2.1 Cognitive dissonance and gender role attitude change

In Germany and most other European countries, gender roles have become more egalitarian in the late 20^th and early 21^st centuries (Boring & Moroni 2022). Two major drivers of change were expanding education and women’s entry into the labor market in the 1960s (Nitsche & Grunow 2016). The pandemic-related constraints on outsourcing housework and childcare, the additional family labor responsibilities, and the disruption in the labor market could reorganize the division of family labor, and “might cause gender relations to shift at a societal scale” (Reichelt et al. 2021: 229).

Previous research has shown that life events or exogenous shocks can affect the division of family labor and might consequently influence GRAs to match the behavior (Danzer et al. 2021; Schober & Scott 2012). An example for this is the transition into parenthood: Previously more egalitarian couples distribute labor more traditionally after the first childbirth (Kühhirt 2012) and also shift their attitudes toward a less egalitarian gender ideology (Baxter et al. 2015; Grinza et al. 2022).

Also, lived realities that were affected by policies reforms can influence GRAs: For instance, in Estonia, mothers and fathers who were affected by a policy reform that extended fathers’ parental leave threefold stated more gender egalitarian views due to the more gender equal sharing of childcare resulting from the paternity leave (Tavits et al. 2023). Similarly, in West Germany, Zoch and Schober (2018) found changes towards more egalitarian GRAs among mothers related to the childcare expansion. The latter studies give support to the notion that lived realities can alter GRAs also in the short-term (Zoch & Schober 2018).

While empirical evidence shows that individuals do change their attitudes over the life course, the theoretical and psychological mechanisms remain unclear. One theory explaining a shift in attitudes following a change in behavior is the cognitive dissonance theory (Festinger 1957). Thereby, a discrepancy between behavior and attitudes can produce a state of discomfort which individuals are motivated to resolve. Often, attitudes are more easily adaptable than behavior (Schober & Scott 2012). This is especially pronounced if the behavior is more resistant to change due to external constraints and outer circumstances (Harmon-Jones & Harmon-Jones 2007), like produced by the COVID-19 pandemic and its containment measures. Concluding, cognitive dissonance theory states that personal experiences can promote attitudinal change, mainly when dissonant cognitions based on one’s attitudes and behavior arise.

I follow Schober and Scott (2012), who connect cognitive dissonance to GRAs by referring to literature on gender identity (West & Zimmermann 1987) that suggests “that male and female roles as partners or parents serve as core standards on which to evaluate gendered aspects of paid work and childcare arrangements” (Schober & Scott 2012: 516). Schober and Scott link GRAs to gender identity and assume that discrepancies between gendered standards and actual behavior likely lead to changes in GRAs to match behavior. However, like Schober and Scott I acknowledge that cognitive dissonance most likely is not the only motivator and theoretical mechanism behind intra-individual attitudinal change.

While most studies on GRA change only examine differences between GRAs before and after a specific life event or policy reform, only few studies actually consider the effect of attitude and behavior discrepancy on GRA change. For instance, Schober and Scott (2012) and Grinza et al. (2022) find that attitudinal changes after the transition to parenthood are more likely among men and women for whom postnatal behavior clashes with prenatal attitudes, supporting the expectations of cognitive dissonance theory.

However, research about life events, for example the transition to parenthood, is likely affected by endogeneity, because people usually self-select into becoming parents. In contrast, the COVID-19 pandemic can be seen as an exogenous macro-level shock that hit people unexpectedly (Vandecasteele et al. 2022). Exploiting the COVID-19 pandemic as an exogeneous shock and framing it as a ‘natural experiment’ hence allows to reduce endogeneity issues that affect other research on GRA change (Hudde et al. 2022).

Despite this advantage, it should be kept in mind that the pandemic was an extraordinary situation, and pandemic behavior does not necessarily translate into post-pandemic patterns (Boll et al. 2021). It might be that individuals only reacted to this extreme situation in a pragmatic way, without consequently adjusting their attitudes (like assumed by Hiekel & Kühn 2023) because they do not expect the situation to last. However, some policy interventions like paternity leaves also just affect a smaller period of time and can be related to GRA change (Tavits et al. 2023). Furthermore, it can be argued that the pandemic most likely will have long-term effects on gender relations (Danzer et al. 2021) not only because pandemic measures lasted for three years but also because pandemic practices like home office arrangements and flexible working hours are prolonged beyond the pandemic and because gender relations were prominently discussed throughout the pandemic.

