Study design
We conducted a cohort study of articles reporting studies evaluating treatments in the field of cancer and published in high-impact-factor journals.
Identification of articles
Search strategy
We screened the highest impact factor journals in the following categories: 50 in “Oncology”, 25 in “Medicine, General and Internal” and 25 in “Medicine, Research and Experimental” (Journal citation report 2013, Thomson Reuters). We selected the journals that were publishing clinical studies or systematic reviews of clinical studies or observational studies evaluating the effect of interventions on humans and identified 24 journals from “Oncology”, 17 from “Medicine, General and Internal” and 6 from “Medicine, Research and Experimental”. We then searched MEDLINE via PubMed on March 1, 2015 for articles published from January 1, 2014 to June 30, 2014 in the selected journals by using the following search strategy: “name of the journal” in the journal search field; “cancer” in title and abstract field; article type “randomized controlled trials”, “clinical trials”, “observational studies”, “meta-analysis” or “systematic reviews” and text availability “abstract”.
Eligibility criteria
We included all studies evaluating an intervention to improve the health of patients with any type of cancer, regardless of study design. These interventions could concern chemotherapy, targeted therapy, radiotherapy, surgery, hormone therapy, immunotherapy and supportive care (e.g., analgesics, antibiotics, antiretroviral, dietary supplements, multivitamins, vaccination). We excluded studies of diagnostics, screening, prognostic factors, biomarkers, correlation and gene, molecular and protein expression that did not evaluate any treatment. We also excluded animal studies and narrative reviews.
Data extraction
An online data extraction form was developed and preliminarily tested on a sample of 30 articles. The following data were collected: journal type (i.e., cancer or general medical), study design (systematic reviews/meta-analyses (SRs/MAs), randomized controlled trials (RCTs), phase I/II non-randomized trials and observational studies), sample size and funding source (i.e., for profit, non-profit, both and not reported). The types of cancer and type of cancer treatments were classified according to the US National Cancer Institute” [18].
We determined whether the abstract conclusion favoured the study treatment, did not favour the study treatment or was neutral [19]. We checked whether there was an open access to the article on PubMed and recorded the online publication date on PubMed. Finally, we also checked whether the published article had issued a press release or not. For this purpose, we searched EurekAlert (online free database for science press releases: http://www.eurekalert.org/) using keywords from PubMed, online or journal publication date, journal name, authors’ first and last names and title.
Two researchers (RH, LG) with expertise in clinical epidemiology independently screened the titles and abstracts for 25% of the citations retrieved and extracted specific information. The reproducibility was very good (kappa > 0.9 for all items) (Additional file 1). Then, the remaining citations were divided among the two researchers for further screening and data extraction. The full text was retrieved to record the funding source when not reported in the abstract.
Online media attention measured by Altmetric score
The primary outcome was the online media attention measured by the Altmetric score. The Altmetric Web-based application tracks the attention scholarly articles receive online by using data from three main categories of sources: social media (i.e., Twitter, Facebook, Google+, Pinterest and blogs); traditional media (i.e., mainstream, such as The Guardian, New York Times, and science-specific, such as New Scientist and Scientific American) and online reference managers (i.e., Mendeley and CiteULike) [20]. This score, providing a quantitative measure of attention a scholarly article receives online, is derived from an automated algorithm. The score is weighted by the relative coverage of each published research article in each type of source (e.g., news, Twitter) [9]. For example, an average newspaper story is more likely to bring attention to the research article than an average tweet [9]. Additional file 2 provides details on how the Altmetric score is calculated.
The effect of time is important in exposure of media attention to the article [11]. In general, the published article receives maximum online attention within 6 months of its publication. Each mention of an article on online sources affects the Altmetric score. Therefore, we chose a delay of at least 10 months from the last publication date (June 30, 2014) to the Altmetric search date (May 1, 2015) to allow for sufficient exposure for a stable Altmetric score.
We searched Altmetric Explorer [7] by using the PubMed unique identifier (PMID) for the selected articles (Altmetric search date: May 1, 2015). Then, we downloaded the Altmetric score and number of news items, science blogs, tweets, Facebook posts, Google+ posts, Mendeley readers, CiteULike and some other sources where the published article was mentioned.
Statistical analysis
Qualitative variables are described with frequencies and percentages (%). Quantitative variables are described with medians [Q1–Q3]. We used the negative binomial GEE model to study the association of explanatory variables and Altmetric score. Regression coefficients represent the logarithm of the ratio of mean (RoM) values of the Altmetric score per unit change in the covariate. We chose this model to explain the wide dispersion of Altmetric score (greater variance than the mean). Using a function “offset”, we adjusted for the duration between online publication dates of articles (or journal publication date if the online publication date was greater than journal publication date) and the search date for Altmetric score (May 1, 2015) to account for the same post-publication exposure period. Clustering due to journals was accounted for by adding an exchangeable correlation structure to the model.
Univariate and multivariate analyses involved the following pre-specified explanatory variables: (1) journal impact factor, (2) study design in four classes (i.e., SR/MA, RCT, phase I/II non-randomized trial and observational study[as a referent group]), (3) abstract conclusion (in favour of study treatment (yes vs no [not in favour of study treatment and neutral]), (4) funding source (for profit [profit, both (profit and non-profit)] vs non-profit [non-profit, none and not reported]), (5) open access to the article (yes vs no) and (6) presence of a press release (yes vs no). All these variables were entered in the multivariate model to assess the association of each variable with high Altmetric score (controlling for the other variables in the model). Results are expressed as RoMs with 95% confidence intervals (95%CIs) for both univariate and multivariate analysis. Statistical analysis involved use of SAS for Windows 9.4 (SAS Inst., Cary, NC).