Methods
We used data from the ComPaRe long covid cohort to emulate a target trial evaluating the effect of the first injection of the covid-19 vaccine in patients who already have long covid, on the severity of symptoms and on the impact of the disease on their lives.14
Data sources
The ComPaRe long covid cohort is an ongoing nationwide e-cohort (ie, a cohort where recruitment and follow-up are performed online) of patients with long covid in France, nested in the ComPaRe research programme (www.compare.aphp.fr), an umbrella e-cohort of patients with chronic conditions.15 Participation opened in November 2020 and is ongoing. The cohort includes adult patients who have reported a SARS-CoV-2 infection (whether or not confirmed by a positive polymerase chain reaction (PCR) test result or serological assay, or both) and have symptoms persisting for more than three weeks after the original infection. Recruitment took place from calls on social and general media, by partner patient associations, on the official French contact tracing app TousAntiCOVID, and by a snowball sampling method where participants were encouraged to invite people who had covid-19 and persisting symptoms to enrol.16
Participants in the ComPaRe long covid cohort are contacted for follow-up every 60 days by email with links to an online questionnaire. At each observation (eg, T0=cohort enrolment, T1, T2), patients are asked if they still have symptoms related to covid-19. Those who report persisting symptoms complete the long covid symptom tool and impact tool, a pair of validated patient reported instruments assessing, respectively, 53 long covid symptoms and six dimensions of patients' lives that can be affected by the disease.17 Those reporting no symptoms are asked to specify the date when they first noticed the absence of symptoms.
Since 11 May 2021, every 45 days patients have been self-reporting their covid-19 vaccination status in a different online questionnaire. Those who have been vaccinated report the vaccine received, the date or dates of vaccination, and any adverse effects. In September 2021, all patients who had not previously reported being vaccinated were contacted by email to confirm their vaccination status.
Eligibility criteria
Our analyses used data from patients enrolled in the ComPaRe long covid cohort before 1 May 2021. We included adult patients (aged ≥18 years) with a confirmed or suspected SARS-CoV-2 infection, diagnosed by a health professional, whose symptoms persisted for >3 weeks after the original infection, and who reported at least one symptom attributable to long covid at baseline. We excluded patients whose date of first symptoms was <3 months before baseline because at the time of the study the recommendation in France was to delay vaccination for three months for patients who had recently been infected with the SARS-CoV-2 virus.18 We also excluded patients reporting a history of severe allergy in ComPaRe because of the likelihood of a history of anaphylaxis, a contraindication to vaccination at the time of the study.
Outcomes
The primary outcome was the score on the long covid symptom tool, a validated patient reported outcome developed from patients' lived experience of long covid, assessing the number of symptoms of long covid (online supplemental material 1). The symptom tool score ranges from 0 (ie, remission of disease) to 53, and has been shown to correlate with patients' quality of life and functional limitations. Reproducibility of the score was excellent (intraclass correlation coefficient 0.83, 95% confidence interval 0.80 to 0.86).17 Because the long covid symptom tool is a symptom count score, any change in the score relates to an objective change in patients' perception of their symptoms (ie, at least one symptom has disappeared or appeared between the two measurements). We also investigated the rate of remission of the disease (ie, complete disappearance of symptoms).online supplemental material 1
We used the long covid impact tool, a second validated patient reported outcome, to assess the effect of the disease on patients’ social, professional, and family lives. The impact tool score ranges from 0 (no impact) to 60 (maximal impact) and has been shown to be highly correlated with patients' quality of life and patients' perceived severity of their disease. The impact tool score showed excellent reproducibility (intraclass correlation coefficient 0.84, 95% confidence interval 0.80 to 0.87).17 We also analysed the score on the impact tool classified according to its Patient Acceptable Symptom State, which represents the level of a continuous outcome measure below which patients consider themselves well.19 In a previous study, we estimated that the Patient Acceptable Symptom State for the impact tool was 30/60.17 All outcomes were assessed at 120 days after baseline.
Adverse events after vaccination were analysed by one investigator (V-TT) from participants’ open text answers to the related questions in the online questionnaire. Adverse events were categorised as serious based on the definition of the US Food and Drug Administration: adverse events resulting in death, life threatening, requiring admission to hospital or prolonging an existing stay in hospital, resulting in persistent or substantial disability, or requiring a specific intervention to prevent permanent impairment or damage.20
Study groups and follow-up
To define a vaccinated group and a matched unvaccinated control group in a population where most patients were vaccinated against covid-19, we used the cohort data to emulate a sequence of three trials which were then pooled.21 In the first trial, we identified all patients who met the eligibility criteria when they were enrolled in the ComPaRe long covid cohort (ie, their first observation point, T0). Patients who received their first covid-19 vaccine injection with the ChAdOx1 (AstraZeneca), BNT162b2 mRNA (Pfizer-BioNTech), Ad26.COV2.S (Johnson & Johnson), or mRNA-1273 (Moderna) vaccines between baseline and 60 days (ie, their second observation point, T1), were classified as the vaccinated group and matched in a 1:1 ratio to patients who did not receive the vaccine in the same period (T0 to T1), classified as the control group. Patients were followed up for 120 days (ie, their third observation point, T2, and endpoint of the first trial). Unvaccinated controls who were vaccinated before T2 were censored at the date of vaccination.