A few studies have examined changes in GRAs during the pandemic; however, none of them focuses specifically on changes in the division of family labor as a predictor for GRA change. Both Reichelt et al. (2021) and Vandecasteele et al. (2022) use a form of the cognitive dissonance theory to explain the relationship between labor market transitions and GRAs.

Reichelt et al. (2021) showed that for the US, Germany, and Singapore, changes in the employment constellation of heterosexual couples affected the GRAs of men and women when changes in employment clashed with gendered norms. Thereby, “men express more egalitarian gender-role attitudes if they became unemployed but their partners remained employed, while women express more traditional attitudes if they became unemployed and their partners remained employed” (Reichelt et al. 2021: 228). These findings reinforce the theoretical assumption of pandemic-related behavior influencing attitudes to match lived realities. However, using a cross-sectional design, Reichelt et al. (2021) are not able to examine actual changes in GRAs.

For the Netherlands, Vandecasteele et al. (2022) examined the effect of the COVID-19 pandemic on GRAs regarding maternal employment. Vandecasteele et al. found that individuals experiencing a change in paid workload that clashes with traditional gender norms expressed more gender egalitarian attitudes concerning maternal employment than before the pandemic. For women living with children in couple households, Vandecasteele et al. found a halt in their progression toward more egalitarian attitudes.

Building on the literature on GRA change both during and before the pandemic by following the cognitive dissonance theory and only considering respondents who reported no attitude-behavior discrepancy (i. e., no cognitive dissonance) before the pandemic, I test the following hypotheses:

If women take over more family labor responsibilities during the pandemic, and thus, the division of labor becomes more inegalitarian than before the pandemic, cognitive dissonance emerges and GRAs become less egalitarian to match the changed behavior (H1). Likewise, if men spend more time on family labor during the pandemic and hence, the division of labor becomes more egalitarian than before the pandemic, cognitive dissonance emerges and GRAs become more egalitarian to match the changed behavior (H2). Because of gendered norms and their greater behavioral implications for women (Grinza et al. 2022; Zoch & Schober 2018), women’s attitudes are more strongly expected to change (H3).

2.2 The division of family labor during the COVID-19 pandemic

Because the hypotheses based on the cognitive dissonance theory build on a change in behavior (i. e., the division of family labor) due to pandemic constraints, in the following I summarize the numerous research findings on heterosexual couples’ pandemic behavior. To gain as much insight as possible on gendered behavior and attitudes, when analyzing family labor, housework and childcare responsibilities need to be distinguished. Compared to childcare, housework is often viewed as more feminine, less enjoyable, postponable, and more flexibly performable (Sullivan 2013).

As a consequence of the pandemic, housework responsibilities increased because external services like restaurants or cleaners were no longer available. Additionally, people spent more time at home due to social distancing and curfews, increasing household labor (Reichelt et al. 2021). Childcare responsibilities especially increased due to the closures of childcare facilities and schools and the obligation to quarantine in case of a COVID-19-contact or -infection. Additionally, informal childcare arrangements became inaccessible in times of strict social distancing restrictions when social contacts were limited to members of the same household.

Consequently, housework, childcare, and additional responsibilities like homeschooling were primarily performed by parents (Kohlrausch & Zucco 2020). Due to the constraints on outsourcing family labor and the additional tasks resulting from the pandemic situation, most likely an increase in the amount but also possibly a change in the division of family labor took place.

