Atherosclerosis 143 (1999) 275–284

Triglyceride as a risk factor for ischaemic heart disease in British men:effect of adjusting for measurement error

Matthias Egger a,b,*, George Davey Smith a, Dominik Pfluger b, Ekkehardt Altpeter b,Peter C. Elwood a,c

a Department of Social Medicine, Canynge Hall, Whiteladies Road, Uni6ersity of Bristol, Bristol BS8 2PR, UKb Department of Social and Pre6enti6e Medicine, Uni6ersity of Berne, Berne, Switzerland

c MRC Epidemiology Unit, Llandough Hospital, Penarth, South Glamorgan, UK

Received 19 February 1998; received in revised form 15 September 1998; accepted 3 November 1998

Abstract

Objecti6e: To assess the influence of differential precision in the measurement of the correlated variables total cholesterol andhigh density lipoprotein (HDL) cholesterol on the estimates of the risk of ischaemic heart disease (IHD) associated with plasmatriglyceride levels. Design, setting and participants: The Caerphilly Heart Disease Study (CHDS), a prospective cohort study of2512 middle-aged men living in the town of Caerphilly, south Wales, UK. The results from two sub-studies were used to estimatethe degree of measurement imprecision (laboratory error and within-person variation) in triglycerides, total cholesterol and HDLcholesterol. Main outcome measures: Multivariable risk estimates for major IHD calculated from logistic regression analysis,adjusted and not adjusted for measurement imprecision. Major IHD events were defined as death from IHD, clinical non-fatalmyocardial infarction or electrocardiographic myocardial infarction. Results: There were 261 men with major IHD events duringfollow-up. In age-adjusted analyses, taking measurement imprecision into account strengthened associations with IHD for all lipidfactors. The odds ratio (OR) for one S.D. increase in triglycerides, ignoring measurement imprecision, was 1.36 (95% confidenceinterval [95% CI] 1.20–1.55) but 1.57 (95% CI 1.30–1.89) when taking imprecision into account. The standardised odds ratio fortriglycerides adjusted for measurement imprecision and the two other lipid factors was 1.35 (95% CI 1.09–1.69). In this model,the triglyceride level showed a stronger association than total cholesterol (OR 1.28; 95% CI 1.05–1.56) and HDL cholesterol (ORfor one S.D. decrease 1.20; 95% CI 0.97–1.49). When adding fasting blood glucose and diastolic blood pressure, however, theeffect of triglycerides was reduced and ceased to be statistically significant (OR 1.19; 95% CI 0.95–1.49). This was furtherattenuated upon inclusion of body mass index, smoking status and history of pre-existing IHD. Total cholesterol remained astatistically significant (PB0.05) risk factor in all models. Conclusions: In contrast to other cohort studies, triglycerideconcentration in the CHDS shows an association with the risk of IHD which is independent of total and HDL cholesterol. Thiseffect was pronounced after adjustment for measurement imprecision. It was reduced, however, when adjusted for other factors.While hypertriglyceridaemia may exert an influence independent of other lipid factors, insulin resistance is probably the underlyingmetabolic disturbance. © 1999 Elsevier Science Ireland Ltd. All rights reserved.

Keywords: Lipids; Coronary disease; Men; Measurement error; Statistical models

1. Introduction

Multivariable statistical models have been widelyused in cardiovascular disease epidemiology to identify‘independent’ risk factors, i.e. the factors for whichassociation with cardiovascular disease is not due to

* Corresponding author. Present address: Department of SocialMedicine, Canynge Hall, Whiteladies Road, Bristol BS8 2PR, UK.Tel.: +44-117-9287-387; fax: +41-117-9287-325.

E-mail address: m.egger@bristol.ac.uk (M. Egger)

0021-9150/99/$ - see front matter © 1999 Elsevier Science Ireland Ltd. All rights reserved.PII: S 0 0 2 1 -9150 (98 )00300 -1

M. Egger et al. / Atherosclerosis 143 (1999) 275–284276

correlation with other, confounding, variables. Infer-ences about causality and subsequent recommendationsfor public health measures have been based on theresults from such models. Univariable analyses haveconsistently found triglyceride levels to show an associ-ation with coronary heart disease. In many studies, thisassociation was, however, greatly attenuated when ad-justed for other cardiovascular risk factors. Hulley etal., in an influential logistic regression analysis of datafrom the Western Collaborative Group Study [1],showed that the effect of triglycerides disappeared whentotal and high density lipoprotein (HDL) cholesterolwere added to the model. Based on these results, theyconcluded that screening for and treatment of hyper-triglyceridaemia for prevention of coronary heart dis-ease cannot be justified.

Error in the measurement of exposures is of centralimportance in epidemiological studies that aim to disen-tangle the independent contributions of each factor tothe risk of disease [2–4]. In prospective epidemiologicalstudies, most exposures are in fact measured with somedegree of random error which is independent of diseasestatus [5]. Such ‘non-differential’ random error mayarise both from technical errors in measurement andfrom true biological variation of the variable in ques-tion. It has long been recognised that in univariableanalysis, the effect of non-differential measurement er-ror is to bias estimates of causal associations towardthe null value, an effect which has been referred to as‘attenuation’ [6] or, more lately, as ‘regression dilutionbias’ [7]. More recently, it has become clear that inmultivariable analysis, imprecision in the measurementof confounding variables is of particular concern be-cause it can lead to underestimation or overestimationof the effect estimate [3,8,9]. This is a potential problemin analyses of the relative importance of triglycerides,total cholesterol and HDL cholesterol. Measurementsof triglycerides correlate positively with cholesterol andinversely with HDL cholesterol. Also, triglycerides aremore variable and usual levels of triglycerides willtherefore be less well indexed by a single measure thanusual levels of HDL cholesterol and total cholesterol[10].

