J Clin Epidemiol Vol. 52, No. 1, pp. 49–55, 1999Copyright © 1999 Elsevier Science Inc. All rights reserved.

0895-4356/99/ $–see front matterPII S0895-4356(98)00139-5

Identifying Adults at Low Risk for Significant Hyperlipidemia: A Validated Clinical Index

Steven A. Grover,

1,2,3,4,5,

* Carey Levinton,

1

and Steeve Paquet

1

1

Centre for the Analysis of Cost-Effective Care, and

2

the Divisions of General Internal Medicine and

3

Clinical Epidemiology, The Montreal General Hospital,

4

Departments of Medicine and

5

Epidemiologyand Biostatistics, McGill University, Montreal, Quebec, Canada

ABSTRACT.

The objective of this study was to develop and validate a simple clinical index to identifyindividuals at increased risk of an elevated CHL/HDL ratio. Using recursive partitioning, factors associated withan elevated CHL/HDL ratio were identified among 1993 men and 1631 women in the Lipid Research ClinicPrevalence Study. These factors were weighted using logistic regression analyses to develop a clinical index thatwas validated on 486 men and 484 women reported in the Santé Québec cardiovascular health survey. A highCHL/HDL ratio was defined as

$

5 for women and

$

6 for men which approximates the 75th percentiles reportedin the second United States National Health and Nutrition Survey. In the Lipid Research Clinics cohort, 307men (15.4%) and 188 women (11.5%) had an elevated CHL/HDL ratio. Using separate clinical indices for menand women, significant variables included body mass index, alcohol consumption, age, smoking status, systolicblood pressure, physical activity status, and the presence of diabetes, the study identified 88% of the men and82% of the women with elevated ratios. External validation using the Santé Québec data set demonstrated testsensitivities of 81% for men and 94% for women. Overall, 12% of those with a high CHL/HDL ratio weremisclassified as low risk. The ratio of total plasma cholesterol to HDL cholesterol has been shown to be one of thebest lipid predictors of increased coronary risk. Readily available clinical data can be used to identify 88% ofthose individuals most likely to benefit from lipid screening while obviating the need for such screening in one

quarter of otherwise healthy adults.

J

CLIN

EPIDEMIOL

52;1:49–55, 1999. © 1999 Elsevier Science Inc.

KEY WORDS.

Hyperlipidemia, clinical index, coronary disease, risk prediction

INTRODUCTION

Identifying and treating individuals with risk factors forcoronary heart disease (CHD), such as hypertension or hy-perlipidemia, is increasingly recognized as an importantCHD prevention strategy. While the cost of measuringblood pressure is negligible when performed during routineoffice examinations, serum lipid measurements require sub-stantial health service resources if all adults are to be testedaccording to current expert guidelines [1,2].

Garber

et al.

projected that in the United States, by theyear 1995, the cost to screen over 33 million people aged 65and older would amount to between 50 and 62.2 milliondollars annually [3]. Similarly, Grover

et al.

estimated thatthe initial costs to implement a nationwide cholesterolscreening program among Canadians would be 432 to 560million dollars [4].

Given the increasing concerns surrounding limited healthcare dollars, it would be useful to identify a more selective

screening strategy that obviates the need for lipid testingamong those individuals who are unlikely to require inter-vention. To this end, we have developed a clinical algo-rithm for identifying individuals at low risk of significantserum lipid abnormalities.

Among commonly available serum lipid measurements,the ratio of cholesterol to high density lipoprotein choles-terol (CHL/HDL) has been found to be one of the most ac-curate predictors of future coronary events [5–10]. Accord-ingly, we have used recursive partitioning techniques ontwo representative adult populations to develop and vali-date a simple clinical decision rule for identifying individu-als most likely to have elevated CHL/HDL levels [11,12].The results of these analyses demonstrate that it is feasibleto identify a substantial portion of the North Americanpopulation as being at low risk of an elevated CHL/HDL ra-tio thereby restricting lipid testing and reducing the cost ofscreening.

