The internet has been increasingly relevant as a popular source of information about science (Brossard and Scheufele, 2013). Every day, regular citizens go online to obtain information from immediate issues, such as food safety (Ma et al., 2017) to big picture ones, such as climate change (Fletcher, 2016); from new technologies (Anderson et al., 2010) to health issues (Fox and Duggan, 2013). However, not all information is trustworthy: misinformation of all kinds abound in the cyberspace (Wardle and Derakhshan, 2017). Misinformation associated with science includes conspiracy theories, anti-science propaganda, rumors and straight-up fabricated news about science and scientists (Scheufele and Krause, 2019). In this context, social platforms have a paradoxical role: they both allow for engaging public science communication and are also hotspots for the spread of misinformation. Recently, YouTube has also been flagged as a hotbed of misinformation: until not long ago, videos with conspiratory, racist and pornographic content were being monetized (Mostrous, 2017), and in 2019 investigations reported that YouTube's recommendation mechanism, responsible for 70% of all the watch time the website receives (Solsman, 2018), tended to exhibit videos that were increasingly more right-wing, conspiratory and radicalized in tone (Roose, 2018).

YouTube is a particularly relevant platform because of its enormous reach: it is the second most accessed website worldwide (Alexa, 2020), where 2 billion registered users watch videos monthly (Cooper, 2019). Nonetheless, research on public science communication (and on misinformation about science) on YouTube is still in its infancy (Allgaier, 2019). Some studies have tried to compose a typology of science videos on the platform, focusing, for example, on editing and narrative techniques (Morcillo et al., 2016) and on the difference between online videos produced either television or for the internet (De Lara et al., 2017). Other relevant research themes are the accuracy of the scientific content of the video and its relationship with video engagement and popularity (Keelan et al., 2007; Garg et al., 2015) and what answers people get when they query about politically charged keywords on science (Allgaier, 2016; Shapiro and Park, 2018). However, with the exception of Welbourne and Grant's (2015) work, it seems there are so far no other studies detailing which and how video features and video metrics affect and correlate to the popularity of science videos on the platform. Welbourne and Grant (2015) were mostly worried about pointing out differences in popularity measures (views, comments, subscriptions, number of shares and total rating) and content factors (gender and number of presenters, video pacing and length) between user-generated and professionally-generated science videos.

In this study, which we see as complementary to theirs and also inspired by it, we try to fill this gap in the literature by focusing on another set of video features and platform metrics. In other words, we analyzed to what extent a selection of factors (video theme, video format, number of editing features, video length, number of likes, number of comments, video age, channel productivity and channel) are correlated to the popularity of UGC (User-generated content) science videos on YouTube. To accomplish that, we sampled 441 videos from the ScienceVlogs Brasil project – a group of Brazilian science communication channels on YouTube –, collected data on each video or channel regarding nine factors (see Design and Methods) and performed descriptive and inferential analyses on this dataset, in order to verify the correlation of each factor with the dependent variable “popularity,” measured as the number of views per days since publication (views/day). Additionally, we strive to find which of those factors were causally related to video popularity.

In this article, we first introduce readers to the ScienceVlogs Brasil project; then, we briskly review the literature on the popularity of videos on the platform, focusing on user-generated videos. In “Design and Methods” we give more detail about the data collection, the sampling method, and the variables of interest. In the “Results” section we describe the most important findings and outline the construction of the regression model, and in the following section “Discussion” we interpret the main findings of this work, highlighting key conclusions. Finally, we point to some flaws of this research and promising paths forward in the section “Limitations and directions for future research.”

All the science videos were sampled from the ScienceVlogs Brasil project – an alliance of independent YouTube channels committed to communicating science to the general public, founded in May of 2016. The project, launched with 18 channels focusing mostly on Biological and Exact Sciences, quickly evolved to presently incorporate around 60 channels with a broad variety of themes. The videos were all produced in Portuguese, Brazil's official language. For a list of the channels that participated in this research and other project-relevant information, see Supplementary Material.

