We combined evidence of SNP – diabetic retinopathy associations through inverse-variance weighted fixed-effects meta-analyses, both within and across genetic ancestry groups (Table 1). We identified SNPs corresponding to nine unique genome-wide significant loci across the combined- (N_loci = 9), Non-Hispanic African (NH-AFR-) (N_loci = 1), and Non-Hispanic European (NH-EUR) -ancestry (N_loci = 1) meta-analyses (Figure 1, Table 2, Supplemental Tables 1 – 3). These loci corresponded to nine genes based on proximity (<1Mb): MFSD4A, TENM2, GRHL2, TCF7L2, FBXW8, COL4A1, SLC16A3, TMPRSS6, and G6PD. The most significant SNP in the combined- and NH-AFR-ancestry meta-analyses was rs1050828-T (Val98Met) (odds ratio [OR] = 1.48, 95% confidence interval [CI] = (1.45 – 1.51), p = 1.99 × 10⁹⁰) in G6PD. The most significant SNP in the NH-EUR ancestry analysis was rs7903146-T (OR = 1.08 (1.07 – 1.09), p = 3.14 × 10⁻¹³), intronic in TCF7L2, corresponding to the same lead variant of the strongest type 2 diabetes risk association.²⁴ Additionally, rs1050828-T was significant in the combined- and NH-AFR-ancestry meta-analyses of PDR cases against diabetic retinopathy controls (OR = 1.36 (1.31 – 1.42), p = 3.28 × 10⁻¹⁵) (Supplemental Table 4 – 6). We did not identify any conditionally independent SNPs at this or any locus.

We examined the genomic inflation factor (λ_GC), linkage disequilibrium score regression intercept, and heritability on the liability scale (h²) of the meta-analyses as well as the genetic correlation of diabetic retinopathy between NH-AFR- and NH-EUR-ancestries in the MVP with linkage disequilibrium score regression (LDSC) (Supplemental Table 7). There was little evidence for confounding due to population stratification in ancestry specific analyses as suggested by λ_GC and LDSC intercept values for NH-AFR (λ_GC: 1.0225; LDSC intercept: 0.9817 (0.0059)) and NH-EUR ancestry (λ_GC: 1.0405; LDSC intercept: 0.9841 (0.0077)) analyses. Heritability (h²) estimates were larger for NH-AFR ancestry (8.0% (5.8% – 10.2%)) than for NH-EUR ancestry (3.6% (3.0% – 4.2%)), though were similar to previously published SNP-based heritability estimates (7.0%).¹³ Additionally, there was a high correlation between the NH-EUR- and NH-AFR-ancestry analyses (0.76 ± 0.11, p = 1.72×10⁻¹²).

We estimated the association between genetically predicted gene expression (GPGE) and diabetic retinopathy in 49 tissues from GTEx v8 with S-PrediXcan.²⁵ We identified 12 statistically significant gene-tissue pairs associated with diabetic retinopathy in the combined-ancestry analyses and 13 pairs in the NH-AFR ancestry analyses (Supplemental Tables 8 – 9). These gene-tissue combinations corresponded to nine unique genes (including G6PD), none of which were previously associated with diabetic retinopathy as reported in the NHGRI-EBI catalog of human GWAS at the time of publication. We identified seven statistically significant gene-tissue pairs associated with PDR in the combined-ancestry analyses and 14 pairs in the NH-AFR ancestry analyses (Supplemental Tables 10 – 11). Four genes were statistically significant in both diabetic retinopathy and PDR: FLNA, GAB3, G6PD, and RENBP, though we did not detect any gene-tissue pairs where there were statistically significant GPGE results and high posterior probability (PP.H4>0.9) of colocalization. We did not identify any statistically significant GPGE associations in the NH-EUR ancestry analyses.

Because rs1050828 is located on the X chromosome we investigated the effect of the variant on diabetic retinopathy risk separately for men and women of NH-AFR ancestry. We observed similar effect sizes for the variant on diabetic retinopathy risk for men (OR = 1.46 (1.40 – 1.52), p = 7.52 × 10⁻⁶⁷) and women (OR = 1.54 (1.27 – 1.88), p = 1.38 × 10⁻⁰⁵) (Supplemental Table 12) although the variant did not reach genome-wide significance for women due to reduced sample size. We did not detect additional associations on the X chromosome in a conditional analysis adjusting for rs1050828 genotypes.

