No, Google Scholar Shouldn’t be Used Alone for Systematic Review Searching

9 07 2013

Several papers have addressed the usefulness of Google Scholar as a source for systematic review searching. Unfortunately the quality of those papers is often well below the mark.

In 2010 I already [1]  (in the words of Isla Kuhn [2]) “robustly rebutted” the Anders’ paper PubMed versus Google Scholar for Retrieving Evidence” [3] at this blog.

But earlier this year another controversial paper was published [4]:

“Is the coverage of google scholar enough to be used alone for systematic reviews?

It is one of the highly accessed papers of BMC Medical Informatics and Decision Making and has been welcomed in (for instance) the Twittosphere.

Researchers seem  to blindly accept the conclusions of the paper:

But don’t rush  and assume you can now forget about PubMed, MEDLINE, Cochrane and EMBASE for your systematic review search and just do a simple Google Scholar (GS) search instead.

You might  throw the baby out with the bath water….

… As has been immediately recognized by many librarians, either at their blogs (see blogs of Dean Giustini [5], Patricia Anderson [6] and Isla Kuhn [1]) or as direct comments to the paper (by Tuulevi OvaskaMichelle Fiander and Alison Weightman [7].

In their paper, Jean-François Gehanno et al examined whether GS was able to retrieve all the 738 original studies included in 29 Cochrane and JAMA systematic reviews.

And YES! GS had a coverage of 100%!

WOW!

All those fools at the Cochrane who do exhaustive searches in multiple databases using controlled vocabulary and a lot of synonyms when a simple search in GS could have sufficed…

But it is a logical fallacy to conclude from their findings that GS alone will suffice for SR-searching.

Firstly, as Tuulevi [7] rightly points out :

“Of course GS will find what you already know exists”

Or in the words of one of the official reviewers [8]:

What the authors show is only that if one knows what studies should be identified, then one can go to GS, search for them one by one, and find out that they are indexed. But, if a researcher already knows the studies that should be included in a systematic review, why bother to also check whether those studies are indexed in GS?

Right!

Secondly, it is also the precision that counts.

As Dean explains at his blog a 100% recall with a precision of 0,1% (and it can be worse!) means that in order to find 36 relevant papers you have to go through  ~36,700 items.

Dean:

Are the authors suggesting that researchers consider a precision level of 0.1% acceptable for the SR? Who has time to sift through that amount of information?

It is like searching for needles in a haystack.  Correction: It is like searching for particular hay stalks in a hay stack. It is very difficult to find them if they are hidden among other hay stalks. Suppose the hay stalks were all labeled (title), and I would have a powerful haystalk magnet (“title search”)  it would be a piece of cake to retrieve them. This is what we call “known item search”. But would you even consider going through the haystack and check the stalks one by one? Because that is what we have to do if we use Google Scholar as a one stop search tool for systematic reviews.

Another main point of criticism is that the authors have a grave and worrisome lack of understanding of the systematic review methodology [6] and don’t grasp the importance of the search interface and knowledge of indexing which are both integral to searching for systematic reviews.[7]

One wonders why the paper even passed the peer review, as one of the two reviewers (Miguel Garcia-Perez [8]) already smashed the paper to pieces.

The authors’ method is inadequate and their conclusion is not logically connected to their results. No revision (major, minor, or discretionary) will save this work. (…)

Miguel’s well funded criticism was not well addressed by the authors [9]. Apparently the editors didn’t see through and relied on the second peer reviewer [10], who merely said it was a “great job” etcetera, but that recall should not be written with a capital R.
(and that was about the only revision the authors made)

Perhaps it needs another paper to convince Gehanno et al and the uncritical readers of their manuscript.

Such a paper might just have been published [11]. It is written by Dean Giustini and Maged Kamel Boulos and is entitled:

Google Scholar is not enough to be used alone for systematic reviews

It is a simple and straightforward paper, but it makes its points clearly.

Giustini and Kamel Boulos looked for a recent SR in their own area of expertise (Chou et al [12]), that included a comparable number of references as that of Gehanno et al. Next they test GS’ ability to locate these references.

Although most papers cited by Chou et al. (n=476/506;  ~95%) were ultimately found in GS, numerous iterative searches were required to find the references and each citation had to be managed once at a time. Thus GS was not able to locate all references found by Chou et al. and the whole exercise was rather cumbersome.

As expected, trying to find the papers by a “real-life” GS search was almost impossible. Because due to its rudimentary structure, GS did not understand the expert search strings and was unable to translate them. Thus using Chou et al.’s original search strategy and keywords yielded unmanageable results of approximately >750,000 items.

Giustini and Kamel Boulos note that GS’ ability to search into the full-text of papers combined with its PageRank’s algorithm can be useful.

On the other hand GS’ changing content, unknown updating practices and poor reliability make it an inappropriate sole choice for systematic reviewers:

As searchers, we were often uncertain that results found one day in GS had not changed a day later and trying to replicate searches with date delimiters in GS did not help. Papers found today in GS did not mean they would be there tomorrow.

But most importantly, not all known items could be found and the search process and selection are too cumbersome.

Thus shall we now for once and for all conclude that GS is NOT sufficient to be used alone for SR searching?

We don’t need another bad paper addressing this.

But I would really welcome a well performed paper looking at the additional value of a GS in SR-searching. For I am sure that GS may be valuable for some questions and some topics in some respects. We have to find out which.

