As your quest for relevant data continues, we’ll discuss how creating and maintaining a data inventory can be an effort that saves you heartache and greatly accelerates many of the technical facets of your project. Similarly, we’ll provide you some tools and best practices for making sure that your meetings are impactful, necessary, and accessible to all participants so that meeting with the group doesn’t cost you any momentum!
Research in management, organizational behavior, and psychology has long focused on the performance of teams–often in military, healthcare and industrial contexts. While many aspects of this work are also relevant to scientific teams, there are some key differences having to do with differences in context, leadership, and incentives. In the early 2000’s–as collaboration in science increased–the need for empirical research into the workings of science teams became apparent. The field of “science of team science” or SiTS was launched in 2006 with a conference at the National Institutes of Health (Stokols et al. 2008).
National Academies studies in 2015 (Enhancing the Effectiveness of Team Science, NRC) and 2025 (Research and Application in Team Science, NASEM) assembled the existing evidence base into how team composition, coordination, support, and organizational context could improve outcomes for science teams. Our goal here is not to review the whole field, but to provide a framework for thinking about the team functioning and process and to identify some key team science practices that are supported by both research and practical experience.
Creating a team is not just a matter of putting a bunch of people in a room together. Social scientists have identified consistent patterns in the evolution of teams (Tuckman 1965, Tuckman and Jenson 1977). Knowing that this is a process nearly every team experiences may make it (at least somewhat) more comfortable.
Teams that are assembled from across organizations must agree to adopt a common set of norms and processes in order to progress from storming to performing. This can feel like a detour from the science, but a modest early investment in developing shared practices pays off in the long run.
There is good evidence that collaborative teams produce research that is more novel and has higher impact than work produced by individuals or smaller more homogeneous groups (Lee at al. 2015, Hong and Page 2024, Yang et al. 2022). Woolley et al (2010) found evidence for a “collective intelligence” in teams, which is not strongly correlated with the average or maximum individual intelligence of group members but is correlated with the average social sensitivity of group members, the equality in distribution of conversational turn-taking, and the proportion of females in the group.
Similarly, in a study of 6.6 million medical research papers, Yang et al. found that mixed gender teams consistently produced more novel and more impactful products. In another bibliographic analysis Abbasi and Jaafari (2013) found that inter-institute and inter-university collaborations resulted in higher-impact publications. Interestingly, the result was much weaker for international collaborations.
It seems reasonable to expect that the effects of cultural and economic diversity on teams would be similar to that of gender diversity, but those factors remain harder to parse at this scale. In any case, the bump in creativity or publishing impact is only a happy side effect of assembling a diverse team. 
The real reason to do so is that it allows us to tackle bigger questions, makes our findings more relevant, our science more fun, and our world more fair. What it does not do (at least in our experience) is make the process faster!
The paradox of team science is that the very factors that slow progress may be exactly the factors that generate new insight – Milliken and Martins’ (1996) double-edged sword. The pressing question becomes not: “Does diversity improve team performance?” but rather: “How and when does diversity improve team performance?”
The categorization-elaboration model (CEM, van Knippenberg et al. 2004) proposed that information elaboration—-that is, the exchange, discussion, and integration of task-relevant information and perspectives, was responsible for many of the benefits attributed to diverse groups. But later researchers found there were a few necessary conditions for cognitive elaboration to take place and for groups to reap the benefits. Only when team members brought a learning goal orientation to their work and when they remained open to revising their original ideas (Nederveen Pieterse 2013) did diversity improve team performance.
We are all familiar with the “we’ve always done it this way” effect that can happen when a group of people have been working together for a while. By introducing people from new fields, laboratories, or cultures, that complacent thinking is disrupted. Often, the very act of justifying why we do something the way we do can invite a rethinking and improvement.
Metacognition, or “thinking about thinking” requires individuals to reflect and articulate their process for achieving new knowledge. What information goes in? Is information missing? How should it be analyzed and interpreted? Are those conclusions justified?
In the setting of a group from different backgrounds, that process happens out loud, with multiple opportunities to challenge, re-think, and improve.
A science team may include members from different research disciplines, sectors, geographies or cultures. Along each of those axes, team members will have different personal networks and be more (or less) familiar with different literatures, models, communities, tools, and solutions. Collectively, the group has a much broader range of information to draw on…but only if group members feel empowered to contribute.
“Many hands make light work” the saying goes. Think of a meta-analysis where 10 group members can each read 30 papers instead of 1 individual reading 300 papers. Dividing the workload can speed up the process, but only if there is an efficient way to manage dividing the work and then bringing the results back together again. Conversely, relying on a few skilled coders can be much more efficient than each individual writing their own code, but unless the group has a mechanism for getting broad input on key decisions, they will lose the value created by bringing together a larger group.
As shown in the graphic below, visualization can be valuable throughout the lifecycle of a synthesis project, albeit in different ways at different phases of a project.
Exploratory data visualization is an important part of any scientific project. Before launching into analysis it is valuable to make some simple plots to scan the contents. These plots may reveal any number of issues, such as typos, sensor calibration problems or differences in the protocol over time.
