Article Text

Extended report
Genetic risk scores and number of autoantibodies in patients with rheumatoid arthritis
  1. Marthe T Maehlen1,2,
  2. Inge C Olsen1,
  3. Bettina K Andreassen3,4,
  4. Marte K Viken2,
  5. Xia Jiang5,
  6. Lars Alfredsson5,
  7. Henrik Källberg5,
  8. Boel Brynedal5,
  9. Fina Kurreeman6,
  10. Nina Daha6,
  11. Rene Toes6,
  12. Alexandra Zhernakova6,7,
  13. Javier Gutierrez-Achury7,
  14. Paul I W de Bakker8,
  15. Javier Martin9,
  16. María Teruel9,
  17. Miguel A Gonzalez-Gay10,
  18. Luis Rodríguez-Rodríguez11,
  19. Alejandro Balsa12,
  20. Till Uhlig1,
  21. Tore K Kvien1,
  22. Benedicte A Lie2
  1. 1Department of Rheumatology, Diakonhjemmet Hospital, Oslo, Norway
  2. 2Department of Medical Genetics, University of Oslo and Oslo University Hospital, Ullevål, Oslo, Norway
  3. 3Department of EpiGen, Institute of Clinical Medicine, University of Oslo, Oslo, Norway
  4. 4Department of Biostatistics, Institute of Basic Medical Sciences, University of Oslo, Oslo, Norway
  5. 5Institute of Environmental Medicine, Karolinska Institute, Stockholm, Sweden.
  6. 6Department of Rheumatology, Leiden University Medical Centre, Leiden, The Netherlands
  7. 7Department of Genetics, University Medical Center Groningen and University of Groningen, Groningen, The Netherlands
  8. 8Departments of Medical Genetics and of Epidemiology, University Medical Center Utrecht, Utrecht, The Netherlands
  9. 9Instituto de Parasitologia y Biomedicina Lopez-Neyra, CSIC, Granada, Spain
  10. 10Department of Rheumatology, Hospital Marques de Valdecilla, Santander, Spain
  11. 11Department of Rheumatology, Hospital Clinico San Carlos, Madrid, Spain
  12. 12Department of Rheumatology, Hospital La Paz, Madrid, Spain
  1. Correspondence to Dr Marthe T Maehlen, Department of Rheumatology, Diakonhjemmet Hospital, Oslo 0235, Norway; marthemaehlen{at}


Objective Certain HLA-DRB1 alleles and single-nucleotide polymorphisms (SNPs) are associated with rheumatoid arthritis (RA). Our objective was to examine the combined effect of these associated variants, calculated as a cumulative genetic risk score (GRS) on RA predisposition, as well as the number of autoantibodies (none, one or two present).

Method We calculated four GRSs in 4956 patients and 4983 controls from four European countries. All four scores contained data on 22 non-HLA-risk SNPs, and three scores also contained HLA-DRB1 genotypes but had different HLA typing resolution. Most patients had data on both rheumatoid factor (RF) and anti-citrullinated proteins antibodies (ACPA). The GRSs were standardised (std.GRS) to account for population heterogeneity. Discrimination between patients and controls was examined by receiveroperating characteristics curves, and the four std.GRSs were compared across subgroups according to autoantibody status.

Results The std.GRS improved its discriminatory ability between patients and controls when HLA-DRB1 data of higher resolution were added to the combined score. Patients had higher mean std.GRS than controls (p=7.9×10−156), and this score was significantly higher in patients with autoantibodies (shown for both RF and ACPA). Mean std.GRS was also higher in those with two versus one autoantibody (p=3.7×10−23) but was similar in patients without autoantibodies and controls (p=0.12).

Conclusions The GRS was associated with the number of autoantibodies and to both RF and ACPA positivity. ACPA play a more important role than RF with regards to the genetic risk profile, but stratification of patients according to both RF and ACPA may optimise future genetic studies.

