Ethics

The study presented encompasses two elements. The first of these does not require specific ethical approval, as it focuses on public health/surveillance questions that make use of sequence data and other metadata that is already shared with the wider world as part of the activities of the COG-UK consortium (https://www.cogconsortium.uk/). COG-UK data is released and is publicly available via the ENA, GISAID and the COG-UK website. The element of the work that would/could require ethical approval is the specific examination of outcome data. The use of named patient data in the investigation of communicable disease outbreaks and surveillance of notifiable disease is permitted under Public Health Wales’ Establishment Order. Data were held and processed under Public Health Wales’ information governance arrangements, in compliance with the Data Protection Act, Caldicott Principles and Public Health Wales guidance on the release of small numbers. No data identifying protected characteristics of an individual were released outside Public Health Wales. The use of the genomic dataset for research purposes is also covered as part of the COG-UK project protocol which was approved by the Public Health England Research Support and Governance Office (RSGO) following review by the PHE Research Ethics and Governance Group (REGG).

Data

The Genomic Epidemiology Team within the Public Health Wales Communicable Disease Surveillance Centre maintain a dataset that combines testing data from individuals identified as infected with SARS-CoV-2 with information about the variant implicated based on subsequent genomic sequencing or genotyping. This record-level surveillance dataset, which necessarily contains information enabling identification of individual cases post data-collection, was accessed by the research team on 09/06/2022 for the purposes of the retrospective study reported below.

All SARS-CoV-2 real-time PCR testing was conducted in line with local guidelines. Samples came from two sources. One set of samples (N = 65262, 25.97%) came from the NHS Wales laboratory network. These samples were predominantly hospital and staff samples, as well as some community cases. Of these samples, Reflex PCR testing was undertaken on 11036 samples (16.91%) to identify key spike mutations associated with VOC/VUIs using the Allplex SARS-CoV-2 Variant PCR assays (Seegene, Seoul, Korea) and subsequently sent for confirmatory WGS. The second set (N = 183924, 73.18%) of samples came from the lighthouse lab based in Newport, Wales, operated by Perkin Elmer, and are predominantly community cases, tested as part of the national Test, Trace and Protect programme. Lighthouse lab diagnostic residual samples were collected by PHW, and the Public Health Wales Pathogen Genomics team undertook cherry-picking of positive samples for sequencing. The origin of a further 2143 (0.85%) of samples was not readily coded but these were also available for analysis.

Regardless of sample origin, the Public Health Wales Pathogen Genomics Unit aimed to sequence all samples that were of sufficient quality, as measured by a low qPCR cycle threshold (CT). The threshold to accept a sample for sequencing was generally ≤ 29.99 [26]. However, this cutoff varied throughout the study period due to changes in testing demand and capacity. PCR amplicons were generated using the Illimina COVID-Seq kit with ARTIC primers sourced from IDT. Samples were then loaded onto Illumina NextSeq 550 systems using a 300 cycle medium or high output kit. Once sequences were generated, these were processed by the Nextflow ARTIC nCoV pipeline (see https://github.com/connor-lab/ncov2019-artic-nf). Variants were identified using the software aln2type (see https://github.com/connor-lab/aln2type) using the UKHSA variant definitions (see https://github.com/ukhsa-collaboration/variant_definitions). Assembled genomes and bams were then filtered for human reads and uploaded with metadata to CLIMB COVID, the UK SARS-CoV-2 data analysis platform established by COG-UK [27, 28] to allow UK-wide collation, management and processing. Sequences were subsequently analysed using the CLIMB COVID phylogenetics pipeline which assigns cases to a PANGO lineage [28] and a putative ‘UK transmission group’ using ancestral state reconstruction [29, 30]. PANGO lineages were mapped to WHO variant of concern labels (see https://github.com/phe-genomics/variant_definitions; accessed 12/04/2023). Samples designated as ‘Wildtype’ according to this WHO labels classification scheme were relabeled as ‘Wuhan-Hu-1’. The results from the analysis were then synthesized into a genomics line list, and reported to the Public Health Wales Communicable Disease Surveillance Centre, for integration into a larger combined dataset.

The Wales Immunisation System was used to ascertain vaccination status of confirmed cases. An individual was considered to be “vaccinated” [with 2 doses] if they had had two doses of vaccine 14 days prior to their sample date. ICNet was used to identify case admissions to hospital. ICNet is a hospital infection prevention case management and reporting system used across Wales by infection prevention and control (IPC) teams and for systematic surveillance by Public Health Wales. An admission was classified as an individual with a positive PCR result for COVID-19, who was admitted to hospital on or one day before the day of their first positive test, or in the 28 days following a positive test.

