Quality of reporting in systematic reviews of adverse events: systematic review.

OBJECTIVES: To examine the quality of reporting of harms in systematic reviews, and to determine the need for a reporting guideline specific for reviews of harms.
DESIGN: Systematic review. DATA SOURCES: Cochrane Database of Systematic Reviews (CDSR) and Database of Abstracts of Reviews of Effects (DARE). REVIEW
METHODS: Databases were searched for systematic reviews having an adverse event as the main outcome, published from January 2008 to April 2011. Adverse events included an adverse reaction, harms, or complications associated with any healthcare intervention. Articles with a primary aim to investigate the complete safety profile of an intervention were also included. We developed a list of 37 items to measure the quality of reporting on harms in each review; data were collected as dichotomous outcomes ("yes" or "no" for each item).
RESULTS: Of 4644 reviews identified, 309 were systematic reviews or meta-analyses primarily assessing harms (13 from CDSR; 296 from DARE). Despite a short time interval, the comparison between the years of 2008 and 2010-11 showed no difference on the quality of reporting over time (P=0.079). Titles in fewer than half the reviews (proportion of reviews 0.46 (95% confidence interval 0.40 to 0.52)) did not mention any harm related terms. Almost one third of DARE reviews (0.26 (0.22 to 0.31)) did not clearly define the adverse events reviewed, nor did they specify the study designs selected for inclusion in their methods section. Almost half of reviews (n=170) did not consider patient risk factors or length of follow-up when reviewing harms of an intervention. Of 67 reviews of complications related to surgery or other procedures, only four (0.05 (0.01 to 0.14)) reported professional qualifications of the individuals involved. The overall, unweighted, proportion of reviews with good reporting was 0.56 (0.55 to 0.57); corresponding proportions were 0.55 (0.53 to 0.57) in 2008, 0.55 (0.54 to 0.57) in 2009, and 0.57 (0.55 to 0.58) in 2010-11.
CONCLUSION: Systematic reviews compound the poor reporting of harms data in primary studies by failing to report on harms or doing so inadequately. Improving reporting of adverse events in systematic reviews is an important step towards a balanced assessment of an intervention.

Introduction

A balanced assessment of interventions requires analysis of both benefits and harms. Systematic reviews or meta-analyses of randomised controlled trials are the preferred method to synthesise evidence in a comprehensive, transparent, and reproducible manner. Randomised controlled trials rarely assess harms as their primary outcome; therefore, they typically lack the power to detect differences in harms between groups (table 1). Usually designed to evaluate treatment efficacy or effectiveness, randomised controlled trials are often done over a short period of time, with a relatively small number of participants. These trials are known to be poor at identifying and reporting harms, which can lead to a misconception that a given intervention is safe, when its safety is actually unknown.1 2 3 4 5 6 7 8 Systematic reviews with a primary objective to assess harms represent fewer than 10% of all systematic reviews published yearly.9 10 Systematic reviews of harms can provide valuable information to describe adverse events (frequency, nature, seriousness), but they are hampered by a lack of standardised methods to report these events and the fact that harms are not usually the primary outcome of included studies.9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28

Table 1

 Glossary of terms

Term	Definition
Adverse drug reaction29	An adverse effect specific to a drug
Adverse effect25,29	An unfavourable outcome that occurs during or after the use of a drug or other intervention but is not necessarily caused by it
Adverse event25,29	An unfavourable outcome that occurs during or after the use of a drug or other intervention but is not necessarily caused by it
Complication29	An adverse event or effect following surgical and other invasive interventions
Harm8	The totality of possible adverse consequences of an intervention or therapy; they are the direct opposite of benefits
Safety8	Substantive evidence of an absence of harm. The term is often misused when there is simply absence of evidence of harm
Side effect29	Any unintended effect, adverse or beneficial, of a drug that occurs at doses normally used for treatment
Toxicity8	Describes drug related harms. The term may be most appropriate for laboratory determined measurements, although it is also used in relation to clinical events. The disadvantage of the term “toxicity” is that it implies causality. If authors cannot prove causality, the terms “abnormal laboratory measurements” or ‘laboratory abnormalities’ are more appropriate to use

