We used linked data from Early Pandemic Evaluation and Enhanced Surveillance of COVID-19 (EAVE II), a Scotland-wide cohort of 5.4 million people ( ~ 99% of the Scottish population), consisting of primary and secondary care, COVID-19 vaccination, SARS-CoV-2 testing, and mortality data39. Additionally, we used SARS-CoV-2 antibody testing data from the Enhanced Surveillance of COVID-19 in Scotland (ESoCiS), a national serological surveillance programme that sampled residual blood from multiple sources, including antenatal, blood donor, paediatric and primary care (i.e. general practice) settings, and tested these for the presence of SARS-CoV-2 IgG antibodies to the spike protein40,41.

Serology measurements used in this study were obtained from testing residual blood from people attending primary care settings for routine purposes. Samples were obtained from biochemistry laboratories (700 weekly samples) covering 11 regional health authorities, which represent >90% of the Scottish population. Additionally, primary care samples related to individuals aged 6 years and older and were chosen according to an age/sex/geographical sampling frame proportional to the Scottish general population. The serology results for primary care samples taken between April 20, 2020 and March 28, 2022 were linked with EAVE II data. All analyses were restricted to individuals who had received at least two doses of a COVID-19 vaccine prior to the serology sample date, unless otherwise specified.

We also undertook supplementary analyses of serology measurements from blood donors, similar to those applied to the primary care cohort. Methods and results of these analyses are described in Supplementary Note 1.

This study involves human participants. The Public Benefit and Privacy Panel Committees of Public Health Scotland and Scottish Government approved the linkage and analysis of the deidentified datasets for this project (2021-0115).

We defined testing IgG negative, to be a negative SARS-CoV-2 IgG test result obtained post-vaccination (at least 14 days after completion of the primary vaccine schedule). Assay manufacturers define a negative IgG test by quantitative IgG levels and stated cut-offs: <33.8 binding antibody units (BAU) per ml (tested using the Diasorin SARS-CoV-2 TrimericS IgG assay for antibodies to the S1/S2 protein)42. The outcome of a test result with undetectable IgG levels, as determined by the assay manufacturers, using a cut-off of <4.8 [BAU/ml], is presented in the sensitivity analysis in Supplementary Note 3. An arbitrary cut-off of <100 [BAU/ml] was also used in a sensitivity analysis to study a looser definition of a low post-vaccination IgG response.

For the analysis of the association between SARS-CoV-2 IgG antibody levels (hereafter referred to as “IgG levels”) and subsequent SARS-CoV-2 infection or severe COVID-19 outcomes, we considered SARS-CoV-2 infections, hospitalisations and deaths that occurred after the serology test date. SARS-CoV-2 infections were defined as positive RT-PCR test results; for multiple positive tests, the first test after the serology test was selected. COVID-19 hospitalisations were sourced from the Scottish Morbidity Record and were defined as hospital admissions where COVID-19 was listed as the primary reason for admission (International Classification of Diseases [ICD-10] codes U07.1 and U07.2); COVID-19 deaths were sourced from National Records of Scotland and defined as deaths where COVID-19 was mentioned on the death certificate (also ICD-10 codes U07.1 and U07.2).

For the analysis of factors associated with a negative IgG test, we used 26 comorbidity-based risk groupings43. We used either the total number of risk groups (0, 1, 2, 3-4, ≥5) for all individuals, or the specific risk groups as separate variables (Supplementary Table 5).

For the analysis of IgG levels and subsequent COVID-19 outcomes, the exposure was defined as either a positive or negative IgG test result, or as quantiles of the IgG. Quantiles were defined by analysis of the distribution of all measurements (Supplementary Fig. 1) as: very high (≥2000 BAU/ml), high (230–1999 BAU/ml), low (33.8–229 BAU/ml), very low (4.8–33.7 BAU/ml), and undetectable (< 4.8 BAU/ml).

In all analyses, we accounted for potential confounding due to: biological sex; age; BMI; SARS-CoV-2 infection (i.e. prior to the serology sample date); vaccine dose and vaccine product last administered ( ≥ 14 days prior to the serology sample date); number of days between the last vaccine dose and the serology sample date; number of days since the start of the pandemic (March 1, 2020); classification variables – derived from vaccine records and patient data held by the national public health body relating to whether the individual was a care home resident, immunosuppressed, severely immunosuppressed, and/or on the Scottish shielded patient list; and socio-economic status measured in quintiles of SIMD.

