Friday, May 3, 2013

Using social network analysis to find the impact of teacher turnover

This painting from a Washington D.C. tapas restaurant is not social network analysis, but social network analysis can help us unwind some mysteries about how the world is arranged.

Much has been made about the fact that America is coming up short when teaching children core concepts in STEM, or Science, Technology, Engineering, and Mathematics.

How does America perform, exactly? In 2009, the country's students came in 23rd place in science the Program for International Student Assessment (PISA). That's below Belgium and Hungary. The top three performers, from first to third, were China, Finland and Hong Kong (tested separately than the rest of China).

When the results came out, U.S. Education Secretary Arne Duncan called it "an absolute wake-up call for America." The only country to land people on the moon was now in the middle of the pack for teaching children about science.

It made the news. Town hall meetings sprung up. Companies like Exxon Mobil developed marketing and outreach campaigns. And so the public became aware of the STEM crisis.

There's another crisis schools are faced with, and it's much less publicized: the turnover crisis. To some degree, these two problems are related.

A year before PISA, Dr. Sharif Shakrani, Co-director of Michigan State's Education Policy Center, warned about teachers giving up the profession in droves. In my home state, Illinois, turnover cost the state $224 million. The state with the highest turnover, Texas, spent more than half a billion dollars from turnover.

Nationally, school districts and taxpayers spend $2.2 billion each year to replace teachers who have dropped out of the profession. While every penny matters in times of government austerity, the cost isn't only in dollars and cents.

"An inordinate amount of their capital—both human and financial—is consumed by a constant process of hiring and replacing beginning teachers who leave before they have mastered the ability to create a successful learning culture for their students," Shakrani wrote. "There is a growing consensus among researchers and educators that the single most important indicator in determining student academic performance is the quality of instruction provided by teachers."

I don't usually talk about my full-time work on this website, but I work for a National Science Foundation grant to improve the quality of STEM education. I won't speak to the politics of teacher turnover, as that's neither the purpose of this website or the grant, but part of my work consists of analyzing teacher networks and measuring how turnover and collaboration changes those networks.

What is social network analysis?

A six-node social network graph, from the Wikimedia commons.

To quote Wasserman and Faust, social network analysis is the "study of relationships among social entities with a focus on patterns and implications of relationships."

But what does that really mean? In short, social network analysis is concerned with how humans exchange things: information, learning, friendship, diseases, money, tweets, likes, power, weapons, etc. But it's not just concerned with how one or two people exchange these things, but how whole networks of people exchange these things, and how these networks are arranged.

Some of the earliest work analyzing social networks was applied to industrial engineering, where companies obsess over the shortest, most efficient method of mass-producing and distributing products. More recently, social network analysis has been applied to virology to head off the deadly H5N1 flu virus, or to combat networks of terrorists.

Oddly enough, social network analysis was one of those skills I picked up as a graduate student in journalism at the University of Illinois. Investigative Editors and Reporters (IRE) has some great resources on how to use this tool, along with examples of how it's been used before.

Popular applications in investigative journalism, as one would imagine, involve tracking the flow of bribes, power and political influence to find movers and shakers. Sometimes these individuals are hidden until exposed by social network analysis.

Social network analysis is relatively new to journalism, and it's especially new to education, and STEM education specifically.

Social network analysis is useful for STEM education for a number of reasons. My grant, EnLiST (Entrepreneurial Leadership in STEM Teaching and learning) focuses on developing "entrepreneurial teacher leaders," or teachers who have the content knowledge and entrepreneurial skills to seek out resources and collaborations that could fundamentally change STEM education in their districts.

To give you an example of what an entrepreneurial teacher leader does in STEM education, some of the teachers we've trained go on to start student-run biofuel plants. Others get students involved with underwater robotics competitions. Yet others start inter-generational STEM initiatives, where high school students take on the role of STEM role models as they teach elementary students lessons in entomology, technology, chemistry, and physics.

All these initiatives rely on networks of teachers. We know from research that reform in a school district is much more likely to succeed from the ground-up. It also relies on a combination of "weak" and "strong ties," as we rely on trusted individuals for vital information, but we also need to bring us novel information.

As a grant that trains teacher-leader to produce classroom innovations, we're very interested in tracking these individuals to see where their influence reaches in teacher networks. More than that, we're interested in social network analysis as a diagnostic tool to find out the state of information exchange in schools, before any initiatives have a chance to start. It gives us a baseline from which to measure impact of our efforts.

Per our IRB (Institutional Review Board), I am not allowed to divulge any information that would identify specific teachers, schools, or districts. But we are, of course, permitted to present and publish our findings (a list of some of these presentations are in my bio for reference).

What our findings show, as one would expect, is that teaching and learning networks suffer when teachers leave the school. When those networks decline, it gets harder to participate in reform efforts to improve STEM education.

But there is, however also a potential antidote: if you have a conscious effort to bring teachers together in a focused collaboration, teachers will talk, connections will grow, and the potential for reform returns.