These “fitness for use” visualizations are even more critical for synthesis projects. In synthesis, we are often re-purposing publicly available datasets to answer questions that differ from the original motivations for data collection. As a result, the metadata included with a published dataset may be insufficient to assess whether the data are useful for your group’s question. Datasets may not have been carefully quality-controlled prior to publication and could include any number of ‘warts’ that can complicate analyses or bias results. Some of these idiosyncrasies you may be able to anticipate in advance (e.g. spelling errors in taxonomy) and we encourage you to explicitly test for those and rectify them during the data harmonization process. Others may come as a surprise.
During the early stages of a synthesis project, you will want to gain skill to quickly scan through large volumes of data. The figures you make will typically be for internal use only, and therefore have low emphasis on aesthetics.
Specific applications of exploratory data visualization include identifying:
This may include temporal, spatial, or taxonomic coverage, to name a few.
For example, the metadata might indicate a dataset covers the period 2010-2020. That could mean one data point in 2010 and one in 2020! This may not be useful for a time-series analysis.
Do the units “make sense” with the figure? Typos in metadata do occur, so if you find yourself with elephants weighing only a few grams, it may be necessary to reach out to the dataset contact.
Consider the following:
In addition to those other uses of exploratory visualization, you may also find:
Identifying all of these pieces of information is an important precursor to the data harmonization stage, where you will process the datasets you have selected into an analysis-ready format. Visualization is a–relatively–easy way of doing these checks!
It may sound facile, but one of the best things you can do to make your life easier when you ‘actually’ start making graphs is to draw your hypothetical graph. You will likely find that creating a small sketch of your ideal figure will help you make some critical decisions. This can save you time so you don’t laboriously code a graph that–once created–doesn’t meet your expectations. You may still change your mind once you’ve seen the graph your code produces (and that is okay!) but you’ll likely make some of the necessary larger decisions with pen and paper and streamline your process in generating a visualization that perfectly meets your needs.
ggplot2You may already be familiar with the ggplot2 package in R but if you are not, it is a popular graphing library based on The Grammar of Graphics. Every ggplot is composed of four elements:
ggplot function callNote that the theme component may be implicit in some graphs because there is a suite of default theme elements that applies unless otherwise specified.
This module will use example data to demonstrate these tools but as we work through these topics you should feel free to substitute a dataset of your choosing! If you don’t have one in mind, you can use the example dataset shown in the code chunks throughout this module. This dataset comes from the lterdatasampler R package and the data are about fiddler crabs (Minuca pugnax) at the Plum Island Ecosystems (PIE) LTER site.
# Load needed libraries
library(tidyverse); library(lterdatasampler)
# Load the fiddler crab dataset
data(pie_crab)
# Check its structure
str(pie_crab)tibble [392 × 9] (S3: tbl_df/tbl/data.frame)
$ date : Date[1:392], format: "2016-07-24" "2016-07-24" ...
$ latitude : num [1:392] 30 30 30 30 30 30 30 30 30 30 ...
$ site : chr [1:392] "GTM" "GTM" "GTM" "GTM" ...
$ size : num [1:392] 12.4 14.2 14.5 12.9 12.4 ...
$ air_temp : num [1:392] 21.8 21.8 21.8 21.8 21.8 ...
$ air_temp_sd : num [1:392] 6.39 6.39 6.39 6.39 6.39 ...
$ water_temp : num [1:392] 24.5 24.5 24.5 24.5 24.5 ...
$ water_temp_sd: num [1:392] 6.12 6.12 6.12 6.12 6.12 ...
$ name : chr [1:392] "Guana Tolomoto Matanzas NERR" "Guana Tolomoto Matanzas NERR" "Guana Tolomoto Matanzas NERR" "Guana Tolomoto Matanzas NERR" ...With this dataset in hand, let’s make a series of increasingly customized graphs to demonstrate some of the tools in ggplot2.
Let’s begin with a scatterplot of crab size on the Y-axis with latitude on the X. We’ll forgo doing anything to the theme elements at this point to focus on the other three elements.
ggplot(data = pie_crab, mapping = aes(x = latitude, y = size, fill = site)) +
geom_point(pch = 21, size = 2, alpha = 0.5)ggplot() function call above, this geometry can assume those mappings without re-specifying
We can improve on this graph by tweaking theme elements to make it use fewer of the default settings.
ggplot(data = pie_crab, mapping = aes(x = latitude, y = size, fill = site)) +
geom_point(pch = 21, size = 2, alpha = 0.5) +
theme(legend.title = element_blank(),
panel.background = element_blank(),
axis.line = element_line(color = "black"))element_... helper functions. element_blank removes theme elements but otherwise you’ll need to use the helper function that corresponds to the type of theme element (e.g., element_text for theme elements affecting graph text)
We can further modify ggplot2 graphs by adding multiple geometries if you find it valuable to do so. Note however that geometry order matters! Geometries added later will be “in front of” those added earlier. Also, adding too much data to a plot will begin to make it difficult for others to understand the central take-away of the graph so you may want to be careful about the level of information density in each graph. Let’s add boxplots behind the points to characterize the distribution of points more quantitatively.
ggplot(data = pie_crab, mapping = aes(x = latitude, y = size, fill = site)) +
geom_boxplot(pch = 21) +
geom_point(pch = 21, size = 2, alpha = 0.5) +
theme(legend.title = element_blank(),
panel.background = element_blank(),
axis.line = element_line(color = "black"))
ggplot2 also supports adding more than one data object to the same graph! While this module doesn’t cover map creation, maps are a common example of a graph with more than one data object. Another common use would be to include both the full dataset and some summarized facet of it in the same plot.
