6th October 2017, Volume 130 Number 1463

Thomas Robinson, Rod Jackson, Susan Wells, Andrew Kerr, Roger Marshall

Many international guidelines and national health policies for the prevention of cardiovascular disease (CVD) advocate using estimates of patients’ absolute CVD risk, generated from multifactor risk prediction equations, to inform management decisions.1–6 However, there remains uncertainty about how clinicians use risk assessment scores when they are available.

In New Zealand, treatment based on CVD risk assessment scores was first advocated in 1993.7 The New Zealand Guidelines Group introduced guidelines for risk assessment and management on the basis of risk in 2003, and these have been widely disseminated and updated regularly.5,8 Initially, dissemination was via paper charts, but a web-based assessment and decision support tool (PREDICT) was integrated into some general practitioners’ electronic practice management systems from 2002.9 For primary prevention of CVD, guidelines recommend treatment should be informed primarily by a patient’s estimated five-year CVD risk, calculated using a modified Framingham Heart Study risk equation. The risk factors included in the equation are: age, gender, blood pressure, smoking and diabetes status, and Total/HDL cholesterol ratio. The score is then adjusted to take into account several factors, including ethnic differences in risk that occur in New Zealand, and a family history of premature ischaemic CVD (the ‘New Zealand-adjusted Framingham risk score’).

Although uptake of CVD risk assessment was initially slow, it has been strongly supported by health care organisations at all levels, and CVD risk assessment is one of New Zealand’s six national health targets. By February 2016, 90% of New Zealanders in the target groups had received a CVD risk assessment within the last five years.10 New Zealand therefore provides the opportunity to study CVD risk assessment in a ‘mature’ programme.

The aim of this study was to investigate how primary care clinicians use CVD risk scores to inform their treatment decisions. Specifically, we tested the hypotheses that clinicians make decisions on initiating statins based on one, or some, of the following factors:

  1. Whether the patient falls above or below the two five-year estimated CVD risk thresholds recommended by guidelines.
  2. Absolute CVD risk, but used as a continuous variable.
  3. Single risk factors relevant to the medication (ie, lipid measurements).

We hope that this information can inform future policy decisions about how to prevent CVD.

Methods

PREDICT-CVD is the most commonly used CVD risk assessment tool in New Zealand and is available in 35–40% of primary care practices. To calculate estimated five-year CVD risk, clinicians open a CVD risk profile online form that auto-populates with available data from the patient’s electronic record. The doctor or nurse completes the risk profile, which is then sent securely to a central server. The patient’s estimated New Zealand-adjusted Framingham five-year CVD risk score is immediately returned together with, if requested, guideline-based risk management recommendations. Whenever PREDICT is used, an electronic risk profile is stored anonymously. With the permission of health providers, this profile is linked to an encrypted National Health Index number (NHI) and made available to University of Auckland researchers.9 Important recommendations at the time of this study were that people with estimated CVD risk of 15% or more over five years should have lifestyle management, but should be initiated on medications if there were no improvements after three months, and people at a 20% or more risk should normally be initiated on medications at the same time as lifestyle management.

Patients were included in these analyses if they were first risk assessed using PREDICT between 1 July 2007 and 30 June 2011, and met New Zealand Guideline recommendations for risk assessment. We therefore included Māori, Pacific and Indian men aged 35 years and older, other men and Māori, Pacific and Indian women aged 45 years and older, and other women aged 55 years and older.5 Ethnicity was determined from PREDICT and hospital records and, if more than one ethnicity was recorded, it was prioritised in the following order: Māori, Pacific, Indian and New Zealand European/other.

We excluded patients for whom treatment decisions are not recommended on the basis of the estimated CVD risk. This included people with prior CVD, others defined to be at high clinical risk (ie, people with known genetic lipid disorders, people with diabetic nephropathy), and patients with a single high-risk factor (ie, blood pressure consistently ≥170/100mmHg, total cholesterol ≥8.0mmol/L or a total cholesterol to HDL cholesterol ratio ≥8).5 Prior CVD was defined as a history of angina or myocardial infarction, stroke, transient ischaemic attack, peripheral vascular disease, percutaneous coronary intervention or coronary artery bypass graft reported by the clinician at the time of the risk assessment. Finally, because we wished to study statin initiation, we exclude patients who had been dispensed a statin in the six months prior to the time of CVD risk assessment.

