You are here

  • Home
  • Archive
  • Volume 73, Issue 8
  • Analysis and modelling of cholesterol and high-density lipoprotein cholesterol changes across the range of C-reactive protein levels in clinical practice as an aid to better understanding of inflammation–lipid interactions

Article Text

Article menu
Clinical and epidemiological research
Extended report
Analysis and modelling of cholesterol and high-density lipoprotein cholesterol changes across the range of C-reactive protein levels in clinical practice as an aid to better understanding of inflammation–lipid interactions
  1. Hanna Johnsson1,
  2. Maurizio Panarelli1,
  3. Allan Cameron2,
  4. Naveed Sattar3
  1. 1Department of Biochemistry, Glasgow Royal Infirmary, Glasgow, UK
  2. 2Department of Acute Medicine, Glasgow Royal Infirmary, Glasgow, UK
  3. 3Institute of Cardiovascular and Medical Sciences, BHF Glasgow Cardiovascular Research Centre, University of Glasgow, Glasgow, UK
  1. Correspondence to Dr Hanna Johnsson, Department of Biochemistry, Glasgow Royal Infirmary, Glasgow, G4 0SF UK; h.johnsson{at}doctors.org.uk

Abstract

Objectives Raised total cholesterol (TC) and reduced high-density lipoprotein (HDL) cholesterol levels are established cardiovascular disease (CVD) risk factors. However, in autoimmune conditions the lipid–CVD association appears paradoxical, with inflammation as a potential confounding factor. We therefore sought to model the relationship between systemic inflammatory illness and lipid levels using C-reactive protein (CRP) as the prototypical marker of inflammation. Our hypothesis was that there would be an inverse association between raised CRP levels and both TC and HDL-cholesterol levels.

Methods Results from samples analysed simultaneously for CRP and lipids in a 6-month period were collected retrospectively from a large city hospital laboratory database that collates results from both primary and secondary care. The relationships between CRP and lipids were determined using graphical techniques and empirical, non-parametric, best fit models.

Results A total of 11 437 blood samples was included. We identified a significant (p<0.001) biphasic relationship between TC and CRP: TC increased within the healthy CRP range of less than 5 mg/l, but decreased with CRP levels above 10 mg/l. The two effects approximately cancelled each other out in the intermediate CRP range of 5–10 mg/l. There was an inverse relationship between HDL-cholesterol and CRP.

Conclusions Lipid levels change significantly during inflammatory illness in a population with both acute and chronic conditions. These results provide a strong epidemiological basis for the better understanding of lipid changes in inflammatory conditions and with anti-inflammatory therapies.

  • Cardiovascular Disease
  • Inflammation
  • Lipids

https://doi.org/10.1136/annrheumdis-2013-203293

Statistics from Altmetric.com

    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.

    Introduction

    Cardiovascular disease (CVD) remains the leading cause of death in developed countries and is the main cause of premature mortality in patients with rheumatoid arthritis (RA).1 The risk of CVD mortality in patients with RA is increased by approximately 50% compared to the general population,2 and the rate of CVD is also increased in other autoimmune disorders.1 Assessing and treating CVD risk factors are therefore important to reduce the incidence of CVD and associated premature mortality. Increased total cholesterol (TC) and reduced high-density lipoprotein (HDL) cholesterol levels have been shown to be the best lipid markers of CVD risk.3 They are included in CVD risk calculators such as Framingham and SCORE to estimate an individual's CVD risk and thereby determine the need for preventive measures.

    However, lipid–CVD associations are complicated in certain circumstances because low TC is associated with increased all-cause mortality in the elderly,4 in patients on haemodialysis,5 and in those with cancer.6 The same observation has also been noted in patients with autoimmune disorders such as RA.7 In patients with RA, levels of TC and HDL-cholesterol decline when the disease is active, although the TC:HDL-cholesterol ratio appears to be less affected.8 These observations suggest that inflammation may play a confounding role in the association of lipids with CVD risk in groups known to exhibit higher levels of inflammation. This is not well appreciated by many clinicians, so lipids continue to be measured in many patients without regard to their inflammatory burden. Furthermore, an increase in lipids with successful treatment of RA with biological agents has been ‘perceived’ by the US Food and Drug Administration and others as a potential adverse effect of such agents.9 However, if a higher inflammatory burden was causally linked with lower lipid levels, then inflammation suppression might be expected to lead to elevations in lipid levels.