References

  1. PubMed versus Google Scholar for Retrieving Evidence 2010/06 (laikaspoetnik.wordpress.com)
  2. Google scholar for systematic reviews…. hmmmm  2013/01 (ilk21.wordpress.com)
  3. Anders M.E. & Evans D.P. (2010) Comparison of PubMed and Google Scholar literature searches, Respiratory care, May;55(5):578-83  PMID:
  4. Gehanno J.F., Rollin L. & Darmoni S. (2013). Is the coverage of Google Scholar enough to be used alone for systematic reviews., BMC medical informatics and decision making, 13:7  PMID:  (open access)
  5. Is Google scholar enough for SR searching? No. 2013/01 (blogs.ubc.ca/dean)
  6. What’s Wrong With Google Scholar for “Systematic” Review 2013/01 (etechlib.wordpress.com)
  7. Comments at Gehanno’s paper (www.biomedcentral.com)
  8. Official Reviewer’s report of Gehanno’s paper [1]: Miguel Garcia-Perez, 2012/09
  9. Authors response to comments  (www.biomedcentral.com)
  10. Official Reviewer’s report of Gehanno’s paper [2]: Henrik von Wehrden, 2012/10
  11. Giustini D. & Kamel Boulos M.N. (2013). Google Scholar is not enough to be used alone for systematic reviews, Online Journal of Public Health Informatics, 5 (2) DOI:
  12. Chou W.Y.S., Prestin A., Lyons C. & Wen K.Y. (2013). Web 2.0 for Health Promotion: Reviewing the Current Evidence, American Journal of Public Health, 103 (1) e9-e18. DOI:




The Scatter of Medical Research and What to do About it.

18 05 2012

ResearchBlogging.orgPaul Glasziou, GP and professor in Evidence Based Medicine, co-authored a new article in the BMJ [1]. Similar to another paper [2] I discussed before [3] this paper deals with the difficulty for clinicians of staying up-to-date with the literature. But where the previous paper [2,3] highlighted the mere increase in number of research articles over time, the current paper looks at the scatter of randomized clinical trials (RCTs) and systematic reviews (SR’s) accross different journals cited in one year (2009) in PubMed.

Hofmann et al analyzed 7 specialties and 9 sub-specialties, that are considered the leading contributions to the burden of disease in high income countries.

They followed a relative straightforward method for identifying the publications. Each search string consisted of a MeSH term (controlled  term) to identify the selected disease or disorders, a publication type [pt] to identify the type of study, and the year of publication. For example, the search strategy for randomized trials in cardiology was: “heart diseases”[MeSH] AND randomized controlled trial[pt] AND 2009[dp]. (when searching “heart diseases” as a MeSH, narrower terms are also searched.) Meta-analysis[pt] was used to identify systematic reviews.

Using this approach Hofmann et al found 14 343 RCTs and 3214 SR’s published in 2009 in the field of the selected (sub)specialties. There was a clear scatter across journals, but this scatter varied considerably among specialties:

“Otolaryngology had the least scatter (363 trials across 167 journals) and neurology the most (2770 trials across 896 journals). In only three subspecialties (lung cancer, chronic obstructive pulmonary disease, hearing loss) were 10 or fewer journals needed to locate 50% of trials. The scatter was less for systematic reviews: hearing loss had the least scatter (10 reviews across nine journals) and cancer the most (670 reviews across 279 journals). For some specialties and subspecialties the papers were concentrated in specialty journals; whereas for others, few of the top 10 journals were a specialty journal for that area.
Generally, little overlap occurred between the top 10 journals publishing trials and those publishing systematic reviews. The number of journals required to find all trials or reviews was highly correlated (r=0.97) with the number of papers for each specialty/ subspecialty.”

Previous work already suggested that this scatter of research has a long tail. Half of the publications is in a minority of papers, whereas the remaining articles are scattered among many journals (see Fig below).

Click to enlarge en see legends at BMJ 2012;344:e3223 [CC]

The good news is that SRs are less scattered and that general journals appear more often in the top 10 journals publishing SRs. Indeed for 6 of the 7 specialties and 4 of the 9 subspecialties, the Cochrane Database of Systematic Reviews had published the highest number of systematic reviews, publishing between 6% and 18% of all the systematic reviews published in each area in 2009. The bad news is that even keeping up to date with SRs seems a huge, if not impossible, challenge.

In other words, it is not sufficient for clinicians to rely on personal subscriptions to a few journals in their specialty (which is common practice). Hoffmann et al suggest several solutions to help clinicians cope with the increasing volume and scatter of research publications.

  • a central library of systematic reviews (but apparently the Cochrane Library fails to fulfill such a role according to the authors, because many reviews are out of date and are perceived as less clinically relevant)
  • registry of planned and completed systematic reviews, such as prospero. (this makes it easier to locate SRs and reduces bias)
  • Synthesis of Evidence and synopses, like the ACP-Jounal Club which summarizes the best evidence in internal medicine
  • Specialised databases that collate and critically appraise randomized trials and systematic reviews, like www.pedro.org.au for physical therapy. In my personal experience, however, this database is often out of date and not comprehensive
  • Journal scanning services like EvidenceUpdates from mcmaster.ca), which scans over 120 journals, filters articles on the basis of quality, has practising clinicians rate them for relevance and newsworthiness, and makes them available as email alerts and in a searchable database. I use this service too, but besides that not all specialties are covered, the rating of evidence may not always be objective (see previous post [4])
  • The use of social media tools to alert clinicians to important new research.

Most of these solutions are (long) existing solutions that do not or only partly help to solve the information overload.

I was surprised that the authors didn’t propose the use of personalized alerts. PubMed’s My NCBI feature allows to create automatic email alerts on a topic and to subscribe to electronic tables of contents (which could include ACP journal Club). Suppose that a physician browses 10 journals roughly covering 25% of the trials. He/she does not need to read all the other journals from cover to cover to avoid missing one potentially relevant trial. Instead it is far more efficient to perform a topic search to filter relevant studies from journals that seldom publish trials on the topic of interest. One could even use the search of Hoffmann et al to achieve this.* Although in reality, most clinical researchers will have narrower fields of interest than all studies about endocrinology and neurology.

At our library we are working at creating deduplicated, easy to read, alerts that collate table of contents of certain journals with topic (and author) searches in PubMed, EMBASE and other databases. There are existing tools that do the same.

