Abstract
Protection against SARS-CoV-2 wanes over time, and booster uptake has been low, in part because of concern about side effects. We examined the relationships between local and systemic symptoms, biometric changes, and neutralizing antibodies (nAB) after mRNA vaccination. Data were collected from adults (n = 364) who received two doses of either BNT162b2 or mRNA-1273. Serum nAB concentration was measured at 1 and 6 months post-vaccination. Daily symptom surveys were completed for six days starting on the day of each dose. Concurrently, objective biometric measurements, including skin temperature, heart rate, heart rate variability, and respiratory rate, were collected. We found that certain symptoms (chills, tiredness, feeling unwell, and headache) after the second dose were associated with increases in nAB at 1 and 6 months post-vaccination, to roughly 140-160% the level of individuals without each symptom. Each additional symptom predicted a 1.1-fold nAB increase. Greater increases in skin temperature and heart rate after the second dose predicted higher nAB levels at both time points, but skin temperature change was more predictive of durable (6 month) nAB response than of short-term (1 month) nAB response. In the context of low ongoing vaccine uptake, our convergent symptom and biometric findings suggest that public health messaging could seek to reframe systemic symptoms after vaccination as desirable.
Introduction
Vaccination against SARS-CoV-2 has been repeatedly shown to reduce infections, hospitalizations, and mortality, but protection wanes considerably over time for all of these outcomes, even following booster vaccination.1 Moreover, uptake of booster vaccinations has been low, with only 17% of the US population having received the bivalent booster as of May 2023, despite the vaccine having been widely available for over six months at that time.2 Among individuals who received at least one dose of a Covid-19 vaccine, the most commonly reported reasons for not having received a booster were: first, a perception of low added benefit in protection from illness, given a personal history of prior vaccination or SARS-CoV-2 infection, and second, worry about side effects.3,4
Recent evidence has suggested that greater systemic symptoms following SARS-CoV-2 vaccination may reflect a more potent immune response.5–7 A deeper understanding of this relationship may help to address low rates of vaccine uptake. If the association is clinically meaningful, public health messaging might aid uptake by clearly reframing some post-vaccination symptoms as positive indications that the vaccine is likely to be working rather than undesirable side effects.
Although there are several reports suggesting that SARS-CoV-2 vaccine reactogenicity (i.e., resulting symptom burden or physiological perturbation) predicts higher subsequent anti-spike immunoglobulin level,5–7 only a small number of studies have specifically measured neutralizing antibodies (nAB), and results are inconsistent.8–10 Quantifying functional antibody activity (i.e., nAB) is important because although they are correlated, vaccine effects on nAB and absolute anti-spike IgG are dissociable, and nAB specifically appear critical in conferring protection from Covid-19. Only approximately 50% of the variability in nAB is predictable from anti-spike IgG,11 and nAB has been reported to have a larger effect size (i.e., lower hazard ratio per 10-fold increase) than anti-spike IgG in predicting subsequent Covid-19 incidence.12 It has been demonstrated that providing animals with neutralizing antibodies alone confers protection against disease even after high-dose SARS-CoV-2 exposure,13 and in one study in humans, nAB level was estimated to mediate over two-thirds of vaccine efficacy.12 A recent meta-analysis14 and large pooled cohort analysis15 have estimated the rank correlation between nAB level and protection against symptomatic infection to be 0.81 and 0.91 respectively.
Using data from a cohort of adults who received the initial two-dose series of BNT126b2 or mRNA-1273, we used convergent self-report symptom and objective biometric measurements to examine the association between physiological perturbation following vaccination and subsequent nAB concentration in serum at 1 and 6 months after full vaccination. We examined symptom predictors for each dose that included the presence or absence of 13 individual symptoms and total symptom burden. Objective measurements we assessed included vaccination-induced change in skin temperature, heart rate, heart rate variability, and respiratory rate, which were measured during sleep using a wearable device (the Oura ring).
Methods
Subjects
Subjects were participants in the Building Optimal Antibodies Study,16 a large observational study designed to identify psychosocial, behavioral, and biological predictors of immune responses to COVID-19 vaccination. Participants were adults aged 18 years and above who did not have a previous history of immune-related diseases and were not currently undergoing treatment with medications known to impact the immune system. Ethical approval was provided by the Institutional Review Board at the University of California, San Francisco, and all study participants provided written informed consent.
