Comparison of family history measures used to identify high risk of coronary heart disease

344 Silberberg et al. Genetic Epidemiology 16:344–355 (1999)

© 1999 Wiley-Liss, Inc.

Comparison of Family History MeasuresUsed to Identify High Risk of CoronaryHeart Disease

Jonathan Silberberg, 1,3* Jayne Fryer, 1 John Wlodarczyk, 1

Randall Robertson, 1 and Keith Dear 2

1Discipline of Medicine, University of Newcastle, Callaghan, NSW, Australia2Department of Statistics, University of Newcastle, Callaghan, NSW, Australia3Department of Cardiovascular Medicine, John Hunter Hospital, Newcastle,NSW, Australia

We examined 15 published continuous family history measures (scores) as well astwo new formulations in terms of several desirable properties. We applied the scoresto sample pedigrees and found that some systematically increase with family size. Incontrast to aggregate scores, non-aggregate scores are sensitive to the age, sex, andcovariate status of individual relatives but are unstable when the families are small.We also applied these scores to our own population case-control data, characterisedby a high proportion of missing and false-negative responses. In these small fami-lies, all scores provided significant discrimination between CHD cases and controlsbeyond the usual categorical definition of positive family history, but appeared nobetter than detailed categorical definitions or even simple counts. Our new formula-tions offer no solution to the problems of few data; most scores apply asymptoticapproximations to differences between observed and expected number of affectedrelatives and are not suited to small families. All scores would be improved by rul-ing out families with only one affected relative, as is being done in the NHLBIFamily Heart Study. We recommend that researchers, when using a family historymeasure, consider the number of informative families and other characteristics oftheir data prior to choosing any particular formulation. Genet. Epidemiol. 16:344–355, 1999. © 1999 Wiley-Liss, Inc.

*Correspondence to: Jonathan Silberberg, Department of Medicine, John Hunter Hospital, Locked Bag1, Hunter Region Mail Centre, Newcastle, New South Wales, 2310, Australia. E-mail:[email protected]

Contract grant sponsor: Commonwealth Department of Human Services and Health Research and De-velopment; Contract grant number: HS 335; Contract grant sponsor: University of Newcastle SenateResearch Committee.

Received 21 April 1997; Revised 20 November 1997; Accepted 16 February 1998

Comparison of Family History Measures 345

Key words: familial aggregation; family history score; coronary heart disease

INTRODUCTION

There has been a large body of literature on family history measures (or scores).In the case of coronary heart disease (CHD), many investigators use dichotomousmeasures that do not consider the family size, number of affected relatives, or rela-tives’ age and risk factor profile [Hunt et al., 1986; Silberberg et al., 1998b]. A con-tinuous score developed in Utah [Williams et al., 1984] is being used in the NHLBIFamily Heart Study [Higgins et al., 1996], a genetic epidemiology study of CHD.

This paper reviews family history scores in the medical literature. We discussdesirable properties for most purposes, and present two new score formulations forconsideration. We applied all scores to sample pedigrees and also tested the dis-crimination afforded between cases and controls in the Newcastle Family HistoryStudy [Silberberg et al, 1998a,b].

DESIRABLE PROPERTIES OF A FAMILY HISTORY SCORE FORCOMMON DISEASE

Figure 1 describes several pedigrees to illustrate the following features:

• a score should consider the risk profile of a family, taking into account therisk profile of each relative in terms of relevant covariates (age, sex, and smok-ing status). Because early onset is characteristic of familial cases of disease,such pedigrees should have a higher score (compare Fig. 1a and b). Likewise,since CHD affects women at a later age than men, families with an affectedmother should have a larger score (compare Fig. 1g and h).

• a score should be robust to family size and time at risk, not changing system-atically with the number of relatives. Pedigrees with the same proportion ofrelatives affected should have a similar score (compare Fig. 1c and d); like-wise small pedigrees should be considered uninformative rather than protec-tive (compare Fig. 1e and f).

• a score should consider relationship of relatives; close blood relatives beinggiven more weight.

