Introduction  

The Cochrane Collaboration is a network of researchers, healthcare professionals, patients, carers and people interested in healthcare who largely work for free with altruistic motives to produce systematic reviews of care.1 Cochrane systematic reviews are available in the Cochrane Library (http://www.cochranelibrary.com/) and can be freely accessed in 101 countries, available to residents of a further 14 countries via subscription paid by respective governments or other organisations. It is also available for purchase at a cost or as open access if authors have paid the article processing charge.2 Each full review has a summary and abstract web page accessible to anyone with internet access3 (http://www.cochrane.org/). Cochrane Schizophrenia Group (CSzG) sits within this wider Collaboration.

Twitter is a popular4 free to use microblogging social media platform, which, at the time of the trial, allowed users to send a 140 character message called a tweet to lists of followers. A tweet may contain a hashtag (#), used as a means of searching by topic, and/or a handle (@), which denotes another Twitter user account. Although Twitter (blocked in China since 2009), and some other platforms are not available in China, the usage of social media is exceptionally high with 91% of Chinese residents using some form of social media compared with about 67% of those in the USA.5 The Twitter-like systems of Weibo and Wechat are popular in China.6 7

In 2011, CSzG began to use Twitter. @CochraneSzGroup following remains small but at the time of the study was over 1300. Our randomised trial (tweeting links vs not tweeting links of similar paired reviews8 showed that social media dissemination of reviews nearly tripled the unique page visits to each review and nearly increased threefold the time people spent visiting the Cochrane summaries page when they followed the link via a tweet.9 We subsequently identified a variety of case reports, surveys and case series’ conclusions on when to microblog (table 1). Each study stresses the importance of appreciating weekly behaviour patterns when tweeting and how to increase traffic as more tweets mean more visibility. In contrast to other social media such as Facebook, Twitter activity dies down a bit during weekends.10 For studies that compare 1 day of the week against another, conclusions are conflicting as are comparisons between weekday and weekend tweeting. Four support microblogging (tweeting) at the weekend11–14 and three during the ‘working’ week.15–17 One large survey of 4.8 million tweets reported that early morning tweets received most clicks while late evening tweets had more retweets.18 There is peak tweeting activity between the hours of 11:00 and 13:00. The best days and times to microblog in terms of site traffic are likely to differ according to the target audience. For example, tweets for shoppers may be best at the end of the week when most are shopping, and tweets of work-related information may be better during the working week.

Given the above findings, we considered examined our own @CochraneSZGroup account by analysing data from Twitter Analytics19 for the period 1 June 2015 to 25 October 2015 (figure 1). We found the @CochraneSZGroup engagement rate to be higher during the working week than at the weekend, with the greatest activity on Thursday and Friday.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Average engagement* rate (1 June 2015 to 25 October 2015). *Engagement rate is a broad measure of Twitter activity encompassing CTR and Twitter interactions.

Methods and analysis

Outcomes

The primary outcome was the total number of visits to the relevant Cochrane summary page in the 7 days following the intervention (including the day of tweeting) as reported on Google Analytics. For the primary outcome, repeated views of the same page during a single user session are counted in the total (unique page views are a secondary outcome). The standard free account in Google Analytics produces various reports in real time and data about whom, when and where someone has visited a site as well as how they ‘arrived’ there.9 In this way, we monitored all traffic to the Cochrane summary page and traffic directly from Twitter. The average half-life of a tweet (with a web link), defined as ‘the amount of time at which this link will receive half of the clicks it will ever receive after it’s reached its peak’, has been estimated as 2.8 hours.28 However, to capture any possible cascade effect of tweeting, we extended the monitoring period to 7 days.29 The secondary outcomes are listed in figure 2 and provide other measures of incoming activity and exiting behaviour. For a glossary of Google Analytics terms, please see ref 20. After the final follow-up period, JSF downloaded data from Google Analytics using Supermetrics (business analytics software)30 to produce a spreadsheet for analysis.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Flow diagram.

Statistical power

The sample size for this study is fixed by the number of published Cochrane reviews within the jurisdiction of the CSzG (n=194). As a multiarm trial (five arms) with no natural control arm for comparison and a log-normal continuous primary outcome, analysis of the primary outcome is by analysis of covariance where the effect size is the ratio of between-arm and residual variances. The detectable (standardised) effect size is calculated using standard sample size software after specifying power, alpha, number of arms and number of observations per arm. This represents the ratio of between-arm and residual variances. In a one-way analysis of variance study, sample sizes of 39, 39, 39, 39 and 39 are obtained from the 5 groups whose means are to be compared. The total sample of 194 achieves 90% power to detect differences among the means versus the alternative of equal means using an F test with a 0.05 significance level. The detectable standardised effect size assumes a common SD within a group of 1.00. Therefore, the size of the variation in the means is represented by their SD which is 0.28. After taking logs of baseline number of visits for each review, the actual common SD within a group is 0.93. With a standardised detectable effect size of 0.28% and 90% power, this equates to a between-group SD of approximately 0.26 or variance (=SD²) of approximately 0.07.

