Young adults’ and adolescents’ compliance with public health measures is essential for containing the spread of diseases such as that caused by coronavirus disease 2019 (COVID-19)1,2. Adolescents can be infectious while remaining asymptomatic or showing only mild symptoms3. This group also has a strong need for social connection and hence can be highly effective in spreading the disease4. Unfortunately, adolescents and young adults are also identified as having relatively low levels of compliance with public health measures5,6. Campaigns during the COVID-19 pandemic have thus explicitly targeted this group (for example, and, and proposals have been made to utilize social media ‘influencers’ with a large adolescent following4.

A key factor that shapes compliance is social influence2,4,7,8. In particular, perceived behaviour of those in adolescents’ close circles is shown to influence adolescents’ compliance with public health restrictions8. Social influence in general is thought to amplify both risky and beneficial behaviours, particularly among adolescents4,7.

Social influence on compliance with public health restrictions can be understood from a social norm framework9. When public health authorities announce a certain restriction, they prescribe an injunctive norm (that is, what is ought to be done collectively)10. Injunctive norms are different from, and can even be in conflict with, descriptive norms (that is, what is believed to be practised in reality by most people)10,11. Injunctive norms will be stronger if they are in sync with descriptive norms. Indeed, research has shown that perceptions of others’ compliance with and approval of restrictions are associated with respondents’ own compliance8,12,13,14 and that compliance with restrictions has been stronger in communities with high pre-pandemic levels of social capital and cohesion15. Conversely, people who accept the injunctive self-isolation norms but do not perceive widespread compliance experience a strong loss of trust in others and reduced compliance16,17. Interestingly, while perceptions of others’ compliance have a positive correlation with trust in others, one’s own agreement with social-distancing norms has a negative association with trust16.

The literature of norm compliance during the pandemic is, however, scarce as to perhaps one of the strongest forms of social influence, namely, parental influence. In existing research, a generic inner circle’s norm compliance is studied without a specific reference to parents8,12,17,18. Social science research has shown that parents are highly influential in shaping adolescents’ values, attitudes, beliefs and indeed behaviour19,20,21,22. Socialization theory explains this parental influence via a myriad of channels. Parents can shape their children’s values, behaviour and norms through modelling whereby children observe their parents’ behaviour and expressed attitudes; through formal training whereby parents directly instruct expected behaviours; and through conditioning with rewards and punishment shaping children’s behaviour23. Interestingly, past research has shown a particularly strong influence of mothers and a relatively weak influence of fathers on adolescents’ values and behaviours19,22. This stronger influence of mothers relative to fathers is often explained by the tendency that mothers interact more frequently with offspring than fathers do19 and that mothers are often more involved in their children’s activities outside the home, such as educational and social events22.

Additionally, while transmission of values, beliefs, norms and behaviour predominantly occurs from parents to child24, when the child reaches late adolescence, parent–child relationship dynamics change25. During late adolescence, while parental power decreases, conflict with parents and the influence of peers—both of which tend to increase from early to mid-adolescence—decrease, too26,27. Overall, parental influence tends to rebound and stabilize from age 16 to 19. This means that parental influence persists during and even beyond late adolescence and that as parent–child relationships become more egalitarian, adolescents also start exerting some influence on their parents, albeit a smaller one than parents do on their children25,26,27.

Given the above literature, one expects a strong influence of parents, particularly mothers on adolescents’ compliance with pandemic restrictions, and a weaker influence of adolescents on their parents’ compliance. To my knowledge, however, there is no research on within-family influence on compliance during the COVID-19 pandemic. Moreover, the pandemic presents a novel research set-up, which helps address some other gaps in our knowledge regarding within-family transmission of norms and values. Authorities raised a very strong and novel injunctive public health norm during the pandemic. The extent to which compliance with this norm is transmitted across generations will shed light into the transmission process. For example, while past research studied transmission of values that have deep roots in society, such as religious beliefs20, gender and work norms23,28, and other value orientations such as hedonism, achievement and self-direction29,30, it is not known yet whether a similar transmission process occurs for a specific and novel norm, such as compliance with a particular set of public health measures.