Another way to reduce the individual work (reading) load is to organize journals clubs or even better organize regular CATs (critical appraised topics). In the Netherlands, CATS are a compulsory item for residents. A few doctors do the work for many. Usually they choose topics that are clinically relevant (or for which the evidence is unclear).

The authors shortly mention that their search strategy might have missed  missed some eligible papers and included some that are not truly RCTs or SRs, because they relied on PubMed’s publication type to retrieve RCTs and SRs. For systematic reviews this may be a greater problem than recognized, for the authors have used meta-analyses[pt] to identify systematic reviews. Unfortunately PubMed has no publication type for systematic reviews, but it may be clear that there are many more systematic reviews that meta-analyses. Possibly systematical reviews might even have a different scatter pattern than meta-analyses (i.e. the latter might be preferentially included in core journals).

Furthermore not all meta-analyses and systematic reviews are reviews of RCTs (thus it is not completely fair to compare MAs with RCTs only). On the other hand it is a (not discussed) omission of this study, that only interventions are considered. Nowadays physicians have many other questions than those related to therapy, like questions about prognosis, harm and diagnosis.

I did a little imperfect search just to see whether use of other search terms than meta-analyses[pt] would have any influence on the outcome. I search for (1) meta-analyses [pt] and (2) systematic review [tiab] (title and abstract) of papers about endocrine diseases. Then I subtracted 1 from 2 (to analyse the systematic reviews not indexed as meta-analysis[pt])

Thus:

(ENDOCRINE DISEASES[MESH] AND SYSTEMATIC REVIEW[TIAB] AND 2009[DP]) NOT META-ANALYSIS[PT]

I analyzed the top 10/11 journals publishing these study types.

This little experiment suggests that:

  1. the precise scatter might differ per search: apparently the systematic review[tiab] search yielded different top 10/11 journals (for this sample) than the meta-analysis[pt] search. (partially because Cochrane systematic reviews apparently don’t mention systematic reviews in title and abstract?).
  2. the authors underestimate the numbers of Systematic Reviews: simply searching for systematic review[tiab] already found appr. 50% additional systematic reviews compared to meta-analysis[pt] alone
  3. As expected (by me at last), many of the SR’s en MA’s were NOT dealing with interventions, i.e. see the first 5 hits (out of 108 and 236 respectively).
  4. Together these findings indicate that the true information overload is far greater than shown by Hoffmann et al (not all systematic reviews are found, of all available search designs only RCTs are searched).
  5. On the other hand this indirectly shows that SRs are a better way to keep up-to-date than suggested: SRs  also summarize non-interventional research (the ratio SRs of RCTs: individual RCTs is much lower than suggested)
  6. It also means that the role of the Cochrane Systematic reviews to aggregate RCTs is underestimated by the published graphs (the MA[pt] section is diluted with non-RCT- systematic reviews, thus the proportion of the Cochrane SRs in the interventional MAs becomes larger)

Well anyway, these imperfections do not contradict the main point of this paper: that trials are scattered across hundreds of general and specialty journals and that “systematic reviews” (or meta-analyses really) do reduce the extent of scatter, but are still widely scattered and mostly in different journals to those of randomized trials.

Indeed, personal subscriptions to journals seem insufficient for keeping up to date.
Besides supplementing subscription by  methods such as journal scanning services, I would recommend the use of personalized alerts from PubMed and several prefiltered sources including an EBM search machine like TRIP (www.tripdatabase.com/).

*but I would broaden it to find all aggregate evidence, including ACP, Clinical Evidence, syntheses and synopses, not only meta-analyses.

**I do appreciate that one of the co-authors is a medical librarian: Sarah Thorning.

References

  1. Hoffmann, Tammy, Erueti, Chrissy, Thorning, Sarah, & Glasziou, Paul (2012). The scatter of research: cross sectional comparison of randomised trials and systematic reviews across specialties BMJ, 344 : 10.1136/bmj.e3223
  2. Bastian, H., Glasziou, P., & Chalmers, I. (2010). Seventy-Five Trials and Eleven Systematic Reviews a Day: How Will We Ever Keep Up? PLoS Medicine, 7 (9) DOI: 10.1371/journal.pmed.1000326
  3. How will we ever keep up with 75 trials and 11 systematic reviews a day (laikaspoetnik.wordpress.com)
  4. Experience versus Evidence [1]. Opioid Therapy for Rheumatoid Arthritis Pain. (laikaspoetnik.wordpress.com)




“Pharmacological Action” in PubMed has no True Equivalent in OVID MEDLINE

11 01 2012

Searching for EMBASE Subject Headings (the EMBASE index terms) for drugs is relatively straight forward in EMBASE.

When you want to search for aromatase inhibitors you first search for the Subject Heading mapping to aromatase inhibitors (aromatase inhibitor). Next you explode aromatase inhibitor/ if you are interested in all its narrower terms. If not, you search both for the general term aromatase inhibitor and those specific narrower terms you want to include.
Exploding aromatase inhibitor (exp aromatase inhibitor/) yields 15938 results. That is approximately twice what you get by searching aromatase inhibitor/ alone (not exploded). This yields 7434 hits.

It is different in MEDLINE. If you search for aromatase inhibitors in the MeSH database you get two suggestions.

The first index term “Aromatase Inhibitors” is a Mesh. It has no narrower terms.
Drug-Mesh are generally not arranged by working mechanism, but by chemical structure/type of compound. That is often confusing. Spironolactone for instance belongs to the MeSH Lactones (and Pregnenes) not to the MeSH Aldosterone Antagonists or Androgen Antagonist. Most Clinicians want to search for a group of compounds with the same mechanism of action, not the same biochemical family

The second term “Aromatase Inhibitors” [Pharmacological Action]  however does stand for the working mechanism. It does have narrower terms, including 2 MeSH terms (highlighted) and various substance names, also called Supplementary Concepts. 