The findings regarding the division of family labor during the pandemic are ambiguous: While some found support for a ‘re-traditionalization’ hypothesis (Allmendinger 2020) with women stepping in even more regarding both housework and childcare (e. g., Illing et al. 2022; Kohlrausch & Zucco 2020; Naujoks et al. 2022; Zoch et al. 2021; Zucco & Lott 2021), others found a relative increase in men’s contribution, especially with regard to childcare and hence, more gender equal family labor arrangements (e. g., Bujard et al. 2020; Globisch et al. 2022; Jessen et al. 2021; Kreyenfeld & Zinn 2021; Zinn et al. 2020). Fathers may have had more time availability due to short-time work, reduced hours, home office, or unemployment and were pushed to increase their time spend on childcare, particularly if a reduction of hours in paid employment for fathers was accompanied by mothers working more than 20 hours per week on-site (Globisch et al. 2022). That some studies found that men increased their relative share more than women could also be a ceiling effect: Women often already performed the greater or total share, which is why they could not increase it much further.

Using pairfam data, Hank and Steinbach (2020) found an almost equal split between those couples in which the female share of housework and childcare increased and those in which it decreased. They concluded that there is not one observable trend of how couples reacted to the COVID-19 crisis regarding the division of labor but rather heterogeneous responses (Hank & Steinbach 2020). The heterogeneity in results is also shown by other researchers who found no change in the division of family labor during the pandemic but stability of pre-pandemic arrangements (e. g., Boll et al. 2021; Möhring et al. 2020; Zucco & Lott 2021). All studies agree that both before and during the pandemic women perform the greater absolute share of unpaid family labor.

Who took over the (additional) tasks in heterosexual couples during the pandemic depended on multiple factors. An important aspect was the employment situation of couples before and during the pandemic, as the COVID-19 pandemic led to many employment transitions like unemployment, reduced or increased hours, short-time work, and home office arrangements (Möhring et al. 2020; Kohlrausch & Zucco 2021).

For the UK, Hudde et al. (2021) found no evidence for pre-pandemic GRAs influencing the distribution of housework during the pandemic like it would be expected by gender ideology theory (Davis & Greenstein 2009). In the discussion about which causal order the relationship between attitudes and behavior has, the null findings from Hudde et al. supports the notion of examining the effect of behavior on attitudes. However, the feedback effects between attitudes and behavior should always be kept in mind when researching on them (e. g., Preston 2023).

3.1 Data

The following analyses are based on release 13.0 of the German Family Panel ‘pairfam’ (Brüderl et al. 2022; Huinink et al. 2011). Pairfam is a representative panel study launched in 2008. It has a multi-cohort design, which means that the initially more than 12,000 anchors (i. e., primary respondents) were randomly drawn from four nationally representative birth cohorts: 1971–1973, 1981–1983, 1991–1993, and 2001–2003. A more detailed description of the data can be found in Huinink et al. (2011).

In the analyses, the last pre-pandemic wave (11) and the first full pandemic wave (13) were used. Wave 12 was not included in the analyses as GRA items were only reported every other year. Wave 11 was conducted from mid-October 2018 to the end of May 2019 and contains 9,435 interviews. Wave 13 was conducted from October 2020 to April 2021 and contains 7,009 interviews.

From November till December 16^th, a so-called ‘lockdown light’ was active, followed by a more restrictive lockdown from mid-December till May 2021. The more restrictive lockdown was accompanied by higher infection-numbers and, starting in mid-April, included a curfew from 9:00 pm till 5:00 am. Although schools and childcare facilities were technically open, they were also affected by quarantine regulations and there were time periods in which only emergency caretaking was offered (between December 16^th and January 10^th) (Bundesregierung, 2020). Hence, the pandemic situation was quite tense during the data collection period, with a less and a more restrictive lockdown influencing peoples’ lives.

Wave 11, like all pre-pandemic waves was conducted using only computer-assisted personal interview (CAPI) methods. Because of the pandemic, wave 13 was conducted using a mixed-mode strategy. Respondents and interviewers could choose between computer-assisted telephone interviews (CATI) and CAPI methods. Roughly two thirds of the respondents in wave 13 were interviewed using CATI methods.

3.2 Variables

The dependent variable in the following analyses is the anchor’s GRA. Pairfam contains four GRA items:

1. “Women should be more concerned about their family than about their career.”

2. “Men should participate in housework to the same extent as women.”

