Immune correlates of risk for SARS-CoV-2 infection in children: a prospective, community-based cohort study

The identification of immune markers that correlate with protection from disease after vaccination or infection—referred to as correlates of protection—is a critical component of vaccine science¹. To establish an immune marker as a correlate of protection, it is often first shown to be a correlate of risk, that is, to have a statistical association with a clinical endpoint used to measure vaccine efficacy¹. Research on both correlates of risk and correlates of protection helps to accelerate the development and approval of vaccines, thereby supporting efforts to mitigate the burden of infectious diseases^2,3,4. Since the COVID-19 pandemic began in 2020, there has been an abundance of research quantifying the value of neutralizing antibodies (NAbs) and binding antibodies (BAbs) for SARS-CoV-2 as correlates of protection against symptomatic infection and severe COVID-19 disease in adults^5,6,7,8,9. Both BAbs and NAbs were successfully used in immunobridging studies for licensure of vaccines against variants of concern and for approval of vaccines for populations not included in the original trials for ethical reasons, such as pregnant individuals and children^10,11.

Despite extensive evidence of asymptomatic SARS-CoV-2 infection and transmission^12,13, research on immune correlates associated with SARS-CoV-2 infection—regardless of symptoms—remains underexplored. This research gap reflects the high logistic and financial burden inherent to a study designed to reliably ascertain asymptomatic infections, and requirements of large-scale serologic testing. Additionally, rapid viral evolution complicates efforts to characterize antibody-mediated protection^14,15. The emergence of hybrid immunity, arising from combinations of vaccination and infection, has further diversified immune profiles, making it even more challenging to establish universal correlates of protection^{16,17,18,19,20,21}. Consequently, identifying current correlates of risk and protection requires continual reassessment across different immune markers and population subgroups.

Children represent a particularly understudied subgroup. Although SARS-CoV-2 infection typically presents with milder or asymptomatic courses in the pediatric population^22,23,24,25, the relationship between antibody responses and protection from infection in children remains poorly characterized. Prior research postulates differences in pediatric presentation of SARS-CoV-2 may be caused by age-related differences in both the innate immune response, such as differential generation of inflammatory cytokines²⁶, and the adaptive immune response, such as the generation of NAbs against strain-specific antigens²². It has also been proposed that this difference in infection risk and subsequent immune response may be due to lower levels of angiotensin-converting enzyme II receptors and weaker propensity for binding of these receptors to the SARS-CoV-2 spike protein in children^22,23,27. Viral interference, the notion that greater pre-existing exposure to other viruses may create cross-reactive immunity, is another possible mechanism for differential immune protection in children compared to adults²⁸.

The limited existing pediatric correlates research has shown that younger children have higher antibody titers to SARS-CoV-2 receptor binding domain (RBD)²⁹ and mount a stronger and more immediate mucosal antibody response compared to adults³⁰. One study showed a stronger negative association between nucleocapsid-specific (anti-N; infection-induced) BAb titers and risk of infection compared to spike-specific (anti-S; vaccine- or infection-induced) BAb titers, and postulated that a combination of anti-N and anti-S BAb titers may be a better correlate than either single measure³¹. However, these conclusions are incomplete, because the few pediatric correlates studies on this topic include only short-term follow-up, typically focus on symptomatic infections alone³², and are conducted in either post-vaccination infection-naive cohorts or post-infection prospective cohorts with small sample sizes.

Despite widely available pediatric vaccines, data linking SARS-CoV-2 antibody levels to subsequent infection risk—particularly in non-naive children who are tested regardless of symptoms—remain scarce. This study aimed to quantify the association between SARS-CoV-2 antibody titers and infection risk in a large, prospective, and community-based pediatric cohort with serial PCR-based surveillance. The study cohort is particularly well-suited to identify correlates of risk in a pediatric population because follow-up extends through periods of endemic SARS-CoV-2 circulation. We evaluated whether these associations differ by age, prior infection, or vaccination history, and examined potential antibody thresholds indicative of protection. Finally, we conducted exploratory analyses of antibody kinetics after vaccination or infection to understand the antibody-level dynamics of the pediatric immune response to SARS-CoV-2.

The CASCADIA study recruited a large cohort of healthy children with high adherence to weekly PCR testing for SARS-CoV-2

The CASCADIA study enrolled 1795 participants under the age of 18 years from 1127 households between June 2022 and March 2024³³. Enrollment in the study was continuous. The majority (80%) of participants were enrolled by March 2023, 95% were enrolled by August 2023, and the last participant enrolled on December 3, 2023. A total of 1509 participants had a successful enrollment blood draw at least 7 days before their first positive SARS-CoV-2 test during study enrollment (Supplementary Fig. 1). In this study cohort, there were 84 participants (6%) between 6 months and 1 year of age, 192 (13%) between 2 and 4 years, 745 (49%) between 5 and 12 years, and 486 (32%) between 13 and 17 years at the time of enrollment (Table 1). Participants’ sex was reported to be 49% female and 51% male. Participants were predominantly non-Hispanic (89%) and White (72%) and were from affluent households, with 64% reporting an annual household income of >$100,000. A majority (84%) of participants attended daycare, school, and/or held a part-time job. Most participants (86%) lived in households with 3–5 members and 72% lived in households with 2–3 children. There was only one child in the household for 21% of the participants.

As detailed previously³³, study participants provided a blood draw at enrollment, and then underwent weekly testing for SARS-CoV-2, with additional testing if COVID-like illness (CLI) symptoms were reported. Percentages of weekly swab adherence and the number of participants enrolled and still under follow-up in the time-to-event analyses are shown in Fig. 1A. Overall, testing adherence was high, with a median of 2 swabs missed throughout the primary analysis follow-up.