    The aim of this study of a large biochemistry database was to investigate and model the relationship between systemic inflammatory illness and lipid levels in a general, unselected patient population using CRP as the prototypical marker of inflammation. Our hypothesis was that there would be an inverse association between raised CRP levels and both TC and HDL-cholesterol levels.

    Methods

    Data collection and study population

    We retrieved results retrospectively of samples that had been analysed simultaneously for CRP, TC and HDL-cholesterol in a 6-month period (July to December 2011). The samples were from both primary and secondary care, taken for a variety of clinical indications and analysed at four north Glasgow biochemistry laboratories. In these laboratories, TC, HDL-cholesterol and high sensitivity CRP were determined by commercially available kits on auto analyser Architect c16000 (Abbott, Diagnostics, Abbott Park, Illinois, USA). For CRP the coefficient of variance (CV) was 10% and the limit of quantification was 0.1 mg/l with less than 0.2 mg/l as the lowest reported level. For TC and HDL-cholesterol the CV was 1% and the limit of quantification was 0.08 mmol/l. Values of CRP and HDL-cholesterol below the minimum threshold for detection were assigned to their expected values based on the known distribution of CRP,10 and HDL-cholesterol (internal data).

    We also gathered details on patients’ age and gender but otherwise had no details on body mass index, lifestyle factors or other clinical information. As we were specifically examining the relationship between lipids and inflammation in adults without severe dyslipidaemias we excluded results from patients under the age of 16 years and samples with extreme lipid values, as outlined below. Given our use of anonymised patient data previously collected as part of a departmental audit, our local ethics committee was satisfied that formal ethics application was not required.

    Descriptive statistics

    Data were initially examined using graphical techniques and tested for normality using QQ plots and the d’Agostino test. Median participant characteristics were then tabulated and the significance of gender differences was analysed using the Mann–Whitney U test. Changes in CRP and lipids with age were described graphically. Samples with gender and age data missing were excluded from this analysis but included in the modelling.

    Modelling

    The main aim of this study was to find a parametric model that describes the relationship between CRP and lipids. Using graphical techniques and empirical, non-parametric, best fit models, we determined the pattern of distributions and general relationships between, CRP, TC and HDL-cholesterol. When necessary, we applied appropriate power transformations (see supplementary material, available online only). Subjects with severe dyslipidaemias, defined for this sample size as TC of over 9.9 mmol/l or less than 0.4 mmol/l, were excluded (see supplementary material, available online only).

    The shape of the empirical best fit curves was used to generate hypotheses as to the overall relationships between first CRP and TC, then between CRP and HDL-cholesterol. Once the basic form of the relationship was determined, we selected appropriate regression techniques to find the maximum likelihood parameters of the model.

    The fit of the model was assessed by testing the normality, homoscedasticity and sum of squares of the residuals. The CI for the parameters were calculated from their individual t scores and SE. A generalised additive model was used to correct for age and gender. All statistical analyses were carried out using code written in the statistical programming environment R V.2.13.1.11

    Results

    Descriptive statistics

    A total of 11 437 blood samples was included in the study. Of these, 6044 were from women and 5377 were from men. The gender was not recorded in 16 cases and the age was not recorded in 12 cases. Median lipid levels were significantly different in men and women as described in table 1.

    View this table:
    Table 1

    Median age, lipid levels and CRP of study population

    Median CRP values increased linearly with age; median TC increased with age until approximately 50 years and then decreased again. Median HDL-cholesterol showed a very small increase with age that was statistically but not clinically significant.

    Modelling the relationship between TC and CRP

    We initially plotted the relationship between TC and CRP, and as expected most values were clustered around the lower end of the range of CRP as illustrated in figure 1. To describe the relationship better, we performed a logarithmic transformation to the base e of CRP (log CRP) as a histogram of CRP showed an approximately logarithmic distribution.

    Figure 1
    Figure 1

    The relationship between TC and CRP shows a cluster at the lower range of CRP. Please note that this figure was produced with all raw data in advance of removing extreme TC values. CRP, C-reactive protein; TC, total cholesterol.