Several studies have identified challenges when developing a systematic review of adverse events.9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 These include: the poor quality information on harms reported on original studies, difficulties in identifying relevant studies on adverse events when using standard systematic searches techniques, and the lack of a specific guideline to perform a systematic review of adverse events. The need for better reporting on harms in general1 2 3 4 5 6 7 8—and in systematic reviews9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 in particular—has been voiced. In a previous review28 of systematic reviews from the Cochrane Database of Systematic Reviews (CDSR) and Database of Abstracts of Reviews of Effects (DARE), our team identified a significantly increased number of reviews of adverse events over the past 17 years (P<0.001); however, the proportion of these reviews out of the total number of reviews was unchanged at 5%.11 14 28 Some positive points were noted—for example, the increased number of databases searched per review and the reduction in number of systematic reviews limiting their search strategies by date or language—but appropriate reporting of search strategies was still a problem.28

The PRISMA30 (preferred reporting items for systematic review and meta-analysis) statement was developed to deal with suboptimal reporting in systematic reviews. Thus far, PRISMA has mainly focused on efficacy and not on harms. A reporting guideline specific for systematic reviews of harms is crucial to provide a better assessment of adverse events of interventions. The first step for successful guideline development is to document the quality of reporting in published research articles to justify the need for the guideline.30 The goal of this review was to determine whether there is a need for a guideline specific for reviews of harms30 through assessment of the quality of reporting in systematic reviews of harms published between January 2008 and April 2011, from two major databases.

Methods

Development of the checklist items

To assess the quality of reporting in systematic reviews of harms, we developed a set of items (table 2) to be reported in these reviews.