• a score should not be inflated by a single individual; small families are particu-larly vulnerable to the influence of a single affected member. The simplest solu-tion is to exclude these from score calculation and rely on categorical definitions.

METHODS

Literature Review

We searched MEDLINE (1966–1996) and bibliographies of retrieved manuscriptsfor original studies. The scores identified differed both in the information required foreach relative and the computation of the expected risk of disease. For comparison ofscores, we used the same dataset and the same expected risk of disease.

346 Silberberg et al.

Fig. 1. Sample family trees describing desirable properties of a family history score. Contrast betweenleft-hand side and right-hand side the following features: a, b: age of onset; c, d: family size; e, f: informa-tiveness; g, h: sex of affected relative. m female not affected; q male not affected; l female affected; nmale affected; text time at risk (years); smoking status indicated; è indicates proband.


Fitting Various Scores to CHD Case/Control Data

Cases and controls completed a self-administered questionnaire [Silberberg etal., 1998a]. For the score analysis we excluded second-degree relatives as they addedlittle information due to the large amount of missing data in this study [Silberberg etal., 1998b], relatives aged less than 20 years as reported CHD is most often amisclassification [Hunt et al., 1986], and also half-blood and non-blood relatives.Only those for whom gender, disease status, and time at risk were known were in-cluded. The missing data reduced the information to an average of 3.6 relatives perproband, most families having only one affected relative (mean 1.2, range 0 to 6).

Definitions

Disease status was determined by the reported lifetime CHD status or the re-ported cause of death. The disease status for relative i in the family is referred to asoi, where oi = 1 if affected and 0 if not affected.

Time at risk for each relative was defined as age at diagnosis if diseased, cur-rent age if not diseased and alive, or age at death if unaffected and deceased. Eachindividual contributed only disease-free life experience [Hunt et al., 1986].

The expected risk of disease was calculated using Hunter Region CHD inci-dence rates reported by the MONICA project according to age, gender, and smok-ing-status [Chun et al., 1993]. The expected risk for relative i in the family is referredto as ei [Lubin and Bale, 1987, used pi]. We defined the expected risk for a relative

with t years at risk as ei = 1– y

t

=∏

1(1–qy) where qy is the gender- and smoking-spe-

cific incidence rate at age y. This is constrained to less than one, in contrast to theusual calculation (age-specific time at risk multiplied by the population age-specificincidence rates), which is actually the expected number of events and can exceedone [Lubin and Bale, 1987].

New Formulations

This paper introduces two non-aggregate scores, both of which consider therisk status of affected relatives. The first uses external rates for comparison, derivedfrom the Hunter population [Chun et al., 1993]; the second uses rates derived fromthe study sample.

This first score (FHS-1), is defined as [ ( ) / ] / [ ]w o e e wi i i i ii

n

i

n

−==∑∑

11 with wi =

1 for second-degree and wi = 2 for first-degree relatives. Note that if the score is

based on only first-degree relatives then wi = 2 for all i and so the equation simpli-

fies to [ ( ) / ] /o e e ni i ii

n

−=∑

1 ; where n is the number of relatives. This formulation

differs from that of Schwartz et al. [1988] in dividing by n, thus adjusting for familysize. An affected relative at low risk of CHD contributes a large positive value to thisscore; conversely, high-risk affected relatives make only a small contribution. If un-affected, the negative contribution of high-risk relatives is greater than that of low-risk relatives.

348S

ilberberg et al.

TABLE I. Family History Scores*

Desirable properties

Considers Robust to Considersrisk profile family size relationship

Referencea Score formulationb of family and time at risk of relatives Comments

Slack and Evans [1996] Σoi/Σei Yes Yes N/A Comparable to the standardisedmortality ratio (SMR)

Chakraborty et al. [1984] Σ (oi – ei)2 /[ei(1 – ei)] Yes No N/A Non-aggregate score, discussed

by Lubin and Bale (1987)Williams et al. [1984] [(|Σoi – Σei| – 1/2)* |Σoi–Σei|] / Yes No Yes If only one affected relative, the