In this Article, I study how parents impact their adolescent child and the adolescent child their parents in complying with social-distancing measures. In addition, I investigate how living arrangements, such as whether the child co-resides with their parents during the lockdowns, and relationship quality in the family moderate within-family transmission of compliance. Relationship quality is found to be a key moderator of interpersonal influence; hence, it is important to study how much within-family transmission processes vary by the quality of parent–child relationship23,25. In addition, whether the child co-resides with parents may be another important moderator of within-family transmission of compliance. First, physical proximity is directly related to intergenerational cohesiveness31. Second, sharing the same living space may facilitate socialization, for example, observing parents’ behaviour for modelling, and monitoring and reinforcing children’s behaviour are easier when children and parents live together.

I address the research questions using the longitudinal University College London (UCL) COVID-19 survey, which collects compliance data during the two national lockdowns enacted in the United Kingdom, one in May 2020 and another in February–March 2021. The UCL COVID-19 survey is implemented as an extension of existing cohort studies and most relevantly of the Millennium Cohort Study (MCS)32,33. Data have been collected from the main MCS respondent who was at the age of 19 during the two national lockdowns. In addition, the parents of the main respondent have also been asked to respond to the survey independently. Using the pre-pandemic MCS sweeps, I reconstructed family triplets that comprise the child, the mother and the father (Methods).

Next to addressing the knowledge gaps discussed above, the dataset and the models I use also ameliorate several methodological shortcomings in the literature of social influence and compliance. First, in almost all existing studies, the extent of others’ compliance is based on self-reported measures from the respondent. The link between compliance and expectations of others’ compliance could thus simply reflect a projection of one’s own compliance to others8,34. In the dataset I use, compliance is measured independently from each member of the child, mother and father triplet. Second, most studies on compliance rely on cross-sectional or prospective data8,12,17,18. This shortcoming is particularly important in understanding social influence on compliance, for a link between compliance and compliance of others could reflect similar people self-selecting into similar environments (for example, those who are already complying with the norm are more likely to be friends) rather than genuine social influence (for example, one friend influencing another to comply)34. Those that utilize longitudinal data have explored a plethora of predictors of compliance but did not explore the association between social influence and compliance1. The longitudinal nature of the data I use allows to decompose the causal chain (that is, parental influence on child versus child’s influence on parents). Moreover, having repeated measures on the child, the mother and the father allows to control for ultimately all observed and unobserved time-invariant household-level factors (for example, socioeconomic background, geographic region, ethnic composition of the household and so on) through the use of the so-called household fixed effects. This feature of the data and the modelling strategy give strong leverage to address omitted variable bias35.

Next to addressing knowledge gaps in the literature, a focus on within-family dynamics of compliance with health measures may be beneficial from a public health policy perspective. Media campaigns targeting young adults (for example, and as well as policy research5 during the COVID-19 pandemic portray this group as distinct and sometimes deviant. Adolescents are often stigmatized unfairly as ‘spreaders of the virus’36. Identifying parental influence on adolescents’ compliance may first redistribute some of the responsibility across all generations.

Second, recognizing within-family influences may help develop better public health campaigns. The data I use here are based on recent but past episodes of national lockdowns. At the time of writing this Article, there is no public health restrictions in the United Kingdom. However, how the COVID-19 pandemic will evolve is still unclear. With the emergence of new virus variants, new restrictions may be needed. Even after the COVID-19 pandemic is safely behind us, there will be other epidemics. The results of this study imply that public health campaigns that target not only adolescents and young adults but also their parents and that acknowledge within-family dynamics in compliance with public health measures may have better chances of succeeding than those that ignore such dynamics.