Table 2

 Definitions of reporting items

Item and number	Definition of “yes”
Title
1a) Specifically mentions “harms” or other related term	Should contain a word or phrase related to harm, such as “adverse effect,” “adverse event,” “complications,” and “risk”
1b) Clarifies whether both benefits and harms are examined or only harms	Mentions whether benefits are reviewed or not
1c) Mentions the specific intervention being reviewed	Mentions the intervention being reviewed
1d) Refers to specific patient group or conditions (or both) in which harms have been assessed	Clearly states the specific group of patients or conditions being reviewed
Abstract
2a) Specifically mentions “harms” or other related terms	Should contain a word or phrase related to harm, such as “adverse effect,” “adverse event,” “complications,” and “risk”
2b) Clarifies whether both benefits and harms are examined or only harms	Mentions whether benefits are reviewed or not
2c) Refers to a specific harm assessed	Clearly states the adverse event being reviewed. General descriptions are accepted, for example, “cardiovascular events” or “maternal complications.” Also acceptable if the report is searching for any adverse event and it does not refer to a specific harm
2d) Specifies what type of data was sought	Clearly names the kind of studies searched (for example, randomised controlled trials, cohort studies, case reports only, all types of data)
2e) Specifies what type of data was included	Clearly names the kind of studies included in the analysis or results (that is, randomised controlled trials, cohort studies, case reports, all types of data)
2f) Specifies how each type of data has been appraised	Clearly states the method of quality appraisal for included studies (that is, Jadad score, Cochrane risk of bias tool, Newcastle-Ottawa scale)
Introduction
3a) Explains rationale for addressing specific harm(s), condition(s) and patient group(s)	Reports reasons for proceeding with the systematic review
3b) Clearly defines what events or effects are considered harms in the context of the intervention(s) examined	Clearly states the adverse event being reviewed. General definitions are acceptable (for example, “cardiovascular events” or “maternal complications”). Also acceptable if the report is searching for any adverse events, as in a scoping review, and does not refer to a specific harm
3c) Describes the rationale for type of harms systematic review done: hypothesis generating versus hypothesis testing	Clearly states whether a specific adverse event is being reviewed (hypothesis testing review; for example, cardiovascular deaths) or the authors are searching for any adverse events related to the intervention (hypothesis generating)
3d) Explains rationale for selection of study types or data sources and relevance to focus of the review	Clearly states which study designs are included (for example, randomised controlled trials, cohorts, case reports). Rationale is not necessary for “yes”
Objectives
4a) Provides an explicit statement of questions being asked with reference to harms	Clearly defines, preferably at the end of introduction section, the intervention and the adverse events being reviewed. For this review, it was considered acceptable if intervention and outcome were clearly stated
Methods
Protocol and registration
5a) Describes whether protocol was developed in collaboration with someone with clinical expertise for field or intervention under study	Mentions whether a protocol was developed previously. For this review, it was not necessary to state the clinical expertise of who developed it
Eligibility criteria
6a) Clearly defines what events or effects are considered harms in the context of the intervention(s) examined	Provides a clear definition of the adverse event being reviewed. If a general description was provided previously (that is, “cardiovascular events”), now the specific events need to be defined
6b) Specifies type of studies on harms to be included	Defines which kind of study designs will be included (for example, randomised controlled trials, cohort studies, case-control studies)
Information sources
7a) States whether additional sources for adverse events were searched (for example, regulatory bodies, industry); if so, describes the source, terms used, and dates searched	Reports whether any other sources of adverse events (other than regular peer reviewed journals) were searched. For this review, it was not necessary to have the terms used and dates searched
Search
8a) Presents the full search strategy if additional searches were used to identify adverse events	Reports search strategy used. Adverse events terms and databases searched
Study selection
9a) States what study designs were eligible and provide rationale for their selection	Reports the study designs included in the review (for example, randomised controlled trials, cohort studies, case-control studies). No rationale is required
9b) Defines if studies were screened on the basis of the presence or absence of harms related terms in title or abstract	Reports whether the study screening is based on the report of adverse events or not (for example, review of mortality associated with antiglycaemic drugs; only studies reporting on mortality were included)
Data collection process
10a) Describes method of data extraction for each type of study or report	States whether a data extraction form is used, and how it is done (in duplicate, checked by a second author)
Data items
11a) Lists and defines variables for which data were sought for individual therapies	Reports variable(s) sought for the intervention reviewed (for example, whether the outcome is kidney failure, what variable was used to define kidney failure—creatinine level, creatinine clearance, need of dialysis)
11b) Lists and defines variables for which data were sought for patient underlying risk factors	States whether any potential patient risk factors or confounders are sought (for example, age, sex, comorbidities, previous events)
11c) Lists and defines variables for which data were sought for practitioner training or qualifications	States any variable(s) sought for healthcare personnel (practitioner) (for example, relevant degree, years of experience, other qualifications)
11d) Lists and defines harms for individual therapies	Provides the definition for the harm(s) sought (for example, side effects of propranolol and atenolol defined as severe bradycardia (heart rate ≤45 beats per min) and severe hypotension (systolic blood pressure <80 mm Hg)
Risk of bias in individual studies
12a) For uncontrolled studies, describes whether causality between intervention and adverse event was adjudicated and if so, how	Specifies whether review authors adjudicate if the intervention can cause the harm (for example, use of Bradford Hill criteria for causality32
12b) Describes risk of bias in studies with incomplete or selective report of adverse events	Describes how studies not reporting adverse events are handled regarding risk of bias. For example, adverse events reported in a randomised controlled trial investigating glucose control in type 2 diabetes included hypoglycaemia but not mortality. Did the authors assess whether the event did not occur (zero deaths), was not measured (and may have occurred), or was measured but not reported
Summary measures
13a) If rare outcomes are being investigated, specifies which summary measures will be used (for example, event rate, events or person time)	Defines the summary measures used for rare events
Synthesis of results
14a) Describes statistical methods of handling with the zero events in included studies	Clearly defines how studies with no adverse events reported are reported and analysed (for example, when zero was an outcome for a 2×2 table, reviewers added a 0.5 value to it)
Additional analysis
16a) Describes additional analysis with studies with high risk of bias	Describes whether studies with high risk of bias are analysed separately (for example, subgroup analysis with high and low risk of bias studies analysed separately)
Results
Study selection
17a) Provides process, table, or flow for each type of study design	Provides a table (or text) containing included studies and a clear reason for exclusion of studies
Study characteristics
18a) Reports study characteristics, such as patient demographics or length of follow-up that may have influenced the risk estimates for the adverse outcome of interest	Reports any potential confounder or patient risk factors that can affect the outcome (adverse event)
18b) Describes methods of collecting adverse events in included studies (for example, patient report, active search)	Reports how adverse events are investigated in included studies (for example, voluntary report, active search)
18c) For each primary study, lists and defines each adverse event reported and how it is identified	Clear definition of adverse event under investigation and how it is identified (method of measurement, assessment or identification)
Conclusion
26a) Provides balanced discussion of benefits and harms with emphasis on study limitations, generalisability, and other sources of information on harms	Clearly discusses the benefits of the interventions and the harms identified in the review

The items were originally based on a draft generated from analysis of a systematic review of harms conducted previously.33 During the development of the data extraction form for this current review, several items were added. The wording and content were further refined over telephone meetings, by a group of experts in systematic reviews and guideline development.8 31 32 34 35 36 37 38 39

Not every PRISMA item has a corresponding harms item, and a few have more than one suggestion per item. PRISMA items 15, 19-25, and 27 did not have any specific harms related items.