    The non-parametric model between log CRP and TC showed that TC was not normally distributed around its mean. However, this was corrected by using a square root transformation of TC (sqrt TC). This indicated that if there was a relatively simple parametric model, it was between log CRP and sqrt TC.

    The non-parametric model showed a linear increasing relationship between transformed TC and CRP in patients within the healthy CRP range of less than 5 mg/l, but a linear decrease in transformed cholesterol with CRP levels above 10 mg/l. The two effects approximately cancelled each other out in the intermediate CRP range of 5–10 mg/l. A segmented linear regression in these three ranges was carried out, and produced a model with a highly significant upward slope in the CRP less than 5 mg/l range, a highly significant downward slope in the CRP over 10 mg/l range, and a flat line in the intermediate range. The transition points were continuous and the residuals passed several tests for normality. This model can be seen in figure 2, with full details in the supplementary material (available online only).

    Figure 2
    Figure 2

    Segmental linear regression of sqrt TC and log CRP showed a significant upward slope in the CRP less than 5 mg/l range, a significant downward slope in the CRP over 10 mg/l range, and a flat line in the intermediate range. CRP, C-reactive protein; sqrt TC, square root transformation of total cholesterol.

    However, a segmented linear regression requires arbitrary choices of cut-off points and is highly artificial. To produce a closer, less arbitrary and more natural model, we took as the starting point the shape of the total relationship. The pattern suggested two effects at work: a positive association between TC and CRP at low CRP levels and a negative association at higher levels. These two effects then appeared to cancel out in the mid-range. Assuming this to be the case, we would expect the gradient of the line to follow a logistic function, taking the form:Embedded Image

    Where A is the TC level at CRP=1 (ie, log CRP=0), B is the slope of the linear relationship in the non-inflammatory state, C is the slope of the linear relationship in the inflammatory state, D is the CRP level at the midpoint of the logistic transition from non-inflammatory to inflammatory state.

    This model was then fitted by non-linear least squares, resulting in an excellent fit, with normally distributed residuals, constant variance and a model deviance that was actually slightly lower than the non-parametric model, although our specific model is very close indeed to the non-parametric model (see figure 3). The details of the model fit are shown in the supplementary material (available online only).

    Figure 3
    Figure 3

    Non-linear least squares model of the relationship between sqrt TC and log CRP: sqrt TC=2.209+0.0285×log e CRP−0.108×log eCRP/(1+27.65/CRP)+N(0, 0.075). Where N(0, 0.075) describes the error of the formula; data points are normally distributed around the line with a mean of 0 and variance of 0.075. CRP, C-reactive protein; sqrt TC, square root transformation of total cholesterol.

    The final model was therefore:Embedded Image

    Modelling the relationship between HDL-cholesterol and CRP

    The scatterplot of HDL-cholesterol against log CRP showed that the relationship between the two was in some ways more straightforward than that between cholesterol and CRP, in that HDL-cholesterol was constantly decreasing as CRP increased, and the line was fairly straight. However, HDL-cholesterol was not normally distributed around its mean, and this could not be corrected with a power transformation.

    A robust linear regression was therefore carried out. This effectively deals with the problem of non-normal distribution. This is illustrated in figure 4 and the parameters of the model are shown in the supplementary material (available online only).

    Figure 4
    Figure 4

    Linear regression of the relationship between HDL-cholesterol and log CRP showed that HDL-cholesterol decreases linearly with increasing CRP. The parameters of the model are shown in the supplementary material. CRP, C-reactive protein; HDL, high-density lipoprotein.

    The relationship between CRP and TC:HDL ratio

    The graphical relationship between CRP and the TC:HDL-cholesterol ratio revealed that again there was an increase in TC:HDL-cholesterol at lower values of CRP, a very slight decrease in the ratio at higher levels of CRP, and a levelling off in the middle.

    We therefore again used non-linear regression to derive an easily interpretable four parameter model. Because of the mathematical properties of the TC:HDL-cholesterol ratio, we used its log for the regression. As the ratio was again not normally distributed about its mean, we used a robust non-linear regression. The shape of this relationship is shown in figure 5, and the details of the regression are shown in the supplementary material (available online only).

