Eligibility criteria

Eligible study designs were systematic reviews (SR) with or without meta-analyses (MA), excluding narrative, scoping and non-systematic reviews. In terms of the PICO criteria, the study population included adults aged 18 years and older; reviews including children were eligible if outcomes were stratified by age. Exposure was defined as acute SARS-CoV-2 infection diagnosed 4 weeks or more prior to Long COVID ascertainment, in conformity with the U.S. federal working definition of Long COVID [15]. Comparators (i.e., controls) were defined according to the individual SR reviewed. We considered Long COVID as any or at-least one patient-reported, clinically presented, or administrative (e.g., ICD-10 codes) outcome associated with Long COVID. Studies that exclusively reviewed the prevalence of conditions with preexisting medical definitions (e.g., diabetes) arising post-COVID-19 were excluded, consistent with the WHO consensus definition of post-COVID-19 condition as a diagnosis of exclusion [16]. The relevant context was Long COVID diagnosed and treated in high income countries, thus only peer-reviewed articles in English were considered.

Search strategy and study selection

The following three databases were searched from January 1, 2019, through June 9, 2023: LitCOVID, PubMed, Embase. Full database search strategy can be found in S1 Table. In addition to database searches, secondary searches were performed within Web of Science to identify potential reviews that met the eligibility criteria. We manually screened the reference lists of systematic reviews and searched Google Scholar.

Database search results were imported into a reference manager (EndNote X20; Clarivate Analytics) for deduplication, then uploaded into Covidence (Covidence, Melbourne, Victoria, Australia) screening software to remove additional duplicates. An initial pilot was performed to screen title/abstract and full text articles, and any revisions to the search strategy were recorded. Dual screening of both title/abstract and full text was conducted by two reviewers (MH and OA) independently. Any disagreements were resolved by a third reviewer (DP).

Data extraction and quality assessment

The data extraction template was piloted on a subset of SRs by two independent reviewers (MH and OA). Data extraction was conducted by a single reviewer (MH) into an excel spreadsheet; the extracted data from two SRs were then randomly selected and reviewed by one adjudicator (DP). The SRs’ corresponding authors were contacted no more than two times over the course of two weeks to obtain missing data. Data collected included article identifying information, study type, design characteristics of primary studies, setting, participant characteristics, relevant outcomes and the ROB and limitations. Where possible, characteristics specific to studies/populations informing the subset of relevant outcomes were extracted.

Risk of bias assessment

The AMSTAR 2 critical appraisal tool for SRs that include non-randomized studies was used to determine risk of bias [17]. AMSTAR 2 evaluates SRs through sixteen domains, seven of which critically impact the validity of the review, including protocol registration before commencement of the review (item 2), adequacy of the literature search (item 4), justification for excluding individual studies (item 7), ROB from individual studies included in the review (item 9), appropriateness of meta-analytical methods (item 11), consideration of ROB when interpreting the results of the review (item 13) and assessment of presence and likely impact of publication bias (item 15). Results from studies that have one or more critical weakness will be considered to have low or critically low overall confidence. Studies were assessed by one reviewer (MH) with blinded validation by a second reviewer on one randomly selected study (DP).

Data synthesis

A meta-analysis was not undertaken as the included SRs were not sufficiently homogenous in population characteristics and design. For meta-analyses that reported relevant outcomes of Long COVID, we reported prevalence as a percentage with 95% CI and risk factors as and odds ratios (OR) or hazard ratios (HR) with 95% CI. The Higgins I² estimate of heterogeneity was reported for all outcomes where available. If a meta-analysis was not performed, outcomes were reported as median and interquartile range (IQR) if there were at least 5 studies.

We reported pooled estimates or manually calculated the median and IQR of prevalence for each category of 1) hospitalization status (hospitalized, non-hospitalized, mixed); 2) duration of follow-up (<3 months or ≥ 3 months); 3) use of a COVID-negative control group in the primary studies; 4) vaccination status (completed primary series vs. did not complete primary series); 5) COVID variant (wild-type, alpha/beta/delta, omicron). Risk factor outcomes were reported in accordance with the respective SRs. Due to the large number of risk factors investigated, we only reported pooled outcomes or which included at least 5 studies to calculate median/IQR.

We also conducted a narrative synthesis of the ROB identified by the review of the SR. Summary of the ROB comprised the ROB tool used, the number of included primary studies with high or critical ROB, and most frequent ROBs identified.