Search strategy

We searched the CDSR (via the Cochrane Library) and DARE (via the Centre for Reviews and Dissemination and the Cochrane Library) databases for systematic reviews having an adverse event as the primary outcome measured. DARE is compiled through rigorous weekly searches of bibliographic databases (including Medline, Embase, PsycINFO, PubMed, and the Cumulative Index to Nursing and Allied Health Literature). It also involves less frequent searches of the Allied and Complementary Medicine Database and the Education Resource Information Center, hand searching of key journals, grey literature, and regular searches of the internet. CDSR includes all the systematic reviews published by the Cochrane Collaboration. We selected the combination of these two databases because they are likely to represent the most comprehensive collection of systematic reviews published in healthcare.28

The search was limited to a 40 month period between 1 January 2008 and 25 April 2011. The web appendix shows the search strategy used. The dates were selected to include recent reviews in order to describe the current state of reporting in systematic reviews of harms.

Eligibility criteria

Reviews were selected if the primary outcome investigated was exclusively an unintended effect or effects of an intervention. It could be an adverse event, adverse effect, adverse reaction, harms, or complications (table 1) associated with any healthcare intervention (such as pharmaceutical interventions, diagnostic procedures, surgical interventions, or medical devices). Articles with a primary aim to investigate the complete safety profile of an intervention were included. Reviews were not excluded on the basis of their results or conclusions.

We excluded reviews assessing both beneficial and harmful effects (reviews of both efficacy and harms), reviews of desirable side effects of drugs, or reviews of prevention or reduction of unintended or adverse effects. No limitations on interventions, patient groups, or language were applied.

Screening and data extraction

Relevant studies were screened by title and abstract (when available) independently by two authors (LZ and SG). Any disagreements were resolved by consensus; if consensus could not be reached, disagreements were resolved with a third author (SV). Abstracts that were identified as potentially meeting the DARE criteria but were not assessed were called “provisional abstracts.” The full text was retrieved for these abstracts.

The data extraction was based on the items developed, piloted, and refined (table 2). Each field received a “yes” if the item was reported as defined, or “no” if not reported. The data were extracted by one author (LZ) and verified by a second author (YL). Disagreements were resolved by consensus.

Outcome

The main outcome assessed was the quality of reporting in reviews of harms for each of the 37 items. We also measured the proportion of each “yes” response for each year of the search: 2008, 2009, and 2010-11 (12 reviews published in the first four months of 2011 were combined with the reviews published in 2010). The quality of reporting was compared between the earliest and latest years reviewed (2008 v 2010-11) to assess any improvement in quality of reporting during the study period.

We deliberately decided not to include an intermediate category (such as “unclear”). If the item was not clearly reported, it was considered as a “no” response and the unclear category would simply be duplication.

The present review did not aim to evaluate the reason behind the review author’s decisions to examine harms. Our goal was to measure the quality of reporting on those reviews— answering the question “is the item clearly reported?” The goal was not to judge whether a methodologically appropriate decision was made (for example, statistical tests used, data extraction, data pooling), but to ensure clarity in reporting regarding the choices made. Information on the types of interventions, nature of included study designs, as well as search strategies and databases searched in systematic reviews published between 1994 and 2011 was previously reported by our team.28

This study does not intend to measure the effect of the PRISMA statement, for two reasons. Firstly, PRISMA focuses on efficacy; thus measurement of its effect would reasonably focus on systematic reviews of efficacy and not specifically of harms. Secondly, the 37 items measured in the present review were new items and not those found in PRISMA.

Data analysis

Data were collected as dichotomous outcomes (“yes” or “no”) for each item, and presented as proportions of reviews for each category (reported from 0 to 1) and as proportions divided by the database that the reviews were identified from (CDSR and DARE). We also provided 95% confidence intervals for each proportion based on methods described by Wilson and Newcombe40 41 using a correction for continuity.

An overall reporting quality rate was provided through an unweighted average of proportions of items with good reporting. P≤0.05 was considered statistically significant. We did statistical calculations using StataIC-13.

Results

The search yielded 4644 unique references. After screening and retrieving full text articles, an extra 14 papers were excluded as they did not fulfil the inclusion criteria. A total of 309 reviews were identified as systematic reviews of meta-analyses primarily assessing harms, of which 13 were identified at CDSR and 296 at DARE (fig 1). Disagreements at the inclusion and exclusion criteria were discussed by LZ and SG and consensus was reached after discussion. Three of the included papers were published in Chinese, two in Spanish, and one in Portuguese. Table 3 provides detailed information on the reporting of each item.