  • genetics
  • rheumatoid arthritis
  • ACPA
  • RF
  • antibodies
View Full Text

Statistics from


Autoantibodies play an important role in the diagnosis, prognosis and treatment of patients with rheumatoid arthritis (RA). In the revised American College of Rheumatology (ACR)/European League Against Rheumatism classification criteria from 2010, presence of rheumatoid factor (RF) and anti-citrullinated proteins antibodies (ACPA) weight heavily as both their presence and level may generate points, making a classification of RA more likely.1 Autoantibody-positive RA (presence of RF or ACPA), and especially ACPA-positive RA, is associated with a more destructive disease course characterised by rapidly progressive erosive disease and also with an increased risk of ischaemic heart disease.2–4 ACPA and RF negativity has been associated with a drug-free remission, while ACPA-positive patients responded better to rituximab than ACPA-negative patients.5–7

RA is a heterogeneous disease, and therefore patients are often stratified into more homogenous subsets according to the presence or absence of autoantibodies. RF and ACPA, which are moderately to strongly correlated, are present in 50–80% of RA patients.4 ,8–10 ACPA can be observed up to 10 years before disease onset.9 RA is assumed to be caused by a combination of environmental and genetic factors, where the latter is estimated to account for 60% of the risk.11 ,12 Genome-wide association studies (GWAS) have led to the discovery of several RA-associated single-nucleotide polymorphisms (SNPs).13–15 Interestingly, these risk variants often confer risk to only one subset of RA, providing additional evidence for the differentiation between autoantibody-positive and autoantibody-negative RA.16 Most GWAS have been performed on only ACPA-/RF-positive patients, and thus most risk variants are associated with this disease subset. The main genetic contribution to ACPA-positive RA is conferred by the HLA-DRB1 locus and accounts for approximately one-third of the genetic risk.17 Certain DRB1 alleles encoding a similar amino acid sequence in position 70–74, known as the shared epitope (SE), are strongly associated with ACPA-positive RA and, to a lesser extent, RF-positive RA.18–20

Most genetic studies have looked at each genetic variant individually, and few studies have investigated the combined genetic risk in relation to autoantibody status in RA.21–23 Karlson et al developed a weighted genetic risk score (GRS) that consisted of 8 DRB1 alleles and 14 (and later 31) non-HLA SNPs, all associated with RA. Increasing GRS was associated with both erosive RA and RF-positive RA.21 ,22 This association with autoantibody-positive RA was later replicated in ACPA-positive patients by Kurreeman et al,23 who in addition found that ACPA-negative patients also had an increased GRS compared with controls. However, neither of these studies classified patients by ACPA and RF status. Therefore, it is still unknown whether the genetic risk contribution differs between patients who carry none, one or two autoantibodies.

In the current study, we expanded on Karlson’s GRS and made four different GRS, all containing genetic data on 22 non-HLA RA-associated loci but with different resolution level of HLA-DRB1 data included. Our aim was to explore the ability of these four scores to discriminate between patients and controls (using nearly 10 000 individuals) and to examine associations to autoantibody status, including the number of autoantibodies present.

Materials and methods

Patients and controls

We had clinical and genetic data available from 4956 patients and 4983 controls from Norway, Sweden, the Netherlands and Spain. All patients fulfilled the ACR 1987 criteria for RA.24 The study had been approved by national/regional ethics committee in each country, and all patients had given informed consent.

The Norwegian data consisted of 907 patients who were from four cohorts of RA patients and 1112 controls from the Norwegian bone marrow registry.25 The Swedish patients (n=2562) and controls (n=1968) came from the Epidemiological Investigation of Rheumatoid Arthritis (EIRA) cohort.26 ,27 The Dutch patients (n=648) were from the Leiden Early Arthritis Cohort, while the Dutch controls were from the Groningen cohort (n=1085).13 ,28 ,29 The Spanish patients had been recruited at hospitals across Spain (n=839), and the controls were from the Spanish DNA bank (n=818).13 All Spanish individuals were of white southern European descent. We had data on the presence of ACPA and RF for 4844 and 4793 patients, respectively. Analyses of autoantibodies in the patient cohorts have been described previously.3 ,13 ,20 ,25 ,28 ,30

SNP selection and genotyping

All cohorts had data available for 22 established RA susceptibility SNPs (non-HLA SNPs), which are shown in the online supplementary table S1. SNP genotyping methods varied between cohorts, and details can be found in the online supplementary text file. All SNPs had either reached GWAS significance level (p<5 × 10−8) in previous studies or had been independently validated in replication studies.13–15 ,31–37 The published OR and the OR calculated from our dataset (ACPA-positive patients vs controls) were comparable, with the effect sizes in the same direction (see online supplementary table S1). No linkage disequilibrium was found between the SNPs, except for rs6920220 and rs10499194 (D′=1, R2=0.06), which have been found to exert independent risk effects.34