Full size image

A shows the total number of participants under observation each week in light blue bars, which corresponds to counts on the left y-axis. Values on the right y-axis correspond to percentages. The percentage of participants who are still under observation in the time-to-event analysis and who are positive that week is shown in the dark blue line. The percentage of participants who are still under observation in the time-to-event analysis and who have a successful swab that week is shown in the red line. B shows the Pangolin lineage sequencing results for the n = 342 SARS-CoV-2 infection events included as outcomes in the time-to-event analyses. There were n = 233 infections sequenced, with the greatest number of observed strains being XBB.1.5 (n = 55), EG.5 (n = 31), and XBB.1.16 (n = 30).

The majority of children were recently vaccinated at their time of enrollment, and nearly half were infected with descendants of Omicron during follow-up

Most participants (93%) received at least one COVID-19 vaccine dose before enrollment and 62% had been vaccinated within the 6 months prior to their enrollment blood draw. The median time elapsed since the most recent vaccination was 4 months (IQR 1.8, 7.5). The majority (84%) of participants had received mRNA vaccine(s) manufactured by Pfizer only, compared to Moderna only (10%), or a combination of Pfizer and Moderna (4.7%). The most common vaccine formulations received prior to enrollment were original monovalent only (59%), followed by original monovalent and bivalent (33%).

Cumulative follow-up across participants was 250,013 person-days, with an overall median participant follow-up time of 131 [interquartile range (IQR) 49, 264] days. There were 342 (23%) incident laboratory-confirmed SARS-CoV-2 infections occurring prior to censoring (described below) and therefore included as events in the primary analysis. A total of 908 (60%) participants received a vaccine dose after their enrollment blood draw and were censored at their post-enrollment vaccine administration date. An additional 114 (7.6%) participants withdrew from the study before experiencing a SARS-CoV-2 infection and were censored at their study withdrawal date. Finally, 145 (9.6%) participants were administratively censored at the study end date. Demographic characteristics were similar across follow-up statuses (Supplementary Table 1). A schematic describing the time-to-event analysis is shown in Fig. 2.

Full size image

The x-axis shows study time, and the blue bar indicates the recruitment and enrollment period. Enrollment blood draws occurred between June 10, 2022, and December 3, 2024. Each horizontal black line represents a possible participant timeline. The red drop icon marks a participant’s enrollment blood draw; the green virus icon marks a SARS-CoV-2–positive test; the syringe icon marks receipt of a SARS-CoV-2 vaccine after enrollment; and the child icon marks study withdrawal. The black dashed line denotes the end of the CASCADIA study. For the risk analyses, participants were followed from their enrollment blood draw until they tested positive for SARS-CoV-2. Participants were censored at withdrawal, at the end of the study (March 31, 2024), or if they received a SARS-CoV-2 vaccine before testing positive. Gray dotted lines indicate periods when participants remained enrolled but were no longer included in the risk analysis because they had already experienced the event of interest or were censored. A small proportion of participants began submitting nasal swabs after study enrollment, but before their enrollment blood draw was successfully completed. These timelines, highlighted in light yellow, were excluded because the analysis focused on estimating infection risk in relation to baseline antibody levels.

The percentage positivity across time, among those still under observation in the time-to-event analysis, is shown in Fig. 1A. Of the 342 incident infections, there were 233 (68%) positive swabs with genetic sequencing results. All observed strains among sequenced swabs were descendants of the B.1.1.529 (Omicron) lineage, and the most frequently observed strains were XBB.1.1.5 (n = 55), EG.5 (n = 31), XBB.1.16 (n = 30), and B.Q.1 (n = 26) (Fig. 1B). The majority (77%) of the 342 incident infections used in the primary analysis self-reported at least one CLI symptom within 7 days before or after infection, and 68% of participants reported two or more CLI symptoms.

Variant-specific anti-S BAb titers were highly correlated with each other, moderately correlated with pseudovirus neutralizing antibody titers, and weakly correlated with anti-N BAb titers

All antibody titers—anti-N BAbs, variant-specific anti-S BAbs, and pseudovirus neutralizing antibody titers (pVNT) against Ancestral/WA.1 and Omicron/B.1.1.529—were correlated with each other (Fig. 3). Anti-S BAb titers were highly correlated (Spearman’s correlation coefficients R = 0.92–0.99), suggesting that potential associations between anti-S and risk of SARS-CoV-2 infection may not be strongly dependent on variants of SARS-CoV-2. Anti-S BAb titers exhibited moderate to high correlation with variant-specific pVNT titers (R = 0.70–0.89) and weak correlation with anti-N BAb (R = 0.31–0.37) titers. Anti-N BAb titers were weakly to moderately correlated with pVNT (R = 0.37–0.55). All correlations were statistically significant, even after adjusting for multiple comparisons, except for anti-N and anti-S RBD in the 6-month—1-year age group (R = 0.16, adjusted p-value = 0.15). Overall, correlations between antibodies suggested all three assay types could possibly contribute complementary information towards understanding risk of SARS-CoV-2 infection. These correlations were similar across age categories (Fig. 3) and, as defined later in this section, immune history categories (Supplementary Fig. 2).

Full size image

The lower left triangle of the grid shows scatterplots for each antibody pair. The upper right triangle of the grid shows the Spearman’s correlation coefficient for each antibody pair, overall and within each age category. The diagonal shows the marginal distribution of each antibody, overall and within each age group. The colors correspond to age categories: gray = overall, purple = 6 month to 1 year, green = 2–4 years, blue = 5–11 years, yellow = 12–17 years.