In the analysis investigating the association between IgG levels and COVID-19 related outcomes, either the total number of risk groups (0, 1, 2, 3-4, ≥5), or specific risks associated with severe outcomes were included as confounders. We accounted for the administration of additional vaccine doses (or boosters) between the serology sample date and the COVID-19 related outcome. We also adjusted for Urban Rural Classification (URC) and SIMD.

To examine characteristics associated with testing IgG negative post-double vaccination, we fitted two generalised additive models (GAMs) (Fig. 4). To examine multimorbidity, model A included the total number of risk groups (e.g., 1, 2, 3-4, or 5 + ) as an exposure variable; model B included specific risk groups (e.g., asthma, chronic kidney disease, etc.) as exposure variables and were restricted to individuals with at least one risk. For more details of each specific model, please see Supplementary Table 7.

Given input datasets from EAVE-II and serology datasets from ESoCiS we constructed data containing serology measurements linked with individual demographics, GP records, SARs-COV-2 testing and hospitalisation data for individuals who had received at least two doses of any COVID-19 vaccine. Model A and Model B are Generalised Additive Models (GAMs) where the number of comorbidity-based risks is used as the exposure for the outcome of testing IgG negative after vaccination. Similarly Model B instead used individual risks as the exposure variable for all individuals with at least one risk. Models C, D and E are Cox Proportional Hazard Regression Models (CPHRMs) where the survival is defined as SARs-CoV-2 infection for Model C and hospitalisation/death for Models D and E. Testing negative for IgG is the exposure for models C, and D whereas quantiles of quantitative IgG measurements are used as the exposure for model E.

GAMs were used to account for the influence of multiple confounding variables, some of which exhibited non-linear relationships. Smoothed spline functions were used to address such confounding due to age, the time interval between the last vaccine administration and the serology sample date, as well as the date of the serology sample, to account for residual effects due to the period of the pandemic. Age was stratified by number of vaccine doses, type of product and product combinations, motivated by the timeline and prioritisation of vaccine rollout in Scotland.

Additional risk factors were chosen for inclusion in the models if the p-value calculated in a univariate model was 0.1, or if there was a clinical rationale for their inclusion (and if the group had five or more individuals with the outcome). We calculated adjusted (and unadjusted) ORs from the exponential of the coefficients of the parametric terms of the GAM, using a Wald statistic for the confidence intervals (at 95% CI).

To study the association between IgG levels and subsequent SARS-CoV-2 infection or severe COVID-19 outcomes, we fitted a further two models: one using infection as the outcome (model C), and two with severe COVID-19 as the outcome (models D and E). In model D, the number of risk groups was included as a confounder while in model E, individual risks that were associated with significantly higher risk of severe outcomes were decoupled from all other risk groups and included separately as confounders. In models C and D, we used a binary exposure variable (positive/negative IgG test result) with the number of risk groups as a confounder. In Model E, we used quantiles of IgG titres as the exposure to further investigate how IgG levels, rather than just positive/negative IgG tests, affected the HRs of severe outcomes. These models are also summarised in Supplementary Table 7.

The number of vaccinations post-serology sample date were adjusted for as a time-dependent covariate. Age, sex, and BMI were adjusted for in all models (C, D, and E), with the survival period defined with respect to the start of the pandemic, accounting for the severity of SARS-CoV-2 variants, virus abundance in the general population, lockdowns, and other changes in social relations. Further adjustments were made for care home residency, SIMD, and urban/rural classification in Model C.

The follow-up period extended up to three months after the last recorded IgG measurement (i.e., up to June 28, 2022). Individuals were censored if they died from non-COVID-19 reasons before the end of the study period. Hazard ratios (HR) were calculated from the exponent of the fitted coefficients of the Cox models, with 95% CIs obtained using Wald tests.

All statistical analyses were performed using R version 3.6.1. We followed a pre-specified study analysis plan (available from the authors on request). Results are reported according to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines (Supplementary Table 6).

Further information on research design is available in the Nature Research Reporting Summary linked to this article.