At the time of data collection (May to July of 2018), 36 channels belonged to the ScienceVlogs Brasil project. We decided to exclude channels that produced videos that did not correspond to our definition of science communication: the communication of science-related topics to non-specialized audiences using in a simple and non-academic language. We rejected three channels: one that was focused on explaining math exercises to students; another identified as an entertainment channel, focused on recording situations using slow-motion effects, and another that was the channel of the ScienceVlogs group, whose communicators used to send messages to the audiences of the project. We considered the remaining 33 channels for our analysis. All of them are user-generated, except one (channel Zoa), which posts both content produced informally by the host, but also snippets of footage of a tv show about science, presented by the same host for a local tv channel in the Northeast of Brazil. We chose to keep videos from Zoa in our sample because they are presented in a very colloquial and relaxed manner, and the editing is not sophisticated, which makes this channel not unlike the other user-generated channels. We selected an average of 10 videos from each channel, which is proportional to the number of original videos in each channel (i.e., we considered a stratified random sampling with proportional allocation. Our final sample could not exceed a certain number of videos since the analysis would be performed manually by only one researcher and within a restricted deadline. The datasets generated by the aforementioned process can be obtained upon request to the corresponding author.

To perform the video sampling, we assigned a number for each of them the oldest to the newest, and we applied the function random. Sample from the Python software (v.3). We selected the video corresponding to the number given by the function; then we recounted the videos excluding the selected videos and performed the function again. If the selected video was not directly related to science communication – e.g., if it was a social or political commentary without a research background, or if it was a video about the presenter's personal life, or if the video was not authored by the presenter, we excluded it from the counting process and begin again. Using this method, we sampled 441 videos from the 33 channels. For budgetary reasons and time constraints, only one researcher was responsible for coding and reviewing the data of all videos.

We manually collected data from each video according to the variables:

As discussed in the literature review, all of these factors could potentially affect video popularity on YouTube. Many other such factors (number of keywords, initial number of views, number of channel subscribers, thumbnail attractiveness) could also be important; however, time constraints and practicality guided our decision for this selection.

These nine variables served as covariates to the dependent variable “popularity of the video,” measured in the number of views of the video divided by the number of days since it was posted (views/day). We chose this index for popularity since the number of views alone can be highly influenced by the video publishing date (older videos have time to accumulate views), and we wished to minimize this effect. In this work, we considered popularity as a function of the video alone, and thus we did not consider the number of people subscribed to the video's channel, for example (it also would not be possible for us to track the number of channel subscriptions at the time of the video release).

The descriptive and inferential analyses were performed using the software R and went from September 2018 to February 2019. We interpreted the strength of the correlation index according to the parameters stated in Mukaka (2012). We performed a logarithmic transformation in the response variable in all analyses to obtain a normal and homoscedastic model. We then built a multiple linear regression model for the relationship between ln(Yi) and the dependable variables:

in which Yi is the number of views in the i-th video. We established the variable β0, which is the expected value of the natural logarithm for a “benchmark video” (β0) – a video in vlog format, with zero likes and comments, minimal size, minimal channel productivity and minimal video age and length, zero editing features and Biological Sciences as the theme. This is a base value for the model, to which the effects of all the other variables will be added. This is a base value for the model, to which the effects of all the other variables will be added. The coefficient β1 represents the increment (positive or negative) on the expected value of the natural log of the dependent variable for the increase in one unit in the number of likes; β2 represents the increment (positive or negative) on the expected value of the natural log of the number of views for the increase in one unit in the productivity variable; β3 represents the increment (positive or negative) on the expected value of the natural log of the number of views for the increase in one unit in the age variable; β4 represents the increment (positive or negative) on the expected value of the natural log of the number of views for the increase in one unit in the length variable. The coefficients λl, l = 1.0.0.10 represent the increments (positives or negatives), associated to the number of editing features of the video, on the expected value of the natural log of the number of views. The coefficients αj, j = 1.7 represent the increments (positives or negatives), according to the video format associated, on the expected value of the natural logarithm of the number of views with vlog format. The coefficients δk, k = 1.7 represent the increments (positives or negatives), according to the video theme associated, on the expected value of the natural logarithm of the number of views with Biological Sciences theme. Finally, we assumed that ϵi ~ Normal (0, σ2) are mutually interdependent errors.

