Demographic Research: Volume 46, Article 21
https://www.demographic-research.org 629
children.
7
The set of explanatory variables also included gender, birth cohort (born pre-
1945, between 1945–1955, and post-1955), education level (primary, secondary, tertiary
education), employment status (retired, working, other), union duration (measured as a
continuous variable), previous divorce experiences
8
(has never divorced versus has
already divorced at least once), homeownership (yes or no), perceived financial distress
(household makes ends meet with great difficulty, with some difficulty, fairly easily,
easily), number of limitations in daily activities (scored from 0 to 6), depression level (a
0-to-12 scale based on the EURO-D depression scale, where 0 is ‘not depressed’ and 12
is ‘very depressed’) (see Prince et al. 1999), country of residence, and the wave in which
the respondent entered the observation. Descriptive statistics are reported in Table 1.
Unfortunately, two of our dataset’s variables were characterised by a non-negligible
number of missing values: union duration (approximately 10% of the sample) and
homeownership (roughly 3%). Importantly, missing information about union duration
showed no specific pattern by gender, country, birth cohort, or socioeconomic status,
whereas it was slightly more frequent in more recent waves. As union duration and
homeownership are key variables in our analysis, eliminating such a large share of
respondents would have remarkably reduced the final number of observations.
Consequently, we decided to retain them in the sample after having imputed the missing
information – which we achieved through multiple imputations by chained equations
(MICE) using STATA (see Lee and Carlin 2010). This technique allows each variable to
be imputed using its own conditional distribution and specifying different models.
Accordingly, we imputed union duration (a continuous variable) using a linear regression
model and a logistic regression for homeownership (a dummy variable). Multiple
imputation estimates several values for each missing data point, thus introducing the
uncertainty associated with the missing data into the model. We then used these values
in the analysis and combined the results following Rubin’s (1987) rule.
7
The variable about the frequency of contact is available in SHARE with the following categories: ‘coresiding
child(ren),’ ‘daily,’ ‘several times a week,’ ‘about once a week,’ and ‘rarely.’ Contact is considered either
personally, by phone, mail, email, or any other electronic mean during the previous 12 months. We used this
information for the dichotomisation made between those with grandchildren and report having coresiding
children or daily contact with them (‘frequent contact,’ about 60% of the sample), and those who have
grandchildren and report having contact with children several times a week or less (‘weak contact,’ about 40%
of the sample). Nevertheless, given the relatively small number of events, a finer division of the variable into
more categories would have resulted in imprecise effect estimates, leading to inconclusive findings.
8
We considered only previous divorces and not previous union dissolutions as SHARE collects only
information on the former.