We utilized EHR data from the MVP to compare mean hemoglobin A1c (HbA1c %) and outpatient random plasma glucose (mg/dL) values post diabetes diagnosis for men of NH-AFR ancestry with and without the G6PDdef risk allele (rs1050828-T) stratified by diabetic retinopathy status. Individuals with the G6PDdef risk allele with diabetic retinopathy had similar levels of HbA1c (7.27% vs. 7.29%; p = 0.46), yet substantially higher levels of plasma glucose (185 mg/dL vs. 145 mg/dL; p = 4.93 × 10⁻¹⁸⁶) compared to individuals without the G6PDdef risk allele and without diabetic retinopathy (Figure 2a – 2b, Supplemental Table 13).

We compared HbA1c and glucose levels at different times during the natural history of diabetes in men of NH-AFR ancestry in the MVP with and without the G6PDdef risk allele. In each participant, we determined the mean outpatient random plasma/serum glucose (mg/dL) and mean HbA1c (%) a year prior to diabetes diagnosis as well as a year prior to insulin prescription. During the period prior to diabetes diagnosis, individuals with the G6PDdef risk allele had a significantly lower HbA1c level than those without (6.70% vs. 7.06%; p = 6.05 × 10⁻¹²), but a significantly higher plasma glucose level (165 mg/dL vs. 135 mg/dL; p = 1.12 × 10⁻⁴⁶) (Supplemental Table 14). HbA1c and glucose levels were higher, but patterns were similar during the year prior to the initial outpatient prescription of insulin (Supplemental Table 14).

We compared the difference between mean HbA1c and predicted HbA1c (generated by regressing HbA1c on outpatient random plasma glucose) for all individuals with diabetes in the MVP (Figure 2c). Of all the individuals of NH-AFR ancestry, 76% were concentrated in the top and middle tertile, suggesting an HbA1c either higher than or similar to the predicted HbA1c (Figure 2d). However, 80% of individuals of NH-AFR ancestry with the G6PDdef risk allele were in the bottom tertile, suggesting an HbA1c lower than the predicted HbA1c (Figure 2e).

We demonstrate that individuals with the G6PDdef risk allele have a higher predicted odds of diabetic retinopathy compared to individuals without the G6PDdef risk allele at lower ranges of HbA1c, with risk of diabetic retinopathy converging near an HbA1c of 10% (Figure 3, Supplemental Figure 1). To investigate whether the effect of the G6PDdef risk allele on diabetic retinopathy risk could be attributed to differences in plasma glucose levels, we performed a mediation analysis in individuals of NH-AFR ancestry in the MVP. For men, we estimated a 2.39% (total effect, p < 2.0 × 10⁻¹⁶) increase in diabetic retinopathy risk due to the G6PDdef risk allele. However, the indirect effect of plasma glucose completely mediated the effect of the G6Pddef risk allele on diabetic retinopathy (average indirect effect = 2.37%, p < 2.0 × 10⁻¹⁶, average direct effect = 0.03%, p = 0.95) (Supplemental Table 15). We identified similar results in women (Supplemental Table 16). As the effect of the G6Pddef risk allele was mediated by increased plasma glucose, the increase in risk of diabetic retinopathy due to the G6Pddef risk allele is most likely due to suboptimal control of glycemia.

We further investigate if the G6Pddef risk allele increases risk of diabetic macular edema, a severe subtype of diabetic retinopathy, as well as other diabetes complications identified with an EHR based algorithm in men of NH-AFR ancestry. We identified associations between the G6PDdef risk allele and diabetes complications including diabetic macular edema (OR = 1.55 (1.45 – 1.65) p = 1.31 × 10⁻⁴¹) and diabetic kidney disease (OR = 1.24 (1.18 – 1.31) p = 2.59 × 10⁻¹⁵). To evaluate whether the effect of the G6PDdef risk allele on diabetic retinopathy is driven by the association with diabetic macular edema, we performed a separate analysis for diabetic retinopathy excluding individuals with diabetic macular edema. The G6PDdef risk allele was still significantly associated with diabetic retinopathy without the presence of diabetic macular edema (OR = 1.33 (1.27 – 1.39) p = 5.7 × 10⁻³⁵).