Fig 1 PRISMA flow diagram

Table 3

 Summary of findings

Item and number	Proportion (%) and No of reviews with “yes” responses
	Total reviews (n=309)	CDSR reviews (n=13)	DARE reviews (n=296)
Title
1a) Specifically mentions “harms” or other related term	53 (n=165)	54 (n=7)	53 (n=158)
1b) Clarifies whether both benefits and harms are examined or only harms	76 (n=234)	92 (n=12)	75 (n=222)
1c) Mentions the specific intervention being reviewed	98 (n=303)	100 (n=13)	98 (n=290)
1d) Refers to specific patient group or conditions (or both) in which harms have been assessed	58 (n=178)	92 (n=12)	56 (n=166)
Abstract
2a) Specifically mentions “harms” or other related terms	84 (n=261)	85 (n=11)	85 (n=250)
2b) Clarifies whether both benefits and harms are examined or only harms	92 (n=285)	92 (n=12)	92 (n=273)
2c) Refers to a specific harm assessed	97 (n=299)	100 (n=13)	97 (n=289)
2d) Specifies what type of data was sought	50 (n=156)	92 (n=12)	49 (n=144)
2e) Specifies what type of data was included	49 (n=151)	92 (n=12)	47 (n=139)
2f) Specifies how each type of data has been appraised	6 (n=19)	23 (n=3)	5 (n=16)
Introduction
3a) Explains rationale for addressing specific harm(s), condition(s) and patient group(s)	99 (n=307)	100 (n=13)	99 (n=294)
3b) Clearly defines what events or effects are considered harms in the context of the intervention(s) examined	94 (n=289)	92 (n=12)	94 (n=277)
3c) Describes the rationale for type of harms systematic review done: hypothesis generating versus hypothesis testing	99 (n=306)	100 (n= 13)	99 (n=293)
3d) Explains rationale for selection of study types or data sources and relevance to focus of the review	39 (n=120)	54 (n=7)	38 (n=113)
Objectives
4a) Provides an explicit statement of questions being asked with reference to harms	86 (n=265)	100 (n=13)	85(n=252)
Methods
Protocol and registration
5a) Describes if protocol was developed	11 (n=35)	100(n=13)	7.4 (n=22)
Eligibility criteria
6a) Clearly defines what events or effects are considered harms in the context of the intervention(s) examined	75 (n=231)	92 (n=12)	74 (n=219)
6b) Specifies type of studies on harms to be included	74 (n=230)	92 (n=12)	74 (n=218)
Information sources
7a) States whether additional sources for adverse events were searched (for example, regulatory bodies, industry)	17 (n=54)	54 (n=7)	16 (n=47)
Search
8a) Presents the full search strategy if additional searches were used to identify adverse events	69 (n=214)	100 (n=13)	68 (n=201)
Study selection
9a) States what study designs were eligible	74 (n=229)	100 (n=13)	73 (n=216)
9b) Defines if studies were screened on the basis of the presence or absence of adverse events	71 (n=218)	92 (n=12)	70 (n=206)
Data collection process
10a) Describes method of data extraction for each type of study or report	64 (n=199)	100 (n=13)	63 (n=186)
Data items
11a) Lists and defines variables for which data were sought for individual therapies	50 (n=154)	23 (n=3)	51 (n=151)
11b) Lists and defines variables for which data were sought for patient underlying risk factors	42 (n=129)	23 (n=3)	43 (n=126)
11c) Lists and defines variables for which data were sought for practitioner training or qualifications	1 (n=4)	0 (n=0)	1 (n=4)
11d) Lists and defines harms for individual therapies	68 (n=210)	69 (n=9)	68 (n=201)
Risk of bias in individual studies
12a) For uncontrolled studies, describes whether causality between intervention and adverse event was adjudicated	3 (n=10)	23 (n=3)	2 (n=7)
12b) Describes risk of bias in studies with incomplete or selective report of adverse events	2 (n=6)	0 (n=0)	2 (n=6)
Summary measures
13a) If rare outcomes are being investigated, specifies which summary measures will be used (for example, event rate, events or person time)	73 (n=227)	100 (n=13)	72 (n=214)
Synthesis of results
14a) Describes statistical methods of handling with the zero events in included studies	13 (n=41)	46 (n=6)	12 (n=35)
Additional analysis
16a) Describes additional analysis with studies with high risk of bias	20 (n=61)	77 (n=10)	17 (n=51)
Results
Study selection
17a) Provides process, table, or flow for each type of study design	79 (n=244)	92 (n=12)	78 (n=232)
Study characteristics
18a) Reports study characteristics, such as patient demographics or length of follow-up that may have influenced the risk estimates for the adverse outcome of interest	55 (n=171)	85 (n=11)	54 (n=160)
18b) Describes methods of collecting adverse events in included studies (for example, patient report, active search)	62 (n=190)	85 (n=11)	61 (n=179)
18c) For each primary study, lists and defines each adverse event reported and how it is identified	22 (n=68)	23 (n=3)	22 (n=65)
Discussion
26a) Provides balanced discussion of benefits and harms with emphasis on study limitations, generalisability, and other sources of information on harms	83 (n=257)	92 (n=12)	83 (n=245)