3. “A child under 6 will suffer from having a working mother.”

4. “Children often suffer because their fathers spend too much time at work.”

For all items, respondents could indicate on a 5-point-scale whether they 1 “completely disagree” or 5 “completely agree”. For comparability, I poled all variables in the same direction so that higher values represent less egalitarian attitudes.

I calculated Cronbach’s alpha to test whether creating an additive index is justifiable, however, it was too low to do so for all or various combinations of the items (alpha ~ 0.5). This is in line with previous findings regarding gender ideology: That GRAs towards housework do not necessarily reflect attitudes about childcare is explained by the multidimensionality of gender ideology (Grunow et al. 2018; Knight & Brinton 2017).

Hence, additionally to analyzing the division of housework and childcare separately, I analyze GRA items separately. Following Nitsche and Grunow (2016) I use item 2) to operationalize GRAs concerning the division of housework. To operationalize GRAs towards childcare, following Zoch and Schober (2018) I use item 1) and item 3), as they both concern attitudes about maternal employment. Because the alpha-value is too low to combine both items, I assume that they each cover a distinct area of maternal employment attitudes and analyze both items separately.

Following Nitsche and Grunow (2018), I neglected item 4) in all analyses, because egalitarian and inegalitarian persons could likewise agree with the statement.

The main independent variable to test the hypotheses is the division of family labor, specifically the division of housework (i. e., doing laundry, cooking, and cleaning) and childcare. For both kinds of family labor, respondents were asked to which extent, on a 5-point scale, they or their partner perform those tasks. Possible answers were: “(almost) completely my partner”, “for the most part my partner”, “split about 50/50”, “for the most part me”, and “(almost) completely me”. Respondents who claimed another person was responsible for those tasks or for whom none of the answers were applicable were removed from the sample.

The independent variables including information on how couples were sharing family labor, are structured as follows: Anchors were assumed to have an ‘inegalitarian’ arrangement if a female respondent claimed to do most or all family labor or if a male respondent claimed his partner was doing most or all work. Because the share of couples in which the man did most or all family labor is very small, they were combined with respondents who claimed to share labor about 50/50 and labeled as having an ‘egalitarian’ arrangement.

My assumptions about the relationship between changes in behavior and changes in attitudes are based on the cognitive dissonance theory. I assume that a pandemic-related emergence of cognitive dissonance can be associated with a change in GRAs. Hence, to test my hypotheses it is relevant to only examine respondents who had no attitude-behavior discrepancy before the pandemic and thus, are considered to not have experienced pre-pandemic cognitive dissonance.

To ease the operationalization of pre-pandemic cognitive dissonance, I created three dummy variables based on the GRA items, indicating ‘egalitarian’ or ‘inegalitarian’ GRAs. Given the skewness of the GRA items in pairfam towards egalitarianism and following Hiekel and Kühn (2023), I categorized respondents answering 1 or 2 as ‘egalitarian’ and those answering 3 or more as ‘inegalitarian’.

Separated by all three GRA items and both, the division of housework and of childcare, respondents who in wave 11 were egalitarian in their GRAs and their family labor division or were inegalitarian in their GRAs and their family labor division were categorized as having no pre-pandemic attitude-behavior discrepancy and hence, ‘no cognitive dissonance’. Those who had an egalitarian GRA but distributed family labor in an inegalitarian way or had an inegalitarian GRA but distributed family labor in an egalitarian way were categorized as having ‘cognitive dissonance’.

3.3 Sample selection

In the analyses, I use balanced panel samples of respondents who participated in waves 11 and 13. Because the gendered division of family labor is the main independent variable, I only kept individuals living in a cohabitating, heterosexual relationship (married and unmarried) who were 18 years or older in wave 11.

I analyze three sub-samples of individuals experiencing ‘no cognitive dissonance’ from two separate superordinated samples, one in which all individuals in cohabitating relationships are included (‘cohabitating sample’), and one in which only parents living with at least one child (‘parents sample’) are included. The influence of changes in the division of housework is analyzed in both samples, because of the assumed differences between parents and non-parents and the influence of changes in the division of childcare only for parents.