Anti-N BAb concentration may be an important proxy for a recent SARS-CoV-2 infection³⁴. To account for baseline anti-N BAb levels in our primary risk analyses, we categorized anti-N BAb levels into 3 groups based on manufacturer guidance (positive by the manufacturer guidance, i.e., high, >5000 AU/mL; negative by the manufacturer guidance, i.e., low, 5000 to 125 AU/mL; and negative by the manufacturer guidance and below the manufacturer’s lower limit of quantitation, i.e., very low, <125 AU/mL). Participant characteristics stratified by anti-N BAb seropositivity status at enrollment are shown in Supplementary Table 2. While those with very low anti-N BAbs were younger (median age 7.1 years, IQR [3.4, 10.7]) compared to those with quantifiable anti-N BAbs (10.6 [7.2, 13.6] for high and 10.0 [6.5, 13.3] for low), these groups did not differ noticeably in other baseline demographic characteristics. There was large titer variability for other antibody types within all three anti-N statuses.

There was significant overlap in immune status across age categories, despite differences in vaccine rollout dates

Age-specific enrollment is shown across time and stratified by number of prior vaccine doses in Fig. 4A. Higher rates of vaccination and number of vaccine doses were observed among those 5–17 years (Supplementary Table 3), reflecting the earlier vaccine recommendation dates for these older children compared to children under 5 years of age. The distribution of antibody titers by age and number of vaccine doses (0, 1, 2, 3, or more) is shown in Fig. 4B. Except for anti-N BAbs, after controlling for age, antibody titers were consistently highest for children with more vaccine doses. The distribution of enrollment antibody titers by age and recency of vaccination (>6 months vs. ≤6 months since last dose) in Fig. 4C showed high overlap in the antibody levels of children with and without recent vaccination.

Full size image

A Distribution of participant enrollment timing by age category and vaccination status (gray indicates 3+ doses, yellow indicates 1–2 doses, and turquoise indicates 0 doses at the time of enrollment). B Distributions of enrollment antibody titer levels by age and number of vaccine doses at enrollment. The lines through points are locally estimated scatterplot smoothing (LOESS) regressions, and the ribbon around each line is the estimated standard error. C Distribution of enrollment antibody titer levels by age and most recent vaccination dose at time of enrollment (blue indicates ≤ 6 months since the participant’s most recent vaccine; red indicates > 6 months since the participant’s most recent vaccine) among vaccinated only participants.

When combining vaccination history with enrollment anti-N status (positive or negative), most participants (58%) had vaccine-only immunity (vaccination history, negative anti-N), 35% had hybrid immunity (vaccination history, positive anti-N), 3.4% had prior infection immunity (no vaccination history, positive anti-N), and 3.2% were SARS-CoV-2 naïve (no vaccination history, negative anti-N). Distributions of enrollment antibody titers stratified by immune status are presented in Supplementary Fig. 3A. Titers were on average highest for those with hybrid immune status, and second highest for those with vaccine-induced immunity, with the exception of anti-N. Individuals who were suspected to be naive for both SARS-CoV-2 infection and vaccination at enrollment were observed to have the lowest overall levels of antibodies. When the antibodies were further stratified by participant age (Supplementary Fig. 3B), the differences by immune status remained, despite differences in vaccine rollout dates and transmission levels by age.

Both pVNT and BAb titers were negatively associated with risk of infection after accounting for possible confounders

The primary analysis including all children showed that after adjusting for age, anti-N BAb status, daycare/school/job attendance, and calendar period of enrollment, for each considered baseline antibody titers except B.1.1.529 anti-S BAb, antibody levels were significantly associated with risk of a SARS-CoV-2 infection overall (Fig. 5A). Estimated adjusted hazard ratios, which ranged from 0.60 to 0.87 per positive unit difference in log₁₀-fold antibody level (AU/mL), compare any two subpopulations of children alike in age, anti-N BAb status, daycare/school/job attendance, household size, and calendar period of enrollment, but differing in their antibody titers by tenfold. Age-specific hazard ratio estimates did not reveal any obvious pattern in the adjusted association between baseline titers and risk of infection, though conclusions were limited by large uncertainty in all age groups except 5–11 years (Fig. 5A). While point estimates for children 2–4 years of age suggested a positive association for some of the titers, null or positive associations were within the corresponding confidence intervals (Supplementary Table 4).

Full size image

A Estimated adjusted hazard ratios for a SARS-CoV-2 infection, overall and by each pediatric age category. Each adjusted hazard ratio corresponds to the difference in risk of infection for a log-tenfold higher antibody measurement, holding all other covariates constant. Overall estimates are adjusted for age category, daycare/school/job attendance, household size, anti-N category, and 30-day enrollment time period in PH models. Age-stratified estimates are adjusted for daycare/school/job attendance, household size, and anti-N category. Corresponding numbers are shown in Supplementary Table 4. B Estimated adjusted hazard ratios for a positive SARS-CoV-2 test for each quintile compared to the first quintile (i.e., lowest antibody level), adjusting for age category, anti-N status, daycare/school/job attendance, and enrollment interval. Within each antibody group, participants were divided into equally sized quintiles. Labels of antibodies correspond to pseudovirus neutralizing antibody titers (pVNT) against WA.1 and B.1.1.529 variants, anti-N BAb, and each variant-specific Anti-S BAb. For both panels, square symbols represent the hazard ratio estimates, and the error bars represent 95% confidence intervals computed using robust standard errors. Arrows at the top of the error bars indicate the confidence interval extended further than the y-axis of the graph. Hazard ratio estimates below 1 (shaded in green) represent an association between higher titers and lower risk of infection. Hazard ratio estimates above 1 (shaded in red) represent an association between higher titers and higher risk of infection.

