Study outline
We conducted two nationwide matched cohort studies (Study 1, Study 2) using administrative data (Fig 1A). In Study 1, we defined three COVID-19 outpatient cohorts (Fig 1A): Cohort A – patients suffering conditions that required antidepressants/anxiolytics and were exposed to fluvoxamine around the index COVID-19 diagnosis; Cohort B – patients suffering such conditions but were not exposed to fluvoxamine; and Cohort C- patients free of psychiatric disorders and treatments around the index COVID-19 diagnosis (Table 1 for details). The three cohorts were mutually exactly matched on a range of pre-COVID-19 characteristics: (i) comparison between Cohort A (treatment) and Cohort B (control) was of primary interest; (ii) comparisons of Cohorts A or B with Cohort C were considered supportive, and informed about the “joint” effect of two co-exposures (antidepressant/anxiolytic + underlying conditions). We expected a limited number of people prescribed fluvoxmine (unlike most of the other antidpressants/anxiolytics, it is only partly reimbursed), hence to improve precision of the Cohort A vs. Cohort B estimates, we conducted network meta-analysis of results in matched sets A vs. B, A vs. C and B vs. C. Although derived from the same pool of original patients, matched contrasts were based on different pseudopopulations (by selection and weighting).
Study 2 also aimed to estimate effect of specifically fluvoxamine exposure (treatment), but through a contrast to a specific “other” antidepressant/anxiolytic – paroxetine (control). Fluvoxamine and Paroxetine Cohort subjects (definitions in Table 1) were exactly matched on the same covariates as in Study 1 (Table 2).
Prescriptions for fluvoxamine and paroxetine are repeatable – one issued prescription can cover a maximum of 12 months of treatment. We reasoned that prescriptions that would pertain to a period shorter than 3 months were not likely, and also that prescriptions for a period much longer than 3 months were not very likely since the treated conditions require medical follow-up and reconsideration of treatment. Therefore (Table 1), we considered that subjects were prescribed and (presumably) exposed to treatment if at least 1 prescription was issued over a period of time between 3 months prior to- and 7 days after the index COVID-19 diagnosis; and were not prescribed (exposed) if no prescriptions were issued within 6 months prior to- and up to 21 days after the index COVID-19 diagnosis.
We used anonymzed data routinely managed by the Croatian Institute for Public Health (CIPH), who prepared the initial raw dataset from databases on: (i) COVID-19 laboratory test results (polymerase chain reaction [PCR]-based or rapid antigen tests [RAT]) and COVID-19 patients diagnosed on clinical/epidemiological criteria; (ii) COVID-19 vaccinations; (iii) all hospitalizations; (iv) deceased individuals; (v) Central Health Information System (CEZIH) - primary healthcare database maintained by the Ministry of Health (Fig 1B). It included all subjects diagnosed with COVID-19 at points of mass outpatient testing (managed by CIPH) or by their general practitioners (i.e., we omitted subjects first diagnosed when seeking hospital assistance for any reason) between February 25, 2020 (first recorded case in Croatia) and October 15, 2021. Each individual was linked to her/his data on: date and mode of COVID-19 diagnosis; demographics and COVID-19 vaccination status at diagnosis; medical histories from January 1 2019 to October 31 2021, including comorbidities (International Classification of Diseases [ICD-10] codes), all issued prescriptions (Anatomical Therapeutic Chemical codes, ATC) and other medical care, hospital admissions and diagnoses and dates and causes of death (Fig 1B). We received a merged database and a) excluded subjects <16 years of age and those for whom data on sex, date of birth, COVID-19 testing date/result/date of diagnosis, or vaccination status/dates were missing or were erroneously entered; b) identified subjects with more than one COVID-19 episode: we considered that positive PCR/RAT tests or ICD-10 code U07.1/U07.2 entries or hospitalizations related to COVID-19 that were ≥30 days apart indicated two separate COVID-19 episodes. Only the first documented COVID-19 episode for each subject was included in the analysis; c) we set the cut-off date for COVID-19 diagnosis at August 15 2021, to allow for a follow-up period long-enough for outcomes to occur (until October 31) (Fig 1A). Finally, we identified patient subsets of interest (Table 1), their outcomes and their matching covariates (Table 2).