Let’s calculate some summary statistics of crab size to include that in our plot.
# Load the supportR library
library(supportR)
# Summarize crab size within latitude groups
crab_summary <- supportR::summary_table(data = pie_crab, groups = c("site", "latitude"),
response = "size", drop_na = TRUE)
# Check the structure
str(crab_summary)'data.frame': 13 obs. of 6 variables:
$ site : chr "BC" "CC" "CT" "DB" ...
$ latitude : num 42.2 41.9 41.3 39.1 30 39.6 41.6 33.3 42.7 34.7 ...
$ mean : num 16.2 16.8 14.7 15.6 12.4 ...
$ std_dev : num 4.81 2.05 2.36 2.12 1.8 2.72 2.29 2.42 2.3 2.34 ...
$ sample_size: int 37 27 33 30 28 30 29 30 28 25 ...
$ std_error : num 0.79 0.39 0.41 0.39 0.34 0.5 0.43 0.44 0.43 0.47 ...With this data object in-hand, we can make a graph that includes both this and the original, un-summarized crab data. To better focus on the ‘multiple data objects’ bit of this example we’ll pare down on the actual graph code.
ggplot() +
geom_point(pie_crab, mapping = aes(x = latitude, y = size, fill = site),
pch = 21, size = 2, alpha = 0.2) +
geom_errorbar(crab_summary, mapping = aes(x = latitude,
ymax = mean + std_error,
ymin = mean - std_error),
width = 0.2) +
geom_point(crab_summary, mapping = aes(x = latitude, y = mean, fill = site),
pch = 23, size = 3) +
theme(legend.title = element_blank(),
panel.background = element_blank(),
axis.line = element_line(color = "black"))ggplot2 graph you need to leave this top level ggplot() call empty! Otherwise you’ll get weird errors with aesthetics later in the graph
It is sometimes the case that you want to make a single graph file that has multiple panels. For many of us, we might default to creating the separate graphs that we want, exporting them, and then using software like Microsoft PowerPoint to stitch those panels into the single image we had in mind from the start. However, as all of us who have used this method know, this is hugely cumbersome when your advisor/committee/reviewers ask for edits and you now have to redo all of the manual work behind your multi-panel graph.
Fortunately, there are two nice–entirely scripted–alternatives that you might consider: Faceted graphs and Plot grids. See below for more information on both.
In a faceted graph, every panel of the graph has the same aesthetics. These are often used when you want to show the relationship between two (or more) variables but separated by some other variable. In synthesis work, you might show the relationship between your core response and explanatory variables but facet by the original study. This would leave you with one panel per study where each would show the relationship only at that particular study.
Let’s check out an example.
ggplot(pie_crab, aes(x = date, y = size, color = site))+
geom_point(size = 2) +
facet_wrap(. ~ site) +
theme_bw() +
theme(legend.position = "none")ggplot2 function that assumes you want panels laid out in a regular grid. There are other facet_... alternatives that let you specify row versus column arrangement. You could also facet by multiple variables by putting something to the left of the tilde
In a plot grid, each panel is completely independent of all others. These are often used in publications where you want to highlight several different relationships that have some thematic connection. In synthesis work, your hypotheses may be more complicated than in primary research and such a plot grid would then be necessary to put all visual evidence for a hypothesis in the same location. On a practical note, plot grids are also a common way of circumventing figure number limits enforced by journals.
Let’s check out an example that relies on the cowplot library.
# Load a needed library
library(cowplot)
# Create the first graph
crab_p1 <- ggplot(pie_crab, aes(x = site, y = size, fill = site)) +
geom_violin() +
coord_flip() +
theme_bw() +
theme(legend.position = "none")
# Create the second
crab_p2 <- ggplot(pie_crab, aes(x = air_temp, y = water_temp)) +
geom_errorbar(aes(ymax = water_temp + water_temp_sd, ymin = water_temp - water_temp_sd),
width = 0.1) +
geom_errorbarh(aes(xmax = air_temp + air_temp_sd, xmin = air_temp - air_temp_sd),
width = 0.1) +
geom_point(aes(fill = site), pch = 23, size = 3) +
theme_bw()
# Assemble into a plot grid
cowplot::plot_grid(crab_p1, crab_p2, labels = "AUTO", nrow = 1)labels = "AUTO" argument means that each panel of the plot grid gets the next sequential capital letter. You could also substitute that for a vector with labels of your choosing