Serum was collected from study participants before they received a COVID-19 vaccine and again 1 and 6 months after they completed their initial two-dose series of BNT162b2 or mRNA-1273. Participants independently arranged to be vaccinated in the community, and vaccination date and type was later determined using official records. History of SARS-CoV-2 infection was examined by measuring levels of anti-spike IgG antibodies at baseline and anti-nucleocapsid IgG antibodies at 6 months. Participants with a positive result on either test were excluded from analyses. Participants were also excluded if they received a third vaccine dose before the 6-month follow-up, or if no outcome measurement was available. Sample characteristics are provided in Table 1.
Demographic characteristics are provided for subjects included in at least one symptom analysis (left) or biometric analysis (right). Biometric wearable devices were offered to a subset of participants who were over 50 years of age and had a compatible smartphone. A total of 364 subjects contributed to analyses, with one individual providing data for the biometric but not symptom analyses.
Outcome
Neutralizing antibodies against SARS-CoV-2 were measured via pseudovirus assay at 1 and 6 months following vaccination as described previously.16 In brief, serum from each participant was serially diluted and incubated with pseudovirus expressing full-length SARS-CoV-2 protein (Wuhan/D614G strain), permitting virion binding and neutralization by host antibodies. Serum-virus mixtures were then incubated with susceptible cells, allowed remaining functional pseudovirus to deliver a luciferase reporter gene intracellularly. After 66-72 hours, the medium was removed, lysis buffer and luciferase substrate were added, and luciferase activity was measured as luminescence. Neutralizing antibody titers were expressed as the inhibitory dose 50 (ID50), defined as the serum dilution at which relative light units (RLU) would be reduced by 50% compared to serum-free control wells.
Daily symptom surveys
Subjects were sent links to surveys each evening for six days, beginning on the date they anticipated receiving each dose of a SARS-CoV-2 vaccine. The survey included the question, “Did you experience any of the following physical symptoms today? (Check all that apply.)”. The following options were provided, in order: Tiredness; Headache; Muscle pain; Chills; Joint pain; Fever; Nausea / vomiting; Feeling unwell; Tender or swollen lymph nodes (lymphadenopathy); Injection site pain, redness or swelling; Pain or swelling in the arm that did not get the vaccination; Other allergic reactions (difficulty breathing, swelling of face/throat, rash); Stomachache; None of the above. For each survey entry, vaccine dose dates were used to calculate calendar days since receipt of either dose one or two. For each symptom (or group of symptoms), data were collapsed to reflect either presence on any of the six days or absence on all six days.
Biometric collection and analysis
Heart rate (HR), heart rate variability (HRV), respiratory rate (RR), and skin temperature (ST) data were collected from a subset of participants using a biometric wearable device, the Oura ring. Except for one individual, all individuals who provided biometric data were aged over 50 years, because only these individuals were actively offered devices. Collected data were exported from the web application Oura Teams in March 2022. HR and HRV were exported for each 5-minute interval of the longest sleep period per 24 hours. HRV was provided in the form of the root mean square of successive differences (RMSSD).17 Skin temperature was recorded for every 1 minute that the device was worn; these data were filtered to only those recordings that occurred during sleep.18
To test the hypotheses that short-term effects of vaccination on nighttime HR, HRV, RR, and ST are predictive of subsequent neutralizing antibody response, it was necessary to derive summary values of vaccination-induced change in each of these physiological domains for each subject. For this purpose, for each domain, we used a multi-step procedure to identify the summarization approach that best indexed vaccination-induced change;19,20 see the eMethods in Supplement 1 for details. Ultimately, nightly time series were first summarized into single nightly values for each subject for ST and HRV by taking the 99th percentile (i.e., the “stable maximum”), and for HR by taking the 1st percentile (i.e., the “stable minimum”); RR was only available in the form of a nightly average. Then, one summary value was derived for each subject for each physiological domain at each vaccine dose, by taking the peak deviation (maximum for HR, RR, and ST; minimum for HRV) from each subject’s individualized norm (or “baseline”) over the first two nights following receipt of a given vaccine dose. Descriptive and test statistics for candidate summary variables of vaccination-induced change are provided in eTable 2 in Supplement 1. Spearman correlations between final summary variables and symptom count are presented in eFigure 1 in Supplement 1.