Results suggest differential thresholds for risk of infection across antibody markers

An analysis based on antibody titer quintiles suggested a protective association over a wide range of values, most prominently for B.1.1.529 spike pVNT and anti-N BAbs (Fig. 5B). For all antibody titers, the highest quintile was significantly associated with a lower risk of infection compared to the lowest quintile. Similarly, marginalized cumulative incidence curves for antibody titer tertiles showed cumulative incidences of 45–73% over 500 days of follow-up, with a decreasing stepwise relationship seen between all titer tertiles and adjusted risk (Supplementary Fig. 4). Both analyses investigating the dose-response relationship between antibody titers and risk of infection suggested possibly differential thresholds of risk associated with each antibody type.

A negative association between antibody titers and risk of infection was seen in children with high or low anti-N BAb titers, but not in those without quantifiable anti-N BAb titers

Estimated anti-N BAb status-specific hazard ratios and corresponding confidence intervals are shown in Fig. 6A. These stratified results showed all considered antibody titers to be negatively associated with risk of infection in children with anti-N BAb titers above the lower limit of quantitation, without a meaningful qualitative difference in this association in children with high vs. low anti-N BAb titers. In children with anti-N BAb titers below the lower limit of quantitation, there was no evidence for an association between the various non-anti-N antibody titers considered and risk of infection. Titer-specific hazard ratios were found to significantly differ across anti-N BAb statuses for B.1.1.529 spike pVNT (p = 0.014) and anti-N BAbs (p = 0.021) but not other antibodies; after adjusting for multiple comparisons, these differences were no longer statistically significant. The distributions of observed antibody titers after stratification by anti-N BAb categories are shown in Fig. 6B. Although anti-N BAb correlated with several other antibody types, substantial overlap remained in the distributions of non–anti-N antibody titers across anti-N BAb categories. Similar results were obtained when restricting the analysis to participants with negative anti-N BAb status (inclusive of “very low” and “low” categories) (Supplementary Fig. 5).

Full size image

A Adjusted hazard ratios and 95% CIs for each anti-N BAb category, estimated from Cox proportional hazards (PH) regression with an interaction term for anti-N BAb status and the antibody of interest, and adjusting for age category, daycare/school/job attendance, household size, and 30-day enrollment time periodage. Square symbols represent the hazard ratio estimate for a log tenfold higher antibody titer within that antibody strata and the error bars represent 95% Wald confidence intervals computed using robust standard errors. Arrows at the top of the error bars indicate the confidence interval extended further than the y-axis of the graph. Hazard ratio estimates below 1 (shaded in green) represent an association between higher titers and lower risk of infection. Hazard ratio estimates above 1 (shaded in red) represent an association between higher titers and higher risk of infection. Wald tests for the interaction terms equaling zero showed the anti-N BAb category only modified the effect of the antibody for pVNT B.1.1.529 and anti-N BAb. B Distributions of antibody log-titers by anti-N BAb status. The horizontal center line of the boxplot indicates the median, boxes indicate the interquartile range, and whiskers extend to observations within 1.5 × IQR. The gray points show the observed log-fold biomarker values. Sample sizes for each anti-N category shown in both panels are n = 234 (very low), n = 695 (low), and n = 580 (high).

When studying immune status-stratified adjusted hazard ratios (Supplementary Fig. 3C), some differences in estimates were noted. First, the estimates among naive immune status participants exhibited large variance and appeared to be unstable, likely due to low overall variability in antibody titers and small sample sizes. While the prior infection-only immune subgroup was also small in number, these participants’ antibodies covered a wider range of values, so variance inflation was less prominent. Because differences in associations across immune categories were not found to be statistically significant, these antibodies may well prove to be valuable predictors of infection risk even across differing immune profiles.

There were also no statistically significant differences found in stratum-specific associations for vaccine manufacturer type (Pfizer-only versus Moderna-only) among vaccinated children (Supplementary Fig. 6A). Evaluating vaccine formulation type (original monovalent versus original monovalent and bivalent) also did not reveal any difference in associations between antibody levels and risk of infection (Supplementary Fig. 6B). The estimated hazard ratios for an outcome of symptomatic (CLI-1) infection were attenuated for some antibodies, but the directionality of all associations was maintained (Supplementary Fig. 7).

A variable importance analysis highlighted the potential of antibody titers to predict SARS-CoV-2 infection within 180 days

A variable importance analysis reinforced the primary findings, showing anti-N BAbs, variant-specific pVNTs, and several anti-S BAb titers all improved the ability to predict infection events within 180 days as measured by the area (AUC) under the corresponding receiver-operator characteristic curve (Fig. 7). A baseline prediction model including age, daycare, school, job attendance, and household size led to an estimated AUC of 0.556. The addition of a single antibody titer to this baseline prediction model improved the AUC up to an estimated 0.637. The largest gain in AUC occurred when adding Omicron/B.1.1.529 spike pVNT (0.081). The additional inclusion of anti-N BAbs as a predictor into a model with baseline predictors and one antibody titer led to estimates of further AUC improvement of up to 0.076.

Full size image

The y-axis shows each antibody titer considered and the x-axis shows estimated AUC for predicting infection within 180 days. The orange bar shows the estimated AUC of a prediction model based only on baseline predictors: participant age, daycare, school, or job attendance, and household size. The dark blue bar shows the estimated AUC gain when also using antibody titer of interest to predict infections. The light blue bar shows the estimated additional AUC gain when also including anti-N titer for prediction.