Study and participant characteristics

All SRs were published between May 26, 2021 to June 8, 2023, with the most recent primary literature search inclusive of February 10, 2023 [18]. The prevalence SRs included 6 SR/MA, 2 SR, and 1 umbrella review with evidence synthesis of a selection of the primary literature. Many SRs did not provide a complete list of citations of the included studies, so we were unable to disambiguate a unique set of primary studies even after contacting the corresponding authors.

Prevalence SRs included 5 to 196 studies and 1643 to 1,289,044 participants. Among the five risk factor SRs, 6–41 studies of 7170 to 860,783 participants were included. Four conducted meta-analyses for relevant outcomes. The most common study designs of primary studies were cohort studies, followed by cross-sectional, and case-control studies. More information on publication and study design can be found in Table 1 and S3 Table.

Download:

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

Table 1. Study population characteristics, extracted outcomes and assessed limitations and Risks of Bias (ROBs).

https://doi.org/10.1371/journal.pone.0302408.t001

Among the prevalence SRs, 3 included adults only [18–20], 5 presented age-stratified outcomes [6, 7, 21–23], and 1 reported median or mean ages ≥47 for all relevant primary studies [24]. Diagnosis of SARS-CoV-2 infection was laboratory-based in most SRs that reported exposure ascertainment, but self-reported COVID-19 diagnosis was considered eligible by at least one SR [6]. Timing of follow-up ranged from 28 to 730 days from time zero, which varied as the point of COVID-19 diagnosis, symptom onset, hospital admission or hospital discharge [20, 24]. Of the risk factor SRs, one was restricted to adults [25]; the rest reported mean/median ages of 40–69. Most primary studies in 4 out of 5 risk factor studies confirmed COVID-19 exposure by laboratory methods. For more details on participant characteristics, see Table 1.

Prevalence outcomes

Long COVID prevalence or cumulative incidence in the SRs ranged from 21% (IQR: 8.9%-35%) to 74.5% (95% CI: 55.6%-78.0%). Hospitalization status was reported by 6 SRs, 3 at the study level (i.e., whether a study included hospitalized patients) and 3 at the individual level for at least one relevant outcome. Of the latter, the percentage of hospitalized patients ranged from 17.4% to 98.2% (see S4 Table). Two SRs conducted random-effects meta-analysis by hospitalization status at the study level [6, 7]. Four provided outcomes that could be stratified by duration of follow-up at 3 months, two of which calculated the pooled prevalence of Long COVID by duration of follow-up [7, 20]. For the rest, we calculated the median prevalence and inter-quartile ranges stratified by 3 months/12 weeks.

The inclusion of COVID-19 negative comparator groups was noted in three SRs; only one estimated the prevalence of Long COVID stratified by the use of control groups [21]. Three SRs reported the SARS-CoV-2 variant of concern assumed responsible for most infections [19, 22, 23], one of which reported Long COVID prevalence by variant of concern without conducting a meta-analysis [19]. Two SRs commented on vaccination status reported in primary studies [7, 19], neither of which reported prevalence estimates by vaccination status. See S5 Table for more information on comparator groups and other subgroups.

Outcomes with pooled estimates are shown in Fig 2. The I² measure of heterogeneity was over 90% in all reporting pooled prevalence estimates but an I² was not reported for 50% of pooled estimates. Outcomes with median and IQR are shown in Fig 3.

Download:

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

Fig 2. Meta-analytic estimates of prevalence extracted from Long COVID systematic reviews.

N, number of studies; CI, confidence interval; NR, not reported.

https://doi.org/10.1371/journal.pone.0302408.g002

Download:

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

Fig 3. Median and interquartile ranges of prevalence estimates extracted from Long COVID systematic reviews.

N, number of studies; IQR, interquartile range.

https://doi.org/10.1371/journal.pone.0302408.g003

Risk factor outcomes

Up to fourteen different risk/protective factors were extracted from the five risk factor SRs reviewed. Three examined the associations between vaccination and Long COVID risk and found that two or more pre-infection doses of COVID-19 vaccine significantly decreased the risk of Long COVID compared to no or 1 vaccine, but a single dose did not significantly mitigate Long COVID risk [25–27]. Three SRs reported on risks associated with age, sex, acute COVID-19 severity and other sociodemographic and clinical risk factors [25, 28, 29]. All three found female sex to be a significant risk factor for Long COVID. One SR found that categorical age of 40 or greater posed increased risk of Long COVID (OR: 1.21, 95% CI: 1.11–1.33) [25], while another did not find evidence of age-associated risks [28]. The latter also found elevated risks of non-recovery from severe or critical acute COVID-19 with moderate certainty, although associations with hospitalization status were non-significant [28]. Both SRs investigated associations between Long COVID and comorbidities, finding significant risk associations. Full statistical outcomes are summarized in S4 Table.