It is thus conceivable that in multivariable models,triglycerides are simply displaced by the highly corre-lated, but more precisely measured, cholesterol vari-ables [11]. The results could thus be different if thedegree of measurement precision was taken into ac-count. Indeed, there continues to be some debate re-garding the importance of triglycerides [12–14], withthe high variability of triglycerides [15,16] and thecollinearity between triglycerides and HDL cholesterol[17,18] being mentioned as potential reasons for thediscrepant results.

In many epidemiological studies of cardiovasculardisease, the impact of measurement imprecision cannot

be assessed because no data were collected on repro-ducibility of measurements. The Caerphilly Heart Dis-ease Study (CHDS) [19,20] provides an opportunity toexamine this issue. Within the framework of thisprospective cohort study, sub-studies were conductedboth to characterize technical measurement error oftriglycerides, total cholesterol and HDL cholesterol,and to assess the underlying biological variation ofthese variables, which, for simplicity, are also denotedas lipid factors in this report. In the analyses presentedhere, we have examined the influence of measurementimprecision for lipid factors on multivariable risk esti-mates in the CHDS.

2. Methods

2.1. Study population and baseline sur6ey

The CHDS has been described in detail elsewhere[19,20]. Briefly, a total of 2512 men (89% of 2818 menfound to be eligible) were examined between 1979 and1983 in the town of Caerphilly, South Wales. The menwere invited to attend an afternoon or evening clinic. Astandard medical and smoking history was obtained.The Rose chest pain questionnaire [21] was adminis-tered; height, weight, and blood pressure were mea-sured using standardised methods; and a 12-leadelectrocardiogram (ECG) was recorded. The subjectswere then asked to return (after an overnight fast) to anearly morning clinic where a blood sample was takenwith minimal venous stasis and anticoagulated with 1mg/ml disodium ethylenediaminetetraacetate. The samelaboratories were used for all tests. Cholesterol andtriglyceride concentrations were measured with enzy-matic procedures [22,23]. The HDL fraction was iso-lated by precipitation of the low and very low densitylipoproteins (VLDL) with sodium phosphotungstateand magnesium chloride [24]. Glucose determinationswere made on plasma samples with a Beckman Astra 8analyser by the oxygen rate method [25].

2.2. Reproducibility studies

Split duplicate samples were taken at baseline on a5% sub-sample of subjects. Duplicates were presentedblindly in containers labeled with a specially allocatedduplicate number known only at the statistical centre.The split duplicate samples allow assessment of vari-ability due to technical measurement error, which canbe assumed to be random. The samples were selected ina quasi-random fashion by laboratory personnel whowere blind to the men’s characteristics at baseline.Long-term within-person variation was assessed in aconvenience sample of men who were seen for a secondvenesection, on average, 1 year after recruitment. These

M. Egger et al. / Atherosclerosis 143 (1999) 275–284 277

Table 1Baseline characteristics of men with and without incident ischaemic heart diseasea

PMajor incident IHD Odds ratiob (95% CI)No major incident IHDVariable

1938N 2611.40 (1.23–1.61) 0.0001Age (years) 53.4 (4.3) 51.9 (4.5)

Lipid factors1.93 (1.28) 0.00011.33 (1.17–1.51)2.21 (1.24)Triglyceride (mmol/l)

1.28 (1.12–1.46) 0.0002Total cholesterol (mmol/l) 5.98 (1.28) 5.68 (1.11)1.21 (1.06–1.38) 0.004HDL cholesterol (mmol/l) 1.07 (0.30) 1.13 (0.34)

Blood pressure (mmHg)1.38 (1.21–1.58)88.3 (12.5) 0.000192.1 (12.4)Diastolic

140 (18.8) 1.40 (1.23–1.59)Systolic 147 (21.0) 0.0001

1.32 (1.19–1.46)4.94 (1.13)5.37 (1.89) 0.0001Glucose (mmol/l)26.1 (3.6) 1.20 (1.05–1.36)Body mass index (kg/m2) 26.7 (3.5) 0.007

Smoking1.0 –333 (17.2)21 (8.0)Never

1.87 (1.13–3.11) 0.02Former 67 (25.7) 568 (29.3)2.28 (1.26–4.10) 0.008Pipe/cigar 29 (11.1) 202 (10.4)

273 (14.1) 2.38 (1.37–4.13)1–14 cigarettes 41 (15.7) 0.002328 (16.9) 2.80 (1.66–4.73) 0.000158 (22.2)15–24 cigarettes234 (12.1) 3.05 (1.77–5.26)25+cigarettes 45 (17.2) 0.0001

Ischaemic heart disease123 (6.3) 2.67 (1.82–3.92)Angina presentc 0.00140 (15.3)151 (7.8) 3.85 (2.77–5.33)History of severe chest painc 64 (24.5) 0.001

3.18 (2.35–4.31)218 (11.2) 0.00175 (28.7)ECG ischaemia

a Means (S.D.) or numbers (%). IHD, ischaemic heart disease; CI, confidence interval; HDL, high density lipoprotein.b For continuous variables, standardised odds ratios are shown which indicate the change in odds for one S.D. decrease for HDL cholesterol,

and for one S.D. increase for all other variables.c From Rose questionnaire [21].

long-term repeat samples assess variability due to bothtechnical errors and biological long-term within-personvariation.

2.3. Follow-up procedures and definition of e6ents

Men were invited to a follow-up clinic 5 and 10 yearsafter the baseline examination. Follow-up examinationsincluded the completion of the LSHTM chest painquestionnaire [21], the recording of an ECG and avenesection.