The variable comments was not part of this model to avoid collinearity issues with the variable likes (ρ = 0.70). After the model was fitted, a residual analysis, to check the goodness of the model fit, was performed. We also did not add the variable channel to the inferential analysis, since it did not match the type of multiple linear regression model that we regarded as best for this context. Subcategories within each factor that did not have a big enough sample size (N ≥ 10 videos) were excluded from the descriptive analysis.

When examined individually, some factors seem to have more influence on video popularity than others. Only three correlation indexes were somewhat strong in the descriptive analysis: number of comments (0.363), number of likes (0.604), and channel (0.746). The inferential analysis showed that only the number of likes, channel productivity, video age and some video formats had predictive effect over video popularity. This difference of relevant factors happens because, in the descriptive analyses, the relationship of each factor with video popularity is analyzed individually and undisturbed by other factors; in the inferential analysis, however, the effects of different variables can potentialize or mask the effect of other factors, and so the effects change. In this section, we briefly discuss the most important results of the study, starting by the ones that could predict video popularity.

Likes and comments showed a strong correlation with popularity. This makes sense, since engagement rates, partially composed by engagement metrics as likes and comments, are directly used by YouTube's searching and ranking algorithms (Covington et al., 2016). This likes-views dependence generates the rich-get-richer effect, that boosts videos with an initial high number of likes. Next in order is the strong correlation of popularity and channel productivity. Channel productivity may be important for beginner channels to reach the 4,000 h of watch time and become profitable; and for all channels to increase watch time stats and become more relevant on the platform. YouTube's official blog recommends that users only post quality content as a way of increasing views and watch time (Woicicki, 2019), but it seems logical that the bigger the amount of videos in a channel the audience can choose from, the bigger the channel's chances to obtain watch time. There is also the novelty factor: novel videos are more guaranteed to obtain attention than older videos (even more so if they are also topical), and channels that are constantly generating novel and topical content can become a recognized source of information or entertainment, a go-to source when the viewer wants information of some sort.

Regarding video age, the third predictive variable, it seems that newer science videos received slightly more attention – were more popular – than old videos. For this, we have two possible explanations. Firstly, some older channels favored unpopular formats and stopped producing videos long ago. Jornal Ciensacional, for example, started producing interview and science news videos in 2012 and posted at an uneven pace until 2017, when the last video was released. Quer que desenhe produced animations about science for 2 years and stopped all production in 2015. Universo Racionalista, launched in 2015, posted hangouts irregularly (one gap in posting lasted about a year) and produced very few videos. Older videos from these channels were not attractive because of the format, and these channels' low productivity probably hindered YouTube's algorithms from recommending them. Another possible explanation is that the newest videos of our sample were produced by long-running channels that posted regularly and had time to build an audience and experience with video popularity to obtain more views, such as Space Today (launched in 2015), Xadrez Verbal (launched in 2013), Ciência Todo Dia (launched in 2012) and Papo de Biólogo (launched in 2014).

We found some video formats to be significant to the general popularity. Videos featuring interviews and hangouts were shown as tending to be to be less popular than vlogs, while animations tended to be more popular. The most often observed formats in this work – vlogs, short documentaries and animations – were also identified as dominant in other works, such as Morcillo et al. (2016) and Ervitti (2018). The vlog is considered a YouTube-native genre, and is by far one of the most used formats in user-generated videos, requiring (but not restricted to) very little editing.

We could also identify trends about factors that did not seem to affect video popularity. For example, it is not difficult to imagine why interdisciplinary videos exhibited more views/day: they have a broader audience than specialized videos; besides, they are also topical, frequently touching on political themes and current affairs. Interdisciplinary videos also presented the biggest variation in popularity. This probably relates to the fact that almost all channels have produced interdisciplinary videos, with more or less success. This means that the variations in popularity do not depend on a specific group of channels that always produces such videos, but on some videos of all channels. We also noted that videos on Earth and Exact Sciences were very popular (second place in general popularity, right after interdisciplinary videos), although they were the fourth most observed videos after the categories Interdisciplinary, Biological Sciences and Humanities. This could reflect a popular preference, but also the fact that most of such videos come from a small group of broadly successful channels that are good either at producing most watched videos (Ciência Todo Dia), or that have a high productivity rate (Space Today). Coincidentally, these channels produce either a good amount or all or all of their videos about Astronomy and space exploration. Health Sciences and Engineering were themes observed in <15 videos each. Various channels did produce videos on Health Sciences, but Engineering videos came mostly from the channel Peixe Babel, hosted by two women. The little popularity of Engineering videos could be an effect of the small number of videos in the sample, but also to the fact that female science video hosts generally receive less views on YouTube, a phenomenon already referenced elsewhere (Thelwall and Mas-Bleda, 2018).