What can we learn from a social network graph of a school district?

First, a little methodology. In 2011, at a small school district in the Midwest, teachers were given a survey that asked them about who they learned from within the school district. The teachers were provided with a list of all the other teachers in the district, so they didn't need to remember the name of someone they learned from.

On this survey, we asked the teachers bio-data (how long they had been in the district, where they were placed in the district, along with any leadership and teaching roles), and more specific information about what they learned from other teachers. We asked teachers to classify what they learned, according to four categories: science teaching and learning, teaching and learning in non-science subjects, classroom management, and career advice. We also asked the teachers to grade the frequency of this interaction on a 1 to 5 scale ("strength").

Results came in the form of a CSV file. I wrote VBA (Visual Basic for Applications) macros to reformat and apply error-correcting heuristics to the data, and then placed the results into an Access database. Querying the database allows me to produce edge lists for individual schools, learning categories, strengths and more.

I imported the edge list queries from this database into an Excel plugin called NodeXL. It's a handy program for generating social network graphs and metrics (Pajek, Gephi, and R also are excellent options).

Above is the network graph of that entire school district, and all the learning that teachers reported. Each dot, or in social network analysis terminology "node" or "actor," represents a single teacher in this school district. The color of that node is representative of the school where the teacher is based. Each connecting line represents some kind of learning taking place between two teachers in the school district.

The graph is pretty and all, but to unpack its meaning, we have to delve into the graph theory behind it. First off, this isn't a Cartesian coordinate system, so where any given node is placed in the X or Y dimensions is not actually significant. Something in the upper right does not have greater or lesser value that something in the lower left.

What matters is how the node is positioned relative to other nodes.

You see, this is what's known as a "force-directed" graph. Each node is placed, and is given an equal "charge," much like a charged particle. All the nodes have the same kind of charge, which you might remember from chemistry and physics, means they want to repel each other. That repellant force puts distance between the nodes.

That is, until nodes make connections between one another. These connections bring nodes closer.

From: Spring Embedders and Force Directed Graph Drawing Algorithms, Kobourov, 2012
Above is an illustration that shows how the graphing algorithm arranges the nodes. Connections are treated as springs, which draw nodes closer. In fact, the algorithm I used, Harel-Koren Fast Multiscale, actually has Hooke's Law built-in. That's the physical law that says the force necessary to extend or contract a spring is proportional to the distance the spring is extended or contracted.

Note that a series of nodes that have connections in common will be drawn towards each other. We call this behavior "clustering."

With that in mind, scroll back up and take a fresh look at the original graph. What do you see?

Notice four distinct clusters, each of which tends to be dominated by a single color. In these graphs, the color corresponds to the school where the teacher or administrator is stationed. Red is the elementary school, blue is the junior high, green is the high school, and black corresponds to the district office.

But also notice that the majority of nodes in the center of the graph are not black, but a little less than half of the black nodes are in the center. The center doesn't necessarily represent a cluster of individuals from the district office, but rather individuals who have nearly an equal number of ties to the three other clusters.

What does this tell us? Given that clusters tend to be segregated by the school type, we can tell that little if any collaboration is occurring between schools. A situation like this can make it difficult to institute district-wide reform, because there is not as many connections for information or influence to transfer between teachers.

What did turnover do to teaching and learning networks?

It's nice to get a snapshot extant teaching and learning networks. But if we repeat the survey at a later date, we gain an understanding of how and where connections grow, disappear, and shift over time. This kind of longitudinal data is the holy grail of social network analysis.

We conducted the survey again in 2012. This time around, the teachers seemed even farther apart than before, and clusters seemed more segregated.

What exactly changed in the network to make it look like this? Let's identify the teachers who left the district between the 2011 and 2012 surveys. They're highlighted in yellow here on the 2011 graph.

It appears as if the teachers who are leaving the district serve as some fairly important links between clusters. This would help explain why clusters seem farther apart in 2012.

Teachers leaving the district is only half of turnover story. There's also the matter of new teachers coming into the district. How are they situated in the network?

This is the 2012 graph, this time with the new teachers highlighted in yellow. Notice that these teachers typically are situated in periphery of the clusters, and don't generally serve as bridges between clusters. None of the new teachers are positioned in the center of the graph, although several departing teachers were in the center in 2011.

It's what one might typically expect to see in a school district -- teachers leaving the school district likely are taking with them not just experience, but links between classrooms and schools. Teachers entering the district will not likely be able to replicate those links in a short period of time.

These observations indicate that teaching and learning networks are weakened by teacher turnover.

There's more to this story. Fortunately we don't only have graph interpretations to generate an analysis. A thorough social network study also will take a look at metrics which quantify the density of connections, the distance between nodes, and the connectedness of neighbors, among other things.

A look at these metrics provide some interesting insights for this district. The numbers tell the story of a school district that has weathered turnover, but also has brought teachers together around a common goal.

There will be more analysis of this school district's teaching and learning network in the future.