The population is a subset of individuals included in RELINK53, a new register-based dataset collected in 2017/18, consisting of all individuals born in 1953 who lived in Sweden in 1960, 1965 and/or 1968, as well as their ascendant, contemporaneous and descendant family members.Footnote 2 From this dataset we defined the parental generation as all individuals born in 1953 and their siblings. Biological family linkages (both biological parents) were identified through the Swedish multigenerational register and convictions data, and mortality and migration data were merged to all family members. We included all children to the parental generation (1) whose father had been convicted for a final time ten years prior to birth through age 14; (2) who were at risk of being convicted for a crime at age 15 (i.e. who were alive and had not migrated at age 15); (3) could be followed at least until age 25; and (4) for whom both parents could be identified through the multigenerational register—a total of 96774 children nested in 58980 families. This includes 39679 same-sex full siblings who were discordant in the yearly timing of a final paternal conviction within the age window and thereby comprise the sample for the sibling analyses. The fathers had reached a median age of 64 at follow-up, and 99 percent had reached at least age 50, and none of the children had a father who had been convicted in the outcome period at age 15 to 25. We may therefore with good reliability measure the final paternal conviction in the children’s lives, and also have good control over the time ordering between exposure and outcome, thereby avoiding the risk for reverse causation (see Besemer et al. 2017).
Swedish Conviction Data
The current study may be viewed as part of a tradition of Nordic criminological research that utilizes administrative data to measure criminal behavior and criminal sanctions at the individual level (see Lyngstad and Skardhamar 2011). In this context, conviction records may generally be said to capture offenders who have committed several offenses, and/or serious offenses, and/or traditional person- and property-oriented offenses (Kyvsgaard 2002). Although a main limitation with official crime data is that a large share of criminal offenses are undetected, Swedish conviction data has been characterized as an official data source that has a relatively high degree of coverage with respect to criminal offenses (von Hofer 2014). This is in part because Swedish police and prosecutors are required to report all offenses that come to their attention, and that there are no legal grounds for the use of discretion in the handling of the criminal offenses during this phase of the criminal justice process (von Hofer 2014).Footnote 3 Furthermore, if a prison sentence is not the typical outcome on the basis of the penal code, the prosecutor can choose to issue a summary sanction order or a waiver of prosecution. These types of sanctions are also included in the conviction register.Footnote 4 Waivers of prosecution and summary sanction orders together constitute around half of all convictions in the convictions register from 1973 to present day.
It is important to note that longitudinal criminological studies that use official data do not only measure the etiology of criminal behavior, but also the behavior of the criminal justice system (Bushway and Tahamont 2016). In Sweden, a conviction is a quite heterogeneous measure of exposure as it may indicate a wide variety of rule-breaking behavior, which may also yield in a number of different custodial or non-custodial sanctions. As reviewed above, theories about the intergenerational continuity in crime range from mechanisms related to the exposure to parental criminal offending and to criminal justice processing respectively. Given our use of conviction data, we cannot separate between these mechanisms but have to assume that a parental conviction likely captures both.
Aside from the overall conviction risk in children, we analyze the subcategories of violent and property crime, since these may be more or less tied to familial mechanisms and/or other social causes. Whereas violent crime has been shown to be more strongly tied to familial confounds (Frisell et al. 2011; Kendler et al. 2015), property crime make up a large part of adolescence-limited crime and may, generally, be more tied to social processes outside of the family, such as those related to peers and routine activities. Because the age of criminal responsibility in Sweden is 15, we lack coverage of criminal behavior prior to this age. Still, our follow-up, covering a ten year period during which crime typically peaks and starts to decline, ought to be a rather reliable method to measure the prevalence of persons who commit crime more frequently. As argued by von Hofer (2014), these individuals are likely to be “arrested and convicted—sooner or later—irrespective of short-term trends in the clearance rate and the likelihood of arrests leading to convictions” (p. 171).
A limitation with the sibling design is that these models restrict the study population to families who have at least two full siblings with different birth years and birth orders. Aside from the threat to external validity (e.g. single-child families are excluded), we are limited in our capacity to control for these variables as they are collinear with the exposure to a final paternal conviction (i.e. it is always the older sibling who is exposed). Given the decline in convictions across successive cohorts during recent decades in Sweden (von Hofer 2014), the estimates in the sibling model are likely overestimated, to some degree. However, we also present the results for children who were born within different lengths of time after the final paternal conviction occurred. Comparing these groups of unexposed children are informative with regard to the extent by which our results may be driven by period effects.
When it comes to birth order, there are reasons to believe that our estimates are underestimated due to our inability to control for this variable. Previous studies indicate that later-born siblings are generally more disadvantaged than are first-borns (e.g. Modin 2002), and a recent study provides solid evidence for this also being the case with delinquency, in both a Nordic and Anglo-Saxon context (Breining et al. 2020). Moreover, any degree of sibling carryover effects, where the exposure and outcome of one sibling affect the exposure and outcome of another sibling (Sjölander et al. 2016), would primarily result from the older sibling influencing the behavior of the younger (Farrington 2002; Kendler et al. 2015; Mikkonen et al. 2020). Given such a direction of sibling influence, the effects in our exposure-discordant design would most likely be downwardly biased as it is always the older sibling who is exposed to a paternal conviction.
From the conviction register we extracted information on crime type, type of sanction, the date of conviction, and the date of the crime. Our main outcome variables are any crime, violent crime, and property crime. Any crime is defined as all acts described in Swedish penal law or special laws that result in a criminal sanction. Violent crime includes homicide, manslaughter, assault, robbery, kidnapping, threat, unlawful coercion, unlawful threat, rape, molestation, assault or threat to a public servant, and violent resistance. Property crime includes theft, grand theft auto, and shoplifting.