Participants completed daily surveys on the day of and five days following each vaccination dose, reporting the presence or absence of specific symptoms each day. For each symptom (or group of symptoms), data were collapsed to reflect either presence on any of the six days or absence on all six days. Neutralizing antibodies were measured at 1 month and 6 months following completion of the initial vaccine series. Each symptom at each dose was individually tested as a potential predictor of neutralizing antibody levels at both outcome time points simultaneously, in multivariable mixed-effects models that controlled for other variables (i.e., vaccine, sex, age, and body mass index). After correcting for multiple comparisons, four symptoms remained statistically significant predictors of neutralizing antibodies, although only when measured at dose two: chills, tiredness, feeling unwell, and headache. Injection site symptoms are included in the figure for comparison. Blue lines represent the marginal means +/- 95% confidence intervals. There was no interaction between the effect of each symptom and vaccine or time, nor was the three-way interaction between these variables significant. Therefore, marginal means are not stratified by these variables, and instead represent the average effect across both vaccines and both outcome time points.
Data analysis
Data processing, graphing, and statistical analysis were performed in R v4.2.2. For all analyses, mixed-effects models were fit to nAB data collected at 1 month and 6 months following completion of the initial vaccine series. All models included a core set of terms, including a time point × vaccine interaction and main effects of sex, age, and body mass index. The statistical significance of these terms has been previously described.16 Here, for each vaccine dose, 18 variables were examined as predictors of subsequent nAB level: the presence or absence of 13 symptom categories, the total count of reported symptom categories (excluding injection site symptoms), and the levels of 4 biometric measurements. For each variable, a model was created by adding to the core model structure the following terms: a main effect, an interaction with vaccine, an interaction with time point, and the three-way interaction between these variables. Thus, four hypotheses of interest were tested in each model, with the exception of five models where one or more interaction terms were removed to resolve multicollinearity (see eMethods in Supplement 1 for more detail and eTable 3 in Supplement 1 for the full list of tested hypotheses and associated statistics). Predictor significance was tested using F statistics. Ultimately, 126 p-values (1 to 4 per model) were drawn from 36 models; these were consolidated and corrected for multiple comparisons using the Benjamini-Hochberg method. Statistical significance was defined as corrected p < 0.05; significant F-statistics were followed by post-hoc t-tests without further correction. All presented results represent marginal effects, i.e., effects adjusted for the other terms in the model. Thus, where results are presented without respect to outcome time point, these represent effects averaged across both time points. For statistically significant continuous predictors, the partial correlation (rp) was provided alongside absolute effect sizes. Visualizations represent marginal means +/- 95% confidence intervals (CI) along with partial residuals. Detailed information can be found in the eMethods in Supplement 1.
Results
Symptom predictors of neutralizing antibodies
Each symptom was examined as a predictor of subsequent nAB using separate multivariable models. The frequency of each symptom at each vaccine dose is provided in eTable 1. Following correction for multiple comparisons, no statistically significant effects of symptom presence or absence at dose one were identified. For dose two, no interactions were significant, while main effects were significant for four of 13 symptoms (Figure 1): chills (F1,338.0 = 17.78, p = 0.001), tiredness (F1,339.3 = 14.38, p = 0.004), feeling unwell (F1,338.7 = 15.75, p = 0.003), and headache (F1,338.9 = 14.56, p = 0.004). nAB were higher for subjects reporting vs. not reporting the following symptoms at dose two: chills (1.6 fold higher; increase in log10 ID50 = 0.21, CI 0.12 to 0.30), tiredness (1.5 fold higher; increase in log10 ID50 = 0.17, CI 0.069 to 0.26), feeling unwell (1.5 fold higher; increase in log10 ID50 = 0.17, CI 0.088 to 0.25), and headache (1.4 fold higher; increase in log10 ID50 = 0.15, CI 0.075 to 0.24). Because symptom presence did not interact with outcome time point or vaccine for any symptom, all results are presented without regard to outcome time point and vaccine. Specifically, results represent differences in marginal means (also known as least-squares means); i.e., values represent the average effect of each symptom across both vaccines and both time points, weighting these according to their representation in the sample.
Symptom count as a predictor of neutralizing antibodies
Symptom count was intended as a continuous index of systemic symptom burden, so injection site symptoms were excluded from counting. For dose one, there were no main or interaction effects involving symptom count. For dose two, no interactions were significant, but there was a main effect of symptom count (F1,337.9 = 24.50, p < 0.001; Figure 2), involving a 1.1-fold change in nAB (ID50) per additional symptom (change in log10 ID50 per additional symptom: 0.042, CI 0.026 to 0.058; partial correlation, rp = 0.27, CI 0.17 to 0.36).