[√Σei *(Σoi–Σei)] score is reset to 0.99; reset to zeroif |Σoi – Σei|<=1/2

2:1 first:second-degree weightFain and Goldgar [1986] (Σ oi – Σei)/√[Σei(1 – ei)] Yes Yes N/ALynch et al. [1986] (Σ oi – Σei)/Σei Yes Yes NoReed et al. [1986] (a) (Σoi – Σei)/√Σei Yes No No ei based on age, sex, and calendar-

specific rates(b) Wdeg1*(Σoi – Σei)/√Σei + Yes No Yes Separate scores for first- andWdeg2(Σoi– Σei)/√Σei second-degree relatives and

weighted 2:1Schartz et al. [1988] (a) (Σoi –Σei)/√Σei Yes No No Same as Reed et al. [1986]

(b) Σ(oi – ei)/√ei Yes No No Similar to FHS-1, not adjusted forfamily size

Groeneveld and 1/n * Σ(oi – ei)log[(1 – ei)/ei] Yes Yes NoHitzeroth [1991]

Kee et al. [1991] (Σ oi – Σei)/[Σei√Σei] Yes No N/AMukherjee et al. [1994] 0 = none affected; . . . . ; No No Yes Rank of 0 to 9Van Esch et al. [1994] (a) Σoi/n No Yes N/A Simplest account for family size

(b) Σoi/Σei Yes Yes N/A Same as Slack and Evans [1966]Kerber [1995] (a) Σoi/Σti Yes Yes No

(b) Σwioi/Σwiti Yes Yes Yes 2:1 first:second-degree(c) Σwioi/Σwiei Yes Yes Yes

aArranged in order of publication/bNotation for relative i: oi = disease status (zero, if not affected, or one if affected); ei = expected risk of disease; ti = time at risk; wi=weight; Σ =

i

n

−∑

1where n is the

number of relatives in the score.*N/A = not applicable if score was only applied to first degree relatives.


Fig. 2. Contribution to score made by relatives; non-aggregate measures only.


The second score (FHS-2) is based on the residuals of a logistic regressionmodel, in which we combined all relatives in the dataset and fitted relative’s diseasestatus as a function of their age, sex, and smoking status. The residuals were thenaveraged over each family to form a score [FHS-2 j = Σ(πij - π^ ij) /n; where πij refersto the i th individual in the j th family].

Statistical Methods

We tested the additional information afforded by each score beyond the simpledefinition “one or more first degree relatives affected before 60 years of age” in aseries of nested models [Miettinen, 1985]. Our reference model had disease status asthe outcome variable and the simple family history as the explanatory variable, withcovariates age, sex, and smoking status. We calculated the improvement in model fitas each of the scores were added, i.e., the importance of the score was examined bycomparing the model, which included both the score and the simple definition withthat including the simple definition alone.

RESULTSLiterature Review

Twelve publications reported fifteen unique scores. Most are formulated as ag-gregate measures with numerators of the form (observed minus expected number ofaffected relatives) and differing in the denominator: e.g., Σei, √Σei, (Σei × √Σei). Ag-gregate measures, by definition, do not relate the disease status for each relative tothat relative’s expected risk and so are insensitive to which relative has the disease.The alternative, non-aggregate measures, are sensitive to the risk factor status of anaffected relative, but tend to be unstable when family size is small.

We assessed how well the scores satisfied the previously noted desirable prop-erties (Table I).

• Risk profile of family: Most aggregate scores consider this in some part, de-riving the expected risk of disease (ei) from among age, sex, calendar, andrace-specific rates [Schwartz et al., 1988].

• Robust to family size and time at risk: Some aggregate scores [Williams etal., 1984; Reed et al., 1986; Kee et al., 1991] return higher values with largerfamilies. Of more significance is the impact of a single affected individual,since high scores can result when the family is small. The formulation ofWilliams et al. [1984] resets the score for such pedigrees to 0.99. All non-aggregate scores allow an affected “low risk” relative to overwhelm the scorewhen the number of relatives is small. Figure 2 shows a comparison betweenChakraborty et al. [1984], Groeneveld et al. [1991], Schwartz et al. [1988],and FHS-1. The former two weight affected and unaffected relatives approxi-mately equally; the latter two weight an affected relative at low risk greaterthan an unaffected relative at high risk.