Data on HLA-DRB1 alleles were available for all cohorts. The DRB1 SE alleles were defined as 01:01, 01:02, 04:01, 04:04, 04:05, 04:08, 10:01 and 14:02. As DRB1*13:01 had been found to have a protective effect in these four cohorts, the *13:01 allele was also included.38 The DRB1 genotyping method (including imputation) and resolution levels varied between cohorts, and care was taken to define the alleles at the same resolution level so that the allele groups were comparable across datasets.20 ,38–41 Details concerning the genotyping/imputation methods are given in the online supplementary text.

Genetic risk score

We developed four different weighted GRSs in line with the method described by Karlson et al.21 This method uses the published allelic OR for each SNP. In brief, our cumulative GRS was calculated by taking the number of minor alleles an individual possessed (ie, 0, 1 or 2) for a given SNP or DRB1 allele and multiplying it with the natural log of the published OR for the minor allele (see online supplementary tables S1 and S2). Published ORs were not available for all DRB1 groups/alleles, in which case we used our own estimates (see online supplementary table S2). The risk (or protective) contribution for each genetic marker was then added together to give a weighted combined GRS for each individual.

The four different GRSs differed according to the amount and level of resolution of the HLA-DRB1 data included. The first score, GRS-22, consisted of 22 non-HLA SNPs mentioned previously (see online supplementary table S1). The second score, GRS-23, consisted of 22 non-HLA SNPs and the number of SE alleles (0, 1 or 2) (see online supplementary tables S1 and S2). The third score, GRS-25, consisted of 22 non-HLA SNPs and the 3 main DRB1-allele groups; *01 (*01:01 and *01:02), *04 (*04:01, *04:02, *04:04, *04:05 and *0408) and *10 (*10:01) alleles (see online supplementary tables S1and S2). The fourth GRS, GRS-31, consisted of 22 non-HLA SNPs and all 8 individual SE alleles in addition to the protective effect of 13:01 (see online supplementary tables S1 and S2). As the Swedish cohort did not have complete four-digit DRB1 allele resolution, we were unable to calculate GRS-31 for this cohort.

We set a cut-off of individual genotype success rate to 85%; therefore, individuals who lacked more than three non-HLA genotypes were excluded from the analyses (n=67). For the remaining individuals, missing genotypes were imputed by calculating the expected allele count as described by Karlson et al.21 For the three GRS that included DRB1 data, patients and controls would be excluded if DRB1 data were missing (missing DRB1 data for patients n=525, controls n=568). In patients where the DRB1 genotype was ambiguous (eg, a genotype could be *0101 or *0402), we weighted the alleles according to the frequency of the two alleles in that specific population.

Overcoming population heterogeneity

The mean GRS in patients and controls differed significantly between countries, indicating population heterogeneity. To overcome the issue of population heterogeneity, we standardised the GRS (std.GRS) for each country separately and then pooled the standardised scores. Each individual's GRS was standardised according to the mean (and SD) in the population controls from each country separately. The standardisation procedure is described in detail in the online supplementary text file.


The distribution of GRS and std.GRS between patients and controls is presented as graphs, mean values with SD or as mean difference (Δ mean) with 95% CIs. Student t test was used to test for differences of the std.GRS between patients and controls.

We further wanted to explore the relationships between std.GRS and autoantibody status according to RF-positive/-negative, ACPA-positive/-negative and subgroups according to the number of autoantibodies carried (double-negatives, single-positives, double-positives). We examined the ability of the four standardised GRSs to discriminate between patients (all and divided into subgroups according to autoantibody status) and controls by using receiver operating characteristics (ROC) curves adjusted for gender.42 For the comparison of std.GRS across subgroups based on autoantibody status, we continued with two scores (std.GRS-22 and std.GRS-23) where we had data for all four cohorts. Analysis of variance was used to test overall differences followed by pairwise comparisons across two groups if the overall p value was <0.05. For data analyses, we used Statistical Analysis Software (SAS) V.9.3 (SAS Institute, Cary, North Carolina, USA) and R: A language and environment for statistical computing (R Foundation for Statistical Computing, Vienna, Austria).