For complete results you have to search for both MeSH and Pharmacological action: “Aromatase Inhibitors”[Mesh] yields 3930 records, whereas (“Aromatase Inhibitors”[Mesh]) OR “Aromatase Inhibitors” [Pharmacological Action] yields 6045. That is a lot more.

I usually don’t search PubMed, but OVID MEDLINE.

I know that Pharmacological Action-subheadings are important, so I tried to find the equivalent in OVID .

I found the MeSH Aromatase Inhibitors, but -unlike PubMed- OVID showed only two narrower Drug Terms (called Non-MeSH here versus MeSH in PubMed).

I found that odd.

I reasoned “Pharmacological action” might perhaps be combined with the MESH in OVID MEDLINE. This was later confirmed by Melissa Rethlefsen (see Twitter discussion below)

In Ovid MEDLINE I got 3937 hits with Aromatase Inhibitors/ and 5219 with exp Aromatase Inhibitors/ (thus including aminogluthemide or Fadrozole)

At this point I checked PubMed (shown above). Here I found  that “Aromatase Inhibitors”[Mesh] OR “Aromatase Inhibitors” [Pharmacological Action] yielded 6045 hits in PubMed, against 5219 in OVID MEDLINE for exp Aromatase Inhibitors/

The specific aromatase inhibitors Aminogluthemide/and Fadrozole/ [set 60] accounted fully for the difference  between exploded [set 59] and non-exploded Aromatase Inhibitors[set 58].

But what explained the gap of approximately 800 records between “Aromatase Inhibitors”[Mesh] OR “Aromatase Inhibitors”[Pharmacological Action]* in PubMed and exp aromatase inhibitors/ in OVID MEDLINE?

Could it be the substance names, mentioned under “Aromatase Inhibitors”[Pharmacological Action], I wondered?

Thus I added all the individual substance names in OVID MEDLINE (code= .rn.). See search set 61 below.

Indeed these accounted fully for the difference (set 62= 59 or 61 : the total number of hits in PubMed is similar)

It obviously is a mistake of OVID MEDLINE and I will inform them.

For the meanwhile, take care to add the individual substance names when you search for drug terms that have a pharmacological action-equivalent in PubMed. The substance names are not automatically searched when exploding the MeSH-term in OVID MEDLINE.

——–

For more info on Pharmacological action, see: http://www.nlm.nih.gov/bsd/disted/mesh/paterms.html

Twitter Discussion between me and Melissa Rethlefsen about the discrepancy between PubMed and OVID MEDLINE (again showing how helpful Twitter can be for immediate discussions and exchange of thoughts)

[read from bottom to top]





Things to Keep in Mind when Searching OVID MEDLINE instead of PubMed

25 11 2011

When I search extensively for systematic reviews I prefer OVID MEDLINE to PubMed for several reasons. Among them, it is easier to build a systematic search in OVID, the search history has a more structured format that is easy to edit, the search features are more advanced giving you more control over the search and translation of the a search to OVID EMBASE, PSYCHINFO and the Cochrane Library is “peanuts”, relatively speaking.

However, there are at least two things to keep in mind when searching OVID MEDLINE instead of PubMed.

1. You may miss publications, most notably recent papers.

PubMed doesn’t only provide access to MEDLINE, but also contains some other citations, including in-process citations which provide a record for an article before it is indexed with MeSH and added to MEDLINE.

As previously mentioned, I once missed a crucial RCT that was available in PubMed, but not yet available in OVID/MEDLINE.

A few weeks ago one of my clients said that she found 3 important papers with a simple PubMed search that were not retrieved by my exhaustive OVID MEDLINE (Doh!).
All articles were recent ones [Epub ahead of print, PubMed - as supplied by publisher]. I checked that these articles were indeed not yet included in OVID MEDLINE, and they weren’t.

As said, PubMed doesn’t have all search features of OVID MEDLINE and I felt a certain reluctance to make a completely new exhaustive search in PubMed. I would probably retrieve many irrelevant papers which I had tried to avoid by searching OVID*. I therefore decided to roughly translate the OVID search using textwords only (the missed articles had no MESH attached). It was a matter of copy-pasting the single textwords from the OVID MEDLINE search (and omitting adjacency operators) and adding the command [tiab], which means that terms are searched as textwords (in title and abstract) in PubMed (#2, only part of the long search string is shown).

To see whether all articles missed in OVID were in the non-MEDLINE set, I added the command: NOT MEDLINE[sb] (#3). Of the 332 records (#2), 28 belonged to the non-MEDLINE subset. All 3 relevant articles, not found in OVID MEDLINE, were in this set.

In total, there were 15 unique records not present in the OVID MEDLINE and EMBASE search. This additional search in PubMed was certainly worth the effort as it yielded more than 3 new relevant papers. (Apparently there was a boom in relevant papers on the topic, recently)

In conclusion, when doing an exhaustive search in OVID MEDLINE it is worth doing an additional search in PubMed to find the non-MEDLINE papers. Regularly these are very relevant papers that you wouldn’t like to have missed. Dependent on your aim you can suffice with a simpler, broader search for only textwords and limit by using NOT MEDLINE[sb].**

From now on, I will always include this PubMed step in my exhaustive searches. 

2. OVID MEDLINE contains duplicate records

I use Reference Manager to deduplicate the records retrieved from all databases  and I share the final database with my client. I keep track of the number of hits in each database and of the number of duplicates to facilitate the reporting of the search procedure later on (using the PRISM flowchart, see above). During this procedure, I noticed that I always got LESS records in Reference Manager when I imported records from OVID MEDLINE, but not when I imported records from the other databases. Thus it appears that OVID MEDLINE contains duplicate records.

For me it was just a fact that there were duplicate records in OVID MEDLINE. But others were surprised to hear this.

Where everyone just wrote down the number of total number hits in OVID MEDLINE, I always used the number of hits after deduplication in Reference Manager. But this is a quite a detour and not easy to explain in the PRISM-flowchart.