We used a simple dichotomous outcome; whether the patient was recorded as being dispensed a statin in the six months following the CVD risk assessment. Dispensing was identified by anonymously linking the PREDICT database to the national Pharmaceutical Collection (PHARMS), using encrypted NHI numbers. PHARMS is a data warehouse that is jointly administered by the Ministry of Health and the Pharmaceutical Management Agency, and collects data on government-subsidised medications dispensed by community pharmacies. In 2009, 96% of dispensing episodes were reliably identifiable by NHI numbers.11

To explore the relationship between estimated CVD risk and subsequent statin prescribing we plotted the proportion of patients dispensed a statin in the six months following a CVD risk assessment against levels of estimated CVD risk. If increasing CVD risk were impacting on statin prescribing we would expect to see some form of continuous relationship. If the guideline thresholds were having an important effect on clinician treatment decisions, we expected a sudden jump (or discontinuity) in the proportion of patients dispensed a statin at the thresholds. We explored whether this was the case using a regression discontinuity design. This study design is useful for deciding whether an outcome is caused by an intervention when that intervention is assigned according to a threshold in a continuous assignment variable.12,13 In this study, the outcome was statin dispensing, the assignment variable was estimated CVD risk and the thresholds were the guideline 15% or 20% five-year CVD risk cut-offs. We plotted the proportion dispensed statins using a localised polynomial plot. The statistical significance of any discontinuity can be tested using either parametric regression models or non-parametric (local linear) regression models.14 Although regression discontinuity has good internal validity, it is susceptible to misspecification of the modelling of the relationship between the assignment variable and the outcome, so we compared a number of models for consistency.15,16 For the non-parametric models we used a localised linear regression Stata module by Austin.17 In the parametric analyses we tested logistic regression models with CVD risk scores included as linear, quadratic and cubic terms, and with interaction with a threshold dummy variable.

To gain an understanding of the other two hypothesised clinician decision making processes (consideration of absolute CVD estimated risk or a focus on single risk factors), we used a combination of graphical methods and logistic regression modelling to examine which factors, or combinations of factors, provided the best explanation of statin dispensing. We again show localised polynomial plots of proportions of people initiated on statins against estimated CVD risk scores, but with the total population split into two subpopulations by binary categorisation of different risk factors. If clinicians were only using the estimated New Zealand-adjusted Framingham risk score to determine treatment, we would not expect these other risk factors to provide additional information on dispensing, and therefore the plots of the two subpopulations would be superimposed.

We also developed four logistic regression models to see which best predicted dispensing using i) New Zealand-adjusted Framingham risk score only, ii) TC/HDL cholesterol ratio only, iii) both New Zealand-adjusted Framingham score and TC/HDL cholesterol ratio, and iv) all the individual risk factors that make up the New Zealand-adjusted Framingham risk score. For each of the four final models we tested a number of different specifications of model to obtain the best model in each case.

In all analyses the best parametric model was chosen by beginning with models, which included CVD risk as a linear, quadratic and cubic power and dropping terms which were not significant (backward elimination). We also compared models using Aikake Information Criteria, area under the curve statistics, and by comparing pairs of models using likelihood ratio tests.

All analyses were done using the Stata 13.1 statistical package (StataCorp, College Station, TX, USA).

The cohort study and research process was approved by the Northern Region Ethics Committee Y in 2003 (AKY/03/12/314) with subsequent annual approval by the National Multi Region Ethics Committee since 2007 (MEC/07/19/EXP).

Results

There were 162,518 people with a first PREDICT CVD risk assessment between 1 July 2007 and 30 June 2011. 14,661 of these patients were 75 years and older and 19,521 were younger than the guideline target groups, leaving 126,336 people. Of these, 30,233 had a reason not to be managed primarily on their estimated CVD risk (18,575 had prior CVD, 3,551 had a genetic lipid disorder or diabetic nephropathy, and 10,106 had a single high-risk factor). Of the remaining 96,103 participants, 76,571 (80%) were not on a statin at baseline, and these were included in the study.

