We detail our construction of TwoRavens, a graphical user interface for quantitative analysis that allows users at all levels of statistical expertise to explore their data, describe their substantive understanding of the data, and appropriately construct and interpret statistical models. The interface is a browser-based, thin client, with the data remaining in an online repository, and the statistical modeling occurring on a remote server. In our implementation, we integrate with tens of thousands of datasets from the Dataverse repository, and the large library of statistical models available in the Zelig package for the R statistical language. Our interface is entirely gesture-driven, and so easily used on tablets and phones. This, in combination with being browser-based, makes data exploration and quantitative reasoning easily portable to the classroom with minimal infrastructure or technology overhead.
With the proliferation of open data sources and replicable data repositories comes the promise of increased access to scienti c information and the democratization of quantitative knowledge. However, meaningful analysis of this multitude of data remains outside the grasp of analysts who lack training in statistical modeling or knowledge of and access to statistical software platforms. For those with the necessary expertise, access is still limited by data transfer bottlenecks and installation and hardware overhead. To reduce barriers to statistical and quantitative reasoning and to promote the proliferation of empirical knowledge, we introduce the TwoRavens interface.
TwoRavens is a gesture-driven, Web-based device for the
analysis of statistical models and the exploration of data. It
All code, documentation, and numerous examples for our
open source TwoRavens software project are available at:
http://datascience.iq.harvard.edu/tworavens
is open-source and integrates with Dataverse repositories for
open archiving of data[
For ease of use, the UI is designed to mirror the common quantitative work ow. Moving left to right on the interface, users make intuitive column and row selections or subsets of the data.1 Users may perform transformations on these selections using the input box in the top right. In the center, users specify the intended statistical model in the framework of a directed graph, and have the option to perform transformations on selected columns or tag selections with speci c properties, such as \nominal." Finally, on the right, users select their statistical model and have the option to set covariate values on right-hand side variables to obtain additional bootstrapped simulations of quantities of interest. Upon estimation, statistical results are displayed using a combination of graphs and tables.
In this article we describe the UI and its integration with Dataverse and Zelig. In future research we intend to build upon this foundation in two ways. First, as users supply information about hypothesized relationships in the data, TwoRavens will synthesize their input and automatically suggest appropriate modeling decisions, such as functional form and choice of model. Second, as models are estimated, the results will be cumulatively recorded across users to form a coherent map and meta analysis of the statistical results that exist in Dataverse. 2.
In designing this software, some key goals were set that we believe the next generation of statistical tools will need to achieve. The challenges and solutions to reaching these goals informed some general design principles, which we feel are important to discuss and convey for future researchers. A paramount goal was to keep the interface as a thin client, keeping both the data itself, and the work of statistical 1Depending on the discipline, columns are known as variables, elds, features, etc., and rows are known as observations, cases, units of analysis, etc. processing, remote from the interface. We also required a broadly accessible platform that could be productively used by individuals at all levels of statistical expertise, from novice users with no statistical training, to experienced quantitative researchers. Furthermore, to make data accessible to users with di erent levels of available computational infrastructure, we worked to make this software device independent and feasible not only on traditional computers, which have previously been the sole domain of statistical software, but tablet, mobile, and other smart devices. 2.1
Although we have written an interface for statistical data
analysis, neither the data itself, nor any statistical
analysis, occurs or resides at the client side.2 It is a challenge
to explore data, and to implement statistical models,
without the the immediacy of local interaction. However, having
a thin client with remote data and remote statistical
processing allows some enormous, novel advantages that are
increasingly useful in modern data-intensive science: 1) When
datasets are large, transferring the data to the user for
exploration can be the primary bottleneck for speed, which this
overcomes. 2) Similarly, when datasets are large, the nal
analysis might require distributed processing, which again
requires another transfer of the data. In our architecture,
again, the code for statistical processing is instead placed
close to the original data. 3) When the data has privacy
implications, there are settings where it is allowable to grant
access to summaries and statistical representations of the
data, (or privacy preserving versions of these statistics, such
as those that are provably di erentially private [
Statistical consulting is increasingly a necessary service provided by research universities to aid their research faculty. Most substantive researchers are heavily invested in the collection of their data, understand their variables well, and possess expert substantive knowledge about the plausibility of various relationships. What they may not know is the set of plausible statistical models appropriate to their objectives, the statistical language to describe their goals and aims, or the software knowledge to implement these choices. Consultants often provide that link, allowing experts without statistical training to translate their knowledge and goals into appropriate statistical software routines, and interpreting the results back in a substantively meaningful fashion. Obviously, however, not every user of data has access to a trained statistician. One overarching goal of the TwoRavens project is to provide as much of this service as can be automated. At this stage of the interface, this means providing a way for users to intuitively communicate all their knowledge about the data, so an appropriate model can be constructed for the quantitative task at hand. 2.3
2Or slightly rephrased, we have built a package for statistical analysis of data, that can not handle data, nor analyze any statistics.