    Figure 5
    Figure 5

    Non-linear regression of the relationship between the log of the TC:HDL-cholesterol ratio and log CRP showed an increase in the ratio at low CRP and a small decrease in the ratio at higher CRP. The parameters of the model are shown in the supplementary material (available online only). CRP, C-reactive protein; HDL, high-density lipoprotein; TC, total cholesterol.

    Correcting for age and gender

    We repeated the model fitting process after correcting TC levels for age and gender with a generalised additive model. Although there was a statistically significant change in two of the model parameters, they only amounted to a 0.057 mmol/l higher intercept and a 14.3% decrease in the magnitude of the downward ‘inflammatory slope’ and are thus unlikely to be of any clinical significance.

    Discussion

    In this paper we have presented a model of the relationship between inflammation (using CRP as its most common measure) and lipids. We have shown a significant biphasic relationship between TC and CRP and a linear, inverse relationship between HDL-cholesterol and CRP. The model is built on a retrospective analysis of samples that were analysed simultaneously for CRP and lipids at four city biochemistry laboratories. They were obtained from patients in primary and secondary care for a variety of clinical indications, which are unknown to us. This is not necessarily a limitation, however, but rather ensures that the findings are widely generalisable, particularly as they pertain to the presence of a generalised inflammatory state. Similar to other studies, CRP increased with age,12 ,13 TC increased with age at young ages and decreased with age at older ages,4 ,12 ,14 and expected differences in lipid levels were seen in men and women,15 adding external validity to the findings.

    Our work shows, we believe for the first time, that CRP has a biphasic relationship with TC levels, with a positive relationship at lower levels but an inverse relationship clearly discernible once CRP levels are higher than approximately 10 mg/l. This suggests two separate effects at work: in health, a small increase in TC is associated with a small increase in CRP. This finding is theoretically consistent with studies that have found a modest increase in CRP with rising obesity levels, which are also known to increase TC.16 In inflammatory illness, however, a different mechanism appears to be gradually ‘switched on’, causing a progressive decrease in TC as CRP increases. The hypothesis that an inflammatory effect is ‘switched on’ at higher CRP levels is supported by the degree of agreement with the non-parametric model. The model also quantifies the extent to which TC falls in inflammatory illness throughout its range. For example, a healthy patient with a CRP of 3.6 mg/l and TC of 5 mmol/l who developed a major inflammatory illness leading to a CRP of 200 mg/l would be expected to have a fall of over 1.5 mmol/l in TC to 3.47 mmol/l. This is obviously a clinically important difference of which clinicians should be aware. The findings also hold relevance and potentially help to explain the increase in lipids often seen with a range of anti-inflammatory therapies.8

    Our data show that HDL-cholesterol levels falls continuously as CRP increases. This reduction in both TC and HDL-cholesterol at higher inflammatory levels has previously been alluded to,8 but not before mapped in detail. The mechanism underlying the change is not fully understood but inflammatory cytokines are thought to reduce TC synthesis and increase low-density lipoprotein receptor activity.17 There is also evidence of greater reticuloendothelial system catabolism of lipids,18 and compositional changes in lipid moieties with inflammation,17 and the net effect may be to render lipid particles more atherogenic in inflammatory conditions. However, definitive studies to test this proposition are lacking. There is a need for kinetic studies to confirm whether enhanced peripheral catabolism is the key driver towards lower lipids with higher inflammatory states. Nonetheless, our findings clearly illustrate that lipid levels are significantly altered during inflammatory states, and perhaps even irrespective of the cause, although further large studies would be needed to validate this latter suggestion.

    It has previously been recommended that the TC:HDL ratio should be used to assess CVD risk in autoimmune conditions as it is relatively stable, or perhaps even slightly increased, in inflammatory conditions.19 Our results show that although the ratio changes with inflammation (increasing with CRP to approximately 10 mg/l then plateauing and falling only very gradually at very high levels of CRP) the extent of change is far less than the relative changes in TC and HDL-cholesterol.