I wondered whether this deduplication could be done in OVID MEDLINE directly. I knew you cold deduplicate a multifile search, but would it also be possible to deduplicate a set from one database only? According to OVID help there should be a button somewhere, but I couldn’t find it (curious if you can).

Googling I found another OVID manual saying :

..dedup n = Removes duplicate records from multifile search results. For example, ..dedup 5 removes duplicate records from the multifile results set numbered 5.

Although the manual only talked about “multifile searches”, I tried the comment (..dedup 34) on the final search set (34) in OVID MEDLINE, and voilà, 21 duplicates were found (exactly the same number as removed by Reference manager)

The duplicates had the same PubMed ID (PMID, the .an. command in OVID), and were identical or almost identical.

Differences that I noticed were minimal changes in the MeSH (i.e. one or more MeSH  and/or subheadings changed) and changes in journal format (abbreviation used instead of full title).

Why are these duplicates present in OVID MEDLINE and not in PubMed?

These are the details of the PMID 20846254 in OVID (2 records) and in PubMed (1 record)

The Electronic Date of Publication (PHST)  was September 16th 2010. 2 days later the record was included in PubMed , but MeSH were added 3 months later ((MHDA: 2011/02/12). Around this date records are also entered in OVID MEDLINE. The only difference between the 2 records in OVID MEDLINE is that one record appears to be revised at 2011-10-13, whereas the other is not.

The duplicate records of 18231698 have again the same creation date (20080527) and entry date (20081203), but one is revised 2110-20-09 and updated 2010-12-14, while the other is revised 2011-08-18 and updated 2011-08-19 (thus almost one year later).

Possibly PubMed changes some records, instantaneously replacing the old ones, but OVID only includes the new PubMed records during MEDLINE-updates and doesn’t delete the old version.

Anyway, wouldn’t it be a good thing if OVID deduplicated its MEDLINE records on a daily basis or would replace the old ones when loading  new records from MEDLINE?

In the meantime, I would recommend to apply the deduplicate command yourself to get the exact number of unique records retrieved by your search in OVID MEDLINE.

*mostly because PubMed doesn’t have an adjacency-operator.
** Of course, only if you have already an extensive OVID MEDLINE search.





FUTON Bias. Or Why Limiting to Free Full Text Might not Always be a Good Idea.

8 09 2011

ResearchBlogging.orgA few weeks ago I was discussing possible relevant papers for the Twitter Journal Club  (Hashtag #TwitJC), a succesful initiative on Twitter, that I have discussed previously here and here [7,8].

I proposed an article, that appeared behind a paywall. Annemarie Cunningham (@amcunningham) immediately ran the idea down, stressing that open-access (OA) is a pre-requisite for the TwitJC journal club.

One of the TwitJC organizers, Fi Douglas (@fidouglas on Twitter), argued that using paid-for journals would defeat the objective that  #TwitJC is open to everyone. I can imagine that fee-based articles could set a too high threshold for many doctors. In addition, I sympathize with promoting OA.

However, I disagree with Annemarie that an OA (or rather free) paper is a prerequisite if you really want to talk about what might impact on practice. On the contrary, limiting to free full text (FFT) papers in PubMed might lead to bias: picking “low hanging fruit of convenience” might mean that the paper isn’t representative and/or doesn’t reflect the current best evidence.

But is there evidence for my theory that selecting FFT papers might lead to bias?

Lets first look at the extent of the problem. Which percentage of papers do we miss by limiting for free-access papers?

survey in PLOS by Björk et al [1] found that one in five peer reviewed research papers published in 2008 were freely available on the internet. Overall 8,5% of the articles published in 2008 (and 13,9 % in Medicine) were freely available at the publishers’ sites (gold OA).  For an additional 11,9% free manuscript versions could be found via the green route:  i.e. copies in repositories and web sites (7,8% in Medicine).
As a commenter rightly stated, the lag time is also important, as we would like to have immediate access to recently published research, yet some publishers (37%) impose an access-embargo of 6-12 months or more. (these papers were largely missed as the 2008 OA status was assessed late 2009).

PLOS 2009

The strength of the paper is that it measures  OA prevalence on an article basis, not on calculating the share of journals which are OA: an OA journal generally contains a lower number of articles.
The authors randomly sampled from 1.2 million articles using the advanced search facility of Scopus. They measured what share of OA copies the average researcher would find using Google.

Another paper published in  J Med Libr Assoc (2009) [2], using similar methods as the PLOS survey examined the state of open access (OA) specifically in the biomedical field. Because of its broad coverage and popularity in the biomedical field, PubMed was chosen to collect their target sample of 4,667 articles. Matsubayashi et al used four different databases and search engines to identify full text copies. The authors reported an OA percentage of 26,3 for peer reviewed articles (70% of all articles), which is comparable to the results of Björk et al. More than 70% of the OA articles were provided through journal websites. The percentages of green OA articles from the websites of authors or in institutional repositories was quite low (5.9% and 4.8%, respectively).

In their discussion of the findings of Matsubayashi et al, Björk et al. [1] quickly assessed the OA status in PubMed by using the new “link to Free Full Text” search facility. First they searched for all “journal articles” published in 2005 and then repeated this with the further restrictions of “link to FFT”. The PubMed OA percentages obtained this way were 23,1 for 2005 and 23,3 for 2008.

This proportion of biomedical OA papers is gradually increasing. A chart in Nature’s News Blog [9] shows that the proportion of papers indexed on the PubMed repository each year has increased from 23% in 2005 to above 28% in 2009.
(Methods are not shown, though. The 2008 data are higher than those of Björk et al, who noticed little difference with 2005. The Data for this chart, however, are from David Lipman, NCBI director and driving force behind the digital OA archive PubMed Central).
Again, because of the embargo periods, not all literature is immediately available at the time that it is published.