The records of all men at the National Health ServiceCentral Registry are flagged so that notification ofdeath is automatic and a copy of the death certificate isreceived. Notifications of all admissions to local hospi-tals with a diagnosis coded 410-414 (ischaemic heartdisease (IHD)) on the International Classification ofDiseases (ICD) were received from the hospital activityanalysis. The questionnaires, together with hospital ac-tivity analysis notifications, were used as the basis for asearch of hospital notes for events which satisfied theWorld Health Organisation (WHO) criteria for definitemyocardial infarction. From this information, threetypes of major incident IHD events were defined: (i)death from IHD (cause of death coded as ICD 410-414), (ii) clinical non-fatal myocardial infarction (satis-

fying WHO criteria), (iii) electrocardiographicmyocardial infarction (appearance of major or moder-ate Q or QS waves, Minnesota codes 1-1-1 to 1-2-5 or1-2-7 on a follow-up ECG, when there were no Q orQS waves, Minnesota codes 1-1-any, 1-2-any 1-3-anyon the recruitment ECG). All ECG recordings wereread by one of two highly experienced persons. In menqualifying for more than one category of major IHD,the more serious category was coded.

2.4. Statistical methods

Univariable and multivariable logistic regressionmodels [26] with the occurrence, or not, of a majorincident IHD event as the dependent variable wereused. Men with pre-existing IHD were included. Forcontinuous variables, results are presented as standard-ised odds ratios (OR), defined as the proportionatechange in odds for a one S.D. change in the variablealong with 95% confidence intervals (95% CI). Stan-dardised ORs were computed for an increase in thevariable in question except for HDL cholesterol, forwhich the OR ratio for a one S.D. decrease was calcu-lated. In contrast to the proportional hazards model[27], logistic regression takes no account of the timingof events and ignores duration of follow-up. However,

M. Egger et al. / Atherosclerosis 143 (1999) 275–284278

Table 2Correlations between baseline variables: Pearson correlation coefficients and probability valuesa

BMI Systol. BP GlucoseDiast. BPTotal-C HDL-C Age

0.27 0.15 0.12 0.18TGb 0.35 −0.25 −0.070.0001 0.00010.0001 0.0001 0.002 0.0001 0.0001

0.050.05 0.03Total-C 0.060.06 −0.0080.006 0.05 0.18 0.030.007 0.7

−0.01 −0.03HDL-C 0.04 −0.18 −0.090.20.6 0.00010.00010.07

−0.03 0.01 0.16 0.08Age0.9 0.00010.12 0.0003

0.220.29 0.18BMI0.0001 0.00010.00010.67Diast. BP 0.050.0001 0.03

Syst. BP 0.100.0001

a Probability values are shown in italics.b TG, triglycerides; Total-C, total cholesterol; HDL-C, high density lipoprotein cholesterol; BMI, body mass index; Diast. BP, diastolic blood

pressure; Syst. BP, systolic blood pressure.

in the CHDS, duration of follow-up was very similaracross surviving participants. Furthermore, any modelinvolving time faces the problem that no time of eventis available for electrocardiographically defined myocar-dial infarction.

The bias in multivariable relative risk estimates intro-duced by measurement imprecision was assessed usingthe method described by Rosner et al. [28,29]. Briefly,logistic regression using the main study data is per-formed to estimate the ‘uncorrected’ logistic regressioncoefficients. The reproducibility study is used to esti-mate the within-person covariance matrix, whereas thetotal covariance matrix is estimated from the mainstudy. An estimate of the between-persons covariancematrix is obtained by subtraction of the within-personmatrix from the total covariance matrix. The multivari-able reliability coefficient and the ‘true’ logistic regres-sion coefficient can then be calculated. Confidenceintervals, which take into account both the contribu-tions to the variance due to estimation of the uncor-rected logistic regression coefficient and due toestimation of the reliability coefficient, are alsoobtained.

Repeatability was assessed by calculating Pearsoncorrelation coefficients, by comparing values in pairedt-tests and by plotting the differences between the twomeasurements against their mean [30]. The proportionof variability explained by laboratory variation (splitsamples) or within-person variation (long-term repeatsamples) was calculated from general linear models.Reliability coefficients, defined as the correlation be-tween the unknown ‘true’ value and the measured value,were estimated by the square root of the intraclasscorrelation coefficient between replicate measures on thesame individual [31]. In the case of two replicates, the

intraclass correlation coefficient is equivalent to calcu-lating the usual Pearson correlation coefficient betweentwo measurements, provided each pair of measurementsis counted twice, the second time in reverse order [31].

2.5. Sensiti6ity analyses

Important assumptions [28,29,32] of the method usedfor adjusting logistic regression coefficients for measure-ment error include: (i) that the disease under study israre (5–10% of participants developing disease), (ii) thatmeasurement error is not severe (reliability coefficientabove 0.5), (iii) that the reproducibility study sample isrepresentative for the main study population, (iv) thatthe estimates derived from the reproducibility study areindependent from those obtained from the main study,(v) that within-person variance is independent of themean (bivariate normality), and (vi) that errors are notcorrelated. Natural logarithms were taken to achievenormality in seven variables with skewed distributions:triglyceride, HDL and total cholesterol, serum glucose,systolic and diastolic blood pressure and body massindex (BMI).

In order to examine the importance of the rare diseaseassumption, analyses were repeated considering fatalIHD events only. Similarly, to test the importance ofindependency of reproducibility and main studies,analyses were repeated excluding subjects with splitsamples or long-term repeats from the main study.

The Statistical Analysis System (SAS) software, ver-sion 6.11 was used for analysis [33]. The RELIB macro,programmed in SAS, was used for adjusting logisticregression coefficients for measurement error [29]. In allanalyses, lipid factors were log transformed and trans-formed back.

M. Egger et al. / Atherosclerosis 143 (1999) 275–284 279

Table 3Measures of repeatability, correlation and reliability for 99 split samplesa

Total cholesterol HDL cholesterolTriglyceride

5.63/5.65 1.13/1.14Sample means (mmol/l) 1.72/1.7011.4 8.0Laboratory variation (%) 32.4

0.930.91Pearson correlation coefficient 0.670.95 0.96 0.82Reliability coefficient

a Probability from paired t-tests. Laboratory variation as percentage of total variation.