METHODS

We used the Lipid Research Clinics (LRC) Program Preva-lence Study Cohort to develop the prediction model. The

*Address for correspondence: Dr. Steven A. Grover, Centre for the Anal-ysis of Cost-Effective Care, Montreal General Hospital, 1650 Cedar Ave-nue, Montreal, Quebec, H3G 1A4 Canada.

Accepted for publication on 1 September, 1998.

50

S. A. Grover

et al.

methods and design of the LRC study have been reportedpreviously in detail [13–16]. Specifically, we used the 15%random subsample, including 2510 men and 2316 womenaged 30–74, to support the generalizability of our results.

After deleting those with incomplete clinical and labora-tory data (345 men and 524 women) there remained 2165men and 1792 women. We excluded those who had diag-nosed cardiovascular disease (CVD) including CHD, cere-brovascular disease, or peripheral vascular disease at entry.Among those free of CVD, we also excluded individualstaking prescribed medication for hyperlipidemia and preg-nant individuals. This resulted in 1993 men (92.0%) and1631 women (91.0%) being included in the analysis.

In the LRC study, all men and women were followed upprospectively to provide data on subsequent mortality.Telephone or mail contact began annually in July 1977,and individuals were followed up through June 1987, for anaverage follow-up of 12.4 years. Specific causes of mortalitywere ascertained by death certificate and hospital records,and the vital status of 99% of the participants was estab-lished at least once during the follow-up period [17]. AnyLRC-defined definite or suspected coronary death was in-cluded in the present analyses.

Recursive Partitioning

Men and women were analyzed separately. A high CHL/HDL ratio was defined as

$

5.0 for women and

$

6.0 formen which corresponds approximately to the 75th percen-tile in U.S. population reported in the second NationalHealth and Nutrition Survey (NHANES II) [18].

Beginning with the complete cohort (root node), the pop-ulation was partitioned into two new populations (nodes)based on a splitting rule which maximized the difference inthe proportion of cases with a high CHL/HDL ratio. Eachcandidate variable in turn was examined at all possible cut-points. That cutpoint which produced the “best” split wasthen used to subdivide the root node into two new popula-tions. Each new split created a path or branch from the par-ent node (the population that was partitioned) to its siblingnodes (the populations created from the split). The newlydefined populations were then partitioned using the samecriteria—hence the recursive process. The procedure con-tinued until all subpopulations could no longer be subdi-vided according to a predefined stopping rule. A node orsubpopulation that could no longer be partitioned was la-belled a terminal node. In our model we required all sub-populations to maintain at least 25 individuals.

Once the model had been completed, we eliminated anyspurious population splits by pruning. We pruned the treefrom the bottom up. At each node where the populationhad been subdivided, we tested the null hypothesis thatthere was no difference in high CHL/HDL prevalence be-tween the two populations (P

,

0.01). If the null hypothe-sis was satisfied the branches were recombined to form a

single population and the process was then repeated untilall nodes had been tested.

Finally, we combined different terminal nodes (subpopu-lations) of the tree into classes as defined by an

a priori

clini-cal criterion. In this analysis, we classified as low risk anyterminal node where the subpopulation had at least a three-fold lower risk of an elevated CHL/HDL ratio relative to thetotal population. All other nodes were classified as high risk.

Clinical Prediction Index

To develop a clinical index for predicting a low CHL/HDLlevel, we used logistic regression analysis to determine therelative weights associated with each predictor variableused in the final model. Using a low CHL/HDL ratio as thedependent variable, we forced all significant binary splitsinto the model. An odds ratio for each factor was then cal-culated from its

b

coefficient in the logistic model. Theseodds ratios provided weighted “points” for each of the riskfactors such that a risk index could be created by summingup all the points present in any one individual.