Descriptive results

I first estimate through full information maximum likelihood (FIML) the means, variances and covariances of compliance with social-distancing measures during the two national lockdowns, namely, that in May 2020 and in February–March 2021. These descriptives are given in Table 1. Figure 1 shows the average compliance of the child, the mother and the father at two timepoints, again estimated with FIML. All P values reported below are based on two-sided tests.

Fig. 1: Compliance with social-distancing guidelines.
figure 1

Average compliance among the child, the mother and the father at the two timepoints (first lockdown and the third lockdown). The 95% CI is included. Means are estimated with FIML (N = 6,752).

Table 1 Estimated sample moments

Among both children (χ2(1) = 653.79, P < 0.001) and the mothers (χ2(1) = 8.60, P = 0.003), compliance has been significantly lower in the subsequent lockdown than the first (Fig. 1). The drop in compliance of children has been stronger than that of the mothers (χ2(1) = 343.12, P < 0.001). The fathers’ mean compliance, however, does not change significantly over time (χ2(1) = 0.63, P = 0.426). Overall, mothers have the highest compliance levels, children the lowest and fathers in between but closer to mothers’ than to children’s; the differences in compliance between mothers, fathers and children in the two timepoints are statistically significant (χ2(4) = 1,292.96, P < 0.001). However, recall that the maximum compliance score is 10, and all means including those for children are close to the maximum. Table 1 shows that all correlations, but between fathers’ compliance in the first lockdown and their child’s compliance are large and statistically significant.

Within-family influence

Figure 2 shows the two models fitted to the data. In both versions, in each timepoint, the mother’s and the father’s compliance affect the child’s compliance. These predicted effects from parents to children (that is, respectively, the paths from mother to child in timepoint 1 and 2 that are labelled in Fig. 2 as bmc1, bmc2, and, respectively, the paths from father to child in timepoint 1 and 2 that are labelled as bfc1, bfc2) are based on previous research in socialization and developmental psychology, which shows that parents dominate the transmission process in that parents transmit values to children, whereas adolescents have limited effects on their parents24,25. The paths from a parent to child in the two timepoints are estimated freely. An alternative version would constrain these paths to be the same (that is, bmc1 = bmc2, bfc1 = bfc2), which would impose the extent of parent–child transmission to be the same in the two timepoints. Due to the novelty of the public health restrictions during the pandemic, however, one may conjecture a stronger influence in the first lockdown than the second. I will test this conjecture below. In addition, as mentioned above, research has also shown that adolescent children also exert some limited influence on their parents, a smaller one than parents do on their children25. This possibility is captured by the paths from the child’s compliance in timepoint 1 to parents’ compliance in timepoint 2 (these two paths are labelled in Fig. 2 as bcm (the path from child to mother) and bcf (the path from child to father)). Note that the child’s possible influence on their parents can also be captured as contemporaneous paths (that is, reciprocal paths for bmc and bfc). However, adding those reciprocal paths is complicated methodologically, for they would make the model non-recursive, and without further constraints on the paths, the model would be unidentified37. Given the theoretical expectation that parents dominate the influence process and these methodological reasons, I opted for the specifications in Fig. 2. Nevertheless, I will test below contemporaneous paths from the child to their parents, with additional cross-time equality constraints on parameters to identify the model.

Fig. 2: Longitudinal structural equation models of parental influence.
figure 2

a, Baseline model (model A). b, Extended model with household fixed effects (model B). Comp, compliance with social distancing; subscripts indicate timepoint; e = error (disturbance) term that captures the variance of the outcome variable that is not explained by the predictors in the model.

The models also include lagged direct effects from parents to child (bmcl and bfcl), and an individual’s compliance in timepoint 1 affects their own compliance in timepoint 2. These latter autoregressions capture within-person ‘stability’ in compliance38.