Table 4Measures of repeatability, correlation and reliability for 61 long-term repeat samplesa

Total cholesterol HDL cholesterolTriglyceride

1.97/1.82 5.85/5.68 1.16/1.04Sample means (mmol/l)0.167 0.127Mean difference (mmol/l) 0.151

1.07 0.93S.D. of difference (mmol/l) 0.280.28 0.17Probability 0.0007

45.833.8Within-person and laboratory variation (%) 40.80.590.62Pearson correlation coefficient 0.470.66Reliability coefficient 0.780.68

a Probability from paired t-tests. Within-person and laboratory variation as percentage of total variation.

3. Results

3.1. Data used in analysis

A total of 2512 men were seen at baseline. For thecurrent analysis, 313 (12.5%) men, who were either lostto follow-up or who had missing information in any ofthe baseline variables (Table 1), were excluded. A totalof 2199 men were thus included in the analysis. Therewere 261 men with major IHD events during follow-up(134 (52%) with fatal IHD, 100 (38%) with non-fatalclinical myocardial infarction, and 27 (10%) with elec-trocardiographically defined myocardial infarction).

3.2. Baseline characteristics and correlations

Characteristics at baseline in men with and withoutmajor IHD events are shown in Table 1, along withunivariable ORs and 95% CI. The univariable analysisshows statistically significant associations with all threelipid factors, blood glucose, blood pressure, BMI,smoking and pre-existent IHD.

The correlation matrix for blood lipids and otherbaseline variables is shown in Table 2. Triglyceridelevels are positively correlated with total cholesteroland negatively with HDL cholesterol. Strong positivecorrelations of triglycerides are also noted with BMI,and blood glucose. HDL cholesterol is negatively corre-lated with BMI. BMI is positively correlated with bloodpressure and glucose.

3.3. Measures of repeatability and reliability

For 99 men (4.5%), a split duplicate sample and for61 men (2.8%), a long-term repeat sample was avail-able. Only one man had both a split sample and along-term repeat sample. There were no statisticallysignificant baseline differences in variables shown inTable 1 between men with and without long-term re-peat samples (P\0.15 for all tests). Similarly, therewere no statistically significant differences between menwith and without split samples, except for somewhatlower triglyceride levels in men with split samples (1.72versus 1.98 mmol/l), a difference of marginal statisticalsignificance (P=0.054 by t-test).

Measures of laboratory repeatability as assessed fromsplit samples are shown in Table 3. Laboratory varia-tion was relatively small, and Pearson correlation andreliability coefficients were high for triglyceride andtotal cholesterol levels. A somewhat less stable pictureemerged for HDL cholesterol, with laboratory variationamounting to 32.4% of total variation and values ofboth coefficients below 0.9. The corresponding analysisfor long-term repeat samples is summarised in Table 4.No statistically significant differences were evident fortriglyceride and total cholesterol, respectively, althoughmeans from the second measurement were lower thanthe first for both lipids. For HDL cholesterol, there wasa statistically significant difference in the same direction(0.13 mmol/l, P=0.0007 by paired t-test). These differ-ences may be due to changes in lifestyle and nutrition.

M. Egger et al. / Atherosclerosis 143 (1999) 275–284280

Table 510-year relative risk estimates for ischaemic heart disease from logistic models ignoring and considering measurement errora

Model Independent variables Odds ratio (95% CI)

Considering measurement error in lipid factorsIgnoring measurement error in lipid factors

Laboratory and within-personLaboratory

1.36 (1.20–1.55) 1.41 (1.22–1.63)1 1.57 (1.30–1.89)Triglyceride1.29 (1.13–1.47) 1.43 (1.19–1.73)1.32 (1.14–1.52)Total cholesterol2

1.35 (1.11–1.63)1.36 (1.12–1.65)3 HDL cholesterol 1.22 (1.08–1.39)1.23 (1.03–1.46) 1.35 (1.09–1.69)4 Triglyceride 1.23 (1.06–1.42)

1.20 (1.05–1.39) 1.23 (1.05–1.44)Total cholesterol 1.28 (1.05–1.56)1.18 (1.03–1.35) 1.28 (1.03–1.58) 1.20 (0.97–1.49)HDL cholesterol

1.11 (0.93–1.32) 1.19 (0.95–1.50)5 Triglyceride 1.12 (0.96–1.30)1.31 (1.07–1.59)1.25 (1.07–1.47)1.22 (1.06–1.41)Total cholesterol

1.27 (1.02–1.58) 1.21 (0.98–1.50)HDL cholesterol 1.17 (1.02–1.34)1.23 (1.10–1.37) 1.22 (1.09–1.36)Blood glucose 1.24 (1.11–1.38)

1.34 (1.16–1.54) 1.33 (1.16–1.54)Diastolic BP 1.32 (1.14–1.52)1.07 (0.91–1.25) 1.12 (0.88–1.43)1.06 (0.88–1.28)Triglyceride61.21 (1.04–1.40) 1.24 (1.05–1.46) 1.29 (1.05–1.58)Total cholesterol

1.17 (0.94–1.46) 1.13 (0.91–1.40)HDL cholesterol 1.11 (0.97–1.27)1.22 (1.09–1.37) 1.22 (1.06–1.37)Blood glucose 1.21 (1.08–1.36)1.28 (1.10–1.48) 1.28 (1.10–1.49) 1.27 (1.09–1.48)Diastolic BP1.10 (0.95–1.28) 1.09 (0.94–1.27)Body mass index 1.08 (0.93–1.26)

2.47 (1.51–4.04)2.48 (1.52–4.05)2.49 (1.53–4.07)Current smoking1.44 (0.85–2.44) 1.44 (0.85–2.45)Former smoking 1.44 (0.85–2.44)1.58 (1.03–2.42) 1.57 (1.02–2.40)Angina present 1.56 (1.02–2.39)

2.62 (1.82–3.75)2.65 (1.85–3.79) 2.61 (1.82–3.75)History of MI2.03 (1.45–2.83)2.04 (1.46–2.84)ECG ischaemia 2.05 (1.47–2.85)

a All models adjusted for age. For continuous variables, standardised odds ratios are shown which indicate the change in odds for one S.D.decrease for HDL cholesterol, and for one S.D. increase for all other variables. See Table 1 for the standard deviations used in these calculations.‘Never smoking’ was the reference category for smoking variables.