Validation

To validate the recursive models and clinical indices devel-oped from LRC data, we applied the final results to a sec-ond, independent sample of adults unrelated to the LRCcohort, using the Santé Québec data set [19]. This cross-sectional survey of the Quebec population was undertakenin 1990 to characterize the prevalence of various risk fac-tors for CVD. Among 676 men and 716 women with com-plete data, 486 (72%) and 484 (68%), respectively, were el-igible for model validation after excluding individuals withCVD, those taking lipid lowering drugs, or pregnant women.

RESULTS

Factors Associated with a Low CHL/HDL Ratio

The coronary risk factors of the two populations were simi-lar and are summarized in Table 1.

Using univariate analyses, we identified specific variablesthat were significantly (P

,

0.05) associated with a lowCHL/HDL ratio in the LRC cohort. For males, regularlyengaging in strenuous physical activity at least three timesper week, increasing alcohol consumption, being a non-smoker, the absence of a family history of premature CHD,and a body mass index (BMI)

#

26 kg/m

2

were positively as-sociated with a low CHL/HDL ratio

,

6. Among females,an increased alcohol consumption, non-smoking status, nofamily history of premature CHD, the absence of diabetes,pre-menopausal status, age

#

50, systolic blood pressure

,

120, and BMI

,

24 kg/m

2

were positively associated witha low CHL/HDL ratio

,

5.

Identifying Adults at Low Risk for Significant Hyperlipidemia

51

Recursive Partitioning: Development Data Set

Among men, 15.4% (307/1993) had a CHL/HDL ratio

$

6.0(Figure 1). A BMI of

#

26 kg/m

2

was the strongest predictorof a low CHL/HDL ratio. Other significant factors associ-ated with a low ratio included not being a smoker, moder-ate alcohol consumption (at least one drink per week), andbeing active (engaged in strenuous activity at least threetimes per week). In total, the four low risk subgroups of men(numbered 1–4 in Figure 1), comprised 22.9% (457/1993)of the male cohort. Among this low-risk group, only 3.9%(18/457) had an elevated CHL/HDL ratio

$

6 (Table 2).We completed a similar analysis for 1631 women, among

whom 188 (11.5%) had a high CHL/HDL ratio

$

5.0 (Fig-ure 2). A BMI

,

24 was the best discriminator variable as itidentified 838 individuals among whom only 54 had an ele-vated ratio. Among those with a BMI

$

24, 134/793(16.9%) had an elevated ratio. Within each of the two re-sultant populations, smoking status provided the next bestdiscriminator variable. The partitioning process continueduntil we derived a subpopulation (numbered 1 in Figure 2)with a very low prevalence (0/111) of elevated CHL/HDL.These women had a BMI

,

24, were non-smokers, were

#

50 years of age, and consumed three or more alcoholicbeverages per week. Likewise, five other subgroups wereclassified into the low-risk category, such that the com-bined low-risk subpopulations comprised 31.9% (520/1631)of the female cohort. In this low-risk group, only 1.2% (6/520) had a CHL/HDL ratio

$

5.0 (Table 2).

Model Validation

The final model performed well when applied to the SantéQuébec data set as shown in Table 2. The prevalence of el-

evated CHL/HDL ratios were similar in both the LRC andSanté Québec data sets for males (15.4% vs. 13.8%) and fe-males (11.5% vs. 14.5%). The prevalence of an elevatedCHL/HDL ratio among high-risk LRC and Santé Québecmen was 18.8% versus 15.9% and for women was 16.4%versus 18.3%, respectively. The low-risk groups demon-strated a similar low prevalence of an elevated CHL/HDLratio between study populations: 3.9% for men and 1.2%for women in the LRC cohort versus 7.0% and 3.9%, re-spectively, in the Santé Québec survey.

Clinical Index

Having selected variables with the recursive partitioningmodel, we forced each valuable into a logistic regressionmodel where the outcome of interest was a high CHL/HDLratio. The final model included only variables that were in-dependently associated with a high CHL/HDL ratio at asignificance level of

,

0.05. Accordingly, some factors iden-tified by the recursive model were ultimately dropped fromthe logistic model (i.e., age for men).