Clinical characteristics of all patients and controls eligible for analysis are shown in table 1. The frequencies of positivity of autoantibodies differed according to country, ranging from 52% to 64% for ACPA and 54% to 72% for RF. Data on both RF and ACPA were available for 4741 patients, and the distribution of carrying none, one or two antibodies is shown in table 1.

Table 1

Demographics of RA patients and controls from four cohorts

In online supplementary figure S1, the mean (and SD) GRS-23 (containing 22 risk SNPs and SE) between patients and controls within each cohort is shown on the left, while the difference in means (Δ mean) between patients and controls is shown on the right. Although the difference in means between patients and controls was similar for all countries, there were large differences between the countries in the numerical mean GRS-23 for patients and controls (see online supplementary figure S1). Norwegian patients and controls had the highest mean GRS-23 of 1.77 and 1.18, respectively, while Spanish patients and controls had the lowest means of 1.34 and 0.74, respectively. Due to these population differences, the GRS of each individual was standardised according to the mean (and SD) in the population controls from each country separately (see online supplementary text). The standardised scores from the four cohorts were then pooled together. Population differences were also seen for GRS-22, GRS-25 and GRS-31 (data not shown), and therefore these scores were also standardised before being pooled together.

For all four standardised GRSs, the mean score was significantly higher in patients compared with controls (table 2), as illustrated by the std.GRS-23 distribution curve in figure 1A. We then tested the four scores’ ability to discriminate between autoantibody-positive/autoantibody-negative patients and controls using ROC curve analyses. Our results showed that the std.GRS improved its discriminatory ability when more HLA-DRB1 data were added, with std.GRS-31 having the highest area under the curve for all outcomes (table 3). Further, all four scores could better discriminate between patients who were RF-positive or ACPA-positive than patients who were negative for these antibodies (table 3). The std.GRS discriminated poorly between double-negative patients and controls (table 3).

Table 2

Mean differences between RA patients and controls for the standardised genetic risk scores

Table 3

Discriminatory ability of the four different std.GRS models

Figure 1

Frequency distribution of risk alleles and standardised genetic risk score (std.GRS) in rheumatoid arthritis (RA) patients and controls. (A) Distribution of std.GRS-23 in RA patients and controls. (B) Distribution of risk alleles in controls, ACPA-negative and ACPA-positive patients. (C) Distribution of std.GRS-23 in controls, ACPA-negative and ACPA-positive patients. (D) Distribution of std.GRS-23 in controls and patients stratified according to the number of autoantibodies. Std.GRS-23 consists of 22 non-HLA RA single-nucleotide polymorphisms and shared epitope alleles. ACPA, anti-citrullinated protein antibodies; RF, rheumatoid factor.

For the further comparative analyses of std.GRS across subgroups based on autoantibody status, we continued with only two scores, std.GRS-22 and std.GRS-23. Std.GRS-22 represented the genetic risk outside the HLA while std.GRS-23 included the HLA-DRB1 SE alleles, and such HLA data were available for all four cohorts. The distribution of the number of risk alleles (maximum 46 alleles) and std.GRS-23 is shown in figure 1B and C, respectively. Both graphs demonstrate a shift towards higher genetic risk in ACPA-positive patients compared with controls. Pairwise analyses revealed that both ACPA-positive and RF-positive patients had significantly higher mean std.GRS-22 and std.GRS-23 than controls (table 4). ACPA-positive patients had a mean GRS-23 that was nearly 1 SD higher than controls (Std.GRS-23 Δ mean: 0.92 (0.87–0.97), p=1.8×10−278). Although figure 1C suggests no difference in GRS between ACPA-negative and controls, statistical analyses demonstrated a significantly higher std.GRS-23 in ACPA-negative patients. Similar significant differences were seen in the analyses of RF-negative patients and controls (table 4). We further compared ACPA-positive patients with ACPA-negative patients and RF-positive patients with RF-negative patients, and found higher std.GRS-22 and -23 in patients positive for ACPA and RF, respectively (table 4).