Records for all genotyped and sequenced samples that were collected up to and including February 2022 were extracted for analysis. Note that there is no WHO variant classification for the original wild type SARS-CoV-2 virus as this was not, by definition, a ‘variant’. Therefore, all records without a WHO variant classification and which were collected prior to 22nd March 2021, were presumed to be ‘Wuhan-Hu-1’ infection cases (note that this corresponds to the WHO ‘Wildtype’ classification). Where a variant classification was not available for samples collected after this date, the classification is considered indeterminate or otherwise missing and therefore the corresponding case record was not included in our subsequent analyses. Records that were not dated were also excluded, as were records with no available date of birth.

Probabilistic matching based on the forename, surname, date of birth and postcode fields of each record was then performed using the fastLink package v0.6.0 in R [31], to identify individuals with multiple records in the dataset. The default string match threshold of 0.94 was applied. For the present work, a reinfection was defined as a second positive test record with a corresponding WHO classification, occurring at least 42 days after the first positive test recorded for the same individual. This was consistent with the standard 42 day episode deduplication rule that was used for SARS-CoV-2 surveillance over the study period by Public Health Wales, and is comparable to the interval used in other recent reinfection studies [32]. The matched record pairs that satisfied these criteria were visually inspected and any clearly erroneous matches were removed prior to analysis.

Omicron versus Delta reinfection frequency and inter-infection interval

The time series of the weekly number of positive samples recorded up to 28th February 2022 was first plotted by both week and WHO classification label (Alpha, Beta, Delta, Eta, Gamma, Kappa, Omicron, Zeta, Wuhan-Hu-1, and Unknown), alongside the log-transformed number of cases reported in Wales (case data publicly available at https://coronavirus.data.gov.uk/details/cases?areaType=nation&areaName=Wales, accessed: 10/03/2023). The rolling 7-day average number of samples is also plotted by both sample date and infection type (first versus second infection) to illustrate the temporal distribution of the number of first versus second infections.

Case counts and proportions were then tabulated by first and second infection variant to a) examine the frequency of Omicron versus Delta reinfection, and b) assess whether individuals were more likely to be infected first with the Wuhan-Hu-1, Alpha, Delta or Omicron variants. A chi-square test of independence was applied to the proportions in the resulting 2 (second infection: Delta or Omicron) x 4 (first infection: Wuhan-Hu-1, Alpha, Delta, or Omicron) contingency table to assess whether the first infection variant was independent of the second infection variant. The proportions and chi-square test are also illustrated on a mosaic plot alongside Pearson residuals to indicate the contingency table cells contributing to the overall test statistic. The mosaic plot was created using the VCD package v1.4–10 in R [33]. A Cramer’s V statistic was computed to estimate the Chi-square effect size. Among other assumptions, the Chi-square test assumes that at least 80% of cells contain more than 5 cases and that no cell has an expected value lower than 1. Both assumptions were satisfied in this case although one cell did have a small expected value of approximately 2. A complementary Fisher’s exact test was therefore calculated (not shown), which corroborated the chi-square test result.

The difference in sample collection date for first and second infections was used as a proxy measure for the inter-infection interval. To assess whether there is a difference in reinfection interval between the Delta and Omicron variant, the mean, median, standard deviation, interquartile range, minimum and maximum inter-infection intervals were also tabulated by second infection variant. These descriptive statistics, together with frequency histograms and normal probability plots (not shown) indicated that the reinfection intervals were not normally distributed, a finding confirmed by a formal Shapiro Wilk normality test. A two-tailed Kruskal-Wallis test was therefore used to compare the reinfection intervals of the Delta and Omicron variants.

All inferential statistical tests were two-tailed and conducted using a significance threshold of p ≤ 0.05, unless otherwise stated.

Characteristics and exposures of individuals reinfected with Omicron versus Delta

Descriptive epidemiology of the characteristics of confirmed Omicron reinfection cases was compared with confirmed Delta reinfection cases following tabulation of counts and proportions of both Delta and Omicron reinfection cases with each characteristic of interest. The characteristics considered included Local Health Board of Residence (Aneurin Bevan, Betsi Cadwaldr, Cardiff and Vale, Cwm Taf Morgannwg, Hywel Dda, Powys, or Swansea Bay), Age Group (0–24, 25–59, or 60+), Vaccination Status (Unvaccinated, One Dose, Two Doses, Three Doses, or Unknown), recent international Travel status (Yes or No), Symptom Status (Symptomatic, Asymtomatic or Unknown) and Hospitalisation Status (Hospitalised, Not Hospitalised or Unknown).