Using the logistic regression odds ratios as weights (Ta-bles 3 and 4), clinical indices for males and females werecreated. For each index, points based on each factor’s asso-ciation with an elevated CHL/HDL ratio were derived afterrounding the odds ratio to the nearest integer. Summing upall the points for each individual’s risk factors produces a fi-nal risk score. By inspecting the LRC results, we chose a to-tal score of five or less to classify a woman as low risk and atotal score of two or less for men.

Using these simplified clinical indices, 532 (26.7%) menand 736 (45.1%) women were classified as low risk usingthe LRC data set (Table 5). In men, 7.1% (38/532) of the

TABLE 1.

Population coronary risk factors

Risk factors(

n

5

1993)Mean level (

6

SE)

Lipid research clinics Santé Québec

Males(

n

5

1993)Females

(

n

5

1631)Males

(

n

5

486)Females

(

n

5

484)

Age (years) 45.3 (0.2) 46.1 (0.3) 44.8 (0.7) 45.1 (0.6)Systolic blood

pressure (mm Hg) 124.9 (0.4) 119.8 (0.4) 126.3 (0.7) 119.7 (0.8)Diastolic blood

pressure (mm Hg) 81.0 (0.2) 76.5 (0.2) 78.3 (0.4) 74.2 (0.4)Body mass index

(kg/m

2

) 26.3 (0.1) 24.2 (0.1) 25.6 (0.2) 24.3 (0.2)Total cholesterol

(mg/dL) 206.3 (0.8) 203.6 (0.9) 205.0 (1.7) 200.3 (1.8)HDL cholesterol

(mg/dL) 46.6 (0.3) 60.5 (0.4) 46.9 (0.5) 55.7 (0.6)Prevalence of (%)

Diabetes 3.6 1.8 4.5 4.6Smokers 37.0 33.1 30.7 30.4

52

S. A. Grover

et al.

low-risk group actually had elevated CHL/HDL ratios. Sim-ilarly, for women, the false-negative rate was 4.8% (35/736).Overall, the clinical index misclassified 5.8% (73/1268) ofthe LRC low-risk group.

To validate the clinical indices, we applied them to menand women in the Santé Québec data. The false-negativerate among those with elevated CHL/HDL ratios was 7.7%(12/155) for men and 2.1% (4/187) for women (Table 5).Overall, the clinical index misclassified 4.7% (16/342) of

those with an elevated ratio as low risk in the Santé Québecsurvey.

CHD Mortality in the Lipid Research Clinics

To further assess the sensitivity of this screening strategy,we examined the number of CHD-related deaths that oc-curred over a follow-up period of 12.4 years among LRCparticipants. Of the 520 women classified as being at lowrisk for a CHL/HDL ratio

,

5.0, no CHD deaths occurredcompared with 12 CHD deaths in the high-risk group (0%vs. 1.08%; P

,

0.05). Of the 457 men classified as low riskfor a CHL/HDL ratio

,

6.0, only 2 (0.44%) CHD deathsoccurred, which was significantly less (P

,

0.01) than the42 (2.73%) deaths among 1536 men at higher risk of ele-vated CHL/HDL. These data further confirmed that therisk of not screening low-risk individuals for hyperlipidemiawas extremely small in terms of CHD mortality. Moreover,it demonstrates that lipid abnormalities tend to clusteramong patients with other risk factors such as obesity, hy-pertension, smoking, and a sedentary lifestyle. Accordingly,the clinical index presented herein has identified those atlow risk of both lipid abnormalities and coronary events aswell. Among 56 adults who would eventually die a CHDdeath, 54 (96%) would have undergone lipid screening dueto the presence of factors associated with both coronary dis-ease and a high CHL/HDL ratio.

DISCUSSION

These analyses demonstrate the strong association betweenlifestyle variables and the likelihood of an elevated CHL/HDL ratio. In both men and women, increased CHL/HDLratios were positively associated with smoking and an ele-vated BMI and negatively associated with alcohol consump-tion. These results are consistent with previously reportedfindings of the LRC cohort and other studies [20–24]. On

TABLE 2.