Table 4

Mean differences between patients with antibodies versus controls and between ACPA-positive/RF-positive versus negative patients

Patients with data on both RF and ACPA status were divided into three groups—those having no autoantibodies (double-negative (27%)), one autoantibody (RF-positive or ACPA-positive (20%)) or two autoantibodies (double-positive (n=53%))—and the std.GRS in these groups were compared with controls. Figure 1D demonstrates the distribution of std.GRS-23 in the three patient groups and controls. The graph suggests that double-negative patients had a similar GRS distribution as controls, and pairwise analyses confirmed that there was no significant difference in the mean std.GRS-23 and only a marginal difference for std.GRS-22 (excluding DRB1) (p=0.024, table 5). Patients who were single-positive had significantly higher genetic risk than patients who were double-negative (table 5 and figure 1D). Patients who were double-positive had higher GRS than single-positive patients (table 5 and figure 1D). The GRS was primarily influenced by the presence of ACPA as ACPA-positive–RF-negative patients had similar GRS as double-positive patients. Single-positive patients were divided into two groups—those being ACPA-positive and RF-negative and those who were ACPA-negative and RF-positive—and we found that both subgroups had statistically significant higher std.GRS than controls (table 5).

Table 5

Mean differences between patients and controls, where patients were stratified according to autoantibody status


In this large study, we investigated the association between combined genetic risk in patients with RA compared with controls. We demonstrated that patients who carry no autoantibodies have a similar GRS as healthy controls. Further, we found that patients who carry two autoantibodies have a higher genetic risk compared with patients who carry only one, while patients who carry either ACPA or RF have a significantly higher GRS than patients who are negative for both autoantibodies (table 5).

The role of ACPA and RF in disease development is still not clear; however, both autoantibodies predict the development of RA with ACPA being sovereign when compared with RF.10 However, presence of both RF and ACPA, as opposed to the presence of only one, showed an increased positive predictive value of RA.10 Moreover, the presence of several isotypes of RF or having a broad repertoire of ACPA fine specificities has also been shown to increase the risk of RA development.43–46 Thus, it seems that the greater the autoantibody response, the greater the risk of RA. Our results showed that the combined genetic risk follows the same pattern with an overall association between the GRS and the number of autoantibodies a patient carries. However, not surprisingly, the GRS was similar in patients who were single-positive for ACPA to patients who were RF-positive and ACPA-positive. This reflects the fact that most of the genetic risk variants have been revealed through association with ACPA-positive RA, but nevertheless highlights the overlap in genetic risk factors for RF-positive and ACPA-positive RA. The interaction and interplay between RF and ACPA is unknown. However, Scherer et al proposed that RF might preferentially bind to ACPA as RF favours hypoglycosylated IgG and ACPA has been shown to be hypoglycosylated.47 Therefore, the presence of both RF and ACPA may enhance the immune response through immune complex formation and complement activation.47

Although the majority of known genetic risk factors have been found to be associated with ACPA-positive RA only, the evidence of genetic heritability from twin studies in ACPA-negative RA is comparable to ACPA-positive RA, 66% and 68%, respectively.11 ,48 Results from the Immunochip screen in RA have demonstrated that 11 genetic variants (out of 45) conferred similar risk in ACPA-positive and ACPA-negative patients, though only HLA-DRB1*0301 has shown association exclusively to ACPA-negative RA.13 ,49 ,50 As our GRS contained SNPs and DRB1 alleles primarily associated with ACPA-positive disease, it is not surprising that the GRS performed better at predicting autoantibody-positive disease compared with autoantibody-negative disease. The intriguing result was that there was no association between std.GRS-23 (and only a marginal association with std.GRS-22) and RA patients who were negative for both autoantibodies (table 5), indicating minimal genetic overlap between the two subsets of RA. The ROC analysis demonstrated consistent results since the std.GRS discriminated poorly between the double-negative patients and controls (table 3). Our findings highlight the lack of knowledge concerning the genetic background of ACPA-negative/RF-negative disease. This gap could be a result of the few studies performed in this subset but perhaps also indicates genetic heterogeneity.51 The ACPA-negative/RF-negative subset may potentially consist of several subsets with currently undiscovered genetic risk variants and autoantibodies.

Kurreeman et al23 found that ACPA-negative patients (n=378) had an increased GRS compared with controls. However, patients were not stratified for the presence of RF and therefore the association may be due to the presence of RF-positive patients in the ACPA-negative subset. Our findings emphasise the importance of stratifying for both ACPA and RF in genetic studies as patients who were positive for one autoantibody but negative for the other had a significantly higher std.GRS than patients being negative for both (tables 4 and 5). Although testing for both RF and ACPA will increase cost, the benefit of having distinct phenotypes, such as truly autoantibody-negative patients, will reduce the chance of false-positive/false-negative findings in genetic studies.