The characteristics of the study group are given in Table 1. Nine percent of patients had an estimated five-year CVD risk of 15% or greater, and 3% a risk score 20% or greater. The majority of participants were European, but there were significant numbers of Māori, Pacific, Indian and other Asians.

Table 1: Characteristics of the included population at baseline risk assessment.

Category

n

Percent of total (76,517)

Age (years) 

35–44

8,146

11%

45–54

26,827

35%

55–64

27,678

36%

65–74

13,920

18%

Sex 

Male

46,185

60%

Female

30,386

40%

Ethnicity 

Māori

12,325

16%

Pacific

12,737

17%

Indian

6,672

9%

Other Asian

3,950

5%

European & others

39,685

52%

Smoking 

Non-smoker

63,741

83%

Smoker or quit within 12 months

12,830

17%

Diabetes status

No diabetes

69,792

91%

Diabetes

6,779

9%

Blood pressure 

Less than 140/90mmHg

56,630

74%

Systolic ≥140 or diastolic ≥90mmHg

19,941

26%

Total/HDL cholesterol ratio 

Less than 5

58,643

77%

5 or above

17,928

23%

5-year CVD risk 

<5 %

18,366

24%

5–9.9%

36,925

48%

10–14.9%

13,970

18%

15–19.9%

4,738

6%

20% and above

2,572

3%

Total

76,571

100% 

Figure 1 shows the relationship between the New Zealand-adjusted Framingham CVD risk score and statin dispensing in the six months after the CVD risk assessment. Statin use increases with estimated CVD risk, however, but even among the highest risk patients the majority are not treated. There is no apparent sudden stepwise increase initiation of a statin at either 15% or 20% CVD risk thresholds, as might be expected from guideline recommendations.

Figure 1: Probability of being dispensed a statin in the six months after PREDICT risk assessment (total post) for study participants (ie, statin initiation), against estimated CVD risk.

c

Vertical lines are at the two guideline thresholds.  

Regression discontinuity statistical analyses confirmed this impression, although small changes cannot be discounted. Table 2 presents the best parametric and non-parametric models for the two thresholds. For the non-parametric models we found the null effect (ie, no discontinuity) was consistent across models using a wide range of bandwidths (25% to 300% of optimal bandwidth). We fitted parametric models with the two thresholds either included singly or together, and with the relationship between CVD risk and statin prescribing modelled in a number of ways. The best fitting models used CVD risk in either a quadratic or cubic form, but the finding of null effect was not sensitive to any of the models used.

Table 2: Regression discontinuity analyses of the impact of 15% and 20% risk thresholds on statin initiation in the six months after the risk assessment for those not on a statin at assessment.

 

Coef.

Std. Err

95% CI

P value

Lower

Upper

15% threshold

Non-parametric

0.025

0.025

-0.024

0.073

0.314

Parametric (Cubic)

0.107

0.062

-0.014

0.228

0.082

20% threshold

Non-parametric

0.014

0.041

-0.066

0.095

0.731

Parametric (quadratic)

-0.149

0.077

-0.300

0.002

0.053

15% & 20% thresholds (Parametric - cubic)

15% threshold

0.096

0.064

-0.030

0.223

0.135

20% threshold

-0.052

0.088

-0.223

0.120

0.555 

Figure 2 explores whether important risk factors for CVD affect dispensing over and above being included in the CVD risk equations. Local polynomial plots of the risk score against proportion of people initiated on a statin within six months of assessment are shown for people with and without individual risk factors. People with an elevated total/HDL cholesterol ratio are more likely to have a statin initiated at all levels of CVD risk score. People with diabetes are also more likely to have a statin initiated at all levels of the CVD risk score, but particularly so at lower levels of risk. Of other variables examined there is a small positive effect from having Indian ethnicity, and a small negative effect from being a smoker. The blood pressure level and gender had little effect on statin initiation, other than from being included in the New Zealand-adjusted Framingham risk score.