Respondents who did not answer the items concerning their GRAs in wave 11 or wave 13 were removed. Those who did not report their division of family labor in wave 11 or wave 13 were removed. Eventually, the superordinated ‘cohabitating sample’ consists of 3,169 respondents, and the superordinated ‘parents sample’ consists of 1,933 respondents. The number of cases dropped on the single items is shown in Appendix Table A2.

Because I am interested in a pandemic induced emergence of cognitive dissonance, I only analyze those respondents from the superordinated samples who were categorized as having ‘no cognitive dissonance’ before the pandemic in three sub-samples: For GRA item 2) and the division of housework: n=1,566, for GRA item 1) and the division of childcare: n = 1,096, and for GRA item 3) and the division of childcare: n = 994.

In Appendix Table A5, the distribution of cognitive dissonance for the superordinated samples by GRA item can be seen. Roughly 50 % of the respondents in all three sub-samples report no attitude-behavior discrepancy before the pandemic. If attitude-behavior discrepancies are reported, for housework, most respondents report a more egalitarian attitude than behavior, while for childcare and the two maternal employment GRAs, the picture is more divers even though more respondents report egalitarian GRAs and inegalitarian behavior. Especially women report more discrepancy between attitudes and behavior, and thereby more often report a more egalitarian attitude than behavior. Hence, Table A5 shows that attitude-behavior discrepancies also exist within other framings than the pandemic and therefore, that cognitive dissonance is not the only cause of attitude formation (Schober & Scott 2012). It also emphasizes that I only analyze a selective group of those who had aligned attitudes and behavior prior to the pandemic, which should be kept in mind when interpreting the results.

IV = Independent variable

In Table 1, the summary statistics of the three sub-samples with ‘no cognitive dissonance’ are shown. For comparison, the summary statistics of the whole superordinated samples including weighted means and sex stratified analysis can be found in Appendix Tables A3 and A4.

The GRA item 2) sub-sample is on average younger than the other two sub-samples and has a slightly higher share of highly educated respondents. 34 % of the GRA item 2) sub-sample have no children living in the household. The average household income is lower in the GRA item 2) sub-sample while the average working hours per week are at least 1 hour higher than in both childcare sub-samples.

In comparison to the superordinated samples, the sub-samples with GRA item 2) and item 3) include 5 percentage points more men. In the GRA item 2) sub-sample, there are 7 percentage points more non-parents as compared to the whole cohabitating sample. In all three sub-samples the average GRAs are less egalitarian than the whole superordinated samples. Also, the average change in GRAs is greater in the three sub-samples, especially for GRA item 2) and 3).

Concerning item 2), respondents report on average the most egalitarian attitudes, which is due to almost universal approval to the item possibly resulting from the high social desirability connected to item 2) (Hudde & Engelhardt 2021). Correspondingly, the lowest shares of inegalitarian family labor divisions are reported in the GRA item 2) sub-sample. However, during the pandemic, the share of those reporting an inegalitarian division of family labor increased in the GRA item 2) sub-sample by 14 percentage point, while the share decreased in both other sub-samples considering childcare.

Looking at the share of respondents who reported an inegalitarian division of family labor in wave 11 shows how the combination of aligned attitudes and behavior is distributed in the three sub-samples. While the GRA item 2) sub-sample only includes 29 % of respondents who report inegalitarian attitudes and behavior, for the other sub-samples and considering childcare, 67 % (GRA item 1)) and 49 % (GRA item 3)) report inegalitarian GRAs and behavior. Appendix Table A3 shows, that in the superordinated samples before the pandemic, the share of those reporting an inegalitarian division of family labor is 62 % in both samples.

In all three sub-samples, the average GRA got more egalitarian during the pandemic. The largest change can be found in the sub-sample with GRA item 1) (19 percentage points).

Table 2 shows the summary statistics separated by sex for all three ‘no cognitive dissonance’ sub-samples. In all samples, women are younger than men and have a lower share of high education. Women report a smaller household income than men in both waves. The biggest sex difference can be found when looking at working hours: Women, and even more pronounced, mothers, work substantially fewer hours than men both before and during the pandemic.