The version in Fig. 2b includes household fixed effects. Those household fixed effects, as in conventional multilevel regression framework, can be thought of as dummy variables—one for each household. They capture the effect of all observed and unobserved time-invariant factors defined at the household level (for example, family socioeconomic background, ethnic and educational composition of the household, location of residence and so on). In the structural equation modelling framework, these fixed effects can be included in models as a latent variable. The indicators of this latent fixed effect variable are all endogenous variables whereby the paths from the latent variable to the endogenous variables are constrained to 1 (ref. 39). If we expanded the model with further exogenous time-varying or time-invariant explanatory variables (which I will do in the ‘Additional results’), those variables will be allowed to freely correlate with the latent variable FE (that is, household fixed effects; see Supplementary Methods 1 and Supplementary Fig. 1 for an illustration). In the model in Fig. 2b, thus, every time-invariant variable defined at the household level is controlled for. This provides strong leverage in addressing omitted variable bias and hence makes a causal interpretation more credible32. The interpretation of the paths in model B, however, is different than those in model A. In model B, the paths are about the effect of within-household change in compliance on within-household change in compliance.

Table 2 shows the estimation results for models A and B in Fig. 2 and selected fit measures. Figure 3 plots the path coefficients from the same set of results. Fit measures show that both models A and B fit data reasonably well. A likelihood ratio test (LR χ2(1) = 5.60, P = 0.018) and the Akaike Information Criterion (AIC) suggest that model B has a slightly better fit than model A, while the Bayesian Information Criterion (BIC) suggests that model A has better fit than model B.

Fig. 3: Path coefficients in models A and B.
figure 3

Midpoints represent unstandardized coefficients of the models in Fig. 2a,b (N = 6,752), for example, the first estimate (Comp child (t = 2) on Comp mother (t = 2)) shows the path coefficient from the mother’s compliance at timepoint 2 on the child’s compliance at timepoint 2. The 95% CI is included.

Table 2 Unstandardized structural equation model results

Estimated path coefficients show that the mother’s compliance affects their child’s, particularly during the first lockdown. A unit increase in the mother’s compliance score in timepoint 1 is associated with, depending on model specification, 0.3 (b = 0.29, P < 0.001 in model A) to 0.2 (b = 0.17, P = 0.010 in model B) points increase in the child’s compliance in timepoint 1. Note that an effect of 0.17 in model B is a within-household effect, net of all time-invariant observed and unobserved household-level factors. The mother to child effect is smaller in timepoint 2: 0.13 in model A and 0.06 in model B. A formal test of the differences between the paths bmc1 and bmc2 is statistically significant in model A (Wald χ2(1) = 5.86, P = 0.016) and insignificant at the 0.05 level in model B (Wald χ2(1) = 2.81, P = 0.094). This suggests some reduction in maternal influence in the subsequent lockdown than the first.

The path coefficient from the father to the child, however, is statistically insignificant in both specifications for both timepoints (respectively, in models A and B, b = 0.073, P = 0.177 and b = −0.002, P = 0.973 (timepoint 1); b = 0.045, P = 0.494 and b = −0.026, P = 0.707 (timepoint 2)). But the father’s compliance is correlated significantly with the mother’s compliance (for example, in timepoint 2, covariance between the mother’s and father’s compliance is estimated as 0.278, P < 0.001 (model A) and as 0.141, P = 0.027 (model B)) and hence indirectly with the child’s compliance too particularly in timepoint 2 (correlation between father’s and child’s compliance in timepoint 2 is 0.144, P < 0.001).

The lagged paths from the mother (bmcl) and the father (bfcl) in timepoint 1 to child in timepoint 2 are statistically insignificant. Again, this does not imply that compliance of the parents in timepoint 1 has no effect on the compliance of the child in timepoint 2. In fact, the indirect effect of mother’s compliance in timepoint 1 on their child’s in timepoint 2 through their own compliance in timepoint 2 and the child’s compliance in timepoint 1 are significant (indirect effect = 0.18, P < 0.001 in Fig. 2a; indirect effect = 0.08, P = 0.025 in Fig. 2b).