There are strengths and weaknesses in the current study. First, this study is considerably larger than any of the previous studies investigating combined genetic risk in RA. All patients are from cohorts of well-characterised RA patients with available data on both RF and ACPA. In addition, our GRS included data on DRB1 SE alleles, which explain a large proportion of the genetic risk in RA. We did not have genotype data for all established RA susceptibility loci (46 known loci). However, in a previous study, a GRS encompassing 39 genetic markers was compared with a GRS consisting of 22 markers, and no differences were seen in the ability of the models to discriminate between RF-positive individuals and controls.21 ,22 These results highlight the relatively small contribution of the non-HLA susceptibility loci to the overall risk of RA compared with the classical HLA loci.

GRSs have been developed for a range of complex diseases, but their ability to predict ‘individuals at risk’ has been disappointing.52–54 Results from our ROC analysis also showed poor ability to discriminate patients from controls. Our study therefore supports the view that the GRS lacks potential as a clinically useful screening test, although GRS may possibly play a role in risk prediction in the future when many more susceptibility loci have been identified.

The high GRS in the Norwegian and Swedish population most likely reflect that the RA susceptibility loci included have a higher minor allele frequency in northern compared with southern European populations. This observation is not surprising as most GWAS performed on RA encompass a higher proportion of patients of northern European decent. Both the Leiden and EIRA cohort had participated in some of the original GWAS discovery studies. To overcome the issue of population heterogeneity, we standardised the computed GRS according to the population control within each country. This method allowed us to analyse data on all individuals simultaneously, and meta-analyses with combined data on an individual level are preferred to combining group estimates.55

This large study on combined genetic risk supports the notion of RA consisting of two distinct disease subsets according to the presence of autoantibodies. Although this distinction has increasingly favoured stratification according to the presence or absence of ACPA, we showed that RF also plays a role with regards to genetic risk. Distinct phenotypes characterised according to the presence or absence of both RF and ACPA may help genetic researchers unravel the genetic background of RA.


The Norwegian Bone Marrow Donor Registry, Rikshospitalet, Oslo University Hospital, is acknowledged for providing Norwegian controls.


View Abstract

Supplementary materials

  • Supplementary Data

    This web only file has been produced by the BMJ Publishing Group from an electronic file supplied by the author(s) and has not been edited for content.

    Files in this Data Supplement:


  • Handling editor Dimitrios T Boumpas

  • Contributors MTM, ICO, MKV, BKA, XJ, LA, RT, AZ, JG-A, PIWdB, JM, MT, MAG-G, LR-R, AB, TU, TKK and BAL contributed to the conception and design of the study. MTM, ICO, BKA, XJ, LA, RT, AZ, PIWdB, HK, BB, ND, FK, JM, MT, TKK and BAL contributed to the analysis and interpretation of the data. MTM, ICO, BKA, TKK and BAL drafted the manuscript. All authors have critically revised the manuscript with regards to intellectual content as well as approved the final version of the manuscript.

  • Funding The Norwegian cohort had been funded through grants from the Southern and Eastern Health region of Norway. The EIRA study was supported by grants from the Swedish Medical Research Council, Swedish Council for Working life and Social Research, King Gustaf V 80-year foundation, the Swedish Rheumatism Foundation, Stockholm County Council, the insurance company AFA, The EU-supported AutoCure project, NIH (P60 AR047782) and from the COMBINE (Controlling Chronic Inflammatory Diseases with Combined Efforts) project. The Spanish cohort had been supported by the RETICS program, RD12/0009 (RIER), from the Instituto de Salud Carlos III. This work was supported by grants from the European Union (Seventh Framework Programme integrated project Masterswitch, grant no. 223404) and the IMI JU-funded project BeTheCure (contract no. 115142-2).

  • Competing interests AZ is supported by grants from the Dutch Reumafonds (11-1-101) and from Rosalind Franklin Fellowship, University of Groningen, the Netherlands. PIWdB is the recipient of a VIDI Award from the Netherlands Organization for Scientific Research (NWO project 016.126.354).

  • Ethics approval Ethical boards in Norway, Sweden, the Netherlands and Spain.

  • Provenance and peer review Not commissioned; externally peer reviewed.

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.