To investigate the effect of these characteristics and exposures upon the odds of reinfection with Omicron as opposed to Delta, an additional binary variable denoting Omicron reinfection status was derived (1 = reinfected with Omicron, 0 = reinfected with Delta). This binary response was regressed on local health board, age group, vaccination and travel status, using binary logistic regression with a canonical logit link. Age group, vaccination and travel status were of a priori interest based on previous research highlighting these as infection or reinfection risk factors [34–37]. As health service planning and delivery is the responsibility of seven independent Local Health Boards in Wales [38], cases’ Local Health Board of residence was also included as a predictor to account for any geographical variation in healthcare delivery and outcomes that may be germane to SARS-CoV-2 reinfection risk, and other spatial covariates and confounds which may be difficult to directly measure. Symptom and hospitalisation status were not included as predictors because the corresponding database fields were very sparsely completed.

Model fit statistics including residual deviance and degrees of freedom are reported. Odds ratios (ORs) were also computed alongside the corresponding p-values and 95% confidence intervals (CIs). A statistical significance threshold of p ≤ 0.05 was applied.

We also fit a conditional Poisson regression model [39] to estimate the difference in reinfection rates between the Omicron and Delta variants, accounting for variation in the sample collection date as well as key demographic characteristics and exposures. Reinfection counts were regressed on second infection variant (Omicron versus Delta) using a conditional Poisson model with a canonical log-link, implemented using the gnm package v1.1–2 in R [40]. Rather than including additional model terms to account for other factors and personal characteristics that may influence reinfection (e.g., specimen date and age group), and estimating separate coefficients for those variables as per unconditional Poisson regression, in the conditional Poisson model we can condition on the sum of events occurring in each of the model strata. The strata are determined by the interaction of the other variables, the effects of which are thus ‘conditioned out’. Here, the model strata were implemented as the interaction between the date that the second positive sample was collected, local health board, age group, vaccination and travel status. Note that as the second sample collection date was included in this interaction, overall infection pressure, testing policy etc., were effectively held constant for the Delta and Omicron reinfection cases being compared within the model strata. Again, symptom and hospitalisation status were not included in the model strata because the corresponding database fields were very sparsely completed. The point estimate and 95% CI for the incidence rate ratio is reported, alongside the model residual deviance and degrees of freedom.

Omicron versus Delta reinfection frequency and inter-infection interval

Consistent with other UK nations, Wales experienced a first wave of cases attributable to the original SARS-CoV-2 strain, which peaked in April 2020. A second wave of cases commencing in Autumn 2020 was driven by a variant of the B.1 lineage containing the D614G mutation. Henceforth we refer to both of these early variant groups as ‘Wuhan-Hu-1’, which corresponds to the WHO ‘Wildtype’ classification label. Note that testing capacity, and by extension, whole genome sequencing and genotyping capacity, was limited during these early waves of cases compared to subsequent waves. A third wave of cases coincided with the dominance of the Alpha variant from January 2021. A fourth wave of cases began in June 2021, at which point the Delta variant became dominant until the Omicron variant caused a fifth wave of record infections in December 2021 (Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Weekly positive samples sequenced by WHO classification label (stacked bars) and the log-transformed number of cases in Wales (gray shaded area).

A total of 251,329 positive samples were sequenced or genotyped over the study period.

https://doi.org/10.1371/journal.pone.0309645.g001

Fig 2 shows the 7-day rolling average number of samples collected by infection type (initial infection versus reinfection), highlighting that the vast majority were initial infections. The minority of samples associated with a ‘reinfection’, as defined here, were evidently collected towards the end of the timeseries when the Omicron variant became dominant in the population.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Rolling 7-day average number of samples by sample date and infection type.

https://doi.org/10.1371/journal.pone.0309645.g002

A total of 2488 reinfection cases were thus identified, comprised of 180 Delta reinfections (86 Males, mean age: 35.6 (years); STD: 19.2) and 2308 Omicron reinfections (898 Males, mean age: 31.6 (years); STD: 18.4). For both groups, the majority of the reinfection samples were sourced from the lighthouse lab network, though the proportion was higher for Omicron compared to Delta cases (75% and 62%, respectively; see Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Delta and Omicron reinfection counts and percentages by sample origin.