Figure 2: Plots of probability of statin initiation against CVD risk by category of other risk factors. 

c 

Four different logistic regression models predicting the initiation of statins were developed, as described in the methods section (Appendix). Table 3 shows the model fit measures for the different models. The New Zealand-adjusted Framingham risk score as sole predictor was best fitted with a quadratic model and gave a moderate prediction (AUC of 0.725). This was however, higher than the best model using total/HDL cholesterol ratio as a sole predictor (AUC 0.682). Total/HDL cholesterol ratio was found to be a better predictor than other lipid variables (Total cholesterol and LDL cholesterol, not shown).

Table 3: Fit characteristics for different logistic regression models for predicting statin initiation.

Model

n

Df

AIC

BIC

AUC

New Zealand-adjusted score

76,571

3

39,626

39,654

0.725

TC/HDL ratio

76,571

4

41,078

41,115

0.682

Combined New Zealand-adjusted score & Lipids

76,571

5

38,698

38,744

0.749

Risk factors individually

76,571

10

38,239

38,331

0.767

Abbreviations: Df—degrees of freedom; AICAkaike Information Criteria; BICBayesian Information Criteria, AUCArea Under the Curve. 

The best model we could fit for predicting statin initiation used all the individual risk factors that are included in the New Zealand-adjusted Framingham risk equation (but did not include the CVD risk score). This had better discrimination and measures of fit than did the New Zealand Framingham risk equation model.

Conclusion

We have used a large observational cohort to try to gain an understanding of primary care clinicians’ decisions to prescribe statins after a first CVD risk assessment, estimated using the New Zealand-adjusted Framingham risk score. We believe that while these processes are clearly complex, we can make some statements about the hypotheses stated in the introduction. Firstly, clinicians appear to be using some estimation of patients’ overall CVD risk in preference to only lipid levels to decide on statin treatment. The models that include multiple risk factors, either individually or as a summary CVD risk score, predict new statin dispensing better than the model with only total/HDL cholesterol ratio.

It is noteworthy that statin dispensing appears to be influenced by several specific risk factors over and above their contribution to the risk score calculation. Individual risk factors, particularly higher TC/HDL cholesterol ratios, having diabetes, and Indian ethnicity are associated with increased likelihood of treatment at all levels of the risk score. The finding that smokers are less likely to be initiated on statins than non-smokers is unexpected. However, it is possible that clinicians, in discussion with their patients, initially focus on smoking cessation to lower risk, and this delays statin initiation. It is not possible to be certain from this whether clinicians are using the risk score and adjusting it according to their interpretation of risk and discussions with their patients, or are simply making their own assessment of all the individual risk factors. However, it seems unlikely that clinicians are going to the trouble of calculating CVD risk and then not using the information.

Finally, there is no evidence that guideline thresholds are substantially impacting on treatment. Rather, clinicians appear to be using the guiding principle of national guidelines, that the intensity of risk management should be proportional to the estimated CVD risk, rather than the recommended thresholds, to inform their decisions.

There is a limited international literature on how clinicians use CVD risk scores in practice. A number of studies have used CVD risk assessments presented to physicians as clinical vignettes. Some of these have found that treatment is more influenced by individual risk factors than by estimated CVD risk scores.18–20 However, others suggest that the provision of CVD risk scores leads to more guideline concordant prescribing of medications.21,22 In contrast to our study, a recently published Australian observational study found that there was little evidence of general practitioners using CVD risk scores to determine when to prescribe lipid lowering medications.23 This difference may reflect the fact that Australian pharmaceutical subsidies are available on the basis of single risk factors and New Zealand’s long and strong policy support for CVD risk assessment. It seems very likely that clinician use of CVD risk scores will be highly dependent on local the health system and professional context.

This study has a number of potential weaknesses. Firstly, we use dispensing of statins as our sole treatment outcome. For some patients the ‘treatment’ may have been lifestyle changes including dietary change or a smoking cessation intervention, which we have not measured. So it is possible that a more complete assessment of treatment initiated by risk assessment would show different patterns. Following patients for 9 or 12 months may have allowed more time for statins to be dispensed and shown a different pattern. However, a previous study of these patients have shown much smaller changes in proportions dispensed statins after six months.11 Secondly, the outcome measured is dispensing, which is determined by both clinician and patient behaviour, and we cannot determine the role of each. The higher rates of dispensing of statins to people of Indian ethnicity, for example, may be due to higher rates of adherence. Certainly as an outcome measure of clinician behaviour the use of dispensing data leads to a degree of outcome misclassification. However, a New Zealand study found that 92% of people first prescribed simvastatin had it dispensed within seven days and therefore misclassification is likely to be small.24 Thirdly, the methods used for testing hypothesised causal links are observational rather than experimental. While it is possible to design experiments where clinicians were randomly given access to different sets of clinical information, it is difficult to see how they would be feasible or ethical. Nevertheless, observational methods are always susceptible to internal validity problems. Given the wide range of information available to clinicians for decision making, it could be argued that this process is too complex to be deciphered using even complete and objective observational data.