We further investigated the effect of the G6PDdef risk allele on the risk of diabetes complications in longitudinal data from the Action to Control Cardiovascular Risk in Diabetes (ACCORD) clinical trial in individuals of NH-AFR ancestry (Supplemental Table 17). Two diabetes complications or a composite outcome of diabetic retinopathy or diabetic nephropathy were significantly associated with the G6PDdef risk allele in the ACCORD clinical trial data, regardless of medication regimen. Individuals of NH-AFR ancestry with the G6PDdef risk allele had a higher likelihood of proliferative diabetic retinopathy (HR = 1.78 (1.56 – 2.04) p = 1.95 × 10⁻⁰⁵) (characterized as photocoagulation or vitrectomy), diabetic neuropathy (HR = 1.37 (1.22 – 1.53) p = 5.45 × 10⁻⁰³) (characterized as vibratory sensation loss), and a composite outcome of diabetic retinopathy and diabetic nephropathy (HR = 1.61 (1.41 – 1.82) p = 2.07 × 10⁻⁰⁴) (specified as renal failure, ESRD, serum creatinine>3.3, photocoagulation, or vitrectomy) (Supplemental Table 18 – 19). With the available incidence data, we estimated the excess risk of diabetic complications due to the G6PDef by computing a crude and an adjusted risk difference (RD) for diabetic retinopathy (adjusted RD = 133 (95% CI: 3 – 264) per 1,000 individuals) and diabetic neuropathy (adjusted RD = 245 (95% CI: 84 – 507) per 1,000 individuals) in men of NH-AFR ancestry (Supplemental Table 20). Assuming the RD is an estimate of the attributable risk and that all of the excess risk is due to inadequate management of hyperglycemia due to reliance on HbA1c, we estimated 53 (95% CI: 30 – 254) of individuals with NH-AFR ancestry with diabetes would need to be screened for the G6PDdef variant and adequately managed for hyperglycemia to prevent one additional case of diabetic retinopathy. Finally, we estimated screening 407 (95% CI: 228 –1,948) individuals of NH-AFR ancestry would prevent one case of diabetic retinopathy (Supplementary Tables 21 – 22).

To further investigate the effects of the G6PDdef risk allele on risk of diabetes complications, we conducted a PheWAS to investigate phenotypes associated with the G6PDdef risk allele separately in individuals with and without diabetes of NH-AFR ancestry in MVP and BioVU. We identified 30 PheCodes in individuals with diabetes and 19 PheCodes (higher order definitions using ICD codes) significantly associated with the G6PDdef risk allele for individuals without diabetes (Figure 4, Supplemental Table 23 – 24). Of these PheCodes, ten were significant in both analyses with the same direction of effect between groups, including anemia (other hereditary hemolytic anemia), metabolic (intestinal disaccharidase deficiencies and disaccharide malabsorption), and digestive (jaundice [not of newborn], cholelithiasis and cholecystitis) PheCodes.

The 20 significant PheCodes in individuals with diabetes included those in the infectious disease (septicemia (1.12 (1.09 – 1.15), p = 1.82 × 10⁻⁵)), endocrine/metabolic (diabetic retinopathy (1.12 (1.10 – 1.14), p = 1.36 × 10⁻⁸)), circulatory system (hypertensive heart and/or renal disease (1.10 (1.08 – 1.12) p = 2.75 × 10⁻⁷)), and genitourinary (acute renal failure (1.12 (1.10 – 1.14) p = 2.14 × 10⁻⁹)) PheCode groups. We did not detect any significant PheCodes in the infectious disease, circulatory system, or genitourinary PheCode groups associated with the G6PDdef risk allele in individuals without diabetes.

We identified PheCodes associated with the G6PDdef risk allele in the analysis of individuals without diabetes that were not found in the analysis of individuals with diabetes. Specifically, we identified ‘obesity’ (1.06 (1.05 – 1.07) p = 4.21 × 10⁻⁰⁶), ‘poisoning/allergy of sulfonamides’ (1.48 (1.35 – 1.62) p = 1.85 × 10⁻⁵), ‘other diseases of blood and blood-forming organs’ (1.24 (1.19 – 1.28) p = 3.31 × 10¹⁰), and ‘other deficiencies of circulating enzymes’ (4.29 (3.68 – 4.54) p = 1.22 × 10⁻⁴⁰).

