Study subjects and characteristics of family members.

Study subjects for this study were derived from the Gilda Radner Familial Ovarian Cancer Registry (formerly referred to as the Familial Ovarian Cancer Registry). The Gilda Radner Familial Ovarian Cancer Registry (GRFOCR) is a self-referred Registry of families with two or more ovarian cancer cases in blood relatives. It was established in 1981 at the Roswell Park Cancer Institute by Dr. M. Steven Piver to study the incidence of familial ovarian cancer [26]–[29]. The primary function of the Registry is to receive family cancer information voluntarily contributed throughout the United States by ovarian cancer patients, referring and concerned physicians, concerned women, and patients of the Roswell Park Cancer Institute (RPCI) Gynecologic Oncology Department. The objectives of the Registry include (i) obtaining detailed family histories from individuals who are apparently from families with two or more cases of ovarian cancer or a syndrome possibly related to ovarian cancer; (ii) documenting through medical records and through pathologist review of tissues the occurrence of cancer; (iii) collecting, processing and storing biological samples, when possible, from Registry participants; and (iv) making the information and biological samples available for research under Institutional Review Board approved research protocols.

Recruitment of subjects.

Registry participants are recruited through probands or index persons meeting at least one of the following criteria: (i) family history of two or more cases of ovarian cancer; (ii) family history of one case of ovarian cancer and two cases of cancer at any other site; (iii) family history of at least one female with two or more primary tumors with one of the primaries being ovarian cancer; (iv) family history of two or more cases of cancer with at least one case being ovarian cancer, and the other cancer considered to be of early onset (≤45 years old). In addition to meeting at least one of the preceding criteria, each participant is required to sign a consent form. Individuals unable to give consent as a result of mental, intellectual, or cognitive deficits are excluded, however, such could appear in the Registry through collection of family history data but without collection of bio-samples from them. Subjects for this study consisted of 7669 adult members from 1919 different pedigrees from the GRFOCR. The families included in the present study comprised of 1412 families ascertained to have two or more cases of ovarian cancer; 17 families with one case of ovarian cancer and two cases of cancer at other sites; and 490 families in the category of those having at least one female with more than one primary tumors with one of the primaries being ovarian cancer or having more than one case of cancer with at least one case being ovarian cancer and the other cancer considered to be of early onset.

Data collection.

Data are collected through Family History Forms completed by subjects. In addition, subjects give authorization to release medical records and archival tissues (where available). Both the completed Family History Form and retrieved medical records are reviewed to ascertain eligibility before subjects are entered into the Registry. Information collected on individuals who do not meet inclusion criteria at the time of collection are not entered into the Registry, and are either destroyed, placed in an inactive locked file, or returned to the individuals upon request. Following formal entry into the Registry, the individuals are provided with an epidemiologic survey form for collection of detailed epidemiologic data and a blood donation form for biosample collection. Permission to invite relatives is also requested from Registry participants and letter of introduction sent to relatives for whom permission to invite is granted. Invited relatives who accept to participate are also asked to sign a consent form after which they are asked to complete voluntarily all necessary data and biosample collection forms.

Construction of pedigrees.

Based on information collected through the Family History Form, we established family and pedigree relationship of every subject. Pedigrees used in this study were constructed from relatives ascertained through probands using computer codes written and implemented in SAS [30] as macros and the resulting established pedigree relationships were checked and corrected for possible errors using MADELINE [31] software package. Where necessary, dummy individuals were added to families for the purpose of connecting relatives within pedigrees, and the affection status for such dummy individuals was set to missing and thus they were not used in the analyses. A total of 7669 real pedigrees members and 6647 connecting dummy persons were included in this study.

Statistical analysis.

In the present study, data used included information on (i) sex, (ii) ovarian cancer affection status defined as affected, unaffected or unknown; (iii) information on affection status for other cancer site such as breast, pancreas and uterus/endometrium, also defined as affected, unaffected or unknown; and (iv) family/pedigree relationships. Estimation of the distributions of relationship types and ovarian cancer affection status among relationship pairs were performed using the Statistical Analysis for Genetic Epidemiology (SAGE) program PEDINFO, version 5.2 [32]. Although analyses were constrained to female pedigree members, male relatives had to be included for the purpose of defining pedigree relationships.

To account for proband ascertainment, ascertainment correction was applied in all segregation analyses by conditioning each pedigree's likelihood on the affection status of the proband.

Segregation Analysis.