Finally, the PREDICT data has inherent limitations for exploring how clinicians and patients assess CVD risk and subsequently make treatment decisions. Firstly, data is only available on decisions where a PREDICT risk assessment is made. It would have been very useful to have a comparison group available where similar data on individual risk factors and outcomes was available, but an absolute CVD risk assessment had not been estimated. A much richer set of information on factors that may have influenced decision making would also have been ideal. Despite these limitations PREDICT does provide a very rich set of data for studying the management of CVD risk.

The conclusions of the study rely heavily on complex statistical models that are not easily confirmable by the reader. We hope that the graphical information will provide a more intuitive approach that supports the conclusions provided by the modelling.

The study also has several important strengths. Firstly, it is a study of primary care clinicians’ use of absolute CVD risk assessment in a country where this approach to managing primary CVD prevention has been best practice for a considerable time and has strong organisation support. Secondly, it reports on actual behaviour rather than hypothetical behaviour as in clinical vignette studies,18–22 and may have stronger generalisability. Thirdly, it is a large study including over 76,000 first CVD first assessments, and therefore has sufficient power to investigate hypothesised decision-making processes. Furthermore, the PREDICT tool has facilitated nearly complete (99%) valid risk factor data collection for key variables due to compulsory fields required to calculate CVD risk and built-in range and validity checks at the point of data entry.

We believe that this study provides some useful lessons for organisations implementing absolute CVD risk assessment programmes. While a threshold may provide useful guidance, the reality for clinicians is a complex array of information on a patient that will modify decisions based upon CVD risk. Treatment decisions are made in consultation with a patient, who will bring other information and values to the decision. In New Zealand, the next version of the CVD risk management guidelines are being developed and it is likely that the current thresholds will be replaced by risk score ranges within which treatment should be discussed with patients based upon their individual circumstances. This is perhaps a better reflection of clinical reality and is consistent with the principle underpinning the New Zealand risk-based guidelines, which has always been that the intensity of interventions should be proportional to the estimated CVD risk.

Appendix

 

Model

NZ-adjusted risk score

TC/HDL ratio

Combined risk score and lipids

Risk factors individually

Variables

CVD risk estimate

0.274

 

0.234

 

CVD risk estimate squared

-0.005

 

-0.004

 

TC/HDL ratio

 

-0.433

0.547

0.928

TC/HDL ratio squared

 

0.253

-0.197*

-0.036

TC/HDL ratio cubed

 

-0.020

 

 

Age

 

 

 

0.125

Age squared

 

 

 

-0.001

Female

 

 

 

-0.099*

Indian

 

 

 

0.746

Systolic BP

 

 

 

0.017

Diabetes

 

 

 

1.329

Diabetes *CVD risk estimate

Smoker

 

 

 

0.358

Constant

-4.500

-3.688

-6.206

-12.891

Fit characteristics 

AUC

0.725

0.682

0.749

0.767

AIC

39,626

41,078

38,698

38,239

BIC

39,654

4,115

38,744

38,331

Table logistic regression models of statin initiation in the six months after risk assessment.
*P<0.05, otherwise P<0.001. 

Summary

Ninety percent of New Zealanders who are in the NZ Guidelines target group have had a CVD risk assessment. There is little information on how this information is used by doctors and their patients. The decision to start cholesterol lowering medications (statins) might be driven mainly by cholesterol levels, or (as NZ Guidelines suggest) by a comprehensive assessment of the person’s overall risk. This paper suggests that the latter approach is more likely, but that the decision process is complex and takes into account other factors.