Quality control data did not indicate any importantlaboratory drift. Within-person variability ranged from33.8 to 45.8%. Reliability coefficients were from 0.66for HDL cholesterol to 0.78 for total cholesterol.

3.4. Multi6ariable analyses ignoring and consideringmeasurement error

Table 5 gives the results from multivariable logisticregression models ignoring and considering measure-ment imprecision in the measurement of lipid factors.The back transformed standard errors used in theseanalyses were 1.72 mmol/l for triglycerides, 1.22 mmol/lfor total cholesterol and 1.33 mmol/l for HDL choles-terol. In analyses adjusted for age only, taking measure-ment imprecision into account considerablystrengthened relative risk estimates of all three lipids.For triglyceride and cholesterol, the size of the associa-tion increased when moving from models ignoring mea-surement imprecision to models considering laboratoryerror, and to models considering both laboratory errorand within-person variation. In the case of HDLcholesterol, the strength of the association was similarwhen adjusted for laboratory error or for laboratoryerror and within-person variation.

When including all three lipid factors (model 4 inTable 5) and adjusting for laboratory error and within-person variation, triglyceride level (OR 1.35, P=0.007)showed a stronger association than total cholesterol(OR 1.28, P=0.015) and HDL cholesterol (OR 1.20,P=0.086). In this model, the effect of HDL cholesterolceased to be statistically significant.

Inclusion of blood glucose and diastolic blood pres-sure greatly reduced the effect of triglyceride (OR 1.19,P=0.14 in the model adjusted for laboratory error andwithin-person variation), while estimates for totalcholesterol (OR 1.31, P=0.009) and HDL cholesterol(OR 1.21, P=0.082) remained essentially unchanged.The OR for triglyceride was further reduced whenBMI, smoking and markers for pre-existent diseasewere added. Conversely, the estimate for total choles-terol was little affected and remained statistically sig-nificant (OR 1.29, P=0.015).

The influence of adjusting for measurement error onrisk estimates of lipid factors in different logistic modelsis illustrated in Fig. 1.

3.5. Sensiti6ity analyses

Results from sensitivity analyses are shown in Table6. Standardised ORs were very similar when analyses

M. Egger et al. / Atherosclerosis 143 (1999) 275–284 281

were repeated excluding participants with split samplesor long-term repeats from the main study. Age-adjustedORs tended to be somewhat higher when repeating

analyses on all participants but considering fatal IHDevents only, thus reducing the event rate from 11.8 to6.1%. When adjusting models for the other lipids, bloodglucose and diastolic blood pressure, results were againvery similar to the main analysis. The same holds truefor the analysis both excluding sub-study participantsand considering fatal events only.

4. Discussion

The plasma triglyceride level has long been impli-cated as a risk factor for IHD [34]; however, a NationalInstitutes of Health Panel recently reviewing the evi-dence, concluded that ‘‘the data are mixed; althoughstrong associations are found in some studies, the evi-dence for a causal relationship is still incomplete’’ [35].Indeed, in several studies, the association was reducedand ceased to be statistically significant after adjustingfor other factors, especially HDL cholesterol, in multi-variable analyses [1,36]. These analyses did not, how-ever, consider differences in measurement precision oflipid factors. Triglyceride levels are imprecisely mea-sured because of high intra-individual variation, andthis is exacerbated if non-fasting samples are analysed.Even with fasting samples, the natural variation intriglycerides is generally found to be greater than fortotal cholesterol or HDL cholesterol [10,37]. Thedifficulty in accurately characterising the usual level oftriglyceride concentration will lead to under-estimationof its effect on the risk of IHD, introducing regressiondilution bias [7].

Triglyceride concentration correlates both with totaland HDL cholesterol. Triglyceride levels also correlatewith a number of non-lipid factors, most notably bloodglucose, blood pressure and BMI [1]. In multivariableanalysis, measurement imprecision in correlated expo-sure variables can affect the strength and, in someinstances, the direction of effects in a complex, noteasily predictable manner [11]. Statistical analyses,based on real data that take into account imprecision inthe measurement of key variables, are therefore neededwhen assessing the possible role of plasma triglyceridein IHD. Such analyses have rarely been conducted,mainly because in most cohorts, no sub-studies wereperformed to assess measurement precision.

We used the method described by Rosner et al.[28,29] to adjust logistic regression coefficients for mea-surement error. The method is essentially an extensionof regression calibration to logistic regression, an ap-proach first developed by Prentice for the Cox propor-tional hazards model [38]. The same approach was laterdescribed for use with generalised linear models [39].There are a number of assumptions [28,29,32] that needto be considered when using this method within logisticregression. Most importantly, the outcome should be

Fig. 1. 10-year relative risk estimates for ischaemic heart disease fromfour logistic regression models. Solid lines, results from modelsignoring measurement error; dotted lines, results from models consid-ering technical measurement error; and dashed lines, results frommodels considering laboratory error and within-person variability.