The path from the child to the mother is significant too in model A (b = 0.077, P = 0.002) but attenuates to null once household fixed effects are controlled for in model B (b = 0.01, P = 0.769). Moreover, the path from the child to the father is statistically significant in both specifications (b = 0.187, P < 0.001 in model A, b = 0.102, P = 0.035 in model B). This means that adolescents exert some influence on their parents, particularly on their fathers, while they are influenced mainly by their mothers.

Not surprisingly, the path from one’s own compliance at timepoint 1 to that at timepoint 2 is generally large and statistically significant (the smallest of those autoregression paths is from mother (t = 1) to mother (t = 2), which is estimated as 0.240, P < 0.001 in model B, and the largest is from father (t = 1) to father (t = 2) estimated as 0.51, P < 0.001 in model A).

Additional results

In models A and B, parental influence on the child is contemporaneous, while there is no contemporaneous path from child to parents, in keeping with the finding in the literature that parents dominate the influence process. It is, however, possible to add contemporaneous paths from the child to both the mother and the father (that is, adding paths reciprocal to bmc and bfc in Fig. 1), although to identify the model with those reciprocal effects, one would need to constrain the set of path coefficients from parents to child and that from child to parents in the two timepoints to be equal (that is, bmc1 = bmc2, bfc1 = bfc2, bcm1 = bcm2, bcf1 = bcf2). However, those contemporaneous paths from the child to the parents are statistically insignificant at the 0.05 level (LR χ2(2) = 5.93, P = 0.052 in model A and LR χ2(2) = 3.45, P = 0.179 in model B). Hence, there is not enough evidence in the data for a contemporaneous effect of child on their parents, given the rest of the model.

I then test how living arrangements moderate the paths in Fig. 2, particularly whether the child lived with parents during the lockdowns. To do so, I fit a multiple group version of model A in Fig. 2 (model B failed to converge with multiple groups) whereby the parameters are allowed to differ for families whose children were at home during the lockdowns versus those whose children lived away from their parents. Indeed, the parameters differ significantly across these two groups (LR χ2(25) = 361.04, P < 0.001). Most notable differences are that for children who stayed away from their parents; there was no significant parental influence at timepoint 2 (Wald test of all four paths from parent’s compliance to child’s compliance in timepoint 2 is χ2(4) = 1.41, P = 0.842) and a weak maternal and insignificant paternal influence at timepoint 1 (paths, respectively, from mother’s and father’s compliance to child’s compliance at timepoint 1 are b = 0.19, P = 0.02 and b = 0.11, P = 0.244). Whereas for children who lived with their parents, parental influence was significant in both timepoint 2 (Wald χ2(4) = 40.54, P < 0.001) and timepoint 1 (Wald χ2(2) = 36.60, P < 0.001), and the father had a significant influence on their children at timepoint 2, too (b = 0.15, P = 0.024) (see Fig. 4a for the path coefficients).

Fig. 4: Moderators of within-family influence.
figure 4

Midpoints represent unstandardized coefficients of the models in Fig. 2a fitted with a multiple group framework whereby parameters are allowed to vary across different types of families. a, Comparison of families for which the child was away versus at home during the lockdown. b, Comparison of families that argue more versus less frequently than the sample median. c, Comparison of families with daughters versus sons as the adolescent respondent. The 95% CI is included.

The quality of the relationship between the child and their parents is also shown to facilitate intergenerational transmission of values and norms23,25. To investigate this, as above, I fit a multiple group version of model A in Fig. 2 (model B again failed to converge with multiple groups) whereby the parameters are allowed to differ for families which argued a lot before the pandemic versus which did not as much. Supporting a moderation effect, the parameters differ significantly across these two groups (LR χ2(25) = 152.16, P < 0.001). Most importantly, maternal influence is stronger in families that did not argue as much than in families that argued more, both in the first (path coefficient = 0.64, P < 0.001 versus 0.20, P < 0.001, test of differences between coefficients: χ2(1) = 67.80, P < 0.001) and in the second lockdown, and in fact, in the second lockdown, maternal influence is statistically significant in families that did not argue much (path coefficient = 0.27, P < 0.001) while insignificant in families that argued more (path coefficient = 0.07, P = 0.176, difference in coefficients = 0.20, χ2(1) = 14.56, P < 0.001) (Fig. 4b).

