Cannabis is among the drugs with the highest frequency of (ab)use. About 1 in 5 Europeans aged 15–64 reported to have experimented with cannabis. In the United States the prevalence in ages 16–34 was estimated at 51.6 % (European Monitoring Centre for Drugs and Drug Addiction, 2012). Regular cannabis use has been associated with health problems, including mood and anxiety disorders (e.g., Cheung et al. 2010) and chronic bronchitis (Hall 2015; Joshi et al. 2014). Early onset and regular use during adolescence has possible effects on cognitive functioning (e.g., Crean et al. 2011) and predicts diminished educational (Horwood et al. 2010; Lynskey and Hall 2000) and professional attainment (Fergusson and Boden 2008; Volkow et al. 2014). Furthermore, recent evidence suggests that high-potency cannabis use elevates the risk of developing psychotic disorders (Di Forti et al. 2015, 2014). Namely, the odds of showing psychotic symptoms in individuals who declared to have ever used high-potency cannabis are about three times larger than in individuals who declared to have never used cannabis during their lifetime. The risk of showing psychotic symptoms is further elevated if high-potency cannabis is used daily (i.e., OR = 5.4; P = 0.002; Di Forti et al. 2015). About 9 % of those who initiate cannabis use progress to regular use and abuse (e.g., Volkow et al. 2014; Budney et al. 2007). Given the possible adverse effects on health and lifetime outcomes and given its possible role in triggering first-episode of psychosis, it is important to understand the causes of individual differences in the liability to initiate cannabis use.

Twin and family studies have shown that both genetic and environmental factors (both shared by, and specific to, family members) have an important role in the initiation of cannabis use (Kendler and Prescott 1998; van den Bree et al. 1998; Vink et al. 2010). A meta-analysis of twin studies (Verweij et al. 2010) showed that additive genetic factors explain nearly half the variance in liability to initiate cannabis use (i.e., 48 and 40 % of the variance, in females and males, respectively), while the remaining variance is accounted for—almost equally—by shared and unshared environmental factors (both about 30 %).

Among the several attempts to identify genes that explain the heritability of initiation, a linkage study (Agrawal et al. 2008a) failed to identify statistically significant associated genomic regions, although it did identify several suggestive regions on chromosomes 18 and 1. Likewise, a meta-analysis by Verweij et al. (Verweij et al. 2013) combining the results of two genomewide association studies (GWAS) comprising about 10 000 individuals failed to detect common single nucleotide polymorphisms (SNPs) associated with initiation. It should be noted, however, that the association analysis by Verweij and colleagues was limited to common (i.e., minor allele frequency (MAF) > 5 %) HapMap SNPs (Consortium 2010). With the recent completion of large sequencing projects such as the 1000 Genomes (1000G) (Consortium 2012) and the Genome of the Netherlands (Boomsma et al. 2014; The Genome of the Netherlands 2014), more detailed genotypic information has become available in large GWAS samples. Given the availability of denser SNPs, which are expected to be in high linkage disequilibrium (LD) with the causal variants, we aim to extend the search for genetic variants (GVs) implicated in initiation to previously untagged common GVs, and to other (than common) GVs, such as low-frequency variants (1 % < MAF < 5 %). Such low frequency variants have not typically passed the quality control checks. However, the quality of imputation has been improved by recent advances in imputation techniques (Howie et al. 2012). This opens the door to including such GVs into a genome-wide association study.

Furthermore, to date, the approach for finding genes underlying the heritability of cannabis initiation was to focus on the ‘ever/never used’ dichotomy at the expense of the age at which one initiates (i.e., age at onset). Yet, age at onset is a complex trait (Visscher et al. 2001), subject to the influences of both environmental and genetic factors (Lynskey et al. 2003), and may serve as an important proxy for heavy use. Initiation of cannabis use before age 18 is predictive of both experimentation with other drugs (Agrawal et al. 2006; Lynskey et al. 2006), and of escalated drug use (e.g., Lynskey et al. 2003). Among those initiating in adolescence the risk of progression to symptoms of abuse and dependence is higher relative to the general population (i.e., 17 vs. 9 %, respectively; Volkow et al. 2014). Given its relevance as a predictor for escalated use, our second aim is to perform a genomewide search for GVs that give rise to individual differences in age at onset. To model age at onset as a function of genotype we will apply statistical methods based on survival analysis. This approach utilizes all available information on the age at onset among initiateds and takes into account the censored nature of the observations collected in those who did not initiate at the time they were last seen (i.e., they might initiate at a later time point). The approach is expected to show superior power relative to an analysis of the “ever-never” dichotomy or an analysis restricted to those who initiated (see e.g. Kiefer et al. 2013). To our knowledge, a genomewide survival analysis of age at onset of cannabis use has not yet been reported.

The outline of the paper is as follows. First, we estimate the amount of variance in initiation of cannabis use explained collectively by the currently measured SNPs. The purpose of such analysis is to obtain an indication of the total signal in the measured (and tagged) SNPs without identifying individual SNPs. Second, we conduct SNP-based association analyses of initiation and age at onset. Our primary focus is on identifying genes tagged by the SNPs, relevant to our traits. Therefore, next, we incorporate these SNP-based results in two gene-based analyses. These analyses are exploratory, i.e., conducted genomewide.

All analyses are performed in a sample of Dutch families from the Netherlands Twin Register (NTR). To maximize statistical power, imputation of genotypes in the NTR sample was based on two alternative reference panels: the 1000G Phase 1 project reference panel (Consortium 2012) and the reference panel generated by the Genome of the Netherlands (GoNL) project (Boomsma et al. 2014; The Genome of the Netherlands 2014). The GoNL reference panel was derived by sequencing the whole genome of 250 trio-Dutch families and matches therefore the ancestral background of our sample. The GoNL panel is expected to facilitate imputation of variants which are specific to the Dutch population (Boomsma et al. 2014). Furthermore, the use of the GoNL panel is expected to result in higher imputation accuracy relative to the 1000G panel, especially for low frequency GVs (MAF < 5 %) (The Genome of the Netherlands 2014). Such increased accuracy is expected to increase the statistical power to capture the signal in the measured GVs.