Percentages computed by column.

https://doi.org/10.1371/journal.pone.0309645.t001

The reinfection cases were distributed across the first and second infection variants considered here, as shown in Table 2. For those individuals reinfected with the Delta variant, the most frequent initial infection was Wuhan-Hu-1 SARS-CoV-2 (45.6%) whereas for those reinfected with Omicron, the most frequent initial infection was the Delta variant (55.2%). One of the cases with an initial Wuhan-Hu-1 infection followed by a Delta reinfection, experienced a subsequent third infection with the Omicron variant, confirmed by an additional positive sample.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Delta and Omicron reinfection counts and percentages by first infection variant.

Percentages computed by column.

https://doi.org/10.1371/journal.pone.0309645.t002

A Chi-square independence test confirmed that the proportions of individuals infected with each of the four first infection variants considered here (Wuhan-Hu-1, Alpha, Delta or Omicron), differed between the Delta and Omicron reinfection groups. (X² (3, N = 2488) = 20.387, p ≤ 0.001). This effect is also evidenced in Fig 3, the shading of which indicates that individuals reinfected with Delta were more likely to have had an initial Wuhan-Hu-1 infection compared to those individuals reinfected with Omicron. They were also less likely to have had an initial Delta infection compared to those reinfected with Omicron. A small Cramer’s V statistic, however, indicates that this first by second infection variant effect was small (V = 0.091, 95% CI = [0.057, 0.131]).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Mosaic plot of the number of reinfections by first infection and second infection variant.

The area of each tile is proportional to the corresponding contingency table cell count in Table 1. The vertical height of each tile in the first (second) column reflects the proportion of Delta (Omicron) reinfections that were associated with an initial Wuhan-Hu-1, Alpha, Delta or Omicron infection. The tile shading represents the magnitude and sign of the Pearson residual for the cell of the associated contingency table. Red indicates a cell with fewer cases than expected whereas blue indicates a cell with more cases than expected under the null model.

https://doi.org/10.1371/journal.pone.0309645.g003

Turning to the inter-infection intervals, Table 3 shows the mean, median, standard deviation, standard error, interquartile range, and both the minimum and maximum inter-infection intervals for individuals reinfected with the Delta and Omicron variants. The distribution of these inter-infection intervals is also shown by reinfection variant in Fig 4. The median reinfection interval is shorter for the Omicron variant, consistent with a lower number of days between first and second infection compared to Delta. Note, however, that both the mean and maximum interval are larger for Omicron relative to Delta (as are the standard deviation and interquartile range), indicative of positive skew in the distribution of reinfection intervals for Omicron. This is arguably consistent with the timeline of the emergence of the Delta and Omicron variants. As Omicron emerged later, the maximum possible inter-infection interval is necessarily longer for Omicron compared to Delta reinfections.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. First-to-second COVID infection interval by second infection variant.

Stacked and un-stacked histograms (left and right, respectively).

https://doi.org/10.1371/journal.pone.0309645.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Inter-infection intervals (days) by second infection variant.

https://doi.org/10.1371/journal.pone.0309645.t003

Indeed, formal statistical tests confirmed that a) the inter-infection intervals for both the Delta and Omicron reinfection groups were not normally distributed(W = 0.944, p = 0; W = 0.909, p = 0, respectively), and b) those reinfected with Omicron experienced fewer days between their first and second infection, compared to those reinfected with the Delta variant (χ² (1) = 7.84, p = 0.005).

Characteristics and exposures of individuals reinfected with Omicron versus Delta

A conditional Poisson model was used to estimate the incidence rate ratio for reinfection with the Delta versus Omicron variants based on event counts in the model strata (see Methods). The model had a residual deviance of 2.3 on 11 degrees of freedom, indicating a good fit to the data (χ² (11) = 2.3, p = 0.997).

The estimated incidence rate ratio of 2.17 (95% CI = [1.12, 4.45]), indicates that when the other potentially important characteristics (e.g., sample collection date, age group and vaccination status) also accounted for in the model are held fixed, the Omicron variant was associated with approximately 2.17 times the number of reinfections compared with Delta (i.e., 116.7% more).