We repeated this procedure by emulating two more trials, by considering baseline at 60 days (ie, T1) for the second trial, and T2 for the third; we applied a similar follow-up strategy (ie, follow-up until T3 and T4, respectively). At the baseline of each of the three trials, patients' eligibility criteria were reassessed and those who no longer met the eligibility criteria (eg, because they no longer reported symptoms) were excluded from that trial. Control patients who had since received a covid-19 vaccine were eligible for inclusion in the vaccinated group even though they were in the control group in the first (or second) trial. However, a patient could only be selected once for the control group and once for the vaccinated group. Online supplemental material 2 defines the sequence of trials in more detail.
Statistical analysis
Our causal contrast of interest was the per protocol effect. Within each of the three trials, each patient who was vaccinated was matched to an unvaccinated control according to their probability of getting vaccinated against covid-19 given their baseline covariates (ie, the propensity score). The propensity score was calculated with a non-parsimonious multivariable logistic regression model including variables planned and prespecified before the outcome analyses: sex; age; educational level (≥2 years post-secondary education—higher education v lower); number of comorbidities (self-reported with the International Classification of Primary Care, version 2)22; SARS-CoV-2 infection confirmed by laboratory analysis (yes for patients reporting a positive test result for SARS-CoV-2 infection by PCR test or serological assay, or both, and otherwise no); interval from the start of covid-19 symptoms; history of admission to hospital for covid-19 during the acute phase; score on long covid symptom tool at baseline; and score on long covid impact tool at baseline. Standardised differences were examined to assess balance, with a threshold of 10% indicating a clinically meaningful imbalance.23 Propensity score matching used a calliper width of 0.2 of the pooled standard deviation of the logit of the propensity score.24
For our analysis, we pooled the three trials and estimated the effect of treatment with one model, including a trial covariate, rather than fitting a separate model for each trial. Also, because some people participated in more than one trial, we used a robust variance estimator to estimate conservative 95% confidence intervals.
To correct for the induced time varying selection generated by the artificial censoring of patients in the unvaccinated group at the date of their first vaccine injection, we used inverse probability of censoring weighting with weights proportional to the inverse of the probability of remaining uncensored until each time point, given the baseline covariates. Stabilised weights were obtained by multiplying the weights by the overall probability of being uncensored at each time point.25 To assess the quality of the correction, we compared the number of patients at risk, over time, in the two groups, and the balance of the baseline covariates between the two groups, 120 days after inclusion in the trials (online supplemental material 3).
In the survival analyses, to account for immortal time bias, baseline was considered as the vaccination date for patients in the vaccinated group and the vaccination date of their matched patient for those in the control group. Outcomes were studied in the total population and in a subgroup restricted to participants with a confirmed SARS-CoV-2 infection. Two post hoc subgroup analyses were conducted: onset of symptoms ≤12 months versus >12 months, and vaccine type.
We used the E value to evaluate how the results could be affected by unmeasured confounding. The E value measures the minimum strength of association an unmeasured confounder would need to have with both the intervention and the outcome to fully explain away the treatment effect.26
We performed several sensitivity analyses. Firstly, we restricted the study population to patients who had been included in only one of the three trials (ie, excluding patients included twice in the study, once in the unvaccinated group and then in the vaccinated group) to examine the potential correlation induced by including the same patient in several trials. Secondly, we analysed how our design, based on a sequence of emulated trials, could affect our results by estimating separate treatment effect estimates for each trial and then conducting a meta-analysis with a fixed effect approach. Finally, we conducted a secondary analysis that used standardised mortality ratio weighting as an alternative to propensity score matching.27
Missing baseline and outcome variables were handled by multiple imputation with chained equations that used the other variables available. All statistical analyses were performed with the R statistical package version 4.0.3 (The R Foundation for Statistical Computing, www.R-project.org/).
Patient and public Involvement
The study is a reanalysis of existing data. Patients were involved in design of the questionnaires and measurement tools used in the cohort. In ComPaRe, lay summaries of research results are systematically shared with participants and partner patient associations on the project's homepage.