Home, sweet home? The impact of working from home on the division of unpaid work during the COVID-19 lockdown

A lockdown implies a shift from the public to the private sphere, and from market to non-market production, thereby increasing the volume of unpaid work. Already before the pandemic, unpaid work was disproportionately borne by women. This paper studies the effect of working from home for pay (WFH), due to a lockdown, on the change in the division of housework and childcare within couple households. While previous studies on the effect of WFH on the reconciliation of work and family life and the division of labour within the household suffered from selection bias, we are able to identify this effect by drawing upon the shock of the first COVID-19 lockdown in Austria. The corresponding legal measures left little choice over WFH. In any case, WFH is exogenous, conditional on a small set of individual and household characteristics we control for. We employ data from a survey on the gendered aspects of the lockdown. The dataset includes detailed information on time use during the lockdown and on the quality and experience of WFH. Uniquely, this survey data also includes information on the division, and not only magnitude, of unpaid work within households. Austria is an interesting case in this respect as it is characterized by very conservative gender norms. The results reveal that the probability of men taking on a larger share of housework increases if men are WFH alone or together with their female partner. By contrast, the involvement of men in childcare increased only in the event that the female partner was not able to WFH. Overall, the burden of childcare, and particularly homeschooling, was disproportionately borne by women.

. Average agreement with statements on WFH from couple households without children  Reading example:This radar chart displays the average agreement with different statements on WFH. The smaller the distance on the axis to the centre, the more the respondents disagree with the statement. Blue triangles represent answers from women, pink circles represent men's responses. On average, there is a high degree of similarity in answers to statements of men (pink circles) and women (blue triangles) indicated by the almost overlapping points on the different axis.

Fig. S2. Agreement with statements on reconciliation of family and WFH by gender
Reading example: This bar chart shows the distribution of agreement (from left to right "Strongly agree" (blue), "Rather agree" (green), "Rather disagree" (pink) to "Strongly disagree" (purple)) for four different statements by parents WFH. Statement 1 indicates for example that 42% of all mothers strongly disagree with "easy reconciliation at home", whereas only 25% of all fathers strongly disagree with this statement.

Fig. S3. Division of unpaid work before and during COVID-19 lockdown
Reading example: The two lines in Fig. S3A and Fig. 3B show the overall distribution of unpaid work before (blue and continuous line) and during (grey-purple dashed line) the lockdown, for housework (S3A) and childcare (S3B) respectively. For scale no. 0 (indicating the "woman does everything") we see that the number of households reporting this value increased during the lockdown, i.e. almost doubled.

S.B Descriptive statistics
Tab. S1. Survey sample size and key variables   We asked respondents how they and their partner spent the previous working day during the lockdown, summarized in table S3. The results reveal that women in working couples spent, on average, almost two hours more on unpaid work than men (4h03 compared to 5h58) per day. The average time spent on paid work by women amounts to 6h44 compared to 7h48 for men. It is revealing to look beyond the average time spent on a specific activity by carefully examining differences in the participation rates (i.e. the share of respondents having spent some time on a certain activity) and the average time spent on distinct activities based on this "participating" subsample. We find that 75% of men have participated in housework activities, such as cooking and cleaning on their previous working day, whereas almost all women have done some housework. Comparing men and women who have participated in housework activities, we observe that these women spent around 25 minutes more on these tasks. For childcare activities, we also note large gender differences. Half of the woman participated in physical childcare (feeding, washing and supervision), and these women spent on average 1h50 on this activity. Among men, both the participation rate (44%) and the average time spent on taking care of children (1h29) was slightly lower. The gender difference is much larger for home-schooling related activities. More than one third of of mothers studied with their children, compared to one forth of fathers, and they spent roughly half an hour more on home-schooling than fathers (1h44 compared to 1h15). The mean time spent on childcare, among those having done any childcare activity, amounts to 3h17 for fathers and 4h43 for mothers. Looking at the time spent on childcare, housework and paid work during the lockdown jointly reveals very long overall working days for parents, in particular for mothers.

S.C Robustness tests
We conduct a series of robustness tests to check whether the results presented in the main text are robust in terms of (S.C.i) the sample definition, (S.C.ii) controlling for whether the questions on time use and characteristics of a partner were answered by a male, (S.C.iii) different definitions of the control variables, and (S.C.iv) modified definitions of the dependent variable. In section (S.C.v) we present the results of models presented in the main text based on a linear probability model, instead of a logistic regression. Note that, as we present average marginal effects, this serves as an indirect validation of the average marginal effects. Tab. S10. Definition of age variable: age groups Tab. S11. Definition of children living in the household: age youngest child Tab. S12. Definition of the dependent variable: woman works more Tab. S13. Definition of the dependent variable: more equal division of unpaid work within the household Tab. S14. Linear probability model