• Consider relationship of relatives: Most scores assign a weight of 2 to first-de-gree and 1 to second-degree relatives. Kerber et al. [1995] used a weighting schemebased on the kinship coefficient, which measures the probability that the probandshares a gene with the relative by common descent [Malecot et al., 1969].

Com

parison of Fam

ily History M

easures351

TABLE II. Comparison of Scores Applied to Sample Pedigrees

PedigreeMedian [iqr] in family

1(a) 1(b) 1(c) 1(d) 1(e) 1(f) 1(g) 1(h) history study cases

Slack and Evans [1996] 9.6 3.2 9.6 9.6 0 0 1.25 1.25 3.5 [1.2,7.9]Chakraborty et al. [1984] 161 21 161 323 7.7 0.04 6.8 5.4 11.8 [3.0,42.0]Williams et al. [1984) 3.9 1.6 3.9 6.2 –1.3 0 0 0 0.75 [0,1.9]Fain and Goldgar [1986] 9.8 2.6 9.8 9.8 –2.2 –1.0 0.3 0.3 3.3 [0.3,7.8]Lynch et al. [1986] 8.6 2.2 8.6 8.6 –1.0 –1.0 0.25 0.25 2.5 [0.2,6.9]Reed et al. [1986] 4.8 2.1 4.8 6.8 –1.6 –0.2 0.2 0.2 1.5 [0.2,3.1]Schwartz et al. (b) [1988] 20.8 6.0 20.8 41.6 –4.2 –0.3 0.5 0.2 3.1 [0.25,7.8]Groeneveld and Hitzeroth [1991] 1.85 0.7 1.85 1.85 0.06 0.06 0.06 0.02 0.7 [0.05,1.6]Kee et al. [1991] 15.5 2.2 15.5 10.9 –0.6 –4.8 0.28 0.28 3.9 [0.25,14.6]Van Esch et al. (a) [1994] 0.5 0.5 0.5 0.5 0 0 0.17 0.17 0.4 [0.18,0.67]Kerber (a) [1995] 0.009 0.007 0.009 0.009 0 0 0.003 0.003 0.006 [0.003,0.011]Kerber (b) [1995] 0.009 0.007 0.009 0.009 0 0 0.002 0.002 0.006 [0.002,0.010]Kerber (c) [1995] 11.8 3.2 11.8 10.5 0 0 0.78 0.78 3.5 [1.1,7.6]FHS-1 3.5 0.1 3.5 3.5 –0.5 –0.1 0.09 0.04 1.06 [0.07,2.6]FHS-2 0.55 [–1.03,2.0]


• Effect of missing data: Some studies have classified relatives for whom dis-ease status is unknown as not having the disease. Although random missingdata has been suggested [Hitzeroth et al., 1992], we have found that “posi-tive” reporting is more likely for affected relatives [Silberberg et al., 1998a]and missing data are best excluded.

Sample Pedigrees: How the Scores Perform

The scores were applied to the pedigrees shown in Figure 1 and also to CHDfamilies in the Family History Study (Table II). It can be see that none of the scoressatisfactorily distinguishes between pedigrees in Figure 1g and h, i.e., an affectedmother rather than father, or between pedigrees in Figure 1e and f, i.e., uninformativerather than negative families. Most (11 of 14) return higher values for pedigrees inFigure 1a than b, i.e., are sensitive to younger age of onset when the family is nottoo small. However, four also return higher values for pedigrees in Figure 1d than c,i.e., systematically increase with family size.