I then fit model A in a multiple group framework for boys and girls separately. The results show that the path coefficients (the path from mother to child in timepoint 2 and that from father to child in timepoint 2, as well as that from the child to the mother and from the child to the father in both timepoints) do not differ significantly between boys and girls (LR χ2(7) = 9.37, P = 0.227). However, at timepoint 1, mothers have an influence on daughters (b = 0.4, P < 0.001, 95% confidence interval 0.28–0.51) but no significant influence on sons (b = 0.1, P = 0.152, 95% CI −0.05–0.30) (Fig. 4c).

Robustness checks

Supplementary Results 1 and Supplementary Figs. 2 and 3 present the path coefficients estimated with robust standard errors, clustered at the household level, as well as estimates obtained after applying survey weights. These alternative estimation methods do not alter the estimates and the conclusions in any substantial way.

Models A and B in Fig. 2 control for earlier compliance in estimating effects on later compliance. Moreover, model B includes household fixed effects that account for all time-invariant covariates at the household level. Hence, models A and B should largely address omitted variable bias. Nevertheless, I expand models A and B by including several time-varying covariates, to mainly assess the robustness of parental influence to adjusting for key time-varying covariates. Research has identified several factors that affect compliance. These include political and interpersonal trust, perceived risks and gender1,16,40. Supplementary Results 2 and Supplementary Fig. 4 show the estimated effects of the mother’s and the father’s compliance on the compliance of the child, alongside several covariates: trust in government and interpersonal trust, gender, whether the respondent had COVID-19 and has chronic illness and the mode of the interview in timepoint 2 (phone versus web). These estimates are obtained by expanding models A and B, and in model B, the household fixed effects are allowed to freely correlate with those exogenous variables. The results show that interestingly, taking the paths at timepoint 2, for example, political trust has a positive coefficient (b = 0.10, P < 0.001) while trust in people has a negative coefficient (b = −0.05, P = 0.001), women have higher compliance than men (b = 0.14, P = 0.018), having had COVID-19 strongly reduces compliance (b = −0.54, P < 0.001), having a chronic illness increases compliance insignificantly (b = 0.07, P = 0.219), and those who were interviewed on the phone vis-à-vis web report higher compliance (b = 0.19, P = 0.008). Most importantly, these results show that parental influence estimates are virtually identical after controlling for these covariates (for example, in model A, mother to child path is 0.140, P = 0.001 in timepoint 2 and 0.28, P < 0.001 in timepoint 1; the path from child’s compliance at timepoint 1 to father’s compliance at timepoint 2 is 0.07, P = 0.005).

Finally, Supplementary Results 3 presents the results of analyses that check how much the results are robust to regional differences. Supplementary Fig. 5 shows the path coefficients of model A estimated with a multiple group framework whereby coefficients are allowed to vary in England on the one hand and Scotland, Wales and Northern Ireland on the other hand (separate groups for each country did not converge due to relatively low N per country). While there are some differences between the two regions (for example, mother-to-child path in timepoint 2 is stronger in England than in the other countries), the path coefficients are qualitatively similar between the two regions.


Here I study, using longitudinal data and panel models, compliance with social-distancing measures during the two lockdowns in the United Kingdom (May 2020 and February–March 2021). I do so by analysing triplets that comprise the adolescent child (age 19), their mother and their father. The results show that adolescents have significantly lower levels of compliance with social-distancing measures than their mothers and fathers, while mothers have the highest levels of compliance. Compliance is lower in the subsequent lockdown than the first lockdown for the children and their mothers, while for fathers, there is no change. I should add though that compliance has been generally high for all groups, so the differences are relative.