In summary, we would miss about 70% of biomedical papers by limiting for FFT papers. However, we would miss an even larger proportion of papers if we limit ourselves to recently published ones.

Of course, the key question is whether ignoring relevant studies not available in full text really matters.

Reinhard Wentz of the Imperial College Library and Information Service already argued in a visionary 2002 Lancet letter[3] that the availability of full-text articles on the internet might have created a new form of bias: FUTON bias (Full Text On the Net bias).

Wentz reasoned that FUTON bias will not affect researchers who are used to comprehensive searches of published medical studies, but that it will affect staff and students with limited experience in doing searches and that it might have the same effect in daily clinical practice as publication bias or language bias when doing systematic reviews of published studies.

Wentz also hypothesized that FUTON bias (together with no abstract available (NAA) bias) will affect the visibility and the impact factor of OA journals. He makes a reasonable cause that the NAA-bias will affect publications on new, peripheral, and under-discussion subjects more than established topics covered in substantive reports.

The study of Murali et al [4] published in Mayo Proceedings 2004 confirms that the availability of journals on MEDLINE as FUTON or NAA affects their impact factor.

Of the 324 journals screened by Murali et al. 38.3% were FUTON, 19.1%  NAA and 42.6% had abstracts only. The mean impact factor was 3.24 (±0.32), 1.64 (±0.30), and 0.14 (±0.45), respectively! The authors confirmed this finding by showing a difference in impact factors for journals available in both the pre and the post-Internet era (n=159).

Murali et al informally questioned many physicians and residents at multiple national and international meetings in 2003. These doctors uniformly admitted relying on FUTON articles on the Web to answer a sizable proportion of their questions. A study by Carney et al (2004) [5] showed  that 98% of the US primary care physicians used the Internet as a resource for clinical information at least once a week and mostly used FUTON articles to aid decisions about patient care or patient education and medical student or resident instruction.

Murali et al therefore conclude that failure to consider FUTON bias may not only affect a journal’s impact factor, but could also limit consideration of medical literature by ignoring relevant for-fee articles and thereby influence medical education akin to publication or language bias.

This proposed effect of the FFT limit on citation retrieval for clinical questions, was examined in a  more recent study (2008), published in J Med Libr Assoc [6].

Across all 4 questions based on a research agenda for physical therapy, the FFT limit reduced the number of citations to 11.1% of the total number of citations retrieved without the FFT limit in PubMed.

Even more important, high-quality evidence such as systematic reviews and randomized controlled trials were missed when the FFT limit was used.

For example, when searching without the FFT limit, 10 systematic reviews of RCTs were retrieved against one when the FFT limit was used. Likewise when searching without the FFT limit, 28 RCTs were retrieved and only one was retrieved when the FFT limit was used.

The proportion of missed studies (appr. 90%) is higher than in the studies mentioned above. Possibly this is because real searches have been tested and that only relevant clinical studies  have been considered.

The authors rightly conclude that consistently missing high-quality evidence when searching clinical questions is problematic because it undermines the process of Evicence Based Practice. Krieger et al finally conclude:

“Librarians can educate health care consumers, scientists, and clinicians about the effects that the FFT limit may have on their information retrieval and the ways it ultimately may affect their health care and clinical decision making.”

It is the hope of this librarian that she did a little education in this respect and clarified the point that limiting to free full text might not always be a good idea. Especially if the aim is to critically appraise a topic, to educate or to discuss current best medical practice.

References

  1. Björk, B., Welling, P., Laakso, M., Majlender, P., Hedlund, T., & Guðnason, G. (2010). Open Access to the Scientific Journal Literature: Situation 2009 PLoS ONE, 5 (6) DOI: 10.1371/journal.pone.0011273
  2. Matsubayashi, M., Kurata, K., Sakai, Y., Morioka, T., Kato, S., Mine, S., & Ueda, S. (2009). Status of open access in the biomedical field in 2005 Journal of the Medical Library Association : JMLA, 97 (1), 4-11 DOI: 10.3163/1536-5050.97.1.002
  3. WENTZ, R. (2002). Visibility of research: FUTON bias The Lancet, 360 (9341), 1256-1256 DOI: 10.1016/S0140-6736(02)11264-5
  4. Murali NS, Murali HR, Auethavekiat P, Erwin PJ, Mandrekar JN, Manek NJ, & Ghosh AK (2004). Impact of FUTON and NAA bias on visibility of research. Mayo Clinic proceedings. Mayo Clinic, 79 (8), 1001-6 PMID: 15301326
  5. Carney PA, Poor DA, Schifferdecker KE, Gephart DS, Brooks WB, & Nierenberg DW (2004). Computer use among community-based primary care physician preceptors. Academic medicine : journal of the Association of American Medical Colleges, 79 (6), 580-90 PMID: 15165980
  6. Krieger, M., Richter, R., & Austin, T. (2008). An exploratory analysis of PubMed’s free full-text limit on citation retrieval for clinical questions Journal of the Medical Library Association : JMLA, 96 (4), 351-355 DOI: 10.3163/1536-5050.96.4.010
  7. The #TwitJC Twitter Journal Club, a new Initiative on Twitter. Some Initial Thoughts. (laikaspoetnik.wordpress.com)
  8. The Second #TwitJC Twitter Journal Club (laikaspoetnik.wordpress.com)
  9. How many research papers are freely available? (blogs.nature.com)




PubMed’s Higher Sensitivity than OVID MEDLINE… & other Published Clichés.

21 08 2011

ResearchBlogging.orgIs it just me, or are biomedical papers about searching for a systematic review often of low quality or just too damn obvious? I’m seldom excited about papers dealing with optimal search strategies or peculiarities of PubMed, even though it is my specialty.
It is my impression, that many of the lower quality and/or less relevant papers are written by clinicians/researchers instead of information specialists (or at least no medical librarian as the first author).