Results of a recursive partitioning model to identifyadults with an elevated CHL/HDL ratio (

$

5 for women and

$

6 for men)

Males Females

Risk LRC

a

SQ

b

LRC SQ

Overall 307/1993 67/486 188/1631 70/484(Percent) (15.4) (13.8) (11.5) (14.5)

Low

c

18/457 8/114 6/520 5/129(Percent) (3.9) (7.0) (1.2) (3.9)

High

d

289/1536 59/372 182/1111 65/355(Percent) (18.8) (15.9) (16.4) (18.3)

a

LRC

5

Lipid Research Clinics Follow-up Cohort.

b

SQ

5

Santé Québec Cardiovascular Health Survey.

c

Defined by specific combinations of clinical factors as summarized intext and Figures 1 and 2.

d

Defined as all those who are not classified as low risk.

FIGURE 1. The recursive model for men to identify those at lowrisk of an elevated cholesterol/HDL ratio $6. Each oval noderepresents a decision point where a specific factor separatesthose at low risk (on the right) from those at high risk (on theleft). The square nodes represent terminal nodes 1 through 4which summarize four combinations of variables that identifylow-risk men among whom the prevalence of an elevated choles-terol/HDL ratio is less than 5%. Significant factors includebody mass index (BMI), smoking status, engaging in strenuousactivity at least three times per week (active), age, the numberof alcoholic beverages consumed (drinks), and systolic bloodpressure (SBP).

Identifying Adults at Low Risk for Significant Hyperlipidemia

53

the other hand, some factors were significant only for men,such as physical activity; while age, diabetes, and systolicblood pressure were independent risk factors only forwomen. Although many of these lifestyle factors have been

previously associated with high total cholesterol or lowHDL levels, these gender differences in the present analysisunderscore the importance of separate screening criteria formen and women [25–27].

FIGURE 2. The recursive model for women to identify those at low risk of an elevated cholesterol/HDL ratio $5. Each oval node rep-resents a decision point where a specific factor separates those at low risk (on the right) from those at high risk (on the left). Thesquare nodes represent terminal nodes 1 through 6 which summarize four combinations of variables that identify low-risk womenamong whom the prevalence of an elevated cholesterol/HDL ratio is less than 5%. Significant factors include body mass index (BMI),smoking status, age, the number of alcoholic beverages consumed (drinks), systolic blood pressure (SBP), and diastolic blood pressure (DBP).

TABLE 3.

Factors independently associated with an elevated CHL/HDL ratio

$

5 among men

Risk factor

b

6

(SE) P value Odds ratio Points

Body mass index (kg/m

2

) 0.76 (

6

0.13)

,

0.0001 2.13 2Alcohol consumption

,

1 time per week 0.66 (

6

0.14)

,

0.0001 1.93 2Smoking 0.63 (

6

0.13)

,

0.0001 1.88 2Strenuous physical

activity at least3 times per week 0.48 (

6

0.16) 0.0023 1.62 2

54

S. A. Grover

et al.

The final recursive model revealed that among those inthe LRC cohort without CVD, who were not pregnant orreceiving therapy for hyperlipidemia, 23% (457/1993) ofmen and 32% (520/1631) of women were at low risk for anelevated CHL/HDL ratio. Overall, this represents over onequarter of this healthy population.

Similarly, in the Santé Québec population used formodel validation, 23.5% (114/486) of healthy men and26.7% of healthy women (129/484) were classified as lowrisk. Overall, 25.1% of all Quebec adults aged 30–74 wouldhave been identified as low risk for an elevated CHL/HDLratio thereby reducing the need for lipid testing.

Translating these multivariate analyses into useful diag-nostic tools for clinical decision making requires a clinicalindex that allows for the quick assessment of individual pa-tients during the course of a routine office visit. The result-ing indices are reasonably simple and require informationthat is readily available based on a routine personal historyand very basic clinical exam.

