Statistical analysis

Characteristics of the GenSalt study participants at baseline and during follow-up were summarized as means ± SD or percentages, as appropriate. Due to the longitudinal, family-based GenSalt design, mixed-effect regression models were used to test the associations between SNPs and longitudinal BP changes and hypertension incidence.^28,29 Autoregressive and compound symmetry covariance matrices were used to accommodate the correlations of repeated measurements within individuals and of individuals within families, respectively. For the examination of additive associations between SNPs and longitudinal BP changes, a genotype by follow-up time interaction term and the main effects of these variables were included in the models as fixed effects. Additionally, models were adjusted for the fixed effects of age, gender, and BMI using the PROC MIXED procedure in SAS (version 9.3; SAS Institute, Cary, NC). The following mixed effects model was employed:

where, γ _ijk represents the BP value for the ith individual in the jth family at the kth visit; β ₀ is the mean BP after accounting for covariates, genetic effects, and the interaction term; the terms age _ij, gender _ij, and BMI _ij represent baseline age, gender, and BMI of the ith individual in the jth family, respectively; SNP _ij models the genetic main effect, where the genotype is coded under an additive model; time _ijk is the follow-up time from baseline for the ith individual in the jth family at the kth visit; and the SNP _ij × time _ijk term is the linear interaction between follow-up time and genetic effects. This term was modeled as a fixed effect parameter and used to test the association of each SNP with longitudinal BP change. The random effects terms a _i and b _ij account for the relatedness among individuals in the same family as well as the correlations of repeated measurements among individuals nested within families. The last term represents the residual.

For examination of the additive associations between SNPs and hypertension incidence, a multilevel logistic regression model was employed after excluding 173 participants with hypertension at baseline using the PROC GLIMMIX procedure in SAS.³⁰ Hypertension incidence was modeled as:

where γ _ijk represents hypertension status of the ith individual in the jth family at the kth visit; β ₀ is the log odds of hypertension after accounting for covariates, genetic effects; the terms age _ij, gender _ij, and BMI _ij respresent baseline age, gender, and BMI of the ith individual in the jth family, respectively; SNP _ij models the genetic effect, where the genotype is coded under an additive model; and time _ijk is the follow-up time from baseline for the ith individual in the jth family at the kth visit. P values for β ₄ were used to assess the significance of the association of each SNP with hypertension incidence.

The overall association of each candidate gene with longitudinal BP changes and hypertension incidence was evaluated using the truncated product method (TPM), which combines P values from single-marker association analyses.^31,32 Due to its ability to accommodate correlations among SNPs and increased power to detect gene-based associations compared to other meta-analysis methods, the TPM is preferred for combining P values across genes or genomic loci. To evaluate associations between genes and longitudinal BP changes, the P value for the genotype by follow-up time interaction term was used. For hypertension incidence, the P value for the genotype term was used. The truncation point was set as τ = 0.05, and the P value for the TPM was estimated by 100,000 simulations. Sensitivity analyses were carried out using the TPM after excluding significant SNPs within a gene (identified by single-marker analyses) to examine their influence on the gene-based analysis. Gene-based analysis was conducted using R software (version 3.0.1; http://www.r-project.org. Accessed on 9/17/2016). Additionally, in a sensitivity analysis to evaluate the robustness of findings from the TPM, the Gene-based Association Test that uses Extended Simes procedure (GATES)³³ was also employed.

Bonferroni correction was used to adjust multiple testing in single-marker and gene-based analyses.³⁴ The α-thresholds for single-marker and gene-based analyses were 4.95×10⁻⁴ (0.05/101) and 0.017 (0.05/3), respectively.