S.C.i Sample composition
From the overall sample, we selected 730 heterosexual couples (1,460 adult individuals) in a first step, who are living in the same household, where both partners were either (self-)employed or in short-time work at the point when the survey was answered, and who answered the partner module of the questionnaire or linked their responses via anonymous partner IDs. Due to missing information, mainly in the income variable, the resulting sample corresponding to model (1) of the main text consists of 558 couples. In models (2) and (3), this sample is reduced to the 299 couples with children under 15 years of age.
Model (1), explaining the change in the division of housework, is based on a sample consisting of households with and without children. Model (4) presented in Table S4 is based on a sample of households without children. Thereby, we can check whether the results presented in the main text are driven by households with or without children. The effect of WFH is insignificant in model (4). Thus, we do not find evidence that WFH influences the probability of men increasing their share of unpaid work in childless couple households. In other words, we do not find evidence that WFH influences the probability of men increasing their share of unpaid work within childless households. This also indicates that the effects of model (1), as presented in the main analysis, and based on households with and without children, are driven by households with children, where either men do more housework if both parents are WFH (but not more childcare), or fathers take on more housework (and childcare) if they alone are WFH.
In an additional robustness check (see Table S5), we excluded individuals who worked only partly, and not entirely, from home, which does not have a significant impact on the results.

S.C.ii Controlling for the gender of the survey respondent
This check concerns the fact that 79.6% of the couple questionnaires were filled out by women. Therefore, we test whether the main results change if we control for the gender of the respondent by including a binary variable which takes the value one if the questionnaire was filled out by the male partner (see Table S6). In fact, this variable is highly significant for housework but not for childcare tasks. Moreover, the probability that men take on more housework is no longer significant for the whole sample. This is, however, no surprise as Table S4 already revealed that this effect is driven by households with children.

S.C.iii Specification of the control variables
The variable defining the relative income of the partners presented in the main text is based on categorical income variables. In Table S7 we employ a variable that is based on a subjective assessment of the income difference between partners. Respondents had to report the perceived difference from their partners (low, equal, high). In this robustness test, we make use of this variable. However, the results are not driven by the definition of the income variable and related measurement errors.
Furthermore, we alter the specification of the working hours variable. In one specification (see Table S8), we use continuous working hours instead of a categorical variable. Although the results for each additional hour worked are highly significant, they are small in magnitude. Thus, the effect of each hour is very small, confirming the results obtained by measuring hours worked for pay in categories. In a similar exercise, we vary the definition of part-time work. In the analysis presented in main text, respondents are classified as working part-time in the event that they worked fewer than 20 hours per week for pay. In the models presented in Table S9, those working fewer than 35 hours are classified as working part time. We find that men who work fewer than 35 hours a week without any short-time work arrangement have a significantly higher probability of taking on more housework and childcare during lockdown.
Controlling for age by means of age groups instead of a continuous definition (see Table S10), we detect no major changes in the results.

S.C.iv Specification of the dependent variable
We check the possibility that the results are driven by the definition of the dependent variable. Thus, we change the dependent variable to a dummy variable indicating whether the woman instead of the man within a couple took on more unpaid work during the lockdown. The results are presented in Table S12 and show that the main variable of interest -working from home -is not significant for this specification. Moreover, several other variables having a significant effect on the probability that the male partner within a couple takes on a greater share of unpaid work have no significant effect on the probability that a woman takes on more housework or childcare tasks (such as income and employment status). The only variable that remains highly significant is the pre-lockdown division of unpaid work. We conclude that the unequal division of unpaid work prior to the COVID-19 restrictions and the prevailing gender norms associated with it appear to be the most important predictor.
Furthermore, we changed the dependent variable to a binary variable that becomes one in the event that the division of unpaid work was more equal 1 during the COVID-19 restrictions than before (see Table S13). The results show that only the male partner WFH has a positive effect on the probability that the division of unpaid work becomes more equal, even though the effect for the whole sample is no longer significant (as in Table  1). Also it has a positive and significant effect in all three model specifications if both partners are WFH. In the models presented in the main text (Table 1), the effect of both partners WFH on the probability that a man takes on more childcare tasks is also positive but not significant. This could come from the fact that this dependent variable also responds to the case where the male partner took over a larger proportion of the childcare tasks before the lockdown and the woman increased her share during the COVID-19 restrictions (see Fig. 3 in the main text). If the housework or childcare activities had already been equally distributed before the lockdown, it has a (highly significant) negative effect on the probability that unpaid work was even more equally distributed during the COVID-19 restrictions compared to households where the woman previously did much more unpaid work than her male partner. This finding is in line with the main results. The distribution of income within the couple has a positive significant effect on the division of housework in family households if the male partner earns more (similar to the base model), but is not significant for any other model or category. The results for the remaining explanatory variables are similar to the base model, even though some covariates are no longer significant.

S.C.v Linear probability model
The results presented in the main text and the previous robustness tests are based on a logistic regression, estimated by maximum likelihood. In the corresponding tables, we report average marginal effects. In this section, we estimated the models corresponding to Table 1 based on a linear probability model specification estimated by ordinary least squares. This serves as an indirect test, as the average marginal effects should correspond to the effects of the linear probability model. Table S14 shows that the results do not differ between these model specifications.