The 309 systematic reviews and meta-analyses with harms as a primary outcome focused on the following interventions:

Drugs (223 studies, proportion of reviews 0.72 (95% confidence interval 0.66 to 0.77))

Surgery or other procedures (67 studies, 0.21 (0.17 to 0.26))

Devices (13 studies, 0.04 (0.02 to 0.07))

Blood transfusion (two studies, 0.006 (0.001 to 0.022))

Enteral nutrition (two studies, 0.006 (0.001 to 0.022))

Isolation rooms (one study, 0.003 (0.001 to 0.01))

Surgical versus medical treatment (one study, 0.003 (0.001 to 0.01)).

Titles and abstracts

Titles in close to half of included systematic reviews and meta-analyses of harms did not mention any harm related terms (proportion of reviews 0.46 (95% confidence interval 0.40 to 0.52)) and had no report of a patient population or condition under review (0.42 (0.36 to 0.48)). Twenty five reviews (0.08 (0.05 to 0.11) used the word “safety” to identify a review of harms. In the abstract section, reviews often used harm related terms (0.84 (0.79 to 0.88)). For other terms, one in every 6.5 reviews did not have any harms related word in the abstract, and half of reviews (0.5 (0.44 to 0.56)) did not report the study designs sought or included.

Introduction and rationale

Introductions were well written overall, explaining the rationale for the review and providing information on harms being reviewed. Fifty one reviews (proportion of reviews 0.16 (95% confidence interval 0.12 to 0.21)) were performed to investigate any adverse event associated with an intervention, rather than focusing on a specific event. As per our definition, these reviews provided “an explicit statement of questions being asked with reference to harms” as this broad goal was reported.

Methods

Protocol and registration

Consistent with Cochrane requirements for authors, all included systematic reviews conducted through the Cochrane Collaboration (Cochrane reviews) had a protocol. Cochrane reviews did not refer to a protocol in their full text reviews, but this item was considered “yes” for the reviews for which a protocol could be found. By contrast, only 22 reviews (proportion of reviews 0.07 (95% confidence interval 0.04 to 0.11)) identified through the DARE database reported the use of a protocol. Reporting clinical expertise was not deemed necessary to receive “yes” for this item (table 2).

Eligibility criteria

Almost one third of DARE reviews (proportion of reviews 0.26 (95% confidence interval 0.22 to 0.31)) did not have a clear definition of the adverse events reviewed, nor did they specify the study designs selected for inclusion in their methods section. All Cochrane reviews had a clear report of their search strategy, eligible study designs, and methods of data extraction.

Information sources

Authors did not usually search outside the peer reviewed literature for additional sources of adverse events; for example, only 54 reviews (proportion of reviews 0.17 (95% confidence interval 0.13 to 0.22)) searched databases of regulatory bodies or similar sources. Seven of 13 Cochrane reviews (0.53 (0.25 to 0.80)) searched for data from regulatory bodies or industry, compared with 47 of 296 (0.15 (0.11 to 0.20)) DARE reviews.

Study selection and data items

At the screening phase, most reviews only included studies if the harms being searched were reported (proportion of reviews 0.70 (95% confidence interval 0.65 to 0.75)). Of 67 reviews of complications related to surgery or procedures, only four (0.05 (0.01 to 0.14)) reported professional qualifications of the individuals involved.

