I get asked this question a lot at the beginning stages of a review project. It’s a fair question: researchers want to know what their screening workload is going to look like, and screening abstracts is a tedious process.
There’s some great research happening right now on how to make screening less tedious – using text mining to automate or semi-automate the process, for example. These are promising approaches, but for now, most reviews need a human eye to oversee study selection for at least part of the process.
I used to get flustered when I was asked this question because I was afraid of giving the wrong answer and underestimating the amount of work a researcher would later have to do. On the other hand, if I overestimated the number of citations to screen, a researcher might want to change the search strategy to lower the number of citations or otherwise change the methodology.
I also didn’t have a very good answer. It’s hard to estimate the number of citations to be screened without downloading them from all sources and de-duplicating first. I’ve sometimes estimated the total number of included studies at the end of the project by screening a random sample (such as 100 citations) and calculating the ratio of the sample to the total number of citations to be screened (e.g. if there is one relevant article in a random sample of 100, and 1000 articles total to screen, I would estimate about 10 studies to be included at the end of the project – for more on getting a truly random sample of citations, see my previous blog post on this topic).
Tired of giving bad answers to this question, I’ve crunched the numbers for a few of the review projects I’ve worked on in the last year. For each of the projects, I found the number of citations downloaded from the Ovid MEDLINE search only, then the number of citations left to screen at the title/abstract stage after citations were downloaded from all databases and duplicates removed.
The results are below:
The ratio of MEDLINE search only to total citations to screen ranged from 239% to 1333%. However, the last two columns represent projects that had less of a biomedical focus (social sciences and computer science, respectively). The MEDLINE searches were still relevant in these cases, but we didn’t expect the majority of our studies to come from MEDLINE. Thus, if we exclude the major outlier, with more of a social sciences focus, the results are a little more consistent.
For each of the projects above, I searched 6-7 databases, except project 1, where I searched 14(?!). However, the ratio of citations for project 1 is not exceptionally different than that of the other projects. For now, I can’t see a discernible difference between screening burden and number of databases searched, but possibly more data is needed.
My overall take-away from this exercise is that, for the searches that I run, the screening burden of a systematic review tends to be about 2.5x to 5x that of the original MEDLINE search. In the future, this is the advice that I’ll be giving my researchers to help them better plan their resources and time. I can breathe a small sigh of relief, too, knowing that the information that I give my researchers is just a little more evidence-based than it was the day before.