M. Egger et al. / Atherosclerosis 143 (1999) 275–284282

Table 6Results from sensitivity analyses: odds ratios for increase in triglyceride level by one S.D.a

Sensitivity analysis Model adjusted for Odds ratio (95% CI)

Ignoring measurement error Considering measurement error inlipid factorsin lipid factors

Laboratory andLaboratorywithin-person

Age 1.35 (1.19–1.55) 1.41 (1.21–1.63) 1.55 (1.28–1.88)Excluding sub-study participants1.22 (1.05–1.41) 1.22 (1.02–1.46)+Lipid factors 1.34 (1.07–1.68)

1.10 (0.91–1.32)1.11 (0.95–1.29) 1.17 (0.93–1.48)+Blood glucoseand diastolic BP

1.43 (1.21–1.70) 1.50 (1.24–1.81)Considering fatal events only Age 1.69 (1.31–2.17)1.23 (0.97–1.55) 1.37 (1.02–1.83)+Lipid factors 1.23 (1.02–1.50)

1.10 (0.90–1.34) 1.08 (0.85–1.37)+Blood glucose 1.16 (0.85–1.58)1.43 (1.21–1.70) 1.50 (1.24–1.81) 1.69 (1.31–2.17)AgeExcluding sub-study participants and considering

fatal events only+Lipid factors 1.23 (1.02–1.50) 1.23 (0.97–1.55) 1.37 (1.02–1.83)

1.14 (0.83–1.56)1.08 (0.88–1.33)+ Blood glucose 1.06 (0.83–1.36)and diastolic BP

a See Table 1 for the S.D.s used in these calculations.

relatively rare and the information from sub-studiesand main data should be independent. For the presentstudy, sensitivity analyses were performed to examinethe robustness of results. These analyses showed thatresults were not materially altered when non-fatal IHDevents were excluded, thus reducing the event rate tobelow 10%. Similarly, results were essentially un-changed when excluding the men who participated inlaboratory quality control and long-term repeatabilitysub-studies.

The effect of triglyceride concentration persisted afteradjusting for total and HDL cholesterol even whenignoring measurement imprecision. This is in agreementwith an earlier report which combined 5-year incidencedata from the CHDS with 3-year data from the twinstudy conducted in the town of Speedwell [17]. Takinglaboratory error and within-person variability into ac-count strengthened age-adjusted relative risk estimatesof all lipid factors. When including all lipid factors inthe model, the standardised odds ratio for triglycerideswas reduced from 1.57 to 1.35, but the effect remainedstatistically significant. The association of triglyceridesand IHD may thus partly, but not entirely, be ex-plained by total and HDL cholesterol. This is compat-ible with the complex metabolic interrelations that existbetween lipid factors. The inverse correlation betweenHDL cholesterol and triglyceride level is probably dueto HDL’s role in the clearance of triglyceride-rich parti-cles [40]. Inefficient clearance of triglycerides is charac-terised by low levels of HDL cholesterol which, in turn,will hamper the HDL-mediated reverse cholesteroltransport [41]. Reverse cholesterol transport is consid-ered to be the main factor responsible for the protectiveeffect of high HDL cholesterol levels. Furthermore,

exchange of VLDL triglycerides with LDL cholesterylesters and subsequent hydrolysis of triglycerides leadsto a reduced size of LDL particles which, in turn, willincrease the risk of IHD [42]. Hypertriglyceridaemiamay thus, to some extent, act through affecting HDLcholesterol levels and LDL particle size. On the otherhand there is evidence that hypertriglyceridaemia af-fects both blood coagulation and fibrinolytic function[43,44], a pathway which may be independent of choles-terol factors.

Our results may well underestimate the role oftriglyceride concentration in the aetiology of IHD.First, in patients with very high triglyceride levels, alarge proportion of the total cholesterol is carried byVLDL particles which also contain most of the triglyc-eride [16]. In multivariable analysis, the total triglyce-ride concentration is thus, to some extent, adjusted forthe VLDL triglyceride level, although the degree ofsuch overadjustment has been shown to be only moder-ate [13]. Second, lipid measurements in the CHDS andmany other cohorts [36] were based on fasting plasmasamples. Therefore, if postprandial peaks in triglyceridelevels are more important than fasting levels, the influ-ence of triglycerides will be underestimated. Postpran-dial lipoprotein metabolism has long been postulated toplay a role in atherogenesis [45]. Recent studies haveshown that patients with IHD have more pronouncedpostprandial triglyceride responses and higher levels ofchylomicron remnants than healthy controls [46]. Fur-thermore, postprandial triglyceride response is associ-ated with familial risk of IHD [47]. Third, IHD eventsduring 10 years of follow-up were related to triglyceridedeterminations at baseline. If triglycerides and HDLcholesterol are more important for disease progression

M. Egger et al. / Atherosclerosis 143 (1999) 275–284 283

in later stages, whereas LDL cholesterol is crucial inthe early stages of disease [48], then associations withbaseline measurements will not adequately reflect therole of the different lipid factors. Interestingly, theanalysis of 5-year incidence data showed statisticallysignificant effects of triglycerides and HDL choles-terol, whereas the effect of total cholesterol was weakand did not reach statistical significance [17].

When further adjusted for fasting blood glucose andblood pressure, the standardised odds ratio for triglyc-erides was reduced to 1.19 and became statisticallynon-significant. Previous epidemiological studies of theassociation between triglycerides and IHD have gener-ally not considered fasting blood glucose as a covari-ate [36], although an earlier analysis of the Caerphillyand Speedwell studies had shown the importance ofhyperglycaemia [49]. In our analysis, the model includ-ing lipid factors along with fasting blood glucose andblood pressure will, to some extent, adjust for thepresence or absence of the insulin resistance syndrome(syndrome X) [50]. This consists of a cluster of relatedrisk factors, including hypertriglyceridaemia, low levelsof HDL cholesterol and high blood pressure, a combi-nation which is strongly associated with the risk ofIHD. For example, in the Helsinki Heart Study, thesub-group of patients with high triglyceride and lowHDL cholesterol concentrations and elevated bloodpressure had a nearly fivefold increased risk of IHD[51]. Insulin resistance has been proposed as the un-derlying disturbance with dyslipidaemia and hyperten-sion being a consequence of this basic abnormality[50,52]. Our results and those from an earlier analysisof the CHDS [53] are compatible with this hypothesis,indicating that at least part of the increased IHD riskassociated with triglycerides may be due to the pres-ence of insulin resistance. Insulin sensitivity was not,however, directly measured in the CHDS.