Fitting Scores to CHD Case/Control Data

All unique scores were evaluated except that formulated by Mukherjee [1994],which is based on ranks; fifteen in total. Receiver operating characteristic (ROC)curves (Fig. 3) show that the scores are similar in discriminating between cases andcontrols in this data set. Most improved the fit of the model beyond the simple defi-nition “one or more first-degree relatives affected before 60 years” (Table III). Al-though FHS-2 appears to fit best, the expected values were data-derived and notvalidated against external data. FHS-1, adjusting for family size, improved the fitbetter than Schwartz et al. [1988], the formulations being otherwise identical. The

Fig. 3. Receiver operating characteristic curves for the fifteen scores.


TABLE III. Improvement in Fit Beyond Usual Family History Definition With Addition ofFamily History Score Term*

Improvement in model fit [d.f.=1]

Family history score χ2 p

Slack and Evans [1964] 3.97 0.05Chakraborty et al. [1984] 0.47 0.49Williams et al. [1984] 7.99 0.005Fain and Goldgar [1986] 4.87 0.03Lynch et al. [1986] 3.97 0.05Reed et al. [1986] 9.32 0.002Schwartz et al. (b) [1988] 0.63 0.43Groeneveld and Hitzeroth [1991] 9.69 0.002Kee et al. [1991] 0.49 0.48Van Esch et al. (a) [1994] 10.00 0.002Kerber (a) [1995] 11.54 0.001Kerber (b) [1995] 11.30 0.001Kerber (c) [1995] 2.77 0.10FHS-1 6.01 0.01FHS-2 11.61 0.001

*One or more first-degree relatives affected before 60 years of age.

formulation of Van Esch [1994], a simple count adjusted for family size, performedas well as any with these data.

DISCUSSION

Our main recommendation is that the choice of a particular score should de-pend on the purpose of the study and characteristics of the data. In the Family HeartStudy [Higgins et al., 1996], the formulation of Williams et al. [1984] is in use toidentify high-risk families in three populations. Families with a single affected rela-tive are ruled out and entry is restricted to more informative families, with scoresabove the 90th centile. For limited study samples such as ours, this criterion effec-tively eliminates most of the data.

We found a high prevalence of missing information among relatives (27%),compared to only 4% in Utah [Hunt et al., 1986] and 8% in Detroit [Schwartz et al.,1988]. We were aware that a high proportion of older relatives reported not to haveCHD did have a diagnosis of CHD recorded on the death certificate [Silberberg etal., 1998a]; such false negative reporting in “high-risk” relatives may falsely offsetthe contribution of an affected younger relative. Accordingly, we formulated FHS-1to give less weight to unaffected “high-risk” relatives (Fig. 2).

Using sample pedigrees, we found that published scores differ in their sensitiv-ity to relevant covariates. With three or more affected relatives, most formulationswere responsive to the age and sex of the family overall, with non-aggregate scoresbeing more sensitive to the risk profile of individual relatives. However, with onlyone affected relative, non-aggregate scores tended to return high values that wereextremely sensitive to individual risk factors. No score offers a solution to the prob-lems of few data; all apply asymptotic approximations to the (observed minus ex-pected) differences and return higher values than would exact tests. All formulations


would be improved by ruling out such families since any estimator based on a singleaffected relative is unstable.

In our population, all published scores performed similarly in distinguishingbetween cases and controls. This reflects our small family size and few affectedrelatives, in that all formulations were disadvantaged to the same extent. About 20%of families had two or more affected relatives and the scores tended to capture thisinformation. It is worth noting that the formulation of Van Esch et al. [1994], simplythe proportion of relatives affected, without regard to age or sex, adequately de-scribed these data.

We recommend that other researchers who use family history scores shouldconsider characteristics of their data before selecting any particular formulation. Whenfamilies are small and affected relatives are few, categorical definitions or simplecounts are likely to be adequate. For more informative families, the choice of scoreformulation needs to take into account factors such as the frequency of false-positiveor -negative disease reporting and variation in family size. We have found that nosingle score is “best” for all situations.

ACKNOWLEDGMENTS

We dedicate this work to Roger Williams and thank Steve Hunt for encourage-ment and discussion during the early phases of this project.

