This methodological study of SRs evaluating physical therapy interventions for treating LBP is a secondary analysis of previous research.9 The search was performed in the Physiotherapy Evidence Database (PEDro; www.pedro.org.au) on 10th of January 2018. We chose PEDro because it is one of the most complete indexes in the field of physical therapy.20 Two independent authors performed the screening process. Our search strategy was: “systematic review” for method; “lumbar spine, sacroiliac joint or pelvis” for body part; “pain” for problem; 2015–2017 for year of publication. The time period was selected because it is two years after the publication of the PRISMA-A, therefore authors had a greater opportunity to adhere to these recommendations than in older SRs. Eligibility criteria were: full journal publications (not abstract only); systematic reviews of clinical trials that evaluated the effectiveness of one or more physical therapy interventions for LBP; published between 2015 and 2017; written in English, Spanish, or Portuguese; and with no restrictions for type of analysis (whether the review performed a meta-analysis or not).

Two independent authors evaluated the methodological quality, reporting quality of abstracts, and if abstract reporting was consistent with the full text reporting for the included reviews. They also extracted several review and journal characteristics from the reviews (between March and May 2018). Two authors discussed disagreements until consensus was reached, if consensus could not be achieved a third author provided arbitration.

The following descriptive data were extracted: (1) language of publication (English/ Spanish/ Portuguese); (2) year of publication to calculate the age (in years) of each review by subtracting the year of publication from 2018 (descriptive); (3) review open access on PubMed Central (yes/no); (4) number of collaborating centers based on the number of departments mentioned in the author affiliations (descriptive); (5) which risk of bias tool was used (descriptive); (6) number of studies and types of study designs included in the review (randomized controlled trials/ other types of studies); and, (7) the number of times the review had been cited (descriptive), as downloaded from the Web of Knowledge (4th September 2018), and normalized by dividing by the age of the review. The journal features were: (1) mention of PRISMA-A in the “Instructions to authors” (yes/no); (2) if endorsement of PRISMA for full texts checklist21 was compulsory (yes/no); (3) the word limit for the abstract (descriptive), extracted from the “Instructions to authors”; and, (4) open access journal (yes/no), obtained from the Directory of Open Access Journals or the Journal’s website. The variables extracted as ‘descriptive’ were coded in order of appearance. We also coded the descriptive data that were ‘not reported’ or ‘not applicable’.

The 12-item PRISMA-A checklist was used to evaluate the reporting quality of the abstract of each included SR. Each item was classified as “fully reported” (contained all information specified in the item) or “not reported” (contained some of the information specified in the item, or contained none of the information specified in the item). A score was generated for each review (called the “total number of PRISMA-A fully reported items”) by counting the number of items that were “fully reported”. This score ranged from 0 (low reporting quality) to 12 (high reporting quality) and has been used in previous studies.4,22

Consistency of reporting was evaluated by comparing the reporting in the abstract to the reporting in the full text of the review. To achieve this, the full text was also evaluated with the PRISMA-A so that PRISMA-A results were obtained for the abstract and also for the full text for each systematic review. We then calculated agreement between abstract and full text for each of the 12 items from the PRISMA-A.

The 16-item A MeaSurement Tool to Assess systematic Reviews (AMSTAR-2)23 checklist was used to evaluate the methodological quality of the included reviews. Each item was classified as: “no”, “yes”, “partial yes”, or “not applicable”.

The classification of the methodological quality for each review was based on the AMSTAR-2 scores,23 as follows: “high” (no or one non-critical weakness), “moderate” (more than one non-critical weakness), “low” (one critical flaw with or without non-critical weakness), or “critically low” (more than one critical flaw with or without non-critical weakness). Non-critical weaknesses were related to items 1, 3, 5, 6, 8, 10, 12, 14 and 16; and critical flaws to items 2, 4, 7, 9, 11, 13 and 15.

To investigate if review and journal characteristics were associated with quality of abstract reporting we defined one dependent variable and six independent variables that had been previously identified in the literature.24–26 The dependent variable was the “total number of PRISMA-A fully reported items” (range 0–12 points; described above, under ‘Abstracts reporting’). Three independent variables were categorical: AMSTAR-2 methodological quality score (as dummy variables for ‘high’, ‘moderate’, ‘low’, or ‘critically low’ – described above, under ‘Quality Assessment’); abstract free of spin (categorized as dichotomized variables “yes” of “no”, based on the SPIN 7-item checklist27 described below); and PRISMA for full texts checklist21 endorsed by the publishing journal and extracted from the “Instructions to authors” for each journal or from the PRISMA website28 (“yes” or “no”). Three independent variables were continuous: number of words in each review abstract (counted using the Microsoft Word “Word Count” function), 2017 journal Impact Factor (JIF), downloaded from the InCites Journal Citation Reports’ website (if a journal was not listed in this website, it was coded as not having an JIF, i.e. ‘missing value’), and number of citations, downloaded from the Web of Science Clarivate Analytics (normalized by the number of years since publication).