Looking at GRA item 2) and GRA item 3), men on average report less egalitarian attitudes than women. Women only have less egalitarian attitudes than men when looking at the GRA item 1). Both women and men report more egalitarian attitudes during the pandemic. For all three sub-samples, changes in GRAs are greater for women than for men.

3.4 Methods

I estimate asymmetrical fixed effects models (Allison 2019) to explore changes in GRAs of men and women between the last pre-pandemic wave (i. e., wave 11) and during the pandemic (i. e., wave 13). To estimate asymmetric effects, I use first difference methods (Allison 2019). With two waves, first difference models yield the same estimates as fixed effects models, eliminating all time-invariant heterogeneity by exploiting only within-person variability (i. e., intra-individual changes over time) (Allison 2019). Hence, it can be avoided that potentially unobserved time-invariant confounders bias the estimates. Therefore, in the models, it cannot be controlled for time-invariant variables like birth cohort or education. Because of the first difference approach, it is also not possible to control for age. In the models including control variables it is controlled for income, the number of children living in the household and weekly hours spent in paid employment.

Estimating asymmetric effects means that I am able to distinguish between changes in the division of family labor towards a less egalitarian versus toward a more egalitarian division. To do so, I created two dummy variables: one indicating a change towards a more inegalitarian division and the other one indicating a change towards a more egalitarian division. For both variables, the reference category indicates no change in the division of family labor. For more on the asymmetrical fixed effects approach, see Allison (2019). The dependent variable is the change score of GRAs between wave 13 and wave 11. Hence, the variable ranges from -4 to 4, whereby negative values indicate a change towards more egalitarian GRAs, and positive values indicate a change towards less egalitarian GRAs. 0 indicates no change.

In addition to a model for the overall association between changes in the division of family labor and GRAs, I examined whether the associations differ for men and women by estimating equivalent models separated by sex to test hypothesis H3. Due to the model specification, no sex interaction term can be included in the model. However, performing the analyses in samples separated by sex also allows to examine different tendencies between men and women (e. g., Baxter et al. 2015; Grinza et al. 2022).

The coefficients in all estimated models can be interpreted as the average change in respondents’ GRAs between wave 11 and wave 13 for a one-unit intra-individual change across the two waves on the independent variables. In all models, I estimated panel robust standard errors on the individual level (Brüderl & Ludwig 2014).

Because in pairfam the items on GRAs are captured on a Likert-scale and skewed towards egalitarianism (Hiekel & Kühn 2023), floor effects could occur that might underestimate the effect of the independent variables. To control for floor and ceiling effects I re-estimated all models using Tobit regression (for more on floor and ceiling effects and Tobit regression, see McBee 2010). The Tobit regressions yield very similar results and are shown in the Appendix (Table A6-A8).

4.1 Descriptive Results

In this paper, the association between the emergence of cognitive dissonance due to a pandemic-related change in behavior (i. e., division of family labor) and changes in GRAs is of interest. Table 3 shows changes in GRAs, changes in the division of family labor and the average GRA change score by changes in the division of family labor.

In the GRA item 2) sub-sample roughly 37 % changed their GRA while in both other sub-samples, roughly 50 % changed their GRA from wave 11 to wave 13. In all three sub-samples a greater proportion of those who changed their attitudes report more egalitarian attitudes during the pandemic.

Looking at the changes in the division of family labor from before to during the pandemic in all three sub-samples shows that substantially more respondents report no change in the division of family labor than those who report a change. However, in all three sub-samples, at least 20 % of the respondents report change. Especially the division of housework became less egalitarian, while in both sub-samples considering the division of childcare, a greater proportion of those who changed reported a more egalitarian division of childcare. This matches the finding from the literature that if men stepped in more during the pandemic, it were mostly fathers.