Additional sensitivity analyses conducted to study how the primary analysis results depend on the choice of antibody transformation used (log₁₀ scaling) revealed qualitatively similar results when using either titer ranks (Supplementary Fig. 8A) or standardized log-titers (Supplementary Fig. 8B) instead of log₁₀-titers as the predictor of interest. Results were also similar when restricting analyses to participants with titers measured for all considered antibody types (Supplementary Fig. 8C), suggesting that the slightly different subsets of participants with each titer measured did not meaningfully affect study conclusions. When censoring participants at the time of two or four consecutively missed swabs (Supplementary Fig. 9), the lower event rates affected estimator variability and thus p-values for corresponding tests, but the strength and directionality of the estimated associations remained similar. Results were robust to the exclusion of antibody measurements at the lower or upper limits of the observed range (Supplementary Fig. 8D). Finally, antibody levels did not differ when stratified by sex at birth (Supplementary Fig. 10A), and results of the primary analysis did not qualitatively change when also adjusting for sex at birth (Supplementary Fig. 10B).

Longitudinal serum antibody data suggest that vaccination and infection each boost all antibody titers considered

All CASCADIA participants were prompted to obtain blood draws one year after enrollment, and an optional substudy prompted participants to obtain additional blood draws after SARS-CoV-2 vaccination or infection (see “Methods”). These additional blood draws for a limited subset of participants were used to provide insight into antibody kinetics in a pediatric population. The antibody levels for 170 participants with a blood draw both within 90 days before and 90 days after their first observed SARS-CoV-2 infection, and no SARS-CoV-2 vaccination in between the blood draws, are shown in Fig. 8. The majority of participants were observed to have higher antibody levels after infection, and pre- and post-infection measurements were significantly different for all antibodies (p < 0.001); these differences remained statistically significant after adjustment for multiple testing.

Full size image

Antibody levels before and after participants’ first SARS-CoV-2 infection, for n = 170 participants with at least two blood draws: one within 90 days before and one within 90 days after infection. The horizontal center line of the boxplot indicates the median, boxes indicate the interquartile range, and whiskers extend to observations within 1.5 × IQR. The violins show the distribution of each antibody before and after infection, and the lines connect each participant’s before and after blood draw. The color of the lines indicates the magnitude of the difference in blood draw before or after infection: blue lines indicate the antibody after infection was similar or less than the antibody before infection. Purple and red lines indicate the antibody after infection was larger than the antibody before infection. All p-values were <0.001 for a difference in antibody titers before and after infection using a two-sided Wilcoxon signed rank test for paired data (statistically significant at alpha level 0.05 even after multiple testing correction).

We additionally examined antibody levels measured at least 14 days after SARS-CoV-2 vaccination but before SARS-CoV-2 infection (notably, these are not enrollment antibody levels, but rather antibodies measured after post-enrollment vaccination), and observed patterns of antibody waning across age groups. The majority (86%) of the 629 participants with antibody levels measured more than 14 and less than 200 days after a vaccination that occurred during study follow-up were vaccinated with an updated monovalent mRNA vaccine targeting the XBB lineage of the Omicron variant. When studying by age category, all available antibody level measurements that matched such criteria (Supplementary Fig. 11A), younger age categories exhibited overall linear trends with steeper declines over time for all antibodies except for anti-N. A subset of participants with multiple blood draws (n = 137, each with 2 blood draws) were observed to have a decline in antibody titers between their first and second blood draws post-vaccination, or remain stable in this time period (Supplementary Fig. 11B).

This study assessed a variety of neutralizing and binding antibody titers as potential correlates of risk for SARS-CoV-2 infection in children using data from CASCADIA, a large, prospective, community-based study of a highly vaccinated cohort in Washington and Oregon during endemic circulation. Results suggest that high antibody levels across multiple strains are associated with a lower risk of SARS-CoV-2 infection in children after controlling for age; recent infection history via anti-N BAb status; daycare, school, and/or part-time work attendance; and temporal trends in community transmission levels via calendar period of enrollment. While these results are concordant with findings in adult populations, they underscore pediatric-specific heterogeneity across age and immune profiles, including marker-dependent associations and apparent threshold-like shifts in risk.

The observed negative association between higher baseline antibody titers and infection risk highlights the potentially protective role of pre-existing immunity. The particularly strong association seen for anti-N BAbs and spike pVNTs emphasizes their predictive value in assessing susceptibility to SARS-CoV-2 infection in children. This suggests a broad range of protective antibodies are derived after infection, including strain-specific antibodies not measured in this study. Evaluation of the estimated dose-response relationships based on categorized titer levels suggests a shift in the magnitude of association may occur at lower titer levels for pVNTs and anti-N BAbs. In contrast, a similar shift in association occurs at higher titer levels for most spike BAbs, possibly suggesting qualitatively different thresholds of risk. This pattern supports prior immunological studies that indicate that the quantitative level of antibody titers required for effective protection may vary based on strain-specific antibody and the circulating variant^35,36,37.

The analysis stratified by anti-N BAb status revealed interesting differential relationships. In children with baseline anti-N BAbs above the lower limit of quantitation, consistent inverse associations were seen across all antibody types considered. Conversely, in those with very low anti-N BAbs, no significant association between enrollment antibody titers and infection risk was found. This discrepancy may reflect the inherent differences in immune memory and response following infection versus spike-only vaccination, particularly for prevention of infection^16,18,38. It may additionally reflect a quantitative phenomenon wherein antibody levels in naïve individuals tend to be lower than in non-naïve individuals, and only variability in a sufficiently high range may be relevant for discriminating risk, at least in a highly vaccinated population³⁹. Identifying the threshold of anti-N BAb titers that most accurately reflects a likely prior infection in children remains an open question.