REFERENCES

Chakraborty R, Weiss KM, Majumber PP, Strong LC, Herson J. 1984. A method to detect excess riskof disease in structured data: Cancer in relatives of retinoblastoma patients. Genet Epidemiol1:229–244.

Chun BY, Dobson AJ, Heller RF. 1993. Smoking and the incidence of coronary heart disease in anAustralian population. Med J Aust 159:508–512.

Fain PR, Goldgar DE. 1986. A nonparametric test of heterogeneity of family risk. Genet Epidemiol1:(Suppl)61–66.

Groeneveld H, Hitzeroth H. 1991. Quantification of familial predisposition for disease conditions. SAfr Statist. J 25:45–60.

Higgins M, Province M, Heiss G, Eckfeldt J, Ellison RC, Folsom AR, Rao DC, Sprafka JM, WilliamsR. 1996. NHLBI Family Heart Study: objectives and design. Am J Epidemiol 143:1219–1228.

Hitzeroth HW, De Villiers LS, Groeneveld HT. 1992. Family history as a risk factor for coronary heartdisease in South African families. S Afr Med J 81:7–11.

Hunt SC, Williams RR, Barlow GK. 1986. A comparison of positive family history definitions fordefining risk of future disease. J Chron Dis 39:809–821.

Kee F, Collins BJ, Patterson CC. 1991. Prognosis in familial non-polyposis colorectal cancer. Gut32:513–516.

Kerber RA. 1995. Method for calculating risk associated with family history of a disease. GenetEpidemiol 12:291–301.

Lubin JH, Bale SJ. 1987. On the detection of excess disease risk in family data. Genet Epidemiol4:447–456.

Lynch HT, Kimberling WJ, Biscone KA, Lynch JF, Wagner CA, Brennan K, Mailliard JA, Johnson PS,Soori JS, McKenna PJ. 1986. Familial heterogeneity of colon cancer risk. Cancer 57:2089–2096.

Malecot G. 1969. The mathematics of heredity. San Francisco: WH Freeman.Miettinen OS. 1985. Theoretical epidemiology. New York: John Wiley and Sons, p 222.Mukherjee M, Dastur DK. 1994. Atherogenic markers in the offspring vis-a-vis family history of coro-

nary heart disease. Int J Cardiol 44:65–71.


Reed T, Wagener DK, Donahue RP, Kuller LH. 1986. Young adult cholesterol as a predictor of familialischemic heart disease. Prev Med 15:292–303.

Schwartz AG, Boehnke M, Moll PP. 1988. Family risk index as a measure of familial heterogeneity ofcancer risk. A population-based study in metropolitan Detroit. Am J Epidemiol 128:524–535.

Silberberg JS, Wlodarczyk J, Fryer J, Ray C, Hensley MJ. 1998a. Correction for biases in a population-based study of family history and coronary heart disease: The Newcastle Family History StudyI. Am J Epidemiol 147:1133–1139.

Silberberg JS, Fryer J, Wlodarczyk J, Hensley MJ. 1998b. Risk associated with various definitions of afamily history of coronary heart disease: The Newcastle Family History Study II. Am J Epidemiol147:1123–1132.

Slack J, Evans KA. 1966. The increased risk of death from ischaemic heart disease in first degreerelatives of 121 men and 96 women with ischaemic heart disease. J Med Genet 3:239–257.

Van Esch A, Steyerberg EW, Berger MY, Offringa M, Derksen Lubsen G, Habbema JD. 1994. Familyhistory and recurrence of febrile seizures. Arch Dis Child 70:395–399.

Williams RR, Dadone MM, Hunt SC, Jorde LB, Hopkins PN, Smith JB, Ash KO, Kuida H. 1984. Thegenetic epidemiology of hypertension: a review of past studies and current results for 948 per-sons in 48 Utah pedigrees. In: Rao DC, Elston RC, Kuller LH, Feinleib M, Carter C, Havlik R,editors. Genetic epidemiology of coronary heart disease: past, present and future. New York:Alan R Liss, p 419–442.

Documents

Comparison of family history measures used to identify high risk of coronary heart disease