I can’t help thinking that many of those authors just happen to see an odd feature in PubMed or encounter an unexpected  phenomenon in the process of searching for a systematic review.
They think: “Hey, that’s interesting” or “that’s odd. Lets write a paper about it.” An easy way to boost our scientific output!
What they don’t realize is that the published findings are often common knowledge to the experienced MEDLINE searchers.

Lets give two recent examples of what I think are redundant papers.

The first example is a letter under the heading “Clinical Observation” in Annals of Internal Medicine, entitled:

“Limitations of the MEDLINE Database in Constructing Meta-analyses”.[1]

As the authors rightly state “a thorough literature search is of utmost importance in constructing a meta-analysis. Since the PubMed interface from the National Library of Medicine is a cornerstone of many meta-analysis,  the authors (two MD’s) focused on the freely available PubMed” (with MEDLINE as its largest part).

The objective was:

“To assess the accuracy of MEDLINE’s “human” and “clinical trial” search limits, which are used by authors to focus literature searches on relevant articles.” (emphasis mine)

O.k…. Stop! I know enough. This paper should have be titled: “Limitation of Limits in MEDLINE”.

Limits are NOT DONE, when searching for a systematic review. For the simple reason that most limits (except language and dates) are MESH-terms.
It takes a while before the indexers have assigned a MESH to the papers and not all papers are correctly (or consistently) indexed. Thus, by using limits you will automatically miss recent, not yet, or not correctly indexed papers. Whereas it is your goal (or it should be) to find as many relevant papers as possible for your systematic review. And wouldn’t it be sad if you missed that one important RCT that was published just the other day?

On the other hand, one doesn’t want to drown in irrelevant papers. How can one reduce “noise” while minimizing the risk of loosing relevant papers?

  1. Use both MESH and textwords to “limit” you search, i.e. also search “trial” as textword, i.e. in title and abstract: trial[tiab]
  2. Use more synonyms and truncation (random*[tiab] OR  placebo[tiab])
  3. Don’t actively limit but use double negation. Thus to get rid of animal studies, don’t limit to humans (this is the same as combining with MeSH [mh]) but safely exclude animals as follows: NOT animals[mh] NOT humans[mh] (= exclude papers indexed with “animals” except when these papers are also indexed with “humans”).
  4. Use existing Methodological Filters (ready-made search strategies) designed to help focusing on study types. These filters are based on one or more of the above-mentioned principles (see earlier posts here and here).
    Simple Methodological Filters can be found at the PubMed Clinical Queries. For instance the narrow filter for Therapy not only searches for the Publication Type “Randomized controlled trial” (a limit), but also for randomized, controlled ànd  trial  as textwords.
    Usually broader (more sensitive) filters are used for systematic reviews. The Cochrane handbook proposes to use the following filter maximizing precision and sensitivity to identify randomized trials in PubMed (see http://www.cochrane-handbook.org/):
    (randomized controlled trial [pt] OR controlled clinical trial [pt] OR randomized [tiab] OR placebo [tiab] OR clinical trials as topic [mesh: noexp] OR randomly [tiab] OR trial [ti]) NOT (animals [mh] NOT humans [mh]).
    When few hits are obtained, one can either use a broader filter or no filter at all.

In other words, it is a beginner’s mistake to use limits when searching for a systematic review.
Besides that the authors publish what should be common knowledge (even our medical students learn it) they make many other (little) mistakes, their precise search is difficult to reproduce and far from complete. This is already addressed by Dutch colleagues in a comment [2].

The second paper is:

PubMed had a higher sensitivity than Ovid-MEDLINE in the search for systematic reviews [3], by Katchamart et al.

Again this paper focuses on the usefulness of PubMed to identify RCT’s for a systematic review, but it concentrates on the differences between PubMed and OVID in this respect. The paper starts with  explaining that PubMed:

provides access to bibliographic information in addition to MEDLINE, such as in-process citations (..), some OLDMEDLINE citations (….) citations that precede the date that a journal was selected for MEDLINE indexing, and some additional life science journals that submit full texts to PubMed Central and receive a qualitative review by NLM.

Given these “facts”, am I exaggerating when I am saying that the authors are pushing at an open door when their main conclusion is that PubMed retrieved more citations overall than Ovid-MEDLINE? The one (!) relevant article missed in OVID was a 2005 study published in a Japanese journal that MEDLINE started indexing in 2007. It was therefore in PubMed, but not in OVID MEDLINE.

An important aspect to keep in mind when searching OVID/MEDLINE ( I have earlier discussed here and here). But worth a paper?

Recently, after finishing an exhaustive search in OVID/MEDLINE, we noticed that we missed a RCT in PubMed, that was not yet available in OVID/MEDLINE.  I just added one sentence to the search methods:

Additionally, PubMed was searched for randomized controlled trials ahead of print, not yet included in OVID MEDLINE. 

Of course, I could have devoted a separate article to this finding. But it is so self-evident, that I don’t think it would be worth it.

The authors have expressed their findings in sensitivity (85% for Ovid-MEDLINE vs. 90% for PubMed, 5% is that ONE paper missing), precision and  number to read (comparable for OVID-MEDLINE and PubMed).

If I might venture another opinion: it looks like editors of medical and epidemiology journals quickly fall for “diagnostic parameters” on a topic that they don’t understand very well: library science.

The sensitivity/precision data found have little general value, because:

  • it concerns a single search on a single topic
  • there are few relevant papers (17- 18)
  • useful features of OVID MEDLINE that are not available in PubMed are not used. I.e. Adjacency searching could enhance the retrieval of relevant papers in OVID MEDLINE (adjacency=words searched within a specified maximal distance of each other)
  • the searches are not comparable, nor are the search field commands.

The latter is very important, if one doesn’t wish to compare apples and oranges.