For both males and females in the LRC data set, theseclinical indices provided excellent test sensitivity: 0.88 inmen and 0.82 in females. However, test specificity wasrather poor at 0.29 and 0.51 for men and women, respec-

tively. This reflects our conservative approach, whereby wetried to miss as few high CHL/HDL cases as possible. None-theless, these overall results are similar to other publishedindices to predict perioperative cardiac events, ankle frac-tures, and postoperative delirium [28–30].

A similar study has also been published by Kinlay

et al.

toidentify individuals with high levels of blood cholesterol(

.

6.5 mmol/L) [31]. Independent risk factors included age,a history of hypertension, and a history of heart attack. Thefinal model demonstrated a test sensitivity of 0.77 and spec-ificity of 0.61. However, it is becoming increasingly clearthat an elevated cholesterol level alone is a poor predictorof coronary risk compared to the elevated CHL/HDL ratioused in our present study [5–10].

It should be noted that there remain a number of poten-tial weaknesses with a clinical index based on LRC datafrom the 1970s. Lipid levels have been changing during thepast 20 years and the LRC 15% random sample may not betruly representative of the adult population. Furthermore,the accuracy of patients’ self-reported behaviors may be lessthan the results observed in a carefully conducted surveyfurther undermining the test sensitivity and specificity re-ported herein. Nonetheless, we note that independent vali-dation of the clinical index on the Santé Québec data set,which is representative of the adult Quebec population inthe 1990s, addresses some of these concerns. Applying theclinical index to this contemporary adult population, re-sulted in high test sensitivities of 0.81 for men and 0.94 forwomen and low specificities of 0.34 and 0.44, respectively.Overall, 12% of men and women with a high ratio weremisclassified as low risk.

We also recognize that primary prevention is increas-ingly focussing on identifying those individuals at high ab-solute risk due to multiple-risk factors rather than lipid ab-normalities alone. A multifactorial risk approach may beparticularly attractive if it can be shown to further reducethe false-negative rate observed in our analyses [6].

In conclusion, we have demonstrated that readily avail-able clinical data can be used to identify a substantial groupof individuals at low risk of an elevated CHL/HDL ratio.These individuals are also at very low risk of dying of coro-nary disease over the next decade. Incorporating this simple

TABLE 4.

Factors independently associated with an elevated CHL/HDL ratio

$

5 among women

Risk factor

b 6

(SE) P value Odds Ratio Points

Body mass index (kg/m

2

) 0.96 (

6

0.18)

,

0.0001 2.62 3Age

.

50 0.37 (

6

0.18) 0.038 1.45 2Alcohol consumption

1–4 times per week 0.53 (

6

0.28) 0.063 1.69 2

,

1–2 times per month 1.13 (

6

0.27)

,

0.0001 3.09 3Smoking 1.10 (60.17) ,0.0001 2.99 3Diabetes 1.38 (60.43) 0.001 3.97 4Systolic BP .120 0.45 (60.17) 0.01 1.57 2

TABLE 5. Results of two clinical indices to identify adultswith an elevated CHL/HDL ratio

Males Females

Risk LRCa SQb LRC SQ

Overall 307/1993 67/486 188/1631 70/484(Percent) (15.4) (13.8) (11.5) (14.5)Lowc 38/532 12/155 35/736 4/187(Percent) (7.1) (7.7) (4.8) (2.1)Highd 269/1461 55/331 153/895 66/297(Percent) (18.4) (16.6) (17.1) (22.2)

aLRC 5 Lipid Research Clinics Follow-up Cohort.bSQ 5 Santé Québec Cardiovascular Health Survey.cDefined by specific combinations of clinical factors. For men, low risk

was defined as a score of #2 points using factors summarized in Table 3.For women, a score of #5 points using factors summarized in Table 4 wasdefined as low risk (see text and Figures 1 and 2 for details).

dDefined as all those who are not classified as low risk.