Part of expanding the set of users who have access to quantitative reasoning, means lowering the infrastructure requirements to statistical analysis. While professional researchers may have extensive computational resources, many other settings, especially community colleges and high schools, have limited resources for instructors and none for students. Even free and open source statistical packages like R generally require the expenses of some level of IT support to install in a networked lab. Moreover, even in well-funded universities, substantive classes that use quantitative data may have no classroom lab access. Our goal is to enable a computationally capable access to statistical exploration with very minimal infrastructure. 2.4
In meeting these goals, we developed the following design
principles for this interface: 1) Browser Based: The
ability to run the software entirely through a browser, with no
additional installed software, is possible because of the thin
nature of the interface, and helps facilitate device
independence as it can run on any device that includes a browser
2) Gesture Driven: Common statistical packages are
normally command line, script, or menu driven, but to
facilitate ease of use across devices and without user expertise,
we have tailored a process that allows full exploration and
setup of statistical models by interacting directly with the
data though gesture 3) Graphical Representation:
Directed graphs, also called probabilistic networks, are
increasing used by statisticians as a representation of complex data
generation processes [
The interface is written in Javascript and incorporates
aspects of interactive graphics, Web-based statistics, and
customized R applications. The interactive visualizations use
Data-Driven Documents (D3), which has been in uenced by
tools such as Protovis and Processing [
Furthering its distinction from existing statistical software, it integrates with Dataverse, providing instant access to tens of thousands of datasets by simply launching TwoRavens from any Dataverse repository page, and with Zelig, delivering analysts meaningful yet easily interpretable statistical estimates and graphics. Given the availability of open access data in repositories such as Dataverse, the open source power of R, and the trend towards Web-based, interactive visualizations, such a device ties together many threads of modern quantitative computing. 3.1
In the last decade, the Data Science team at Harvard's
Institute for Quantitative Social Science (IQSS) has developed
software infrastructure and tools to facilitate and enhance
data sharing, preservation, citation, reusability and analysis
[
Dataverse is \an open source data repository, which
allows one to publish, share, reference, extract, and analyze
research data" [
Zelig is a wrapper and interface that allows a large body
of di erent statistical models in the R statistical language
to be used from a uni ed call structure [
Javascript is lightweight, and allows us to run the interface entirely through an internet browser. For ease of use and cross-platform portability, all functionality has gesturedriven capability, so statistical models on datasets stored in online repositories could be run from a tablet or mobile device without a keyboard. This ability expands the set of ways individuals use archived data and quantitative analysis, including bringing real-time data analysis into the classroom without using a computer laboratory setting.
A screenshot of TwoRavens is shown in gure 2. The work ow of developing a statistical analysis moves from left to right, rst examining and selecting variables in the dataset, then constructing a framework of possible relationships, and then choosing and interpreting an appropriate statistical model for that framework. In gure 2, the user has selected a model, tagged ti_cpi as the dependent variable, and graphed an appropriate statistical model. 4.1
Each variable in the dataset can be introduced as a node
in the center space by clicking on the variable name in the
left panel. Users may not be familiar with each variable
in the dataset, so on mouseover information pertaining to
that variable appears, such as primary summary statistics,
graphs, metadata and short descriptions of the variable.3
This information is precalculated by Dataverse when a new
dataset is uploaded, and stored in a json le that is
compatible with the DDI schema [
Although not explicitly in the left panel, TwoRavens includes an option to perform transformations on variables, such as taking the log of a variable or multiplying it by some factor, using the input box in the top right. Users may transform variables in two ways: (1) by manually entering text into the transformation input box; or (2) by clicking on the input box, selecting a variable from the drop down list, and then selecting a transformation from the function list. In this way, the gesture-driven functionality is preserved, while allowing additional exibility for users familiar with R functions.
Users may be interested in subsetting the dataset to examine only observations that have speci c values (for example, only European countries, or only respondents over the age of 60). For such cases, we include a Subset tab that shows the distribution of each variable. These distributions are either a density plot or a bar plot, depending on the variable's level of measurement and its number of unique values. By brushing the plots with the pointer, users select ranges of that variable upon which the data is to be subsetted. The numbers associated with the range are shown so that users may be precise in the ranges they specify. All metadata is remotely recalculated for the new subset using the same Dataverse ingest routines, and a new space that represents the subsetted data is added to the carousel in the center space of the interface. 3For devices not compatible with mouseover, we enable this feature via click and hold.
In the focal, central space, the substantive knowledge of the researcher is easily communicated by drawing a directed graph representation. With a two- nger click, arrows are drawn connecting nodes, depicting the possible relationships between variables. The nodes and arrows are an application of D3's force layout. In the simulated forces that propel the visualization, each node has a gravity, which draws all the nodes together, as well as a charge that repels them from getting too close. The arrows act as simple springs, and are removed when clicked. As the graph between the variables is built up by the researcher, it dynamically rearranges itself by these simulated forces. The researcher can also physically drag the pieces around, which acts as an additional force in the dynamic visualization, or completely turn o the simulated forces, and place every node manually for complete control of the representation (using the force toggle icon, represented by a pin).