To explore the mode of familial transmission of susceptibility to ovarian cancer, we performed complex segregation analysis using the maximum likelihood method to estimate the parameters in each of the hypothesis-based mathematical models examined. Since the presence of a BRCA1 or BRCA2 mutation status does not preclude the presence of additional susceptibility gene(s) which could contribute to disease penetrance, mutation positive patients were therefore not excluded from this analysis. The SEGREG program of SAGE, version 5.2 [32] under Linux operating system was used to fit each model. For each model, we assumed that the presence (or absence) of the putative disease allele influences susceptibility to ovarian cancer, and then applied the regressive multivariate logistic model for binary trait as described by Karunaratne and Elston [33]. This approach enabled us to include available covariates of interest in the fitted models. The fitted models assumed that, conditional on the phenotype and the major type of any individual who belongs to two nuclear families, the likelihoods for those two nuclear families are independent. Therefore, the marginal probability (or susceptibility) that any pedigree member has a particular phenotype is the same for all members who have the same values of any covariates in the model. This susceptibility given by the cumulative logistic function aswhere y_i is the affection status phenotype of the i-th individual; θ_(i) is the logit of the susceptibility for the i-th individual which is defined aswhere β is the baseline parameter; g is the latent genetic “type” [34] or “ousiotype” [35]; and X is the covariate vector. Under a major locus model with two alleles, A and B, A being the susceptibility allele, the three types correspond to genotype g = AA, AB or BB transmitted according to Mendelian mode. The corresponding baseline parameters for susceptibility are then The transmission parameters represented as in each model are the conditional probabilities that a parent of a given genotype transmits the susceptibility allele A to the offspring [36], [37]. The transmission parameters and the allele frequency parameter, for the susceptibility allele are estimated alongside the three baseline parameters for susceptibility in each model depending on the specified assumption. For example, under the assumption of Mendelian inheritance, the transmission parameters are constrained to . The following sporadic, environmental and genetic models were considered in assessing type of familial association and possible evidence of transmission of major effect.

1. Sporadic or no major gene model. In this model, both familial association (FA) (i.e. father-mother (FM), mother–offspring (MD), father–offspring (FO) or sibling (SS)), and transmission of major gene (MG) (i.e. ) are not assumed. There is only one baseline parameter (i.e. ) which is interpreted as the natural logarithm of the odds of susceptibility versus non-susceptibility to ovarian cancer in the absence of other factors.
2. Sporadic model with familial association. This model includes estimation of parameters for familial association (parent-offspring (PO) and sibling) in the absence of transmission of a major gene. Three different models – first with only parent-offspring parameter, second with only sibling parameter, and third with both parent-offspring and sibling parameters, were fitted.
3. Major gene without familial association model. This model assumes the transmission of a major gene but no familial association. Here the susceptibility allele frequency is estimated while the transmission parameters are constrained to the Mendelian mode.
4. Mendelian codominant model with familial association. This assumes transmission of a major gene and familial association. In this and all Mendelian models tested in this study, transmission parameters were constrained to the Mendelian mode as . The allele frequency, familial or residual associations, and baseline susceptibility, were estimated in this and the other Mendelian models tested.
5. Mendelian dominant model which is similar to codominant model above, except that baseline susceptibility parameters for genotypes AA and AB are constrained equal as
6. Mendelian recessive model in which baseline susceptibility parameters for genotypes AB and BB are constrained equal to each other as
7. Mendelian additive model in which baseline susceptibility parameter for genotype AB is constrained to be intermediate of those of AA and BB as .
8. Mendelian decreasing model includes the assumption of decreasing susceptibility with the maximum and minimum susceptibility parameters constrained to genotypes AA and BB, respectively, as .
9. Mendelian increasing model is the reverse of the above decreasing model with the baseline susceptibility parameters constrained as
10. Environmental model in which no transmission of susceptibility allele is assumed. This is a non-Mendelian transmission model in which all the transmission probabilities are set equal to the allele frequency as but all three susceptibility parameters are estimated. This model assumes that the observed familial association and segregation are both due purely to non-transmissible environmental effects and not any major gene.
11. Tau AB free model in which transmission parameters for genotypes AA and BB are constrained to 1 and 0, respectively, while parameter for AB is estimated within 0–1 range as .
12. General non-Mendelian model in which all parameters are estimated. As a result, all other models are nested in the general or full model and thus the general model is used as the baseline to compare all other models in this study.

In the above models 3 to 11 where major gene is assumed, we also assumed that the genotype frequencies are in Hardy-Weinberg proportions (i.e, ). Also, because ascertainment was through probands, we corrected for ascertainment bias in each model by conditioning the likelihood of each pedigree on the affection status of the proband.

For testing the different hypotheses represented by the models, we used the likelihood ratio test (LRT). Since the models are hierarchical, we tested each submodel against the general model by using the test statistic computed as minus twice the difference between the natural log likelihood of the general model and that of the specific submodel. This statistic is asymptotically distributed as chi-square distribution with degree of freedom equal to the difference in the number of parameters estimated in both models. Using this test, a significant chi-square indicates that the submodel tested can be rejected at the given alpha level, which means the hypothesized model does not fit the data. For comparison of non-nested models such as sporadic models or Mendelian models, we used the Akaike Information Criterion (AIC) values [38] to select the most parsimonious model for the data. The AIC for any model is defined as [–2ln(L)+2(number of parameters estimated in the model)]. The model with the smallest AIC is judged the best-fitting model for the data.