Abstract

Aim

Cardiovascular disease (CVD) risk assessment is commonly recommended in guidelines, but there is uncertainty about how clinicians use this information. Our objective was to understand how New Zealand primary care clinicians use CVD risk assessment estimates to inform new statin prescribing.

Method

We used a cohort of patients seen in primary care who have had a CVD risk estimated on the basis of a New Zealand modified Framingham risk equation. These patients were linked to national pharmaceutical dispensing records to determine new statin use in the following six months. Regression discontinuity and logistic regression analysis, and graphical approaches, were used to explore associations between estimated CVD risk and primary clinicians’ decisions to initiate statin treatment.

Results

There were 76,571 patients aged 35 to 75 who were not on a statin, had a first recorded CVD risk assessment between July 2007 and June 2011, and for whom national guidelines recommended management on the basis of estimated CVD risk. Statin dispensing increased with increasing CVD risk. There was no evidence of sudden jumps in the proportions of patients dispensed statins at guideline recommended treatment threshold values of 15% and 20% CVD risk (P=0.314 and 0.731). A logistic regression model using the CVD risk score predicted statin initiation better than models using lipid measures (Area Under the Curve 0.725 versus 0.682). However, further modelling and graphical analysis suggested clinicians were using a range of other information to inform the initiation of statins.

Conclusion

New Zealand primary care clinicians’ statin prescribing decisions appear to be influenced by patients’ predicted CVD risk. However, other factors are associated with increased statin dispensing independent of CVD risk score.

Author Information

Thomas Robinson, Health Systems, School of Population Health, University of Auckland, Auckland;
Rod Jackson, Epidemiology and Biostatistics, School of Population Health, University of Auckland, Auckland; Susan Wells, Section of Epidemiology and Biostatistics, School of Population Health, University of Auckland, Auckland; Andrew Kerr, Cardiology, Middlemore Hospital, Auckland;
Roger Marshall, Epidemiology and Biostatistics, University of Auckland, Auckland.

Acknowledgements

The authors would also like to thank members of the PREDICT Māori Advisory Group and the PREDICT Pacific Advisory Group, as well as general practitioners, practice nurses and patients affiliated with contributing primary health organisations (PHOs). PREDICT was developed by a collaboration of clinical epidemiologists at the University of Auckland, IT specialists at Enigma Publishing (a private provider of online health knowledge systems), primary health care organisations, non-governmental organisations (New Zealand Guidelines Group, National Heart Foundation, Diabetes New Zealand, Diabetes Auckland), several district health boards and the Ministry of Health. The PREDICT software platform is owned by Enigma Publishing (PREDICT is a trademark of Enigma Publishing).

Correspondence

Dr Thomas Robinson, Health Systems, School of Population Health, University of Auckland, 261 Morrin Rd, Auckland 1072.

Correspondence Email

tomrobnz@gmail.com

Competing Interests

Dr Jackson reports grants from Health Research Council of NZ during the conduct of the study. Dr Wells reports grants from Health Research Council of New Zealand and The Stevenson Foundation during the conduct of the study; grants from Roche Diagnostics Ltd, and National Heart Foundation of New Zealand outside the submitted work. Dr Kerr reports grants from HRC during the conduct of the study.