Symptom count was intended as a measure of systemic symptom burden, so injection site symptoms were excluded from counting. Symptom count following the second dose was a statistically significant predictor of subsequent neutralizing antibody level (p < 0.001). For both doses, there was no significant interaction between symptom count and vaccine brand or measurement time point (1 month and 6 months following initial series completion); therefore, results represent the average relationship across both time points and both vaccines (i.e., marginal means +/- 95% confidence intervals). Estimates also represent the average across both sexes and are conditioned on sample means for age and body mass index.
Biometric predictors of neutralizing antibodies
For vaccination-induced change in nightly 99th percentile at dose one, there were no significant main or interaction effects. However, at dose two, there was a significant interaction between outcome time point and vaccination-induced change in nightly 99th percentile skin temperature (F1,127.5 = 13.96, p = 0.005; Figure 3, top right). Post-hoc testing revealed that vaccination-induced change in nightly skin temperature was predictive of nAB at 1-month follow-up (fold change in ID50 per degree Celsius: 1.8; change in log10 ID50 per degree Celsius: 0.26, CI 0.13 to 0.40, p < 0.001; rp = 0.27, CI 0.13 to 0.39) and at 6-month follow-up (fold change in ID50 per degree Celsius: 3.1; change in log10 ID50 per degree Celsius: 0.50, CI 0.35 to 0.64, p < 0.001; rp = 0.45, CI 0.33 to 0.55), with the steeper relationship for the 6-month follow-up being responsible for the interaction.
A subset of study participants wore biometric devices (Oura rings) that collected physiological measurements during sleep. Data were available as nightly time series. The nightly stable maximum (99th percentile) was identified to be most sensitive to vaccination-induced change for ST, while the nightly stable minimum (1st percentile) was most sensitive to vaccination-induced change for HR. Once the most sensitive index of vaccination-induced deviation was identified for each physiological domain, these indices were then tested as potential predictors of subsequent neutralizing antibody responses at 1 month and 6 months simultaneously. Plots depict the relationship between vaccination-induced change in nightly stable-maximum ST (top row) and nightly stable-minimum HR (bottom row) and subsequent neutralizing antibody level at 1 month and 6 months following completion of the initial two-dose series of BNT162b2 and mRNA-1273. Vaccination-induced change in stable-maximum ST at dose two predicted subsequent neutralizing antibody (nAB) level at both 1 month and 6 months later (top right). There was a main effect of vaccination-induced change in stable-minimum HR on subsequent nAB level (bottom right).
For vaccination-induced change in nightly 1st percentile heart rate at dose one, there were no significant main or interaction effects. However, at dose two, a main effect of vaccination-induced change in nightly 1st percentile heart rate was observed (F1,132.0 = 10.20, p = 0.029; Figure 3, bottom right), in the absence of any interaction with outcome time point or vaccine. Averaged across the levels of these and other categorical predictors, nAB levels increased by 1.5-fold for every 10 beats per minute deviation from a subject’s norm (increase in log10 ID50 per additional 10 beats per minute: 0.19, CI 0.071 to 0.31; rp = 0.27, CI 0.10 to 0.41).
Neither vaccination-induced change in nightly 99th percentile HRV nor in average nightly RR was significantly predictive of subsequent nAB via either main or interaction effects, for either vaccine dose.
Discussion
We show here that individuals who reported experiencing chills, tiredness, feeling unwell, or headache following the second dose of a SARS-CoV-2 vaccine subsequently had 140-160% of the neutralizing antibody level of people who did not report these symptoms, at both 1 and 6 months later. We also show that each additional symptom experienced following dose two predicted a 1.1-fold increase in subsequent nAB. This means that, on average, individuals reporting 7 distinct symptoms subsequently had nearly 200% the nAB level of individuals reporting 0 symptoms. Using objective biometric data, we present convergent findings showing that greater vaccination-induced change in body temperature and heart rate, specifically at dose two, predicts greater nAB, especially at 6-month follow-up. Effect sizes were again large, with a difference between individuals of 1 degree Celsius in vaccination-induced change in skin temperature predicting 300% of the nAB level six months later.