Dependent variables, time, cross-sectional, and nominal identi ers are properties that may be tagged, by the user, to a node. Each property is denoted with their own colored halo. To tag a property, a user mouses over a node in the directed graph, at which point buttons appears as arcs around the perimeter of the node. Each arc is colored and labeled, and clicking on the arc associates that property to that node. At this point, the halo is colored appropriately and a legend appears, reminding users what the color of the halo represents. Additionally, the background color of the tagged variable in the left panel is changed to re ect the color of the tagged property.
The portion of the center space that is visible is actually just one element of a carousel, and users have the option to add and remove new workspaces (hereafter, elements). For example, if a user subsets the data, then an additional element is added to the carousel that corresponds to the subsetted data. Users toggle between elements by clicking a chevron or by clicking and swiping in the direction they want the carousel to move. Their current element is shown by highlighting its corresponding dot in the top-center of the center space. By clicking on the plus sign to the right of the dots, users are duplicating the current carousel, and placing its representation at the right of the element array. By clicking the minus sign, users are dropping this element from the carousel. To clear a modeling space but not drop it from the carousel, users may select the Erase icon, represented with a magnet. 4.3
After examining the data and constructing a diagram of relationships, in the right panel users begin to investigate statistical models. The Models tab provides a list of the available statistical models that can be employed by Zelig. On mouseover, users see a brief description of the model so that they have some guidance as to which to select.
The next tab, labeled Set Covar., provides the ability to
interpret any estimated model by means of predicted and
expected values at chosen values of the covariates, as well
as rst di erences created by the changes in the predictions
across changes in the covariates [
When this information is complete, the researcher can estimate the model by clicking the Estimate button. At this time, the information extracted from the user is passed to an instance of an R application hosted on a remote server.4 This remote application rst builds a formula representation of the model from the graph connections the researcher has constructed. In the present version, this is all variables that have a path to the dependent variable, but more complex graphs can include intermediate or post-treatment variables as well as consequences of the dependent variable, which are 4We developed our R application in Rook and host it in rApache. useful for forecasting and imputation, but omitted from the formula for a causally oriented analysis. Using this formula, the R application calls the applicable statistical model from the Zelig library. The results from Zelig are asynchronously returned to the browser interface. As can be seen in gure 4, the plots that Zelig produces, as well as a table of estimates, are available for viewing inside the Results tab.
The Quality of Government (QoG) represents one database
for which the primary bene ts of our tool might be
recognized for two target audiences, the novice user and the
classroom instructor. The QoG is a collection of
countryyear datasets whose variables include data on population,
respect for human rights, and political regime type [
A simple empirical exploration of this question using the QoG data, however, can be fraught with di culties for novice users. Minimally, one needs to be familiar with some statistical software package and to understand enough statistics to be able to analyze the data. The QoG data would have to be downloaded locally, as would the software being used.
Our tool improves accessibility by reducing or removing these initial barriers. Users interact with the data in visual, gesture-based ways, and so the time spent learning how to use TwoRavens is negligible in comparison to what is necessary to analyze data with R, for example. Statistical models are represented visually as a directed graph, and users may instantaneously view each variable's distribution and summary statistics by hovering over a node in the graph. For example, Figure 3 shows a description of the QoG variable bl_asy25f in the left panel. The variables are listed in the left panel, and are added and removed from the modeling space by clicking on the variable name. The modeling space, the center of the gure, shows a representation of a user-de ned statistical model.
For instructional purposes, the QoG represents a type of dataset that may be used for a class project. It contains 746 columns of data in a time-series, cross-sectional format. Many of these columns are substantively interesting dependent variables, while others appear as explanatory variables in many models in quantitative Political Science. For instructors to teach how to use data to study questions in politics, requires some degree of expertise in a statistical software package, a projector, and a computer with the rights to that software. For student projects, each student would have to download the data and some statistical software to their personal computer, and the instructor would have to provide materials and guidance on software usage. Most statistical software is proprietary, so students are often restricted to labs that are equipped with the necessary licenses. TwoRavens removes these barriers, and all that is necessary for the instructor to bring data to the classroom, and for students to analyze the data, is an internet connection and a Web browser. 6.
TwoRavens is a Web-based tool for statistical analysis that is lightweight, broadly accessible, and device independent. It integrates with Dataverse, providing access to the tens of thousands of data repositories that exist there, and leverages the power of Zelig, an R library that provides a common call structure for a large number of statistical models. Entirely gesture-driven and intuitive for users of all levels, TwoRavens reduces barriers to statistical analysis and promotes the proliferation of empirical research.
This article details the design of the user interface and its applicability for use by statistical novices and users not familiar with or who do not have access to statistical software. Although integral and foundational to the TwoRavens project, the UI is only the rst of three layers. In future research, the TwoRavens project will be adding the model \selector" and results \accumulator" layers to provide more automated guidance on model selection and speci cation. The selector synthesizes the user input and automates suggestions, such as potential omitted variables, functional form, and choice of statistical model. The accumulator stores all models that have been estimated on each dataset, and provides users with feedback on existing research using that dataset and datasets judged to be similar.
The authors would like to thank Merce Crosas and Gary King for extensive feedback and ideas, Leonid Andreev, Michael Heppler and Elizabeth Quigley for continued development assistance, and Dwayne Liburd for creating our logo.