Figure 1:

Coefficients with 95 % confidence intervals for a change towards an inegalitarian or egalitarian division of housework. Dependent variable: Change in GRA item 2) (“Men should participate in housework to the same extent as women”). Negative values indicate a change in GRAs in a more egalitarian direction, positive values a change in a less egalitarian direction. pairfam, own calculations

The bottom paragraph of Table 3 shows the average GRA change score by the change in the division of family labor. Matching this paper’s expectations, respondents who experienced emerging cognitive dissonance between pre-pandemic GRAs and pandemic behavior also changed their attitude in the direction of the changed behavior. Interestingly, also respondents who experienced no change in the family labor division changed their attitude on average in a more egalitarian direction. This supports the notion that cognitive dissonance is not the only theoretical mechanism explaining attitudinal change (Schober & Scott 2012). Unfortunately, using the asymmetrical fixed effects approach only allows to analyze the effects of intra-individual changes in the independent variables. Hence, I am unable to examine possible GRA changes for those respondents who experience no change in their division of family labor. This limitation should be kept in mind.

4.2 Asymmetrical Fixed Effects Models

Figures 1, 2 and 3 show the coefficients for changes in the division of housework (Figure 1) and childcare (Figure 2 & 3), derived from the asymmetrical fixed effects regressions. Coefficients are shown for changes towards a more inegalitarian (H1) and towards a more egalitarian (H2) division of family labor. If the intervals do not overlap with zero (dotted line), the coefficients are statistically significant at least on a 5 % level. Coefficients are shown for the whole sub-samples (with and without controls) and separated by sex (without controls). The detailed regression tables can be found in Tables 4–6.

Figure 2:

Coefficients with 95 % confidence intervals for a change towards an egalitarian and inegalitarian division of childcare. Dependent variable: Change in GRA item 1) (“Women should be more concerned about their family than about their career”). Negative values indicate a change in GRAs in a more egalitarian direction, positive values a change in a less egalitarian direction. pairfam, own calculations

4.3 Sensitivity Analyses

Various sensitivity analyzes were conducted to test the robustness of the models shown above (see Tables A5-A11).

To adjust for potential floor and ceiling effects that might affect the GRA items due to their measurement on a 5-point Likert-scale, I replicated all analyzes using Tobit regressions (McBee 2010). All results from the main analyzes are robust.

Also, pandemic-related methodological issues need to be considered and analyzed. In wave 13, panel stability was lower than in previous pairfam waves (Brüderl et al. 2022). This could be a result to the pandemic but also result from lower panel attachment in the refreshment sample that was introduced in wave 11, especially affecting the younger cohorts (Brüderl et al. 2022). To examine who dropped out of the panel from wave 11 to wave 13, I created summary statistics comparing all those who participated in wave 11 but not in wave 13 (26 %) to all those who participated in both waves (Table A9).

Especially younger respondents dropped out of the panel, as well as respondents who were on average more lowly educated, had a lower income, spent more time on the job and had younger children. Hence, it could be that especially time constrained respondents dropped out of the panel. Looking at GRAs, respondents who dropped out had on average less egalitarian attitudes concerning items 1) and 3) than those who participated in waves 11 and 13. Attitudes about item 2) as well as the division of family labor do not vary substantially across samples. This potential selection bias could lead to an underestimation of the real effect and should be kept in mind when interpreting the results.

The mode switch to a mixed-mode design during the pandemic might also bias the results (Groffen et al. 2021). In the case of pairfam, mode-specific selection effects (Fessler et al. 2018) can be neglected, because respondents could self-select into the interview mode. However, mode-specific measurement errors (i. e., respondents answering questions differently across modes) might bias the results. Measurement errors are likely to occur with sensitive questions (Stockdale et al. 2008). Hence, it could be that GRA changes towards less egalitarian attitudes are underestimated in the models because respondents tend to be more biased by social desirability in CATI interviews (Holbrook et al. 2003) and report more moderate attitudes.

To account for potential mode effects, in Tables A9-A11, I replicated the main analyzes while including a control variable for the change in mode. In almost all models, the change in mode has a significant effect and can be associated with a change in GRAs in a more egalitarian direction. Also, the association between a change to an inegalitarian division of family labor and a change toward less egalitarian attitudes become more pronounced. Hence, it could be that respondents who reported more egalitarian attitudes during the pandemic more often self-selected into CATI mode or that attitudes are reported to be more egalitarian in CATI mode. However, because the effects for the main independent variables show the same tendencies, the interpretation of the main analyzes is robust. Nevertheless, it should be noted that it is not possible to fully decompose mode from pandemic effects, and hence, all analyses should be interpreted cautiously (Bozoyan et al. 2021).