Vanderbilt University Medical Center’s DNA Repository (BioVU) is a de-identified database of electronic health records (EHR) that are linked to patient DNA samples.⁴⁷ The Veterans Affairs Million Veteran Program (MVP) is a national research program that incorporates genomic data and health record data, collected from veterans, to investigate how genes, lifestyle, military experiences, and exposures affect health and wellness.^48,49 The United Kingdom Biobank (UKB) is a prospective cohort study with genetic and phenotypic data collected on approximately 500,000 individuals across the United Kingdom between 40 and 69 years of age at recruitment.^50,51 Mass General Brigham Biobank (MGBB; formerly Partners HealthCare Biobank) is a large repository of biospecimens and data linked to extensive electronic health record and survey data.⁵² We utilized the discovery stage summary statistics from Pollack et al.¹⁵ The Action to Control Cardiovascular Risk in Diabetes (ACCORD) study was designed to examine treatment intensity of glycemic control and blood pressure control and the corresponding effects on the prevention of major cardiovascular event in individuals with type 2 diabetes mellitus. In total we identified 66,179 cases and 126,227 controls across the five contributing sources (Table 1).

Race and ethnicity information for Vanderbilt University Medical Center’s BioVU, the MVP, the UKB, and the MGBB have been previously described.^48,53–55 Briefly, contributing sites collected self-reported race and ethnicity information. The MVP utilized the harmonized ancestry and race/ethnicity (HARE) method to merge self-reported and ancestral data.⁵⁶ This resulted in four categories: individuals of predominately non-Hispanic African (NH-AFR)ancestry, individuals of predominately non-Hispanic European (NH-EUR) ancestry, individuals of predominately Asian (ASN) ancestry, and Hispanic individuals.

In BioVU, the MGBB, and the MVP, diabetic retinopathy cases and controls with diabetes were identified with a previously published algorithm.⁵⁷ Briefly, the algorithm utilized EHR data to identify individuals with International Classification of Diseases (ICD) and Current Procedural Terminology (CPT) code-based evidence of diabetic retinopathy, while controls were individuals with ICD evidence of diabetes, no ICD evidence of diabetic retinopathy, as well as a comprehensive ophthalmology exam. Cases and controls were identified in the UK Biobank through the diabetic retinopathy PheCode. Proliferative diabetic retinopathy was identified in a similar manner as diabetic retinopathy in BioVU and the MVP, where cases had ICD and CPT code-based evidence of PDR and controls were individuals with ICD and CPT code-based evidence of diabetic retinopathy but not PDR.

In VUMC and the MVP, age was designated as the age of diabetic retinopathy diagnosis or age at last ophthalmology exam. Sex was classified based on self-reported data and individuals with discordant self-reported sex and genetically inferred sex were excluded from analyses. Duration of diabetes was classified in a similar manner, taking the difference between the date of diabetes mellitus diagnosis and the date of diabetic retinopathy diagnosis or date of the last ophthalmology exam, for cases and controls, respectively. Individuals with a negative or zero duration of diabetes (suggesting the individual received normal care outside of either VUMC or the VA) were removed from the analyses. Three mean plasma glucoses were calculated with an individual’s outpatient plasma glucose tests (excluding point-of-care and inpatient tests), one beginning at the first criteria of diabetes, the second 365 days prior to first criteria of diabetes, and 365 days prior to the first prescription of insulin in MVP. Three mean HbA1c were calculated with all of an individual’s HbA1c tests (excluding point-of-care), one beginning at the first criteria of diabetes, the second 365 days prior to first criteria of diabetes, and 365 days prior to the first prescription of insulin in MVP. We excluded individuals with a mean HbA1c (%) lower than 4% and greater than 20%. We generated predicted HbA1c by regressing mean HbA1c on mean plasma glucose of individuals with diabetes, resulting in the equation: pHbA1c=0.4384+0.01807(PlasmaGlucose). We then calculated the difference in mean HbA1c and predicted HbA1c for each individual.