The analysis also showed that titers for all antibody types contributed to improved prediction of infection risk, reinforcing the importance of variant-specific pVNTs in protection against SARS-CoV-2 infection and highlighting the value of markers of recent infection for determining susceptibility to future infection. A prospective cohort study in the United Kingdom studied anti-N and anti-S BAbs as correlates of SARS-CoV-2 re-infection in 9–13 year-olds during 2022, and found low anti-N titers to be more strongly predictive of re-infection than low anti-S titers³¹. Consistent with other studies, our unit-independent variable importance analysis suggested that neutralizing antibody titers may be stronger correlates of risk than anti-S BAbs³⁵.

Overall, the associations between antibody titers and risk of SARS-CoV-2 infection were generally consistent across subgroups. Although several antibodies and antibody combinations showed negative associations with infection risk, stratified analyses revealed few meaningful differences in association estimates across immune profiles. No statistically significant heterogeneity was observed across immune status categories, suggesting that these antibody markers may retain predictive value regardless of prior infection or vaccination history. Similarly, among vaccinated children, stratification by vaccine manufacturers did not reveal significant differences in the strength of associations between antibody levels and infection risk. Analyses comparing vaccine formulation types (original monovalent versus original monovalent plus bivalent) likewise showed no clear evidence of a differential association. Finally, associations were consistent when restricting the outcome to symptomatic infection, with hazard ratios comparable to those obtained when considering any infection. While these analyses are limited by smaller numbers of events within certain subgroups, together, these findings suggest that the observed antibody titer-infection risk relationships were robust across immune backgrounds, vaccine types, and outcome definitions.

One of the study’s key strengths lies in its large, prospective, community-based cohort, which included children across a wide age range, enabling an age-specific evaluation of correlates of risk. To our knowledge, this represents the largest study to date—of any design—characterizing SARS-CoV-2 antibody responses in a pediatric population. In contrast to most existing studies, the incorporation of weekly participant testing provided a rare opportunity to examine correlates of risk for infection itself, including asymptomatic infection, rather than limiting inference to symptomatic disease. Indeed, a study design that only tested for SARS-CoV-2 if CLI symptoms were reported by participants would have missed 23% of the infection events captured in this analysis. The study period (2022–2024) also captured a critical phase of the pandemic when most children enrolled in this study had been vaccinated and many had acquired hybrid immunity, following the relaxation of school-based mitigation measures in Washington and Oregon. This context differs substantially from earlier correlates analyses conducted within randomized, placebo-controlled trials that focused primarily on short-term protection in SARS-CoV-2–naïve participants^{35,40,41,42,43}. Notably, the Moderna COVE study recently demonstrated that anti-S BA.1 binding and neutralizing antibody titers after a third mRNA-1273 dose were correlates of protection against BA.1 COVID-19 in non-naïve individuals³⁶. As such, correlates of risk findings from this study are complementary to existing findings and contribute to the growing evidence base supporting antibody correlates of risk and correlates of protection in diverse, real-world pediatric populations.

Despite these insights, the study has limitations that must be considered when interpreting results. First, irregular adherence to weekly swabbing by some participants may have resulted in missed asymptomatic infections, which may be especially problematic if the likelihood of an asymptomatic infection is itself modulated by antibody titers. The high overall testing adherence observed in this study nevertheless mitigates potential biases from underreporting of infections. In addition, sensitivity analyses showed that while confidence intervals are wider when participants are censored at the time of missed tests, the associational direction suggested by point estimates remains consistent. Second, hazard ratio estimates may not be directly comparable across antibody types due to differences in measurement scales. To mitigate this challenge, we conducted regression analyses under two different titer standardizations that seek to increase comparability. We also implemented a variable importance analysis that circumvents this issue entirely by directly quantifying the predictive ability of different antibody titers. This variable importance analysis also allows us to investigate the additive predictive value of pairs of antibody titers, rather than one titer alone. Third, the immune status categorizations of vaccine-induced only or naive immunity may be inaccurate, as the manufacturer threshold for determining anti-N positivity may not be optimal for pediatric populations. Further research is needed to clarify this issue. A related limitation is that antibody measurements truncated by the assay’s limits of detection could bias the estimated hazard ratios. However, in sensitivity analyses excluding such antibody titer values, results were qualitatively similar, suggesting that this limitation is unlikely to materially affect our findings.

In the future, it may be fruitful to perform longitudinal assessments of both antibody markers and functional immune responses to better capture the temporal nature of immunity. Although we are able to gain some insight into antibody kinetics, this is a complex question and not a primary objective of the study design or data collection process. Additionally, evaluating the interplay between cell-mediated immunity and humoral responses could provide a more complete understanding of protective immunity in children. Recent research also highlights the need to analyze mucosal and saliva antibodies in addition to serum antibodies^32,44,45, as well as the importance of comparing T cell responses to other less durable antibody markers like anti-N BAbs⁴⁶. Finally, opportunities for research include estimating population susceptibility to specific variants based upon antibody markers and assessing variable risks of community transmission.

In conclusion, the CASCADIA cohort study provides unique and valuable insights into the protective associations between baseline antibody titers and SARS-CoV-2 infection risk in children. The differential effects observed based on prior infection status and the strong predictive value of titers for certain antibody types underscore the complexity of pediatric immune responses to SARS-CoV-2. These findings may inform vaccine strategies and risk assessment in pediatric populations as new viral variants continue to emerge.

Study design

CASCADIA is a community-based prospective cohort study in which participants 6 months to 49 years of age were enrolled from households in Washington and Oregon, USA, starting June 2022, with follow-up ending March 2024³³. Study sites included the University of Washington (UW) (Seattle, WA) and the Kaiser Permanente Northwest (KPNW) Center for Health Research (Portland, OR). Following enrollment, participants completed questionnaires, which included sociodemographic information and vaccination history. Participants were asked to provide a blood sample at enrollment and at one year of follow-up. An optional substudy prompted participants to provide a blood sample 30 days after SARS-CoV-2 vaccination and/or infection. All participants less than 18 years of age at enrollment and who had an enrollment blood sample with anti-N BAb measured at least 7 days prior to their first SARS-CoV-2 positive swab were included in this analysis. This excluded participants who began self-swabbing and tested positive for SARS-CoV-2 on or before a successful enrollment blood draw.