Limitations

The paper has limitations that need to be considered. A conceptual limitation is that reversed causality cannot be ruled out: The causal order of the association between behavior and attitudes is unclear, as they are found to engage in constant feedback (Hudde et al. 2022; Preston 2023). This issue is further enhanced through the fact that attitudes and behavior are measured at the same timepoint in the analyzed data. Hence, further research should analyze data that allow for temporal and causal ordering.

Secondly, I only analyze respondents who reported no pre-pandemic attitude-behavior discrepancy, which I assume can be translated into experiencing no cognitive dissonance. A change in the division of family labor due to (assumingly) external constraints during the pandemic is interpreted as producing cognitive dissonance. However, as the actual feeling of discomfort or dissonance is not asked for in pairfam and a change in family labor not necessarily results from the pandemic, these theoretical mechanisms are only assumptions (Schober & Scott 2012; Vandecasteele et al. 2022).

Considering only respondents who reported no attitude-behavior discrepancy before the pandemic means to only analyzing a very selective group. The mechanisms of changing attitudes for respondents who already experienced attitude-behavior discrepancy before the pandemic hence remain undiscovered. Next to that, due to the first difference method used, mechanisms of changing attitudes when no changes in behavior occur cannot be examined by the analyzes. As the division of family labor is only captured in a relative measure, cognitive dissonance could also emerge due to a change in the absolute amount of family labor, even though the relative division remains the same. These associations are not captured in the model, given the relative measurement of family labor division. Both latter points show that attitudes and behaviors do not always align and support the notion that cognitive dissonance is not the only motivator of attitudinal changes and formation (Schober & Scott 2012). Amongst others, age- and cohort effects might also influence the development of GRAs as well as the division of family labor. Future research should aim to address other mechanisms influencing attitude formation over the life course beyond cognitive dissonance.

Like in other comparable research on GRA change (e. g., Schober & Scott 2012; Zoch & Schober 2018), effect sizes found in the models are moderate. However, it still can provide insights into psychological mechanisms likely influencing attitude formation over the life course.

There are also some important data limitations. Firstly, when only considering respondents who reported no behavior-attitude discrepancy before the pandemic, the sample sizes especially of those who experienced a change in the division of family labor is small, decreasing the statistical power. Other research on cognitive dissonance and GRA change suffers from the same issue (Zoch & Schober 2018). This should be considered when interpreting the results and future research should aim to analyze larger data sets.

Secondly, the items measuring GRAs in pairfam (and many other surveys) are not up to the current stand of research because they are based on a traditional view on gender roles and do not cover variances and differences in egalitarianism (Pérez Sánchez et al. 2021). Future surveys should aim to more validly measure GRAs.

An issue resulting from the data collection is a possible mode effect, as wave 13 was conducted with a mixture of CAPI and CATI methods (Brüderl et al. 2022). Furthermore, pairfam is affected by attrition: A possible selection bias might be that the most time constrained respondents (e. g., parents with young children) dropped out during the pandemic (Bozoyan et al. 2021). Even in light of the sensitivity analyses I conducted, it is not possible to fully separate potential pandemic effects from mode and selection effects. Findings should thus be interpreted with caution.

Keeping the limitations in mind, this paper contributed to both pandemic research as well as research considering the micro-level processes of the behavior-attitude relationship and gender equality in Germany. Analyzing rich longitudinal data allowed for sensitive and detailed analyzes of pandemic-related emergences of cognitive dissonance due to changes in the division of family labor and changes in GRAs during the pandemic, which have not been conducted before. While it cannot be concluded that cognitive dissonance is the only motivator of attitudinal change, aligning with other research, cognitive dissonance shows to be one useful mechanism of explaining heterogenous attitudinal changes across the life course, in a setting in which endogeneity can be reduced.