We defined type 2 diabetes using an algorithm designed specifically for UKB⁵⁸. Type 2 diabetes, diabetic retinopathy, coronary artery disease, and chronic kidney disease were defined using the ICD 9 and 10 codes (Supplementary Table 25). MGBB included a total of 6,623 T2D cases. T2D status was defined based on “curated phenotypes” developed by the MGBB Portal team using structured and unstructured electronic medical record data and clinical, computational, and statistical methods. Natural Language Processing was used to extract data from narrative text. Chart reviews by disease experts helped identify features and variables associated with particular phenotypes and were also used to validate the results of the algorithms. The process produced robust phenotype algorithms that were evaluated using metrics such as positive predictive value (PPV) and negative predictive value (NPV).

We identified diabetic retinopathy – SNP and PDR – SNP associations in the BioVU and MVP with logistic regression using PLINK2, adjusting for age, sex, diabetes duration, mean HbA1c values post diabetes diagnosis, and the top 10 principal components of ancestry, henceforth defined as ‘standard covariates’, separately in individuals of EUR-, AFR-, and ASN-ancestry as well as Hispanic ethnicity. Data for the X chromosome was available in BioVU, MGBB, MVP, and the UKB. In BioVU and MVP, logistic regression analyses were conducted in PLINK2 with the “xchr-model 2” option. We performed whole-genome regression analysis using REGENIE version 2 for UKB and MGBB. We used diabetic retinopathy as a binary outcome and included age, sex, body mass index (BMI) and 10 PCs and imputation batch for each cohort. We used a block size of 1,000 for step 1 and 400 for step 2. All variants with a MAC < 3 for UKB, and MGBB. To provide better-calibrated test statistics, REGENIE supports the option (--firth --approx --firth-se) to use the Firth correction for variants where the p-value from the standard logistic regression test is below a threshold, as defined by pthresh. For UKB and MGBB, we set pthresh to 0.999, causing the Firth correction to be applied. We conducted an inverse-variance weighted fixed effects meta-analysis, within and across ancestral groups with METAL.⁵⁹ We utilized LDSC to calculate the genomic inflation factor (λ_GC), linkage disequilibrium score regression intercept, heritability on the liability scale (h2), and the genetic correlation of diabetic retinopathy between EUR- and AFR-ancestries in the MVP.⁶⁰

We identified G6PDdef risk allele associations with a subtype of diabetic retinopathy, diabetic macular edema (n_cases = 3,812, n_controls = 32,921), and diabetic kidney disease (n_cases = 3,919, n_controls = 49,417), in individuals of NH-AFR ancestry in the MVP. Diabetic macular edema was identified through an ICD code-based method, requiring ≥ 2 ICD code days for diabetic macular edema, while controls were individuals with diabetic retinopathy and no ICD evidence of diabetic macular edema. Diabetic kidney disease was characterized as a >50% decline in eGFR from baseline over, confirmed by a second eGFR value more than 90 days apart with no recovery in between from the baseline kidney function at MVP enrollment or reaching incident end-stage renal disease. Both analyses were conducted with logistic regression using PLINK2, adjusting for age, sex, diabetes duration, mean HbA1c values, and the top 10 principal components of ancestry.

We conducted sensitivity analyses for the diabetic retinopathy – G6PDdef risk allele association in individuals of NH-AFR ancestry. The first was a logistic regression with diabetic retinopathy cases and diabetes mellitus controls, separated by sex (n_men = 33,408, n_women = 3,487) and adjusted for age, sex, diabetes duration, mean HbA1c values, and the top 10 principal components of ancestry in the MVP. The second was a logistic regression in the MVP with diabetic retinopathy cases and diabetes mellitus controls, adjusted for age, sex, diabetes duration, mean HbA1c values, and the top 10 principal components of ancestry as well as the G6PDdef risk allele. The final analysis was conducted for the diabetic macular edema – G6PDdef risk allele association with logistic regression, adjusting for age, sex, diabetes duration, mean HbA1c values, and the top 10 principal components of ancestry. The analysis was restricted to cases with diabetic retinopathy and without diabetic macular edema (n_max = 10,556) and controls with diabetes (n_max = 22,365). The third sensitivity analysis was to investigate the fitted log-odds of diabetic retinopathy with a logistic regression of diabetic retinopathy in men of NH-AFR ancestry in the MVP (n_max = 45,159), separately for individuals with and without the G6PDdef risk allele, adjusting for age, duration of diabetes, and the top two principal components. The fitted odds were generated for men of NH-AFR ancestry of median age (64 years), median duration of diabetes (7.5 years), median principal component 1 (−0.144), and median principal component 2 (−0.024). The final sensitivity analysis utilized the ACCORD clinical trial data to investigate if different medication regimes impacted the association between the G6PDdef risk allele and diabetic retinopathy in individuals of NH-AFR ancestry (n_max = 1,337). We conducted Cox proportional hazard models investigating the association between the G6PDdef risk allele and diabetic retinopathy including standard covariates as well as clinical trial treatment arm, metformin use, sulfonylurea use, thiazolidinedione use, and insulin use (individually and jointly).