    Furthermore, checking whether lipid levels for those on statins are at target in those with elevated CRP (>10 mg/l) will be erroneous. Given that many groups of patients have a significant inflammatory burden, whether acute due to infections or chronic (in particular RA, but also renal failure, etc.), then our findings have clinical relevance in a number of circumstances. The results also help to understand better the apparent paradoxical associations of cholesterol with greater mortality in a number of groups known to exhibit heightened inflammatory burden, for example, RA and other autoimmune conditions,7 and haemodialysis patients.5

    The main limitation of this paper is that it addresses the relationship between CRP and lipids without considering confounding factors other than age and gender. It would have been informative also to consider factors that influence lipid levels and cardiovascular risk, including body mass index, smoking and blood pressure. However, the principal aim was to model the relationship between CRP and lipids across more than a 1000-fold variance in CRP in samples taken for various indications. In this regard, lack of clinical information is far less important than an appreciation of general trends. A proportion of patients in our population will have been on statins that lower both cholesterol, and to a lesser extent, also lower CRP.20 Therefore, statin confounding would work to attenuate the relationships we see rather than necessarily produce them. A further limitation is that we did not consider cardiovascular morbidity and mortality. Nonetheless, there are considerable strengths of our paper in that we have considered a large number of patients from both primary and secondary care and without regard to any specific cause of inflammation, thereby allowing us the power to elucidate the complex relationship between CRP levels and lipids. Furthermore, the statistical analyses were robust and comprehensive.

    In conclusion, we show that CRP has a biphasic relationship with TC, with a positive relationship at lower levels but a clinically significant inverse relationship at CRP levels higher than approximately 10 mg/l. By contrast, HDL-cholesterol levels show a continuous inverse relationship with rising CRP levels. Our data should therefore help physicians (in particular those treating patients with autoimmune conditions such as RA) to appreciate better the apparently ‘paradoxical’ but important inverse relationship between high grade systemic inflammation and lipid levels. It should also help provide a context in which to consider lipid changes with anti-inflammatory therapies better, an area of great current topicality.