The main explanatory variable is the age at the final paternal conviction, constructed by subtracting the date of the last recorded conviction from the birth date of the child. While the date of the conviction was available in exact days, the birth date was available in a monthly format, thus yielding the precision in months. We constructed a dummy variable that took the value of 1 for individuals exposed to a final paternal conviction after birth, and also five-year age dummies to contrast different age periods over the full time window from ten years prior to birth until age 14.
The following control variables were constructed to be included in the population-averaged models: a dummy variable taking one for a final paternal prison sanction; a dummy variable taking one for a final paternal violent conviction; dummy variables for prior paternal conviction frequency categorized into no prior convictions (reference category), 1, 2, 3, 4, and > = 5; dummy variables for accumulated prior paternal prison time categorized into no prison time (reference category), < 3 months, 3–12 months, and > 12 months; a dummy for whether or not the father had a criminal history of violence; a dummy for whether or not the mother was convicted at any time prior to age 15; a dummy for female; dummy variables for birth year; a dummy for single child families; dummy variables for birth order in full families and categorized as 1st born (reference category), 2nd born, 3rd born, and 4th born or higher; and a continuous variable measuring the maternal age at birth.
We employed linear probability models (LPMs) to study the effects of age of paternal conviction on the child's conviction probability. LPMs are consistent estimators for binary outcomes (Angrist and Pischke 2009), and allow for direct comparisons of coefficients across different models (Mood 2010).Footnote 5 We estimated clustered standard errors to account for the fact that observations are not independent but comprise children nested in families (Stock and Watson 2008).Footnote 6
We estimated a set of population-averaged models (in the full sample) and sibling fixed-effects models (in the sibling sample), where the main independent variable was either a dummy variable for exposure after birth (reference category before birth) or a set of dummy variables for age at exposure in 5-year intervals (i.e.− 10 to − 6; − 5 to − 1; 0–4; 5–9; 10–14). Model 1 only included the main independent variable to examine the bivariate association. Model 2 included measured controls for confounding (see Variables above). Finally, in Model 3, we employed sibling fixed-effects. All families with at least two siblings who were discordant in the exposure (before and after birth or age of exposure, depending on model) were included in these analyses.
Since our follow-up of the children’s conviction risk extends over emerging adulthood to age 25, we also estimated non-parametric Kaplan–Meier cumulative probability functions to explore whether or not there were any degree of difference during the ages when the children were convicted (see also D’Onofrio et al. 2010). Kaplan–Meier cumulative probability functions are equivalent to Kaplan–Meier survival curves but measure the failure rate instead of the rate of survival. These functions thus complement the point estimates from the LPMs since they allow us to see if there is any particular stage between ages 15 and 25 when children begin to differ in their conviction risk. We estimated these functions both to show the bivariate association between age at exposure to a parental conviction and the children’s conviction risk in the full sample, and in the sample of full siblings who were exposure-discordant (i.e. before and after birth) at their final paternal conviction. These curves were right-censored at the time of decease or emigration.
Gender has been highlighted as an important variable to control for in sibling designs (D’Onofrio et al. 2013). This is particularly relevant when examining an outcome that typically differs substantially between males and females, such as criminal convictions. To make sure that there would not be any systematic gender differences in exposure-discordant siblings, we made sure that brothers were only compared to their brothers and sisters to their sisters. As a final step, given our focus on the final paternal conviction, we also modeled the association between a final paternal conviction and child conviction risk separately for males/brothers and females/sisters. This analysis basically shows if the effects in the main models are driven by males or females.
Except for the point in time before and after the birth of the child, the five-year age-intervals are arbitrarily chosen. We therefore analyzed the association between the child’s age at paternal conviction and the child’s conviction risk with age in annual and biannual categories, in order to explore if we could justify another categorization of age. Overall, we found a similar age trend, and could not find any particular cut-off where the estimates changed substantially.
We also carried out a set of sensitivity analyses to see if the results were influenced by the left-censored data. First, children born before 1983 are not fully covered on paternal convictions ten years prior to their birth (e.g. the cohort born in 1982 is followed nine years prior to birth, the cohort born in 1981 is followed eight years prior to birth, and so on). We therefore estimated the models for children born 1983 and later. Second, we do not cover the full paternal criminal history among children whose fathers were born before 1958 (in other words, individuals born in 1958 or later were fully covered on criminal history as they turned age 15 in 1973). Since prior paternal conviction frequency is arguably one of the most important confounders when analyzing a timing effect of paternal convictions (e.g. Besemer 2014), we estimated the models on children born to fathers with full criminal history (fathers born in 1958 and later). Third, the transition to parenthood occurred at a wide range of different ages for the fathers in our sample. To make sure that the results were not driven by particularly young fathers, we estimated the models for all families where the father was at least 25 years old at the time of the first-born child. Importantly, in the sibling analyses, we control for all factors that are shared between full siblings by design, among them the full parental history including prior criminal offending and convictions.
Finally, to check the robustness of our results across types of paternal convictions, we carried out subsets of analyses in exposure-discordant children and full siblings by type of final paternal conviction. We separated between those paternal convictions that did, or did not, include a violent crime, as well as between those that did, or did not, result in a prison sanction. These analyses thus yielded a break-down of children and full siblings whose fathers’ last conviction were of a particular type (non-violent, violent, conditional, or prison). The results showed that the bivariate association was somewhat stronger when the final paternal conviction included a violent crime or yielded a custodial sanction, but the substantial results were replicated in the population-averaged multivariate models and sibling FE models.