In addition, I find that mothers, and when the child is living with their parents, the fathers, have a significant influence on their adolescent children’s compliance with social-distancing measures. These effects survive various alternative model set-ups and adjusting for key covariates, including lagged measures of the outcome, household fixed effects and several time-varying covariates.

Compared with mothers, fathers have smaller and mostly insignificant effects on their children’s compliance, in keeping with the earlier results on parental influence19,22. This literature explains a stronger influence of mothers on their children compared with fathers by the propensity that mothers are often more involved in the children’s activities in and beyond the home such as education and social events22 and that mothers tend to interact more frequently with their children than fathers do19. In line with these explanations, I also find that living arrangements of the child moderate parental influence. For children who were at home during the lockdowns, parental influence is stronger and the father too has a significant influence on their children, at least during the third lockdown. Likewise, the quality of relationship between the child and their parents also facilitates within-family influence on compliance.

Children too have significant effects on their father’s and when they live with their parents also on mother’s compliance. This influence from the child to their parents again resonates with past research, which shows that particularly during late adolescence, children have some influence on their parents, though a smaller one than parents have on them25.

A potential limitation is that I rely on self-reported measures of one’s own compliance with social distancing. These self-reported measures may be optimistic or suffer from social desirability bias. However, given the difficulty of obtaining more objective measures of individual-level compliance, it is inevitable to rely, at least until better measures are available, on these self-reported measures. In addition, there is item non-response and attrition in the survey, which is a limitation (Supplementary Table 1). However, missing data are handled through FIML, which is shown to produce unbiased results under certain assumptions. Moreover, robustness checks have been conducted with survey weights, which address non-response in an alternative way, and the results were similar.

The effects I find here correspond to around 10 to 30 percentage points changes in outcomes induced by a unit change in the independent variables, which are relatively modest. Nevertheless, during a pandemic, even small improvements in compliance with public health measures may have strong long-term effects. This is not only because small but consistent behavioural improvements reduce the transmission risk in the long run but also because behaviour can cascade through social networks whereby intergenerational transmission of compliance can reach peers beyond the immediate family through peer-to-peer interactions4,7,8,12.

Here I focus on parental influence on adolescents’ compliance with public health measures during the COVID-19 lockdowns. Adolescents’ and young adults’ compliance with public health measures has been essential for controlling the pandemic. While this group has lower levels of compliance, they are often unfairly portrayed as ‘spreaders of the virus’36. The parental influence on adolescents’ compliance as well as the influence adolescents have on their parents that I document here may redistribute some of the liability across all generations. Moreover, better public health campaigns could be developed by considering these family dynamics for the current pandemic and future epidemics. For example, parents can be reminded of their influence on their children and that young adults can be reminded of their influence on their parents’ compliance with social-distancing guidelines. Such campaigns can produce a powerful message: would one want their possibly vulnerable parents or children to be less protected against the virus due to their own lack of compliance with public health measures?



The data used in this study are publicly available and collected after ethical approval and consent from the participants. The MCS follows a nationally representative sample of nearly 19,000 people born in the United Kingdom in 2000–2002. Since the pandemic has started, a longitudinal UCL COVID-19 survey is implemented with the MCS respondents and their parents. The UCL COVID-19 survey takes place in three waves conducted, respectively, in May 2020, September–October 2020 and February–March 2021. The majority of the surveys are implemented online, but a minority of wave 3 respondents are interviewed via telephone. Waves 1 and 3 correspond with national lockdowns; hence, compliance with social-distancing measures is asked only in waves 1 and 3. I thus use these two waves (henceforth, timepoints 1 and 2).

The parents of the MCS members are invited independently to take part in the survey. No explicit links are made between the parents and the MCS cohort members during invitation or data collection. Details of the UCL COVID-19 survey can be found elsewhere29,30.