Several prior publications have examined the association between systemic symptoms following receipt of a SARS-CoV-2 mRNA vaccine and subsequent nAB level, with inconsistent results. In one report, at dose two of mRNA-1273 (n = 171), none of three local or eight systemic symptoms, nor the presence of any local or any systemic symptom, predicted nAB four weeks later.8 By contrast, there are two reports of a positive association between systemic symptoms and nAB. In one study in 163 subjects, it was reported that for both dose one and two, the presence of at least one systemic symptom predicted higher nAB at 12-19 days after dose two.10 In the other report, the presence of at least one systemic symptom at dose three predicted higher nAB at a median of 54 days later.9
One key methodological distinction that may drive these discrepant results is that the first study did not include any BNT162b2 recipients,8 whereas our study and the other two studies used samples consisting of at least 50% BNT162b2 recipients. We have previously shown that increasing age predicted lower subsequent nAB for BNT162b2 recipients but not mRNA-1273 recipients.16 This may have resulted from the fact that the initial formulation of mRNA-1273 contained over three times the mRNA dose of the initial BNT162b2 formulation,21 and resulted in a humoral immune response that was stronger to such an extent that there was lower variability between subjects, suggesting a ceiling effect.16 It may be that the lower variability between mRNA-1273 recipients in antibody outcomes, and perhaps side effect outcomes, compromises statistical power to detect relationships with side effects, particularly when outcomes are measured shortly after completion of the initial series. Indeed, in the present study, we observed a relationship between vaccination-induced change in skin temperature and nAB that was a stronger predictor of the 6-month than the 1-month outcome. There are a few other reports examining the association between reactogenicity to SARS-CoV-2 vaccination and subsequent nAB, but interpretability is limited due to low samples sizes (4 to 8 per condition),22 the analysis only of nAB trajectories over time,23 which are confounded by absolute levels given that higher decay rates correlate with higher initial levels,11 and the use of a mixed sample of mRNA and adenoviral vector vaccine recipients.24
There are several key strengths of our study compared to the previous large studies.8–10 Firstly, neither of the two studies that included BNT162b2 recipients examined individual symptoms as predictors of nAB. Here, we show that it is specifically chills, tiredness (or fatigue), generally feeling unwell, and headache that drive the predictive relationship between systemic symptoms and nAB. Secondly, these studies all measured nAB within 2 months of receipt of the second dose of an mRNA vaccine, whereas we report on measurements as late as 6 months. This long follow-up is important and relevant given that individuals are unlikely to receive booster vaccination before 6 months or even a year following the initial series. We show that most predictive relationships apply equally for nAB at both 1 month and 6 months following the initial series, but certain biomarker measurements (change in skin temperature and change in heart rate) were more informative for predicting long-term than short-term nAB. Third, in addition to self-report measures that may be vulnerable to psychological or cognitive influences, for example variability in the tendency to notice, emphasize, remember, or report adverse effects, we use objective biometric measurements of physiological perturbation that is not vulnerable to these influences. Using this data, we present findings that neatly concord with our self-report data. Only one prior study has used non-self-report objective biometric data to predict subsequent humoral immune response after SARS-CoV-2 vaccination.25 That study found positive associations between vaccination-induced change in skin temperature and heart rate and subsequent anti-spike immunoglobulin at roughly 1 month later, in a mixed mRNA and adenoviral vector vaccine sample. Here, we extend those findings, demonstrating similar relationships with nAB at 6 months later, in an mRNA vaccine sample. Finally, our study is the first study of reactogenicity and nAB level following mRNA vaccination conducted in the US, enhancing generalizability to that context, and one of the first such studies to be conducted in a general population rather than convenience sample of healthcare workers.
There are several limitations to our study. Firstly, our results are from individuals who received only the initial Covid-19 vaccine series. Estimates based on survey data indicate that as of March 2023, roughly 24% of US adults are unvaccinated, 25% received the initial series but no booster doses, 25% received one booster dose, and 23% received the bivalent booster dose.26 The CDC estimated bivalent booster uptake slightly lower, at 17% as of May 2023.2 It is not clear whether the relationships observed here would apply to individuals undergoing initial vaccination using the updated vaccine formulations.