Summary of ROB assessments of primary studies

A majority of SRs (6 out of 9 prevalence SRs, 2 out of 5 risk factor SRs used the Newcastle-Ottawa Scale (NOS) [30]. Other tools used in the prevalence SRs included the NIH tool (2 studies) [31], Hoy et al. (1 study) [32] and the AHRQ tool (1 study) for cross-sectional studies [33] in a study that also used the NOS for cohort studies. Risk factor SRs also employed ROBINS-I (1 study) [34], JBI checklist for cohort studies (1 study) [35] and QUIPS (1 study) [36].

Six prevalence SRs gave an overall ROB score. Regardless of the tool used, the majority of primary studies were scored as having low or moderate ROB, with only 0–9.9% of studies rated has having a high ROB (e.g., less than five out of nine on the NOS scale). The risk factor SRs reported higher ROB overall, with 0–56% of primary studies rated as having a high or critical ROB.

Out of fourteen SRs, two prevalence SRs [7, 22] and one risk factor SR [25] reported only the overall ROB of the primary studies without a score breakdown. From the 11 SRs that did report a score breakdown, the quality of outcome ascertainment and selection bias were the most frequently top-ranked ROBs; adjustment for confounders, attrition, representative sampling and outcome ascertainment were areas of deficiency in at least two risk factors SRs.

Summary of limitations as described in the SRs

High heterogeneity was identified as a limitation in 8 prevalence SRs, with the exception of the SR on asymptomatic cases [23]. Lack of control/comparator group or representative sampling was noted in 5 prevalence SRs. Lack of standardization in case definition and symptom measurement was noted in 3 prevalence SRs and 4 risk factor SRs. Variable follow-up time, lack of data on age, race/ethnicity, disability and overrepresentation of people hospitalized with COVID-19 were also identified as limitations.

ROB assessment of SRs with AMSTAR 2

All prevalence SRs had weaknesses in at least two critical domains (see Table 2) [17]. Among these, deficiencies on items 9 and 11 are particularly concerning for our aim of identifying biases and limitations in the Long COVID evidence base.

Download:

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

Table 2. AMSTAR 2 assessments of all SRs.

https://doi.org/10.1371/journal.pone.0302408.t002

An adequate score on item 11 required the SR to explicitly justify the decision to perform a meta-analysis based on the compatibility of included studies. None of the prevalence SRs that conducted a meta-analysis included such a statement, while all of the SRs which did not pursue a meta-analysis cited the high degree of heterogeneity in the primary literature as deterrent. We considered this item satisfied if meta-analysis was primarily undertaken by pre-determined subgroups that may create comparable cohorts, such as by hospitalization status or follow-up duration. Nevertheless, the subgroup outcomes in prevalence SRs all reported I² statistics greater than 90%.

ROB assessments conducted by the SRs were often deficient, hence the large number of partial or complete deficiencies on item 9. For a “yes,” the SR had to evaluate primary studies on at least four sources of bias: confounding, selection, measurement of exposures and outcomes, and selective reporting of analyses or outcomes. The NOS rates ROB across the three domains of selection, comparability and outcome assessment without any criteria for selective reporting [30], so SRs that used this tool without modification received a “partial yes” at best on this criterion. On assessing confounding, NOS requires pre-specification of the two most important factors to control to satisfy the criteria of “comparability,” which 6 out of the 8 SRs using the tool failed to specify [7, 20, 22, 23, 25, 27]. Selection bias as well as confounding were equivocally evaluated by most SRs due to ambiguity around the definition of control groups (see Table 1 and S5 Table). Most of the seven prevalence SRs that included the general adult population did not define a control group; one defined the control group as COVID-positive without post-discharge symptoms [20]; one defined it as people without COVID-19 but did not report the number of studies that used a control group or any associated outcomes [22]. This exposes interpretative challenges in how SRs applied the NOS criteria on “selection of non-exposed cohort” and “comparability.” [30].

Risk factor SRs had lower ROB overall, with two SRs with 0–1 deficiency. All had clearly defined non-exposed comparator groups by the PICO criteria (lacking in the risk factor rather than COVID-19 exposure). Nevertheless, a majority were deficient on items 7 and 10, and 40% were deficient on items 9 and 11 (see Table 2).