibility, we created two multiplicative terms (one for each period) and entered them into the fixed- and random-effects solutions after the two-way interactions but before the life course measures. Neither of the two individual age-by-period interactions nor the block was significant in the fixed-effect solution, but both coefficients were significant in the random-effects solution. While the overall age pattern was very similar at all periods, age effects were greater in the earlier than the two later periods. The coefficients for these interaction effects were sharply reduced and became insignificant, however, when the life course measures were added in model 6. We conclude that, rather than signalling a true cohort effect, stronger effects in 1980 reflected a different age pattern of family life course events. This non-cohort interpretation is consistent with the results for the fixed-effects solution.

The large period effect of declining religiosity between 1980 and 1988 and the absence of any effect between 1988 and 1992 argues against any generalized process of seculari- zation. The difference is not attributable to the longer time period between waves one and two: not simply two-thirds but all of the decline in religiosity occurred in the first period (Table 3). We tested whether the period effect differed by church affiliation, and the period effect is not significantly different for Catholics than Protestants. Several scholars using trend data have suggested that period declines observed during the 1980s are due, not to low religiosity in the mid 1980s compared to the mica 1970s, but to unusually high reports of religiosity in 1980 (Chaves 1989; Miller and Nakamura 1993). Because our data do not extend prior to 1980, we cannot say whether it was 1980 or 1988 that was unusual.

The largest effect of age on religiosity was in the younger ages where the curve of increased religiosity was steepest. This effect persisted after controlling for presence and ages of children and other family life course events and was not modified significantly by family life course events. The findings here support a traditional or developmental model of religiosity and age and suggest that aging per se is associated with an increase in religiosity independent of family life course events. All respondents in this sample were married at the start of the panel, which precludes marriage or its assumed prior cause, conventionality, as a cause of the observed growing influence of religion on daily life. It is possible that

433

JOURNAL FOR THE SCIENTIFIC STUDY OF RELIGION

marriage sets in train a sequence of social roles that encourages religiosity. Most obviously, however, this is parenthood, and our research, as did that of Stolzenberg, Blair-Loy, and Waite (1995), shows that the positive effect of age is independent of children. Our sample does not allow us to test the possibility that the age effect we observe here is restricted to or interacts with marriage, but we see no reason to anticipate that the effect is not general. In this case, despite well-founded concerns about cohort and period effects, there does appear to be a simple age effect on religiosity consistent with cross-sectional evidence.

NOTE

Previous versions of this paper have been presented at the 1997 meetings of the Midwest Sociological Society and American Sociological Association in Toronto. We would like to thank Hugh Whitt for comments on an earlier version of this paper. Research was supported in part by Grant 5 RO1 AG04146 from the National Institute on Aging.

REFERENCES

Allison, Paul. 1994 Using panel data to estimate the effects of events. Sociological Methods and Research 23: 174-99. Bahr, Howard. 1970. Aging and religious disaffiliation. Social Forces 49: 60-71. Booth, Alan, Paul Amato, David R. Johnson, and John N. Edwards. 1993. Marital instability over the life course:

Methodology report for fourth wave. Lincoln: University of Nebraska Bureau of Sociological Research. Booth, Alan, and David R. Johnson. 1985. Tracking respondents in a telephone interview panel selected by random

digit dialing. Sociological Methods and Research 14: 53-64. Campbell, Robert A., and James E. Curtis. 1994. Religious involvement across societies: Analyses for alternative

measures in national surveys. Journal for the Scientific Study of Religion 33: 215-29. Chaves, Mark. 1989. Secularization and religious revival: Evidence from U.S. church attendance rates, 1972-1986.

Journal for the Scientific Study of Religion 28: 464-77. . 1991. Family structure and Protestant church attendance: The sociological basis of cohort and age effects.

Journal for the Scientific Study of Religion 30: 501-14. Firebaugh, Glenn, and Brian Harley. 1991. Trends in U.S. church attendance: Secularization and revival, or merely

life cycle effects? Journal for the Scientific Study of Religion 30: 487-500. Glenn, Norval D. 1974. Cohort analysts' futile quest: Statistical attempts to separate age, period, and cohort effects.

American Sociological Review 41: 900-04. Hadaway, C. Kirk, Penny L. Marler, and Mark Chaves. 1993. What the polls don't show: A closer look at U.S. church

attendance. American Sociological Review 58: 741-52. . 1998. Overreporting church attendance in America: Evidence that demands the same verdict. American

Sociological Review 63: 122-30. Hoge, Dean, and Jann Hoge. 1984. Period effects and specific age effects influencing values of alumni in the decade

after college. Social Forces 62: 941-62. Hout, Michael, and Andrew M. Greeley. 1987. The center doesn't hold: Church attendance in the United States, 1940-

1984. American Sociological Review 52: 325-45. . 1990. The cohort doesn't hold: Comment on Chaves (1989). Journal for the Scientific Study of Religion 29: 519-

24. . 1998. What church officials' reports don't show: Another look at church attendance data. American Sociological

Review 63: 113-19. Jasso, Guillermina. 1985. Marital coital frequency and the passage of time: Estimating separate effects of spouses'

ages and marital duration, birth and marriage cohorts, and period influences. American Sociological Review 50: 224-41.

Johnson, David R. 1995. Alternative methods for the quantitative analysis of panel data in family research: Pooled time-series models. Journal of Marriage and the Family 57: 1065-77.

Miller, Alan S., and Takashi Nakamura. 1996. On the stability of church attendance patterns during a time of demographic change: 1965-1988. Journal for the Scientific Study of Religion 35: 275-284.

Ploch, Donald R, and Donald W. Hastings. 1994. Graphic presentations of church attendance using General Social Survey data. Journal for the Scientific Study of Religion 33: 16-33.

Sandomirsky, Sharon, and John Wilson. 1990. Processes of disaffiliation: Religious mobility among men and women. Social Forces 68: 1211-29.

Sasaki, Masamichi, and Tatsuzo Suzuki. 1987. Changes in religious commitment in the United States, Holland, and Japan. American Journal of Sociology 92: 1055-76.

Stolzenberg, Ross M., Mary Blair-Loy, and Linda J. Waite. 1995. Religious participation in early adulthood: Age and family life cycle effects on church membership. American Sociological Review 60: 84-103.

434

AGE AND RELIGIOSITY 435

Williams, David R. 1994. The measurement of religion in epidemiological studies: Problems and prospects. In Religion, aging, and health: Theoretical foundations and methodological frontiers, edited by J. S. Levin, 125-48. Newbury Park: Sage.

Wilson, John, and Darren E. Sherkat. 1994. Returning to the fold. Journal for the Scientific Study of Religion 33: 148- 61.