Study characteristics

Reports of any possible patient related risk factors—such as age, sex, or comorbidities—were sought in less than half of reviews (proportion of reviews 0.41 (0.36 to 0.47)). Furthermore, only 10 reviews adjudicated whether the adverse event could be biologically, pharmacologically, or temporally caused by the intervention, as measured by item 12a.

Results

Fewer than half the reviews (132 of 309; proportion of reviews 0.42 (95% confidence interval 0.37 to 0.48)) only included controlled clinical trials (randomised or not). Only 59 reviews (0.19 (0.14 to 0.23)) included both clinical trials and observational studies. Only observational studies (prospective or retrospective) were included in 109 reviews (0.35 (0.29 to 0.40)), case series or case reports were included in 24 (0.07 (0.05 to 0.11)). Three of 13 Cochrane reviews (0.23 (0.05 to 0.53)) included observational studies compared with 165 of 296 DARE reviews (0.55 (0.49 to 0.61)). After reviewing the full text, nine reviews did not report the designs included anywhere in the text. Length of follow-up or patient demographics were only reported in just over half of reviews (170 of 309 reviews; 0.55 (0.49 to 0.60)).

In a retrospective analysis, we compared the quality of reporting between the years of 2008 and 2010-11 to measure any possible improvement over time on the quality of reporting. There were no significant differences (P=0.079) on the proportion of good reporting between the earlier and later years.

The 13 CDSR reviews had overall better reporting than DARE reviews in the abstracts, methods, and results sections; the other categories had similar levels of reporting quality. Almost half of DARE reviews (proportion of reviews 0.46 (95% confidence interval 0.40 to 0.51)) had poor reporting in the results section. Because the number of reviews was considerably different between databases, we considered it inappropriate to proceed with any formal tests to compare CDSR and DARE reviews.

Figure 2 provides a graphic trend in the proportion of reviews with good reporting over time (2008, 2009, and 2010-11), by each item reviewed. Reviews had a poor report quality on methods and results, with an average of half of items being poorly reported on those sections.

Fig 2 Proportion of good reporting by subheadings

The overall proportion of reviews with good reporting was 0.56 (95% confidence interval 0.55 to 0.57). The corresponding proportions were 0.55 (0.53 to 0.57) in 2008, 0.55 (0.54 to 0.57) in 2009, and 0.57 (0.55 to 0.58) in 2010-11.

Discussion

Principal findings

We conducted this systematic review to assess the quality of reporting in systematic reviews of harms as a primary outcome using a set of proposed reporting items. This is the first step for the development of PRISMA Harms, a reporting guideline specifically designed for reviews of harms.30 There is a substantial difference between the number of systematic reviews measuring an adverse event as the main outcome identified through DARE and those identified through CDSR. CDSR reviews comprised only a small fraction of the total reviews included. This distinction may compromise any direct comparison between the two databases. Reviews of harms as a secondary outcome or a coprimary outcome were not included in this review, which could be one reason for the large dissimilarity. Despite the small number of reviews of harms published in the CDSR, they were better reported overall than DARE reviews, probably owing to the clear guidelines provided by the Cochrane Collaboration29 and the more flexible word limits allowed than those in other peer reviewed journals.

Several items were poorly reported in the included reviews, but a few are especially important when reviewing harms, because the lack of reporting on these could lead to misinterpretations of findings. In a systematic review, the screening phase is crucial and the exclusion of studies due to the absence of harms could overestimate the events and perhaps generate a biased review. Two thirds of reviews of harms only included studies if at least one adverse event was reported in the included studies. In a review of harms, “zero” is an important value, and studies with “no adverse events” are possibly as relevant to the review as those with reported adverse events. Nevertheless, zero events in studies require careful interpretation, because the lack of reported harms may have different reasons: they may not have occurred (that is, a zero event), they may not have been investigated (that is, unknown if zero or no events occurred), or they may have been detected but not reported (that is, unknown if zero or no events occurred). The lack of reporting can be thought of as a measurement bias or reporting bias and should be considered as such.1 2 3 4 5 6 7 8 9 10 11 12 24 25 26

All these scenarios have different implications for readers who need to judge whether an intervention may cause harm. Almost half of the included reviews of harms as a primary outcome did not consider patient risk factors or length of follow-up when reviewing adverse events of an intervention. Readers cannot properly judge whether there is an association between intervention and harms if these critical data are not reported. Most of the poorly reported items were identified at methods and results. The clarity on methods and results are essential to provide a clear picture of author’s intentions and limitations of findings and transparency is warranted.

