Data were drawn from the NCDS, which is a nationally representative study that has followed a cohort of participants all born in a single week in the United Kingdom since 1958. Since birth, they have been followed up a total of 11 times at ages 7, 11, 16, 23, 33, 42, 44, 46, 50 and 55. As data on time spent caring for grandchildren is only available from the most recent interview, all analyses here are cross-sectional, with all women included in the sample being aged either 55 or 56 (depending on whether the interview was conducted in 2013 or 2014) and representing the third generation of women in Fig. 1. The sample was limited to women who had at least one parent alive and at least one grandchild (n = 934). Data from the NCDS are available from the UK Data Service, and the participant characteristics shown in Supplementary Table S1.
Hours spent helping parents per week
Information regarding parental caregiving was included as a count variable. In the most recent interviews, participants were asked whether they ever do various activities for their parents (e.g. shopping for them, helping with basic personal needs, giving them lifts, etc.), and if they do, how many hours on average per week do they spend doing so. Any women who reported not helping their parents do any of the activities were coded as helping their parents for zero hours per week.
Hours spent caring for grandchildren per month
The number of hours spent caring for grandchildren per month was also included as a count variable. Women were asked whether they ever look after their grandchildren without the grandchild’s parents being present, and if they do, at what frequency and for how many hours. Women who stated that they did not care for their grandchildren or did so less often than monthly were coded as caring for their grandchildren for zero hours per month. This measure also includes overnight stays.
Fecundity status at age 55
Fecundity status was derived from information on age, year and reason for last menstrual period, which was collected at ages 44, 50, and 55. Based on this, a binary categorical variable was derived where women were coded as either ‘Still menstruating’ or ‘No longer menstruating’. The latter category comprised of women who were post-menopausal or who had stopped menstruating for another reason, such as a surgical menopause. Women who had stopped menstruating due to menopause or other reasons were grouped together as the direct fitness implications of no longer menstruating are the same, regardless of the reason for it.
Covariates included were selected based on their expected effect on the woman’s ability to help other family members. As a proxy of socioeconomic status, the age at which the woman left education was included. Employment status was utilised to give an indication of the woman’s time constraints (i.e. if she was employed, it can be expected she had less time to care for kin)24, with women being coded as either employed, unemployed, or other, with the latter category including those who are doing something other than formal employment but do not classify themselves as unemployed (e.g. retired, volunteering, studying, etc.). Self-perceived health was used as a measure of how physically able the woman is to help family members25, and number of grandchildren was also included to adjust for how many grandparenting responsibilities a woman had. We also included information on the mortality status of the woman’s parents (i.e. whether she had both parents alive or not), which was derived from interviews at ages 7, 11, 16, 23, 42, 46, 50 and 55. The focal woman’s mother’s and father’s age at birth (collected in the perinatal interview) were also included to control for the amount of help her parents may need, as older parents would expected to be more in need of assistance. Finally, in models predicting hours spent caring for parents, time spent caring for grandchildren was adjusted for, and vice versa for models where hours spent caring for grandchildren was the outcome.
Time spent helping parents and caring for grandchildren were both modelled using zero-inflated negative binomial regression (ZINB). This modelling procedure was selected both due to the over-dispersed nature of the data with excess zeros, and because zero-inflated models allow for zeros to be generated through two distinct processes. Here, the model distinguishes between excess zeroes, which occur when the event could not have happened, and true zeros, which occur when there could have been an event. Therefore, the model estimates a binary outcome (does not care versus does care) and a count outcome (the number of hours spent caring). This method is theoretically appropriate, as there are many different reasons people would offer no care to kin: while some people may choose to invest less, for some people the choice is out of their control, with external factors influencing caring behaviours, such as living far away from kin26. In addition to this, ZINB was found to better fit the data than negative binomial regression (Supplementary Table S2).
Time spent helping parents was first modelled. A ‘base’ model was first made containing the age the woman left education, employment status, marital status, self-perceived health, number of grandchildren, parent mortality status, age of parents, and time spent caring for grandchildren. Fecundity status was subsequently added, and model fitting then carried out on these two models, utilising their Akaike Information Criterion (AIC) value to understand whether a model including fecundity better fit the data than one without. The model with the lowest AIC value is taken to best fit the data. As AIC values penalise models for complexity, it means the model with the most terms will not automatically be selected as the best. The ΔAIC was also calculated, which is the difference between the candidate models AIC and the AIC value of the best fitting candidate model. If the ΔAIC value is ≤ 2, then it indicates that there is still good evidence to support the candidate model, meaning that a candidate model with a ΔAIC of ≤ 2 is almost as good as the best fitting model. A ΔAIC value of between 4 and 7 is taken to indicate the candidate model has considerably less support, and a ΔAIC of greater than 10 indicates there is no support for the candidate model27. The Akaike weights (wi) were also calculated to evaluate model fit, which give the probability that the candidate model is the best among the set of presented candidate models27. The same procedure was then used to model time spent caring for grandchild per month: a model including just the covariates was first made, but this time adjusting for time spent helping parents rather than time caring for grandchildren, with fecundity status then being added, and model fitting was once again carried out using the methods outlined above. All analyses were carried in R using the zeroinfl function with a negative binomial distribution specified28, and model fitting carried out with the package AICcmodavg29. All visualisations were created using ggplot230.
Source: Ecology - nature.com