Our findings are similar to those from a recentanalysis of the Lipid Research Clinics Follow-upStudy [13]. In this analysis, measurement imprecisionof triglyceride concentrations was reduced by takingthe mean of triglyceride measurements from two sepa-rate visits. The relative risk estimate for a natural-logunit increase of triglyceride adjusted for low-densitylipoprotein (LDL) cholesterol, HDL cholesterol andother factors was 1.31, but reduced to 1.09 upon in-clusion of blood glucose into the statistical model.

In conclusion, the controversy surrounding the roleof triglycerides in IHD illustrates the difficulties ofestablishing ‘independent’ effects when risk factorsform a web of interrelations and, to some extent,share pathways in their atherogenic effect. When tak-ing measurement imprecision into account, data fromthe CHDS and the Lipid Research Clinics Follow-upStudy [12] show that the association of triglycerideswith the risk of IHD is statistically independent of

HDL cholesterol levels. However, epidemiologicaldata are unlikely to elucidate the role of triglyceridesin the aetiology of IHD. Further research onmetabolic interrelations, and cellular and molecularmechanisms is needed to develop clearly defined hy-potheses which can be tested in clinical and epidemio-logical studies.

Acknowledgements

We would like to thank everyone who participatedin this study and Peter M. Sweetnam for preparingthe datasets and many helpful comments on earlierdrafts of this paper. The analyses reported here werefunded by a grant from the Swiss National ScienceFoundation. The Department of Social Medicine atUniversity of Bristol is part of the Medical ResearchCouncil Health Services Research Collaboration.

References

[1] Hulley SB, Rosenman RH, Bawol RD, Brand RJ. Epidemiologyas a guide to clinical decisions: The association between triglyce-ride and coronary heart disease. New Engl J Med1980;302:1383–9.

[2] Fuller WA, Hidiroglou MA. Regression estimates after correct-ing for attenuation. J Am Stat Assoc 1978;73:99–104.

[3] Phillips AN, Davey Smith G. Bias in relative odds estimationdue to imprecise measurement of correlated exposures. Stat Med1992;11:953–61.

[4] Davey Smith G, Phillips AN. Confounding in epidemiologicalstudies: why ‘independent’ effects may not be all they seem. BrMed J 1992;305:757–9.

[5] Willett W. An overview of issues related to the correction ofnon-differential exposure measurement error in epidemiologicstudies. Stat Med 1989;8:1031–40.

[6] Spearman C. The proof and measurement of association betweentwo things. Am J Psychol 1904;15:72–101.

[7] McMahon S, Peto R, Cutler J, Collins R, Sorlie P, Neaton J,Abbott R, Godwin J, Dyer A, Stamler J. Blood pressure, strokeand coronary heart disease. Part 1, prolonged differences inblood pressure: prospective epidemiological studies corrected forthe regression dilution bias. Lancet 1990;335:765–74.

[8] Greenland S. The effect of misclassification in the presence ofcovariates. Am J Epidemiol 1980;112:564–9.

[9] Kupper LL. Effects of the use of unreliable surrogate variableson the validity of epidemiologic research studies. Am J Epi-demiol 1984;120:643–8.

[10] Mogadam M, Ahmed SW, Mensch AH, Godwin ID. Within-person fluctuations of serum cholesterol and lipoproteins. ArchIntern Med 1990;150:1645–8.

[11] Phillips AN, Davey Smith G. How independent are ‘indepen-dent’ effects? Relative risk estimation when correlated exposuresare measured imprecisely. J Clin Epidemiol 1991;44:1223–31.

[12] Betteridge DJ. Lipoproteins and coronary heart disease. J HumHypertens 1989;3(suppl 2):13–25.

[13] Criqui MH, Heiss G, Cohn R, Cowan LD, Suchindran CM,Bangdiwala S, Kritchevsky S, Jacobs DR Jr, O’Grady HK,Davis CE. Plasma triglyceride levels and mortality from coro-nary heart disease. New Engl J Med 1993;328:1220–5.

M. Egger et al. / Atherosclerosis 143 (1999) 275–284284

[14] Sattar N, Packard CJ, Petrie JR. The end of triglycerides incardiovascular risk assessment? Br Med J 1998;317:553–4.

[15] Riemersma RA. Coronary artery disease: raised cholesterol ortriglycerides? Int J Cardiol 1984;5:193–4.

[16] Austin MA. Plasma triglycerides as a risk-factor for coronaryheart disease. The epidemiological evidence and beyond. Am JEpidemiol 1989;129:249–59.

[17] Bainton D, Miller NE, Bolton CH, Yarnell JWG, SweetnamPM, Baker IA. Lewis B, Elwood PC. Plasma triglyceride andhigh density lipoprotein cholesterol as predictors of ischaemicheart disease in British men. Br Heart J 1992;68:60–6.

[18] Abbott RD, Carroll RJ. Interpreting multiple logistic regressioncoefficients in prospective observational studies. Am J Epidemiol1984;119:830–6.

[19] Elwood PC, Burr ML, Bainton D, Yarnell JWG, Fehily AM,Baker IA, Sweetnam PM. Epidemiological studies of cardiovas-cular diseases. In: Progress report VII. Cardiff: MRC Epidemiol-ogy Unit, 1991:1991.

[20] The Caerphilly and Speedwell Collaborative Group. Caerphillyand Speedwell collaborative heart disease studies. J EpidemiolCommunity Health 1984;38:259–62.

[21] Rose GA. Chest pain questionnaire. Millbank Memorial FundQuarterly 1965;43:32–9.

[22] Steele BW, Koehler DF, Kuba K, Azar MM. An enzymaticapproach to lipoprotein quantification. Am J Clin Pathol1980;73:75–8.