To identify independently associated SNPs, we used the Genome-wide Complex Traits Analysis (GCTA) software. The GCTA-COJO software performs iterative conditional stepwise selection of independently significant SNPs and joint analysis simultaneously with stepwise model selection.⁶¹ The summary statistics from meta-analyses were used as the input summary data, and we utilized 5,000 unrelated individuals of NH-EUR ancestry from the UK Biobank or 2,217 unrelated individuals of NH-AFR ancestry from BioVU for the linkage disequilibrium structure. A p value threshold of 5 × 10⁻⁸ was used as the selection threshold within GCTA, and the collinearity threshold was set at the default value of 0.9, so that SNPs are not selected if the multiple regression with the current SNPs in the model has r² ≥ 0.9. For any sets of SNPs in LD (r² ≥ 0.1), we selected the most significant SNP with the minimum p value from the GCTA model.

We evaluated genetically predicted gene expression (GPGE) with S-PrediXcan, a gene-level approach that estimates the genetically determined component of gene expression in a given tissue and tests it for association with an outcome using SNP-level summary statistics.⁶² We used the common variants (minor allele frequency > 0.01) from the meta-analyses and the 49 tissues from GTEx V8, incorporating elastic net gene models and covariance matrices developed for either EUR- or AFR-ancestry from 1000 Genomes.⁶³ The Bonferroni-corrected significance threshold was 1.55 × 10⁻⁶ to account for the total number of gene models assessed across the tissues. Coloc, a Bayesian gene-level test, was used to test the hypothesis that a single variant underlies GWAS and eQTL associations at a given locus (i.e. colocalization).⁶² Input included the common variants (MAF >= 1%) from the meta-analyses and eQTL summary statistics corresponding to the gene-expression references used in GPGE analysis, restricting to only variants included in the GTEx V8 models. A statistically significant GPGE result along with a corresponding posterior probability greater than 80% that SNPs were associated with both expression and the outcome was considered as strong evidence of colocalization.

We conducted a mediation analysis to examine if the effect of variant rs1050828-T on diabetic retinopathy was fully or partially mediated through plasma glucose with the ‘mediate’ function, part of the R package ‘Mediation’.⁶⁴ We investigated the effect of the G6PDdef risk allele and plasma glucose on diabetic retinopathy with logistic regression, utilizing the glm() function in R. We utilized the lm() function to investigate the effect of the G6PDdef risk allele on plasma glucose.

We performed a phenome-wide association study (PheWAS) of the G6PDdef risk allele, rs1050828-T, in individuals of NH-AFR ancestry in the MVP (n_diabetes = 33,282, n_no-diabetes = 42,853) and BioVU (n_diabetes = 1,117, n_no-diabetes = 4,427).⁶⁵ We used logistic regression to separately model the ~1,800 PheCode traits as a function of the variant, adjusted for age, sex, and the top 10 principal components of ancestry separately for individuals with and without diabetes, as well as separately for men and women. Interpretation of results were limited to phenotypes with 20 or more cases. We meta-analyzed the PheWAS between sites. Bonferroni-corrected significance threshold was p ≤ 2.75 × 10⁻⁵ to account for the total number of PheCode models assessed in these analyses. All PheWAS analyses were conducted using the PheWAS R package.⁶⁶

We performed a time-to-event analysis with the Cox proportional hazard model in R (‘coxph’ in the ‘survival’ package) using diabetes duration until event with data from the ACCORD trial, including individuals of NH-AFR ancestry (n_max = 1,337) (regardless of trial arm) with 9 diabetic complication outcomes with the G6PDdef risk allele, adjusting for age, sex, diabetes duration, and HbA1c.^67,68 Diabetes complication outcomes were excluded if there were less than or equal to five G6PDdef carriers with an event. The multiple testing correction threshold was 5.60 × 10⁻³ to account for the nine tests conducted.