Secondly, our results are from individuals who did not have any serological evidence of SARS-CoV-2 infection. It is unknown whether in individuals with a prior history of SARS-CoV-2 infection, the same predictive ability of symptoms and vaccination-induced change in biometrics would be observed. However, among individuals receiving a two-dose mRNA vaccination, those with a prior history of SARS-CoV-2 infection have been reported to have both greater subsequent anti-spike IgG concentrations27 and greater reactogenicity,28 suggesting that a predictive association between neutralizing antibodies and reactogenicity may be likely. A third limitation is that our pseudovirus assay used the spike protein from the original Wuhan/D614G strain of SARS-CoV-2, which may limit generalizability of the findings, especially given that omicron variants have comprised virtually all circulating SARS-CoV-2 virus in the US since February 2022 through September 2023.29 Finally, we only address humoral immunity in this study, and although evidence suggests that neutralizing antibodies mediate roughly two-thirds of vaccine efficacy,12 cellular immunity is believed to play an important role in protection from severe disease.30,31
In sum, we show here in a large community sample that systemic symptoms and increases in skin temperature and heart rate following SARS-CoV-2 mRNA vaccination predict higher subsequent nAB level. We show that these relationships are stronger when predicting long-term rather than short-term nAB outcome. Our findings suggest that a reframing of systemic side effects as desirable may help to address the low rate of ongoing vaccine uptake, given that this appears to be at least partly the result of worry about side effects.4
Data Availability
All data produced in the present study are available upon reasonable request to the authors
Author Contributions
JER, SSD, AEM, FMH, AAP, and ESE conceived of and designed the broader study; JER, AEM, AAP, and ESE collected the data; EGD, JER, and AAP accessed and verified the data; EGD designed and conducted the statistical analyses; EGD wrote the first draft of the manuscript; AAP and ESE provided critical feedback on the manuscript. All authors approved the final version for publication.
Competing Interests
AEM has received remuneration from Oura Health for consulting. AAP is an advisor to NeuroGeneces and L-New Co. ESE is on the scientific advisory boards of Meru Health and Oura Health.
eMethods
Biometric data processing
For each nightly sleep period, a time series of measurements was available for three physiological domains (heart rate, HR; heart rate variability, HRV; and skin temperature, ST), while respiratory rate (RR) was provided as an average across the sleep period. For each of these three domains, three candidate methods of summarizing each set of nightly measurements into one nightly summary value were considered, i.e., taking the 1st percentile, the mean, and the 99th percentile. For a given sleep period and domain, these nightly summary values were only calculated where the domain was successfully recorded for the whole sleep period.
For each of the 10 nightly summary variables (3 for HR, 3 for HRV, 3 for ST, and 1 for RR), a subject norm was calculated for each subject by averaging over the available measurements across a 10-night period from nights 7 through 16 after dose one. The start of this window was chosen because vaccination side effects have generally resolved by this time,28,32 and no subjects received a second dose by night 16. Nightly summary variables were then centered on each subject norm.
Nightly statistics were then collapsed across nights to generate per-subject summary statistics of vaccination-induced change. For nights 0 and 1 following each dose, the nightly mean HR and ST was positive on average, while for HRV it was negative on average, indicating that the direction of any effect of vaccination on HR, RR, and ST was an increase, while for HRV it was a decrease. Because subjects received vaccination at varying times of day, between-subject variability in the time to maximal vaccination-induced change was anticipated. Across nights 0 through 3 following each dose, the majority of subjects experienced their peak deviation from baseline (maximum for mean nightly HR, RR, and ST; minimum for mean nightly HRV) on either night 0 or 1, so the larger deviation over these two nights was taken as each subject’s vaccination-induced change in each of the ten variables, for each vaccine dose.
The nightly summary method (i.e., 1st percentile, mean, or 99th percentile) most sensitive to vaccination-induced change in each of HR, HRV, and ST was then identified by using one sample t-tests to compare the dose two summary variables derived from the nightly 1st percentile, mean, and 99th percentile values. Dose two was used for the purpose of variable selection because side effects are more common at dose two and therefore variable comparison was less likely to be influenced by random variability.32 The most sensitive variable was defined as the one with the lowest p-value. Results are provided in eTable 2. The nightly 99th percentile was identified as most sensitive to vaccination-induced change for HRV and ST, while the nightly 1st percentile was most sensitive to vaccination-induced change for HR.