The literature has defined spin as “a misrepresentation of study results, regardless of motive (intentionally or unintentionally) that overemphasizes the beneficial effects of the intervention and overstates safety compared with that shown by the results” [p. 2].29 We used a 7-item checklist27 to detect the presence of spin in an abstract, by comparing the interpretation reported in the abstract conclusion and title to the abstract results: 1) recommendations not supported by the findings; 2) title claims a beneficial effect not supported by the findings; 3) selective reporting of outcomes or analysis; 4) conclusion claims safety based on non-statistically significant results; 5) conclusion claims beneficial effects despite high risk of bias; 6) selective reporting of harm outcomes; and 7) the conclusion extrapolates the review’s findings to a different intervention. Each item was classified as “yes” (i.e., spin was present), “no” (i.e., spin was not present), or “not reported” (i.e., the item evaluated was not reported, or could have been omitted, intentionally or not). We classified the abstract free of spin if none of the seven items was present.

Data were reported for all reviews and stratified into Cochrane and non-Cochrane reviews. We separated the data for Cochrane and non-Cochrane reviews, because previous research indicated that Cochrane reviews of physical therapy interventions have better reporting and methodological quality than non-Cochrane reviews.30 The methodological quality of reviews was reported in frequencies of “high”, “moderate”, “low”, and “critically low” scores (AMSTAR-2). We used IBM SPSS software Version 20.0 for all analyses (Boston, MA, USA).

We calculated the percentage of review abstracts achieving each item of the PRISMA-A, as well as mean and standard deviation (SD) of the “total number of PRISMA-A fully reported items”.

Consistency of reporting was evaluated by Kappa coefficients31 for items 2–12 of the PRISMA-A checklist (item 1 was omitted as it relates to the title). Kappa values were interpreted using the following criteria: <0 “less than chance”; 0.01–20 “slight”; 0.21−0.40 “fair”; 0.41−0.60 “moderate”; 0.61−0.80 “substantial”; and, 0.81−0.99 “almost perfect” agreement.31 An item was considered consistent if Kappa values were 0.61 or higher (i.e., “substantial” to “almost perfect” agreement).

We performed multivariate linear regression analyses to evaluate the association between reporting quality of the abstract and six independent variables. We built a multivariate regression model and included independent variables achieving P < 0.05. Adjusted explained variance (adjusted R2) and Beta coefficient (ß) and its 95% confidence intervals (CIs) were calculated for each possible variable with association. Scatter plots of the “total number of PRISMA-A fully reported items” (dependent variable) with each independent variable were created to asses linear relationship between variables to fit our model.32 Coefficient values of r ≥ 0.7 analyzed by bivariate correlation (multicollinearity) were not included in our linear regression model.33

The mean of the “total number of PRISMA-A fully reported items” was 4.9 ± 2.7 out of 12 for all reviews, with 4.1 ± 2.0 for non-Cochrane reviews and 9.9 ± 1.1 for Cochrane reviews. None of the reviews adhered to all 12 of the PRISMA-A items, and all reviews that adhered to 9, 10, or 11 items were Cochrane reviews (n = 8). The number and percentage of full reporting for each item of PRISMA-A is presented in Table 2. For the total sample, the items with the best reporting were the title (88%), objectives (74%), and information sources (52%). The items with the lowest reporting were eligibility criteria (3%), registration (20%), and synthesis of results (21%). With the exception of the risk of bias item, abstracts of Cochrane reviews had a higher percentage achievement of all items compared to abstracts of non-Cochrane reviews. The item with the largest absolute percentage difference between Cochrane and non-Cochrane reviews was registration.

There was typically inconsistent reporting between abstract and full text with 5 PRISMA-A items having kappa values <0.20 (slight agreement); 3 items with kappa in the range 0.20−0.40 (fair agreement), 1 item in the range 0.40−0.60 (moderate agreement), and 2 in the range 0.60−0.80 (substantial agreement) (Table 3). Additionally, four of the nine Cochrane reviews were republished in other journals.34–37 The republished version of the reviews had lower methodological quality and lower “total number of PRISMA-A fully reported items” than the original Cochrane versions.

Multicollinearity was not identified among the variables in the model. The multivariate model explained 72% of the variance in reporting of abstracts (“total number of PRISMA-A fully reported items”), as presented in Table 4. Higher reporting quality of abstracts was associated with higher JIF (ß 0.20; 95% CI: 0.06, 0.35), greater number of words in the abstract (ß 0.01; 95% CI: 0.00, 0.01), and higher review methodological quality (dummy variable ‘critically low’ with ß -3.06; 95% CI: −5.30, −0.82). Abstract reporting quality was not associated with journal endorsement of the PRISMA for full texts checklist,21 number of citations, and abstracts free of spin of information. In summary, this model explained that in most cases the “total number of PRISMA-A fully reported items” would be around 5.54 (constant) points out of 12. The “total number of PRISMA-A fully reported items” would increase 0.20 points for every additional point in terms of JIF, would increase 0.01 points for each additional word in the abstract, and would decrease 3.06 points for reviews with critically low methodological quality compared to a high methodological quality review.”

The major strengths of our study were the tools used to evaluate reporting and methodological quality, which have been created by well-known and highly respected methodological research groups.3,23 On the other hand, our findings may only apply to SRs evaluating physical therapy interventions for LBP and may not be generalized to abstracts of SRs in general. Also, we searched the journals' instructions to authors between March and May 2018, which could probably have changed from when the authors of the analyzed reviews searched such journals' instructions.