[23] Buccolo G, David H. Quantitative determination of serumtriglycerides by the use of enzymes. Clin Chem 1973;19:476–82.

[24] Lopes-Virella MF, Stone P, Ellis S, Colwell JA. Cholesteroldetermination in high-density lipoproteins separated by threedifferent methods. Clin Chem 1977;23:822–4.

[25] Kadish AH, Little RL, Sternberg JC. A new and rapid methodfor the determination of glucose by measurement of rate ofoxygen consumption. Clin Chem 1968;14:116–31.

[26] Hosmer DW, Lemeshow S. Applied logistic regression. NewYork: John Wiley, 1989.

[27] Cox DR. Regression models and life tables. J R Stat Soc1972;34:187–220.

[28] Rosner B, Spiegelman D, Willett W. Correction of logisticregression relative risk estimates and confidence intervals formeasurement error: the case of multiple covariates measuredwith error. Am J Epidemiol 1990;132:734–45.

[29] Rosner B, Spiegleman D, Willett WC. Correction of logisticregression relative risk estimates and confidence intervals forrandom within-person measurement error. Am J Epidemiol1992;136:1400–13.

[30] Bland JM, Altman DG. Statistical methods for assessing agree-ment between two methods of clinical measurement. Lancet1986;i:307–10.

[31] De Klerk NH, English DR, Armstrong BK. A review of theeffects of random measurement error on relative risk estimates inepidemiological studies. Int J Epidemiol 1989;18:705–12.

[32] Rosner B, Willett WC, Spiegelman D. Correction of logisticregression relative risk estimates and confidence intervals forsystematic within-person measurement error. Stat Med1989;8:1051–69.

[33] SAS Institute. SAS language guide for personal computers,Version 6. Cary, NC: SAS Institute, 1985.

[34] Albrink MJ, Man EB. Serum triglycerides in coronary arterydisease. Arch Intern Med 1959;103:4–8.

[35] NIH Consensus Conference. Triglyceride, high-density lipo-protein, and coronary heart disease. J Am Med Assoc1993;269:505–10.

[36] Hokanson JE, Austin MA. Plasma triglyceride level is a riskfactor for cardiovascular disease independent of high-densitylipoprotein cholesterol level: a meta-analysis of population-basedprospective studies. J Cardiovasc Risk 1997;3:213–9.

[37] Jacobs DR, Barrett-Connor E. Retest reliability of plasmacholesterol and triglyceride. The Lipid Research Clinic Preva-lence Study. Am J Epidemiol 1982;116:878–85.

[38] Prentice RL. Covariate measurement errors and parameter esti-mation in a failure time regression model. Biometrika1982;69:331–42.

[39] Armstrong B. Measurement error in generalized linear models.Commun Stat Ser B 1996;14:529–44.

[40] Patsch JR, Prasad S, Gotto AM, Patsch W. High densitylipoprotein—relationship of the plasma levels of this lipoporteinspecies to its composition, to the magnitude of postprandiallipemia, and to the activities of lipoprotein lipase and hepaticlipase. J Clin Invest 1987;80:341–7.

[41] Glomset JA. The plasma lecithin cholesterol acetyltransferasereaction. J Lipid Res 1968;9:155–67.

[42] Stampfer MJ, Krauss RM, Ma J, Blanche PJ, Holl LG, SacksFM, et al. A prospective study of triglyceride level, low-densitylipoprotein particle diameter, and risk of myocardial infarction.J Am Med Assoc 1996;276:882–8.

[43] Meade TW, Brozovic M, Chakrabarti RR, Haines AP, ImesonJD, Mellows S, Miller GJ, North WR. Haemostatic function andischaemic heart disease: principal results of the northwick parkheart study. Lancet 1986;6:533–7.

[44] Baker IA, Sweetnam PM, Yarnell JWG, Bainton D, Elwood PC.Haemostatic and other risk factors for ischaemic heart diseaseand social class: evidence from the Caerphiily and Speedwellstudies. Int J Epidemiol 1988;17:759–65.

[45] Zilversmit DB. Atherogenesis: a postprandial phenomenon. Cir-culation 1979;60:473–85.

[46] Slyper AH. A fresh look at the atherogenic remnant hypothesis.Lancet 1992;340:289–91.

[47] Uiterwaal CSPM, Grobbee DE, Witteman JCM, van StiphoutW-AHJ, Krauss XH, Havekes LM, de Bruijn AM, van Tol A.Postprandial triglyceride response in young adult men and famil-ial risk for coronary atherosclerosis. Ann Intern Med1994;121:576–83.

[48] Sharrett AR, Patsch W, Sorlie PD, Heiss G, Bond MG, DavisCE. Associations of lipoprotein cholesterols, apolipoproteins A-Iand B, and triglycerides with carotid atherosclerosis and coro-nary heart disease. Arterioscler Thromb 1994;14:1098–104.

[49] Yarnell JW, Pickering JE, Elwood PC, Baker IA, Bainton D,Dawkins C, Phillips DI. Does non-diabetic hyperglycaemia pre-dict future IHD? Evidence from the Caerphilly and Speedwellstudies. J Clin Epidemiol 1994;47:383–8.

[50] Reaven GM. Role of insulin resistance in human disease. Dia-betes 1988;37:1595–607.

[51] Tenkanen L, Pietila K, Manninen V, Manttari M. The triglyce-ride issue revisited. Findings from the Helsinki Heart Study.Arch Intern Med 1994;154:2714–20.

[52] Garg A. Insulin resistance in the pathogenesis of dyslipidemia.Diabetes Care 1996;19:387–9.

[53] Yarnell JWG, Sweetnam PM, Marks V, Teale JD, Bolton CH.Insulin in ischaemic heart disease: are associations explained bytriglyceride concentrations? The Caerphilly prospective study. BrHeart J 1994;71:293–6.

.