Identifying Adults at Low Risk for Significant Hyperlipidemia 55

clinical indices into current lipid screening guidelines couldsignificantly reduce the number of lipid tests ordered amongotherwise healthy adults in our communities.

This work was supported in part by grants from Heath Canada and theDairy Farmers of Canada. Dr. Grover is a senior research scientist(chercheur-boursier) supported by Le fonds de la recherche en santé duQuébec. We thank Dr. Louise Pilote and Mr. Louis Coupal for theirhelpful comments on earlier versions of this manuscript and Ms. NadineBouchard for her excellent secretarial support.

References1. Expert Panel on Detection, Evaluation, and Treatment of

High Blood Cholesterol in Adults. Summary of the SecondReport of the National Cholesterol Education Program (NCEP)Expert Panel on Detection, Evaluation, and Treatment ofHigh Blood Cholesterol in Adults (Adult Treatment PanelII). JAMA 1993; 269: 3015–3023.

2. The Canadian Consensus Conference on Cholesterol: FinalReport: The Canadian Consensus Conference on the pre-vention of heart and vascular disease by altering serum choles-terol and lipoprotein risk factors. Can Med Assoc J 1988;139(Suppl 11): 1–8.

3. Garber AM, Littenberg B, Sox HC. Cost and effectiveness ofcholesterol screening in the elderly. Health Program Office ofTechnology Assessment. Washington, D.C.: U.S. Govern-ment Printing Office; 1989: 35–40.

4. Grover SA, Coupal L, Fakhry R, Suissa S. Screening for hy-percholesterolemia among Canadians: How much will it cost?Can Med Assoc J 1991; 144(2): 161–168.

5. Grover SA, Palmer CS, Coupal L. Serum lipid screening toidentify high-risk individuals for coronary death. Arch InternMed 1994; 154: 679–684.

6. Grover SA, Coupal L, Hu X-P. Identifying adults at increasedrisk of coronary disease: How well do the current cholesterolguidelines work? JAMA 1995; 274(10): 801–806.

7. Stampfer MJ, Sacks FM, Salvini S, Willett WC, Hennekens CH.A prospective study of cholesterol, apolipoproteins, and the riskof myocardial infarction. N Engl J Med 1991; 325(6): 373–381.

8. Castelli WP, Anderson K. A population at risk: Prevalence ofhigh cholesterol levels in hypertensive patients in the Framing-ham Study. Am J Med 1986; 80(Suppl 2A): 23–32.

9. Brunner D, Weisbort J, Meshulam N, et al. Relation of serumtotal cholesterol and high-density lipoprotein cholesterol per-centage to the incidence of definite coronary events: Twenty-year follow-up of the Donolo-Tel Aviv Prospective CoronaryArtery Disease Study. Am J Cardiol 1987; 59: 1271–1276.

10. Goldbourt G, Holtzman E, Neufeld HN. Total and high den-sity lipoprotein cholesterol in the serum and risk of mortality:Evidence of a threshold effect. Br Med J 1985; 290: 1239–1243.

11. Ciampi A, Negassa A, Lou Zihyi. Tree-structured predictionfor censored survival data and the Cox model. J Clin Epide-miol 1995; 48: 675–689.

12. Goulet JR, Mackenzie T, Levinton CM, Hayslett JP, CiampiA, Esdaile JM. The longterm prognosis of lupus nephritis: Theimpact of disease activity. J Rheumatol 1993; 20: 59–65.

13. Central Patient Registry and Coordinating Center for theLipid Research Clinics. Reference Manual for the Lipid Re-search Clinics Prevalence Study, Vols. 1 and 2. Chapel Hill,NC: Dept of Biostatistics, University of North Carolina; Feb-ruary 1974.

14. Lipid Research Clinics Program Epidemiology Committee.Plasma lipid distributions in selected North American popula-

tions: The Lipid Research Clinics Prevalence Study. Circula-tion 1979; 60: 427–439.