The four channels that concentrate most of the views – Ciência Todo Dia (CT), Canal do Pirulla (CP), Papo de Biólogo (PB) and Minuto da Terra (MT) – are good representatives of science channel with well-defined characteristics, that concentrate videos with patterns of theme, length, format, and productivity that were shown to be correlated to popularity. Ciência Todo Dia, for example, is a channel mostly about Astronomy, concepts in physics and space exploration. Almost all videos produced by CT are vlogs and are 7′38″ in length, and video productivity between 1.39 and 1.46 videos per month. Papo de Biólogo, on the other hand, is a channel that produces between 2.08 and 2.38 videos per month and is dedicated to presenting the habits and anatomy of wild and exotic animals in vlogs and small documentaries that are 6′41″ in length. Minute Earth is an interdisciplinary channel by design, producing an average of 2.83 videos per month that answer a variety of questions from a scientific standpoint using animation. The videos are 2′4″ long on average. Canal do Pirulla's host produces vlogs that are a mix of pure Biology videos and (a majority of) well-researched content about current affairs. His videos are fairly long, averaging 22′32″, and he produces between 4.4 and 4.6 videos per month. Taken together, all of these channels produce videos in mostly in successful formats (vlogs, animations and short documentaries) about the most popular themes in the length gap where most views/day are located. The variable productivity showed a big variation here, but it must be mentioned that it was made to reflect the average productivity of the entire productive life of the channel, not necessarily indicating how productive the channel was, say, in the last 3 or 6 months.

On the other hand, channels with the least popular videos (Boteco Behaviorista, Jornal Ciensacional and Canal Zoa) concentrated features that attracted fewer viewers. Boteco Behaviorista is a channel that features hangouts (online conversations) between several psychology researchers and guests both about current events and within the psychology field. The videos are 79′36″ long in average, and the channel produces an average of 1 video per month. Jornal Ciensacional was mostly dedicated to reporting news about science in interviews and short documentaries in videos about 7′02″ long. It has stopped producing videos in 2017, but until then it had produced <1 video per month. Zoa is a special case: as mentioned before, it reposted video footage of a TV show about science, presented by the channel's host, while also producing homemade videos. The videos were 3′04″ on average, and the channel produced around 1.35 videos per month. These channels produced mostly videos on the formats interviews, hangouts and late night TV show (channel Zoa), had low overall productivity, with long gaps in production, and concentrated on producing news. Videos on news become old very quickly, and if the channel does not keep a high productivity rate, they do not fare well. Lastly, video length may also have been an obstacle for popularity: hangouts, if not edited enough, tend to be tiresome to watch, as is the case of BB's videos. The same can be said about medium-sized little edited interviews (as in JS).

This study contains some limitations. Among them are:

Although the scholarship on online science videos has grown in the last few years, many questions regarding video popularity on YouTube are still unanswered (and some, that touch on the functioning of algorithms, will probably remain so). For example, it is not yet entirely clear how the interplay between elements such as number of views, likes and comments functions on YouTube. To which measure is the number of views also causing the numbers of likes and comments to rise? Some other interesting research topics regard the audiences: who are the people who consume science videos? What type of science videos each profile prefers to watch? Also, how much do science videos contribute to different types of science education and to the change of attitudes on scientific issues? These and many other questions will occupy researchers' minds in the years to come.

The raw data can be obtained upon request to the corresponding author.

RV reviewed and selected the factors according to the literature, collected and cleaned the data, and wrote the manuscript and the references. AM performed the inferential and descriptive analysis in consultation with CA, produced all the figures and tables, and reviewed the manuscript. CA offered expertise in the sampling, analysing processes, and also reviewed the manuscript. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.