    References

    1. ↵
      1. Peters MJ,
      2. Symmons DP,
      3. McCarey D,
      4. et al
      . EULAR evidence-based recommendations for cardiovascular risk management in patients with rheumatoid arthritis and other forms of inflammatory arthritis. Ann Rheum Dis 2010;69:325–31.
      OpenUrlAbstract/FREE Full Text
    2. ↵
      1. Aviña-Zubieta JA,
      2. Choi HK,
      3. Sadatsafavi M,
      4. et al
      . Risk of cardiovascular mortality in patients with rheumatoid arthritis: a meta-analysis of observational studies. Arthritis Rheum 2008;59:1690–7.
      OpenUrlCrossRefPubMedWeb of Science
    3. ↵
      Emerging Risk Factors Collaboration, Di Angelantonio E, Sarwar N. Major lipids, apolipoproteins, and risk of vascular disease. JAMA 2009;302:1993–2000.
      OpenUrlCrossRefPubMedWeb of Science
    4. ↵
      1. Schatz IJ,
      2. Masaki K,
      3. Yano K,
      4. et al
      . Cholesterol and all cause mortality in elderly people from the Honolulu Heart Program: a cohort study. Lancet 2001;358:351–5.
      OpenUrlCrossRefPubMedWeb of Science
    5. ↵
      1. Liu Y,
      2. Coresh J,
      3. Eustace JA,
      4. et al
      . Association between cholesterol level and mortality in dialysis patients: role of inflammation and malnutrition. JAMA 2004;291:451–9.
      OpenUrlCrossRefPubMedWeb of Science
    6. ↵
      1. Trompet S,
      2. Jukema JW,
      3. Katan MB,
      4. et al
      . Apolipoprotein e genotype, plasma cholesterol, and cancer: a Mendelian randomization study. Am J Epidemiol 2009;170:1415–21.
      OpenUrlAbstract/FREE Full Text
    7. ↵
      1. Myasoedova E,
      2. Crowson CS,
      3. Kremers HM,
      4. et al
      . Lipid paradox in rheumatoid arthritis: the impact of serum lipid measures and systemic inflammation on the risk of cardiovascular disease. Ann Rheum Dis 2011;70:482–7.
      OpenUrlAbstract/FREE Full Text
    8. ↵
      1. Choy E,
      2. Sattar N
      . Interpreting lipid levels in the context of high-grade inflammatory states with a focus on rheumatoid arthritis: a challenge to conventional cardiovascular risk actions. Ann Rheum Dis 2009;68:460–9.
      OpenUrlAbstract/FREE Full Text
    9. ↵
      1. Rappaport BA
      . Overview of the July 29, 2008 AAC Meeting to Discuss BLA 125276 for tocilizumab for the treatment of moderately to severely active rheumatoid arthritis (RA). FDA, 2008. http://www.fda.gov/ohrms/dockets/ac/08/briefing/2008–4371b1–01-FDA.pdf.
    10. ↵
      1. Wergeland R,
      2. Oyen O,
      3. Bentdal O,
      4. et al
      . MicroCRP: a highly sensitive CRP method applied in the monitoring of renal allograft recipients. Scand J Clin Lab Invest 1999;59:405–16.
      OpenUrlCrossRefPubMedWeb of Science
    11. ↵
      R Development Core Team (2011). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN: 3-900051-07-0. http://www.R-project.org/.
    12. ↵
      1. Bourdel-Marchasson I,
      2. Laksir H,
      3. Puget E
      . Interpreting routine biochemistry in those aged over 65 years: a time for change. Maturitas 2010;66:39–45.
      OpenUrlCrossRefPubMedWeb of Science
    13. ↵
      1. Kushner I
      . C-reactive protein elevation can be caused by conditions other than inflammation and may reflect biologic aging. Cleve Clin J Med 2001;68:535–7.
      OpenUrlAbstract/FREE Full Text
    14. ↵
      1. MacLean DR,
      2. Petrasovits A,
      3. Connelly PW,
      4. et al
      . Plasma lipids and lipoprotein reference values, and the prevalence of dyslipoproteinemia in Canadian adults. Canadian Heart Health Surveys Research Group. Can J Cardiol 1999;15:434–44.
      OpenUrlPubMedWeb of Science
    15. ↵
      1. Pilote L,
      2. Dasgupta K,
      3. Guru V,
      4. et al
      . A comprehensive view of sex-specific issues related to cardiovascular disease. Can Med Assoc J 2007;176:S1–44.
      OpenUrlAbstract/FREE Full Text
    16. ↵
      1. Hak AE,
      2. Stehouwer CD,
      3. Bots ML,
      4. et al
      . Associations of C-reactive protein with measures of obesity, insulin resistance, and subclinical atherosclerosis in healthy, middle-aged women. Arterioscler Thromb Vasc Biol 1999;19:1986–91.
      OpenUrlAbstract/FREE Full Text
    17. ↵
      1. Khovidhunkit W,
      2. Kim MS,
      3. Memon RA,
      4. et al
      . Effects of infection and inflammation on lipid and lipoprotein metabolism: mechanisms and consequences to the host. J Lipid Res 2004;45:1169–96.
      OpenUrlAbstract/FREE Full Text
    18. ↵
      1. Packard CJ,
      2. Slater HR,
      3. Shepherd J
      . The reticuloendothelial system and low density lipoprotein metabolism in the rabbit. Biochim Biophys Acta 1982;712:412–19.
      OpenUrlCrossRefPubMedWeb of Science
    19. ↵
      1. Peters MJ,
      2. Voskuyl AE,
      3. Sattar N,
      4. et al
      . The interplay between inflammation, lipids and cardiovascular risk in rheumatoid arthritis: why ratios may be better. Int J Clin Pract 2010;64:1440–3.
      OpenUrlCrossRefPubMedWeb of Science
    20. ↵
      1. Albert MA,
      2. Danielson E,
      3. Rifai N,
      4. et al
      . RINCE Investigators. Effect of statin therapy on C-reactive protein levels: the pravastatin inflammation/CRP evaluation (PRINCE): a randomized trial and cohort study. JAMA 2001;286:64–70.
      OpenUrlCrossRefPubMedWeb of Science

    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:

    Footnotes

    • Handling editor Tore K Kvien

    • Contributors HJ collated data, performed descriptive statistical analysis, drafted and revised the article. MP designed the project and revised the article. AC performed statistical analysis in R and drafted the article. NS interpreted data and revised the article. All authors have approved the final version.

    • Competing interests NS has consulted for or been on speakers bureau for Roche, Astrazeneca, UCB and MSD. HJ, MP and AC have no competing interests.

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

    • Data sharing statement Full data set available on request from corresponding author.