Table 4 below tabulates the counts and proportions of both Delta and Omicron cases with each level of a set of characteristics and exposures, including Local Health Board of Residence, Age Group, Vaccination Status, Travel Status, Symptom Status and Hospitalisation Status.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Characteristics and exposures of Delta versus Omicron reinfection cases.

https://doi.org/10.1371/journal.pone.0309645.t004

A multivariable logistic regression model was fitted to this data to investigate the effect of particular case characteristics on the odds of reinfection with Omicron as opposed to Delta. The fitted model, which included Local Health Board, Age Group, Vaccination status and Travel as predictors, provided an adequate fit to the observed data, with a residual deviance of 1261.38 on 2473 degrees of freedom (χ² (2473) = 1261.38, p ≃ 1). Note that symptomatic and hospitalisation status were not included as predictor variables owing to the high proportion of records for which this information was missing (see Table 4 above). The computed odds ratios are reported in Table 5 below. For each characteristic (e.g., Local Health Board), the odds ratio reported at a given level (e.g., Powys Teaching Health Board) represents the relative odds of being reinfected with Omicron rather than Delta in the corresponding group of cases (i.e., Powys Teaching Health Board residents), compared to those in the reference group (i.e., Aneurin Bevan UHB residents), whilst all other characteristics included in the model are held fixed. An odds ratio (OR) > 1 indicates increased odds whereas an OR < 1 indicates decreased odds, compared to the reference group. Individuals aged 25–59 were therefore less likely to be reinfected with Omicron than Delta compared to those aged 0–24 years (OR = 0.54, 95% CI = [0.37, 0.78], p = 0.001). Likewise, individuals aged 60+ were also less likely to be reinfected with Omicron than Delta compared to those aged 0–24 years (OR = 0.26, 95% CI = [0.14, 0.47], p < 0.001). Finally, the odds of reinfection with Omicron rather than Delta was higher in individuals who received two doses of vaccine by the time of reinfection, compared to the unvaccinated (OR = 1.96, 95% CI = [1.33, 2.87], p < 0.001). All other odds ratios were not statistically significant.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Characteristics and exposures of Omicron versus Delta reinfection cases—Odds ratios.

https://doi.org/10.1371/journal.pone.0309645.t005

Limitations and future directions

A relatively small number of reinfection cases were available for inclusion in our analyses, particularly as regards Delta reinfections. The number of reinfection cases available for analysis could be substantially increased by pooling genomic data obtained from residents in Wales and other UK nations. Differences in reinfection incidence across the Delta and Omicron groups could then potentially be estimated with greater precision, albeit at the cost of reduced geographical and population specificity. This would also afford inclusion of additional explanatory variables and interactions in the models fitted here. The type of vaccines previously received by reinfection cases could be one such potentially interesting additional factor. However, early evidence supported the efficacy of a ‘mix-and-match’ vaccination schedule, enabling administration of different types of vaccine to individuals across successive doses in the UK [63, 64]. Further studies considering this factor would therefore need to accommodate a potentially large number of vaccine-type combinations.

A related limitation of PCR testing and sequencing data, generally, is that it reflects known and detected infection events, but cannot capture valid infection or reinfection episodes that were not tested and/or reported. There is therefore under-ascertainment of both the original number of cases and the true number of reinfections. There is also no systematic testing for asymptomatic infections, and where asymptomatic testing has occurred it has been in specific cohorts over limited periods of time (e.g., hospitals). Furthermore, different variants may be associated with differential symptomatic infection rates, which could interact with changes in testing policy/practice over time, in turn skewing the representation of particular variants within our sequencing data sample. Taken together, it is possible that the reinfection patterns observed in our sample may not be representative of real-world reinfections. The absence of ground-truth data about the total number real-world infections and the responsible variants, renders this issue challenging to address. Future work could, however, explore the correspondence the number of genomically confirmed reinfections, and antibody/swab positivity results from seroprevalence surveys, as well as the total number of reported cases. This could provide further evidence for the potential for emerging variants to cause further waves of cases and reinfections.

Given the available genomic data, an individual was identified as having had a reinfection if they had a second positive PCR sample recorded 42 days or more after an initial positive sample, both of which were sequenced or genotyped. Due to limited capacity, the number of sequenced and genotyped samples represents a minority of the total number of reported cases (see Fig 1). It is therefore possible that some individuals in the present sample had additional infections prior to or between the two that were here presumed to be their first and second infections. The number of such cases is likely to be small as reinfections were reportedly uncommon during the first two years of the pandemic [18]. Nevertheless, further work could leverage additional testing data to identify and remove such cases.

Finally, the present analyses compared Delta and Omicron reinfections up to February 2022. In principle, a natural extension could be to compare their sublineages. There were, however, relatively few Delta reinfection cases confirmed by either whole genome sequencing or genotyping, rendering further stratification by sublineage unfeasible. Further studies could compare Omicron sublineages though the variant had not diversified substantially by the end of February 2022, so it would be beneficial for such studies to include data from samples received after this date. There were, however, several changes to testing policy in Wales after this date and accounting for these would impact on the number of events in our conditional Poisson model strata.