Data analysis

We compared baseline access activity of the reviews in each of the five trial arms using descriptive statistics. We analysed all outcomes as continuous, with log transformation for outcomes that were positively skewed. We present descriptive statistics, and measures of effect from linear regression models using Monday as a nominal reference group. For the primary outcome and most secondary outcomes, this estimate of effect is the ratio of geometric means and 95% CIs, adjusted for baseline activity. However, as a multiarm study with no pairwise comparisons of particular interest, we derived a single p value for each outcome based on a Wald test that the model coefficients for treatment arm are jointly equal to zero. All data were analysed using Stata V.15 statistical software.

As the allocated interventions were implemented fully as planned, and there were no missing primary outcome data, we did not conduct any sensitivity analyses.

Results

Following randomisation, 194 reviews were tweeted. Table 2 outlines the baseline review activity for the year before the start of the trial. The minimum and maximum correspond to the least and most viewed reviews by allocated day of the week, respectively. There were no differences between the groups.

For this randomised trial, primary and secondary findings were consistent. Tweeting on any day between Monday and Friday was just as good as any other weekday and no clear advantage over another in terms of ratings of activity as recorded in Google Analytics (table 3). Randomised evidence from our previous trial31 suggested an approximately threefold increase in activity in the week after disseminating microblogs with hyperlinks. This study provides before-and-after evidence that this level of increase in activity continues to be realistic.

Discussion

Results from this study demonstrate that there was no statistically significant difference between tweeting on Monday or any day of the week for our primary outcome of ‘all page views’. Equally, there were no clear differences found for any of the secondary outcomes. There are many possibilities why this was so. It is possible that a real effect has failed to be detected in this necessary small trial or the form of the microblogs may not have encouraged differential responses. Perhaps, however, for this type of tweet and Twitter following there is no substantial advantage to microblogging on one working day compared with another.

There was, however, an increase in viewing activity compared with the months prior to the study. This fits with the finding of our previous randomised trial.31 In that 2013–2014 study, the control group had an arithmetic mean of 5.28 ‘all page views’ and this increased in the tweet arm to 13.67. After that trial ended, there was a general increase in baseline activity as evidenced by the data we collected for the year prior to this trial (table 2). Despite this increased baseline activity, tweeting in this study seemed to increase activity by a further threefold. There is a suggestion that interest is increasing and that this has not plateaued. The other comparable variable between the two trials is ‘time spent on page’. In this study, time on page was considerably greater than in our first trial (geometric mean 123 vs 76 s).

Much investment is made to disseminate evidence of the effects of healthcare but the strategies by which this is done are rarely investigated. Increasingly, more funding bodies are prioritising dissemination strategies as part of evaluating successful grant applications and decision-making tools are available to support this.32 33 Medical evidence of enormous impact will disseminate swiftly by every means possible with very little effort beyond the point of disclosure of findings. Maximising impact of most other health evidence does take more effort beyond the publication of the final report. Many choices are available to individuals, institutions and industry and resources limit what can be done but social media is ubiquitous, inexpensive and—as this study continues to suggest—powerful. The first randomised trial in our series illustrated that using a microblog causes nearly a tripling of traffic to a health evidence web page.31 This study supports that finding with the non-randomised before and after hit rates but adds randomised evidence that which day of the week dissemination takes place on does not matter that much. This higher quality evidence does not support the impression we gained form initial routine data analysis (table 1, figure 1) and, again, highlights the value of gaining evidence from randomised trials if possible before embarking on policy based solely on routine data. Perhaps with greater numbers and more power, we would have shown a real difference favouring Monday over the other days as this seemed the only day to have a little more activity than any of the others. This, however, is conjecture and any difference, for this set of evidence, for this following, would be subtle. Our trial tweeted between the hours of 10: 30 and 15: 00 hours local time and this could be a limitation in the local context where it is possible that more professional social media users have sight of it. However, given the global reach of tweets and access across time zones, we believe the impact of this would be minimal. Perhaps further studies could explore this aspect.

There remain many more questions on which type of social media delivers best impact on health behaviour, and on what content, timing and targeting maximally encourages best evidence to be considered and then implemented. Few techniques or tools have been demonstrated to swiftly influence behaviour of those seeking information on health. Social media are accessible tools and it is possible to evaluate how we use them and then use them in a way that does not waste effort or interest.

Conclusions

Use of social media in an area of healthcare that is not particularly high profile, for dissemination of evidence that is highly specialised to a small following continues to have a genuine effect on increasing indicators of interest—but we have no evidence that use of social media on any particular working day of the week matters. This study continues to support the initial finding from the first randomised trial where the use of social media (Twitter/Weibo) almost immediately nearly tripled activity on the reviews’ summary page. For those interested in wider dissemination of evidence, tweet any working day, but do tweet.