Nasal swab collection and SARS-CoV-2 sequencing

Irrespective of symptoms, participants self-collected weekly anterior nasal swabs, with additional swabbing if they experienced at least one acute respiratory illness symptom >72 h after their previous nasal swab. The nasal swabs were tested for SARS-CoV-2 using real-time reverse transcriptase–polymerase chain reaction. Positive swabs with a cycle threshold ≤30 were selected for genetic sequencing. Raw sequencing reads were processed, and alignment files and consensus sequences were generated using a custom reference-based assembly bioinformatics pipeline (https://github.com/seattleflu/assembly). Consensus sequences that met the quality criteria were submitted to NCBI GenBank (Bioproject PRJNA746979) and GISAID. Pango lineage classification was performed using Nextclade⁴⁷. Sequence alignment was carried out with MAFFT, and a phylogenetic tree was constructed with bootstrap support using IQTree⁴⁸. Pairwise sequence distances were calculated with a custom Python script.

Serum sample processing

Serum samples were tested for SARS-CoV-2 antibodies using both binding and neutralization assays. BAbs were measured with a commercial electrochemiluminescence immunoassay on plates pre-coated with SARS-CoV-2 antigens (Meso Scale Diagnostics, Rockville, MD). Each sample was run at both low and high dilutions, as detailed in the Supplementary Materials, and plated on 96-well MSD plates along with controls and reference standard calibrators. Assays were performed on MSD plates coated with anti-N, whole spike protein from multiple variants (Ancestral/WA.1, Alpha/B.1.1.7, Omicron/B.1.1.529, Beta/B.1.351, Delta/B.1.617.2, Gamma/P.1), and Ancestral/WA.1 RBD.

Serum samples were aliquoted and heat-inactivated prior to testing in a lentivirus-based pseudovirus neutralization assay. Spike genes from SARS-CoV-2 variants B.1 D614G, BA.1, and XBB.1.5 were codon-optimized and individually cloned into a pHDM expression vector (BEI NR-53742). Each spike plasmid was co-transfected into 293T cells with the lentiviral backbone pHAGE-CMV-Luc2-IRES-ZsGreen-W (BEI NR-52516) and the required helper plasmids to generate lentiviral particles pseudotyped with the corresponding spike variant.

Measurements that fell outside MSD’s specified assay range were flagged for each antigen. By default, the lower limit of detection was defined as the concentration 2.5 standard deviations above the lowest point on the calibration curve, and the upper limit corresponded to the highest calibrator. Additional laboratory methods are detailed in the Supplementary Materials.

Data source

The primary outcome was time from enrollment blood draw until the first recorded positive SARS-CoV-2 swab result. Participants were censored at the time at which they received any COVID-19 vaccine post-enrollment since their enrollment titers were then no longer directly relevant to their infection risk. Participants were also censored if they withdrew from the study or reached the study end date (March 31, 2024) without experiencing infection.

The exposures of interest were 10 antibody titers measured at the enrollment blood draw. These included (i) pVNT, expressed as the IU/mL required to achieve 80% neutralization, against Ancestral/WA.1 and Omicron/B.1.1.529 variants; (ii) binding antibody (BAb) IgG concentrations against nucleocapsid, the spike receptor-binding domain (RBD), and spike proteins from Ancestral/WA.1, Alpha/B.1.1.7, Beta/B.1.351, Gamma/P.1, Delta/B.1.617.2, and Omicron/B.1.1.529. All antibodies were analyzed in the AU/mL or IU/mL unit scale and log₁₀ transformed for analysis. Each antibody type was analyzed as an individual exposure, and all participants with a measured titer for a specific antibody type were included in analyses involving that antibody type.

Statistical methods

Overall enrollment patterns, participant demographics, enrollment antibody titers, and infection endpoints were described using data visualizations and tabulations. A correlogram was used to display the relationship between different antibody markers. Correlations were quantified using Spearman’s correlation coefficient and corresponding p-values were adjusted for multiple comparisons using the Bonferroni–Holm procedure. Key cohort characteristics were summarized using median and interquartile ranges or counts and proportions.

For the primary analysis, Cox proportional hazards (PH) regression models were used to estimate the adjusted association between enrollment titers and risk of SARS-CoV-2 infection, with each antibody type evaluated separately. Adjustment variables included: daycare/school attendance and/or job (yes; no; or missing); participant age at enrollment (6 months–1 year; 2–4 years; 5–11 years; or 12–17 years), except for age-stratified analyses; household size category (1–2 members, 3–4 members, ≥5 members), and enrollment anti-N BAb status (high: >5000 AU/mL; low or very low: ≤ 5000 AU/mL). The anti-N BAbs were included as a proxy for recent infection and the status categories followed the manufacturer’s threshold for a positive anti-N BAb result (indicating high likelihood of a recent infection). Calendar time, defined in 30-day increments from initiation of study enrollment in June 2022, was used to create separate baseline hazards to flexibly account for temporal trends in SARS-CoV-2 community transmission levels. PH regression models were fitted on the entire study cohort and within each age category separately. Robust standard error estimates were used to account for within-household correlation. Wald confidence intervals (CIs) were constructed on the coefficient scale and then exponentiated to obtain corresponding CIs for hazard ratios.