This study used anonymized administrative data standardly collected through routine procedures, hence ethical approval was waived by the Ethics Committee of the Zagreb University School of Medicine and Croatian Institute for Public Healthy.
Outcomes
We defined three outcomes informing about unfavorable develompents in COVID-19 outpatients. COVID-19-related hospitalization – hospitalization follows within 45 days since the index COVID-19 diagnosis, with U07.1/U07.2 as the leading diagnosis; or hospitalization follows within 30 days since the index COVID-19 diagnosis and U07.1/U07.2 is listed among diagnoses. 30-day all-cause hospitalization – hospitalization follows within 30 days since the index COVID-19 diagnosis. COVID-19-related death – we considered that death was “related to” COVID-19 if meeting any of the following (i) death occurred after the index COVID-19 diagnosis with U07.1/U07.2 as a cause of death, regardless of the time elapsed since the COVID-19 diagnosis (shortest possible period of observation was 77 days - patients diagnosed on August 15, 2021); (ii) death occurred within 14 days since the index COVID-19 diagnosis, regardless of the declared cause; (iii) death occurred in hospital, where hospitalization was COVID-19 related hospitalization (as defined above), regardless of the declared cause of death and time elapsed since the COVID-19 diagnosis.
Identification of exposures, other treatments, comorbidities and vaccination status
Exposure/non-exposure to fluvoxamine or paroxetine was identified based on timing of prescriptions with the respective ATC codes (N06AB08 and N06AB05, respectively) relative to the index COVID-19 diagnosis. Subjects were considered to suffer conditions in which antidepressants/anxiolytics (including fluvoxamine or paroxetine) might have been the main or one of the required treatments if they had at least one entry of any ICD-10 F codes or G30/G31.1 codes between January 1, 2019 up to 7 days after the index COVID-19 diagnosis. Regarding vaccination, patients were classified as “not vaccinated”, or as: a) vaccinated with a single-dose vaccine; b) received 1st dose of a two-dose vaccine; c) received 2nd (full) dose of a two-dose vaccine; and were further sub-classified based on time elapsed between the last vaccine administration and the index COVID-19 diagnosis (<14 days, 14-90 days and >90 days). Online Resource 1 – Supplemental Methods – provides details on indentification of all treatments and (co)morbidities used to identify patients subsets and in covariate matching.
Matching and data analysis
We implemented exact matching using package MatchIt [17] in R [18] with average treatment effect as the estimand (ATE). Outcomes were analyzed by fitting weighted log-binomial models, frequentist (with cluster robust sandwich estimator of the standard error) and Bayesian with three different priors: (i) skeptical prior – moderatly informative neutral prior consistent with an a priori hypothesis of no effect, centered at 0 for the Ln(RR) with standard deviation 0.355. It assigns 95% probability between RR=0.50 and RR=2.0; (ii) optimistic prior – moderately informative prior centered at -0.199 for the Ln(RR), with standard deviation 0.4, i.e., 18% relative risk reduction as seen in an up-dated Bayesian meta-analysis of randomized trials of fluvoxamine in this setting [19], but with 30% probability of an RR >1.0; (iii) pesimistic prior – weakly informative prior centered at 0.199 for Ln(RR) with a standard deviation of 0.77. Although it suggests harm, it leaves 40% probability of an RR <1.0. We used SAS 9.4 for Windows (SAS Inc, Cary, NC) and R package rstanarm [20] . In Study 1, we additonally performed frequentist (R package netmeta [21]) and Bayesian (R packages BUGSnet[22]and gemtc[23], with default priors) network meta-analysis using weighted counts and also the effect measures generated in Bayesian analyses with the skeptical prior.
Sensitivity to unmeasured confounding/bias
We assumed a hypothetical unmeasured confounding that “worked” to diminsh the (presumed) beneficial (risk-reducing) effect of fluvoxamine. Specifically, we assumed that among control subjects (Cohort B in Study 1, Cohort Paroxetine in Study 2), 40% were using some treatment (e.g., other antidepressant/anxiolytic, or any other) that was actually effective against COVID-19 with a marked effect of 30% reduction of the risk of disease deterioration (corresponds to the largest effect reported from RCTs of fluvoxamine [2]), and that only 1% of the fluvoxamine-exposed subjects were co-treated with such a treatment, and we corrected the observed estimates for this bias.