Strengths and limitations

This review was unique by including more than 300 reviews, from two major databases, that looked at harms as a main outcome; each review was evaluated in depth using a novel set of 37 items to measure the quality of reporting. A limitation of this review was the lack of a reporting guideline specific for systematic reviews of harms; different formats of reporting were found, and assessing whether the reporting was adequate was challenging. The reviewers were generous in their assessment, accepting a range of reports as a “yes” to the presence of an item, which could have underestimated the degree of the problem.

Our review was limited exclusively to systematic reviews where harms were the primary focus. We believed that measuring the quality of reporting of adverse events in reviews specifically designed to evaluate adverse events would generate a pure sample focusing on the reporting of such events. The documenting of poor reporting on these reviews would imply poor reporting on adverse events in general. At this stage, we also decided to be inclusive and measure all potentially relevant items. In the future PRISMA Harms extension, we will limit these to the minimum set of essential items for reporting harms in a systematic review.

Comparison with other studies

Hopewell and colleagues14 reviewed a sample of 59 Cochrane reviews, of which only one was a harms review. The remaining 58 reviews focused primarily on benefit, with adverse events as secondary or tertiary considerations. Hopewell and colleagues reported that 32 (54%) reviews had fewer than three paragraphs of information on adverse events, and 11 (19%) had fewer than five sentences on adverse events.

Hammad and colleagues42 reviewed a sample of 27 meta-analyses primarily assessing drug safety and identified that more than 85% of the PRISMA items were reported in the majority of the reviews. However, most reviews did not report the 20 items specifically developed by the authors to address drug safety assessment. Several of the items considered important by Hammad and colleagues where similar to the ones developed by our team. The present review adds to the voice of many authors highlighting the poor reporting in reviews of adverse events.9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 42

Policy implications

Systematic reviews may compound the poor reporting of harms data in primary studies by failing to report on harms or doing so inadequately.9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 We recognise the need to optimise quality of reporting on harms in primary studies, and attempts to enhance it have already been made.6 At this phase, we measured the report quality in systematic reviews. As a future step, we intend to compare the quality of reporting in included studies with the reporting in the systematic review.

Despite their status as the preferred method for knowledge synthesis, systematic reviews can present an incomplete picture to readers by not representing a reliable assessment of a given treatment. Authors of the systematic review have a unique vantage point and can evaluate the entire evidence base under review, including deficits in reporting harms at the primary study level. This vantage point should be used to flag deficits in primary reporting. The goal over time should be to improve the quality and clarity of reporting in systematic reviews as well as the primary studies they evaluate. Although we are glad to see increasing numbers of systematic reviews of harms being published, we are also aware that the validity of the results can be heavily influenced by reviewers’ decisions during conduct of the review—even more so than in reviews of benefit because we are dealing with sparse data and secondary outcomes.43 44 45 Hence, we emphasise here the crucial importance of transparent reporting of the methods used in systematic reviews of harms.

Conclusions

Systematic reviews of interventions should put equal emphasis on efficacy and harms. Improved reporting of adverse events in systematic reviews is one step towards providing a balanced assessment of an intervention. Patients, healthcare professionals, and policymakers should base their decisions not only on the efficacy of an intervention, but also on its risks. Guidance on a minimal set of items to be reported when reviewing harms is needed to improve transparency and informed decision making, thereby greatly enhancing the relevance of systematic reviews to clinical practice.

This review used a set of proposed reporting items to assess reporting, and the findings indicate that specific aspects of harms reporting could be improved. The items will be further refined with the aim of developing a final set of criteria that would constitute the PRISMA Harms. The PRISMA statement30 is a living document, open to criticism and suggestions; these same principles will be shared by the PRISMA Harms. The development of a standardised format for reporting harms in systematic reviews will promote clarity and help ensure that readers have the basic information necessary to make an informed assessment of the intervention under review.

There is room for improved reporting in systematic of harms, including clear definition of the events measured, length of follow-up, and patient risk factors

Comparisons of reviews of harms have shown no improvement in the quality of reporting over time

Lack of detail or transparency in the reporting of systematic reviews of harms could hinder proper assessment of validity of findings

The number of systematic reviews of adverse effects has increased significantly over the past 17 years

Harms are poorly reported in randomised controlled trials, but it is unclear whether there are weaknesses in the reporting of systematic reviews of harms

Although the PRISMA statement aims to provide guidance on transparent reporting in systematic reviews, its recommendations are mainly focused on studies of beneficial outcomes