A number of additional analyses were conducted. First, as a refinement to the primary analysis, PH regression models were implemented with continuous titers replaced by titers categorized into five groups—very high, high, mid, low, very low—based on quintiles of the sample titer distribution. These models allow an assessment of the possibly nonlinear dose-response relationship between enrollment titers and risk of SARS-CoV-2 infection. Monotonicity of this dose-response relationship was enforced with post-hoc regularization using isotonic regression⁴⁹. Similar analyses using antibody tertiles as the exposure were also conducted, and the covariate-adjusted cumulative incidence estimates were plotted against time since enrollment. Sensitivity analyses were then conducted, excluding participants with antibody measurements at the lower or upper limits of the observed range, to assess whether values possibly at the limits of detection influenced results. Finally, sex at birth was added to the adjustment covariate set in additional sensitivity analyses.

Second, to estimate the adjusted association between enrollment titers and risk of infection based on recent infection history, PH regression models including interaction terms between anti-N BAb status and enrollment titers were used. A multivariate Wald test based on robust standard error estimates and inverse-variance weighting was used to assess whether titer-specific hazard ratios differ across anti-N BAb statuses; p-values were adjusted for multiple comparisons using the Bonferroni–Holm procedure.

Third, to render hazard ratios corresponding to different antibody types more directly comparable, the primary PH regression analysis was repeated using, instead of log₁₀-titers, (i) titer ranks: the relative rank of observed titers (i.e., empirical distribution function-transformed titer values) multiplied by 10; or (ii) standardized log-titers: the observed log-titers standardized by their sample standard deviation. The resulting hazard ratio estimates then correspond to a comparison in subpopulations with common values for adjustment variables but differing in log₁₀-titers (i) by 10 percentile points and (ii) by one standard deviation in the sample titer distribution, respectively.

Fourth, to investigate possible biases stemming from missed swabs, the primary analysis was repeated with additional censoring when participants missed two or four consecutive swabs.

Fifth, to disentangle possible mechanistic relationships underlying antibody levels prompted by infection, vaccination, or a combination of the two, the primary analysis was repeated with an interaction term for each immunity profile and log-transformed antibody combination. This allowed for the categorizations of hybrid, vaccine-only, infection-only, and naive immunity; however, estimates for naive immunity were unstable due to small sample sizes and low variation in the antibody levels of naive participants. The primary analysis was also repeated using an interaction term between antibody titers and categorizations of children who received Pfizer-only vaccinations, as well as Moderna-only, Moderna and Pfizer combinations, and no previous vaccinations. This analysis was similarly repeated using interaction terms for the vaccine formulation histories of: original monovalent only, original monovalent and bivalent, another formulation combination, and unvaccinated.

Sixth, the primary analysis was repeated using only symptomatic infections as events. Symptomatic infection was defined as at least one self-reported COVID-like illness (CLI-1) symptom within 7 days before or after their first positive swab. CLI symptoms included: fever or chills, cough, shortness of breath or difficulty breathing, fatigue, muscle or body aches, headache, (new) loss of taste or smell, sore throat, congestion or runny nose, nausea or vomiting, diarrhea, chest pain/persistent pain or pressure in the chest, and pale, gray, or blue-colored skin, lips, or nail beds.

Seventh, a variable importance analysis was performed to compare the extent to which different baseline titers improve infection prediction beyond adjustment variables alone, with AUC used as a measure of predictive ability^50,51. This analysis allows direct comparison of antibody types quantified differently, including in different units and with different distributions. The variable importance analysis was implemented using PH regressions for (i) the baseline model (predictors included age and daycare, school, or job attendance), (ii) an antibody model (baseline predictors plus antibody), and (iii) an antibody plus anti-N model (baseline predictors plus both antibody and anti-N), all fit on the participants with all enrollment antibodies measured. The set of predictors described in (iii) was not assessed for anti-N. Repeated sample splitting across 10 different seeds was used to ensure valid testing of null importance while limiting variance inflation⁵⁰.

Lastly, a subset of participants who enrolled in an optional substudy of CASCADIA, which prompted additional blood draws 30 days after SARS-CoV-2 infection or vaccination, had additional longitudinal antibody measurements available for analysis. Participants with an antibody measurement available both 90 days before infection and 90 days after infection were graphed using box plots, and differences in pre- and post-infection titers were tested using Wilcoxon signed rank tests for paired data, with p-values adjusted for multiple comparisons using the Bonferroni–Holm procedure. Second, all participants with a measured antibody titer between 14 and 200 days post-vaccination were compared across age groups. Linear regression of log₁₀ antibody titers against days since vaccination was used to visualize titer trends within age categories. Then, only participants with paired antibody measurements between 14 and 200 days post-vaccination were graphically assessed to gain insight into possible trends in antibody waning.

All hypothesis tests implemented were two-sided and conducted at significance level 0.05. Analyses were performed in R version 4.3.1⁵² using open-source packages tidyverse 2.0.0⁵³, gtsummary 2.5.0⁵⁴, GGally 2.4.0⁵⁵, survival 3.8–6⁵⁶, and survML 1.2.0⁵⁷. The source data underlying all main text figures and tables, and code to reproduce each figure and table, are stored in a public Code Ocean capsule. Toy datasets and code to run all primary, secondary, and sensitivity analyses are also located in this capsule.

Human subjects

This study was reviewed and approved by the KPNW Region Institutional Review Board (See 45 C.F.R. part 46.114; 21 C.F.R. part 56.114). Informed consent was obtained for all adult participants. For pediatric participants, informed consent was obtained from the minor’s parent or legal guardian. We also obtained assent from children aged 7–17. If a participant turned 18 while participating in the study, we then consented these individuals as adults.

Reporting summary

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