15. Heiss G, Tamir I, Davis CE, et al. Lipoprotein-cholesterol dis-tributions in selected North American populations: The LipidResearch Clinics Program Prevalence Study. Circulation1980; 61: 302–315.

16. Lipid Research Clinics Program. Manual of Laboratory Oper-ations, Vol. 1: Lipid and Lipoprotein Analysis. Bethesda,MD: National Heart, Lung, and Blood Institute, National In-stitutes of Health; 1974. US Dept of Health, Education, andWelfare publication (NIH) 75-628.

17. Jacobs DR Jr, Melbane IL, Bangdiwala SI, Criqui MH, TyrolerHA. Lipid Research Clinics Program. High density lipopro-tein cholesterol as a predictor of cardiovascular disease mor-tality in men and women: The follow-up study of the LipidResearch Clinics Prevalence Study. Am J Epidemiol 1990;131: 32–47.

18. Caroll MC, Sempos C, Briefel R, Gray S, Johnson C. Serumlipids of adults 20–74 years. United States, 1976–80. NationalCentre for Health Statistics. Vital Health Stat 11(242). 1993.

19. Esrey KL, Coupal L, Grover SA. Évaluation du risque de mal-adies cardiovasculaires. Exemple d’application d’une méthodemultifactorielle. In: Santé Québec: Et votre coeur, ça va?Rapport de l’Enquête québécoise sur la santé cardiovascu-laire 1990. Montréal, ministère de la Santé et des Services so-ciaux, gouvernement du Québec; 1994: 95–101.

20. Heiss G, Johnson NJ, Reiland S, Davis CE, Tyroler HA. Theepidemiology of plasma high-density lipoprotein cholesterollevels: The Lipid Research Clinics Program Prevalence Study:Summary. Circulation 1980; 62(Suppl IV): 116–136.

21. Ernst N, Fisher M, Smith W, et al. The association of plasmahigh-density lipoprotein cholesterol with dietary intake andalcohol consumption: The Lipid Research Clinics ProgramPrevalence Study. Circulation 1980; 62(Suppl IV): 41–52.

22. Anderson KM, Wilson PW, Garrison RJ, Castelli WP. Longi-tudinal and secular trends in lipoprotein cholesterol measure-ments in a general population sample: The Framingham Off-spring Study. Atherosclerosis 1987; 68: 59–66.

23. Dwyer T, Calvert GD, Baghurst KI, Leitch DR. Diet, otherlifestyle factors and HDL cholesterol in a population of Aus-tralian male service recruits. Am J Epidemiol 1981; 114: 683–696.

24. Wolfe RN, Grundy SM. Influence of weight reduction onplasma lipoproteins in obese patients. Arteriosclerosis 1983;3: 160–168.

25. Wood PD, Klein H, Lewis S, Haskell WL. The distribution ofplasma lipoproteins in middle aged male runners. Metabolism1976; 25: 1249–1254.

26. Stein RA, Michielli DW, Glantz MD, et al. Effects of differentexercise training intensities on lipoprotein cholesterol frac-tions in healthy middle-aged men. Am Heart J 1990; 119(2):277–283.

27. Howard BV. Lipoprotein metabolism in diabetes mellitus. JLipid Res 1987; 28: 613–628.

28. Goldman L, Caldera DL, Nussbaum SR, et al. Multifactorialindex of cardiac risk in noncardiac surgical procedures. NEngl J Med 1977; 297(16): 845–850.

29. Stiell IG, McKnight RD, Greenberg GH, et al. Implementa-tion of the Ottawa Ankle Rules. JAMA 1994; 271(11): 827–832.

30. Marcantonio ER, Goldman L, Mangione CM, et al. A clinicalprediction rule for delirium after elective noncardiac surgery.JAMA 1994; 271(2): 134–139.

31. Kinlay S, Heller RF. Selective screening for high cholesterolin Australian general practice: The Newcastle CholesterolPrediction Study. J Gen Int Med 1990; 5: 1–8.