Data analysis
Linear mixed-effects models were fit using lme4 using restricted maximum likelihood estimation. Deviation (sum-to-zero) coding was used for categorical predictors. Continuous predictors were mean centered prior to model fitting. Random effects included a random intercept per subject. All models were checked for multicollinearity by examination of variance inflation factors (VIFs) calculated using car. The VIF threshold of 5 was exceeded in the initial model for each of the following predictors of interest: other allergic reactions (dose one), pain or swelling in the non-injected arm (dose one), tender or swollen lymph nodes (dose one), fever (dose one), other allergic reactions (dose two). In all cases, multicollinearity was the result of low variability in these predictors, resulting in the higher-order two- or three-way interaction terms being correlated with vaccine, time point, or both. This was resolved by removing all interaction terms involving each predictor of interest. All final models met assumptions of residual normality, linearity, and equality of variance, as assessed via diagnostic plots. No single observation had an undue influence on model fit, given that all had a Cook’s distance below 1. Significance of model terms was evaluated via F statistics calculated using Type II sums of squares, meaning that the F statistic corresponding to a given term compared the predictions of the full model including that term but without any higher-order interaction terms to the same model without the given term.33 Degrees of freedom were approximated via the Kenward-Roger method.
Along with test statistics, results were described using marginal means (i.e., least-squares means) or using unstandardized or standardized effect size estimates. Unstandardized effect sizes included marginal slopes or the difference between a pair of marginal means or slopes, both calculated using emmeans. The effectsize package was used to convert t or Type II sums of squares F statistics and associated degrees of freedom to standardized effect sizes, specifically the partial correlation (rp) or partial η2. Effect sizes were also presented as fold difference where appropriate. All estimates assumed (i.e., were conditioned on) mean levels of continuous covariates (age and body mass index) and were averaged across the estimates for each level of other categorical predictors (vaccine, sex, time point), weighting each level of these variables proportional to its representation in the sample. Post-hoc testing was performed on these estimates, comparing two using a two-sample t-test, or comparing marginal slopes (simple slopes) to zero using a one-sample t-test. Marginal means were provided in the untransformed (ID50) scale, and fold differences were calculated in this scale. Marginal trends and pairwise contrasts refer to associations with the outcome in the log10 scale. All visualizations were produced using ggplot2 and patchwork, with ggeffects used for calculation of partial residuals.
eTables
Numbers and proportions of subjects reporting each of 13 symptoms are presented, for dose one and dose two. A total of 363 subjects contributed to symptom analyses, with 354 providing data for dose one and 347 providing data for dose two.
Biometric data at the time of either vaccine dose was available for a total of 168 subjects, with 162 providing data at dose one and 145 providing data at dose two. Respiratory rate (RR; breaths per minute) was available only as a nightly average. For heart rate (HR; beats per minute), heart rate variability (HRV; root mean square of successive differences, in milliseconds), and skin temperature (ST; degrees Celsius), one-sample t-tests were used to identify the approach to summarizing the nightly time series that was most sensitive to vaccination-induced change. For HRV and ST, this was the 99th percentile, and for HR, it was the 1st percentile. These summary approaches were then used to test hypotheses regarding the ability of vaccination-induced change in each physiological domain to predict subsequent neutralizing antibody level.
18 variables (13 symptoms or symptom categories, 1 symptom count, and 4 biometric variables) measured following receipt of vaccine doses one and two were examined as potential predictors of subsequent neutralizing antibodies measured at 1 month and 6 months following completion of the initial vaccination series, via 36 total mixed-effects linear models. Where possible, these models included interaction terms involving the variable under investigation with vaccine, outcome time point, and the three-way interaction between these variables. Statistics were extracted from these models and then all 126 p-values were simultaneously corrected for multiple comparisons using the Benjamini-Hochberg method.
Values represent Spearman’s rho. Δ: vaccination-induced change, D1: dose one, D2: dose two, HRV: heart rate variability, HR: heart rate, RR: respiratory rate. ***: p < 0.001, **: p < 0.01, * p < 0.05 (uncorrected).
Acknowledgements
The project was supported by funding from the National Institutes of Health (R24AG048024, 5U24AG066528, and U54CA260581).
Footnotes
↵† Co-Senior Authors
No change to any results. Table and figure legends updated. Discussion updated. Supplementary methods text corrected and condensed. Added eFigure 1. Abstract modified, including changing sample size from symptom analysis n to overall n. Methods moved to before Results. Pseudovirus assay methods clarified. Line spacing changed.
References