Lets take a look at the first part of the search (which is in itself well structured and covers many synonyms).
First part of the search - Click to enlarge
This part of the search deals with the P: patients with rheumatoid arthritis (RA). The authors first search for relevant MeSH (set 1-5) and then for a few textwords. The MeSH are fine. The authors have chosen to use Arthritis, rheumatoid and a few narrower terms (MeSH-tree shown at the right). The authors have taken care to use the MeSH:noexp command in PubMed to prevent the automatic explosion of narrower terms in PubMed (although this is superfluous for MesH terms having no narrow terms, like Caplan syndrome etc.).

But the fields chosen for the free text search (sets 6-9) are not comparable at all.

In OVID the mp. field is used, whereas all fields or even no fields are used in PubMed.

I am not even fond of the uncontrolled use of .mp (I rather search in title and abstract, remember we already have the proper MESH-terms), but all fields is even broader than .mp.

In general a .mp. search looks in the Title, Original Title, Abstract, Subject Heading, Name of Substance, and Registry Word fields. All fields would be .af in OVID not .mp.

Searching for rheumatism in OVID using the .mp field yields 7879 hits against 31390 hits when one searches in the .af field.

Thus 4 times as much. Extra fields searched are for instance the journal and the address field. One finds all articles in the journal Arthritis & Rheumatism for instance [line 6], or papers co-authored by someone of the dept. of rheumatoid surgery [line 9]

Worse, in PubMed the “all fields” command doesn’t prevent the automatic mapping.

In PubMed, Rheumatism[All Fields] is translated as follows:

“rheumatic diseases”[MeSH Terms] OR (“rheumatic”[All Fields] AND “diseases”[All Fields]) OR “rheumatic diseases”[All Fields] OR “rheumatism”[All Fields]

Oops, Rheumatism[All Fields] is searched as the (exploded!) MeSH rheumatic diseases. Thus rheumatic diseases (not included in the MeSH-search) plus all its narrower terms! This makes the entire first part of the PubMed search obsolete (where the authors searched for non-exploded specific terms). It explains the large difference in hits with rheumatism between PubMed and OVID/MEDLINE: 11910 vs 6945.

Not only do the authors use this .mp and [all fields] command instead of the preferred [tiab] field, they also apply this broader field to the existing (optimized) Cochrane filter, that uses [tiab]. Finally they use limits!

Well anyway, I hope that I made my point that useful comparison between strategies can only be made if optimal strategies and comparable  strategies are used. Sensitivity doesn’t mean anything here.

Coming back to my original point. I do think that some conclusions of these papers are “good to know”. As a matter of fact it should be basic knowledge for those planning an exhaustive search for a systematic review. We do not need bad studies to show this.

Perhaps an expert paper (or a series) on this topic, understandable for clinicians, would be of more value.

Or the recognition that such search papers should be designed and written by librarians with ample experience in searching for systematic reviews.

NOTE:
* = truncation=search for different word endings; [tiab] = title and abstract; [ti]=title; mh=mesh; pt=publication type

Photo credit

The image is taken from the Dragonfly-blog; here the Flickr-image Brain Vocab Sketch by labguest was adapted by adding the Pubmed logo.

References

  1. Winchester DE, & Bavry AA (2010). Limitations of the MEDLINE database in constructing meta-analyses. Annals of internal medicine, 153 (5), 347-8 PMID: 20820050
  2. Leclercq E, Kramer B, & Schats W (2011). Limitations of the MEDLINE database in constructing meta-analyses. Annals of internal medicine, 154 (5) PMID: 21357916
  3. Katchamart W, Faulkner A, Feldman B, Tomlinson G, & Bombardier C (2011). PubMed had a higher sensitivity than Ovid-MEDLINE in the search for systematic reviews. Journal of clinical epidemiology, 64 (7), 805-7 PMID: 20926257
  4. Search OVID EMBASE and Get MEDLINE for Free…. without knowing it (laikaspoetnik.wordpress.com 2010/10/19/)
  5. 10 + 1 PubMed Tips for Residents (and their Instructors) (laikaspoetnik.wordpress.com 2009/06/30)
  6. Adding Methodological filters to myncbi (laikaspoetnik.wordpress.com 2009/11/26/)
  7. Search filters 1. An Introduction (laikaspoetnik.wordpress.com 2009/01/22/)




3rd Call for Submissions for “Medical Information Matters”: Tools for Searching the Biomedical Literature

8 05 2011

It takes some doing to breathe life into Medical Information Matters” (blog carnival about medical  information).
A month ago I wrote a 2nd call for submissions post for this blog carnival. Unfortunately the next host, Martin Fenner, didn’t have time to finish a blog post and has come up with a new (interesting) variation on the theme “A Wish list for better medical information”.

Martin asks you to philosophize, blog and/or comment about “Tools for Searching the Biomedical Literature.

You can base your contribution on a recent (editable) survey of 28 different PubMed derivative tools by Zhiyong Lu (NCBI) [1].

Thus, write your thoughts on the various PubMed derivative tools mentioned here or write about your own favorite 3rd party PubMed tool (included or not).

For details, see Martin’s blog post announcing this upcoming edition. The Blog Carnival FAQs are here.

And if you don’t have time to write about this topic, you may still find the survey useful, as well as the views of others on this topic. So check out Martin’s blog Gobbledygook once in a while to see if the blog edition has been posted.

Note [1]: If you have already submitted a post to the carnival, or would like to write about another theme, we will take care that your post (if relevant)  will be included in this or the next edition. You can always submit here.

Note [2]: Would you like to host “Medical Information Matters” at your blog? Please comment here or write to: laika dot spoetnik at gmail dot com. We need hosts for June, July, August and September (submission deadline first Saturday of every month, posting on the next Tuesday)

  1. Lu Z. PubMed and beyond: a survey of web tools for searching biomedical literature. Database. 2011 Jan;2011. doi: http://dx.doi.org/10.1093/database/baq036







Follow

Get every new post delivered to your Inbox.

Join 610 other followers