References

  1. JBS Board. Joint British Societies’ consensus recommendations for the prevention of cardiovascular disease (JBS3). Heart. 2014; 100 Suppl 2:ii1-ii67.
  2. Ferket BS, Colkesen EB, Visser JJ, et al. Systematic review of guidelines on cardiovascular risk assessment: Which recommendations should clinicians follow for a cardiovascular health check? Arch Intern Med. 2010; 170(1):27–40.
  3. Perk J, De Backer G, Gohlke H, et al. European Guidelines on cardiovascular disease prevention in clinical practice (version 2012). The Fifth Joint Task Force of the European Society of Cardiology and Other Societies on Cardiovascular Disease Prevention in Clinical Practice (constituted by representatives of nine societies and by invited experts). Eur Heart J. 2012; 33(13):1635–701.
  4. Stone NJ, Robinson JG, Lichtenstein AH, et al. 2013 ACC/AHA guideline on the treatment of blood cholesterol to reduce atherosclerotic cardiovascular risk in adults: a report of the American College of Cardiology/American Heart Association Task Force on Practice Guidelines. Circulation. 2014; 129(25 Suppl 2):S1–45.
  5. New Zealand Guidelines Group. New Zealand Primary Care Handbook 2012. 3rd ed. Wellington: New Zealand Guidelines Group; 2012.
  6. National Vascular Disease Prevention Alliance. Guidelines for the management of Absolute cardiovascular disease risk. Melbourne: National Stroke Foundation; 2012.
  7. Jackson R, Barham P, Bills J, et al. Management of raised blood pressure in New Zealand: a discussion document. BMJ. 1993; 307(6896):107–10.
  8. New Zealand Guidelines Group. Evidence-based best-practice guideline. The assessment and management of cardiovascular risk. Wellington: New Zealand Guidelines Group; 2003.
  9. Wells S, Riddell T, Kerr A, et al. Cohort Profile: The PREDICT Cardiovascular Disease Cohort in New Zealand Primary Care (PREDICT-CVD 19). Int J Epidemiol. 2015.
  10. Ministry of Health. Health Targets. How is my DHB performing? – 2015/16 Wellington: Ministry of Health; 2014 [Available from: http://www.health.govt.nz/new-zealand-health-system/health-targets/how-my-dhb-performing/how-my-dhb-performing-2015-16/health-targets-2015-16-quarter-2-results-summary
  11. Mehta S, Wells S, Grey C, et al. Initiation and maintenance of cardiovascular medications following cardiovascular risk assessment in a large primary care cohort: PREDICT CVD-16. Eur J Prev Cardiol. 2014; 21(2):192–202.
  12. O’Keeffe AG, Geneletti S, Baio G, et al. Regression discontinuity designs: an approach to the evaluation of treatment efficacy in primary care using observational data. Bmj. 2014; 349:g5293.
  13. Linden A, Adams JL, Roberts N. Evaluating disease management programme effectiveness: an introduction to the regression discontinuity design. Journal of evaluation in clinical practice. 2006; 12(2):124–31.
  14. Nichols A. Causal inference with observational data. Stata Journal. 2007; 7(4):507–41.
  15. Shadish WR, Cook TD, Campbell DT. Experimental and quasi-experimental designs for generalized causal inference. Boston: Houghton Mifflin; 2002. xxi, 623 p. p.
  16. Murnane RJ, Willett JB. Methods matter : improving causal inference in educational and social science research. Oxford ; New York: Oxford University Press; 2011. xv, 397 p. p.
  17. Nichols A. rd 2.0: Revised Stata module for regression discontinuity estimation. Washington: Urban Institute; 2011.
  18. Weiner M, Wells S, Kerse N. Perspectives of general practitioners towards evaluation and treatment of cardiovascular diseases among older people. J Prim Health Care. 2009; 1(3):198–206.
  19. Johansen ME, Gold KJ, Sen A, et al. A national survey of the treatment of hyperlipidemia in primary prevention. JAMA Intern Med. 2013; 173(7):586–8; discussion 8.
  20. Jansen J, Bonner C, McKinn S, et al. General practitioners’ use of absolute risk versus individual risk factors in cardiovascular disease prevention: an experimental study. BMJ Open. 2014; 4(5):e004812.
  21. Persell SD, Zei C, Cameron KA, et al. Potential use of 10-year and lifetime coronary risk information for preventive cardiology prescribing decisions: a primary care physician survey. Arch Intern Med. 2010; 170(5):470–7.
  22. Sekaran NK, Sussman JB, Xu A, Hayward RA. Providing clinicians with a patient’s 10-year cardiovascular risk improves their statin prescribing: a true experiment using clinical vignettes. BMC Cardiovasc Disord. 2013; 13:90.
  23. Schilling C, Mortimer D, Dalziel K, et al. Using Classification and Regression Trees (CART) to Identify Prescribing Thresholds for Cardiovascular Disease. Pharmacoeconomics. 2016; 34(2):195-205.
  24. Mabotuwana T, Warren J, Harrison J, Kenealy T. What can primary care prescribing data tell us about individual adherence to long-term medication?-comparison to pharmacy dispensing data. Pharmacoepidemiol Drug Saf. 2009; 18(10):956–64.