Using pre-pandemic sweeps of MCS, I identify the child, their mother and their father in the COVID-19 survey. In particular, sweep 6 for parents has information on the relationship to the main cohort member, the gender of the parent and a unique within-household identifier of the parent, and sweep 7 includes a unique identifier for the main cohort member using which the parents and the main cohort members are linked with the UCL COVID-19 survey.

I exclude a small minority of individuals (2%, 409 cases) who are a different family member than the child, mother or father participated (for example, a grandparent, brother or another relative or co-residing non-relative). The resulting dataset comprises 6,752 child, mother and father triplets (~20,000 individuals) for whom at least one instance of non-missing compliance data exists.

The original MCS is nationally representative of the cohort. The effective sample of respondents who responded to the COVID-19 survey is diverse too, except for age—the adolescents were all at the age of 19–20 during the survey, and their parents are of similar age due to being at a similar life stage. Of the adolescent respondents, 60% are female, and 20% are of ethnic minority background. Average after-tax weekly total income of the respondent and their partner (if there is any) is reported as £184 by the adolescents and £992 by the parents.

Compliance with social-distancing guidelines for each member of the child, mother and father triplet is measured by the following item. “The next question is about the extent to which you are complying with the social-distancing guidelines issued by the Government. On a scale from 0 to 10, where 0 means that you are ‘not complying at all’ and 10 means you are ‘fully complying’, how much would you say you are complying with the guidelines?” The question does not specify what social-distancing guidelines are at the time of the survey.

Three variables are used as moderators in the ‘Additional results’ in Results, namely, whether the child resided with their parents or away, the pre-pandemic relationship quality with parents and the child’s gender. I now explain how I constructed those variables. Almost all adolescents reported to be living with their parents at timepoint 1 (91%). This ratio is 70% at timepoint 2. A binary time-invariant variable is constructed if the child was staying away from their parents versus home at timepoint 2. If this information is missing for an adolescent, the information on whether the child was living away from their parents versus home at timepoint 1 is substituted. Pre-pandemic relationship quality is constructed as follows. Sweep 6 of MCS asked “Most young people have occasional arguments with their parents. How often do you argue with your [mother | father]?” with answer categories 1 = most days, 2 = more than once a week, 3 = less than once a week, 4 = hardly ever and 5 = never. I compute an average arguing frequency with the mother and the father after reverse coding these items and create a binary variable by median split. Gender of the child is self-reported.

Estimation strategy

The number of observations in the waves of the UCL COVID-19 survey changes due to new respondents joining at the newer waves or attrition (Supplementary Tables 1 and 2). To mitigate the potential effect of missingness on statistical power and bias, I implement FIML estimation in fitting all models, including the model used to estimate descriptive statistics (means, variances and the correlations of the key variables; Table 1). FIML results in unbiased estimates under the assumptions that data are missing at random (that is, missingness is accounted for fully by observed data) and distributed multivariate normally. It is furthermore shown to be rather robust to violations of the latter multivariate normality assumption41. The former assumption of missing at random is also plausible in this case, for the models rely on repeated measures of the outcome variable. In addition, in some model specifications, I include household fixed effects, which capture all observed and unobserved household-level time-invariant confounders. These models make it more plausible to assume missingness to be at random conditional on household fixed effects and earlier or later measures of compliance.

The data are analysed with Stata version 17.0 (all structural equation models) and R version 3.6.1 (Fig. 1).

Statistics and reproducibility

The study is based on the UCL COVID-19 survey, which is a secondary data source collected independently from the author. The survey is conducted as an extension of an existing cohort study (MCS); hence, no statistical method was used to predetermine the sample size. No data were excluded from the analyses, apart from family members who are not the focal child, mother or father. The study does not involve a randomized experiment. The collectors of the data were unaware of the hypotheses tested in this study.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.