Are you talking to me? Interaction between teachers and researchers around evidence, truth, theory, and decision-making

In this blog, I’m talking about computing education research, but I’m not always sure and certainly not always clear about who I’m talking to. That’s a problem, but it’s not just my problem. It’s a general problem of research, and a particular problem of education research. What should we say when we’re talking to researchers, and what should we say when we’re talking to teachers, and where do we need to insert caveats or explain assumptions that may not be obvious to each audience?

From what I know of philosophy of science, I’m a post-positivist. I believe that there is an objective reality, and the best tools that we humans have to understand it are empirical evidence and the scientific method. Observations and experiments have errors and flaws, and our perspectives are biased. All theory should be questioned and may be revised. But that’s not how everyone sees the world, and what I might say in my blog may be perceived as a statement of truth, when the strongest statement I might make is a statement of evidence-supported theory.

It’s hard to bridge the gap between researchers and education. Lauren Margulieux shared on Twitter a recent Educational Researcher article that addresses the issue. It’s not about getting teachers access to journal articles, because those articles aren’t written to speak to nor address teachers’ concerns. There have to be efforts from both directions, to help teachers to grok researchers and researchers to speak to teachers.

I have three examples to concretize the problem.

Recursion and Iteration

I wrote a blog post earlier this month where I stated that iteration should be taught before recursion if one is trying to teach both. For me, this is a well-supported statement of theory. I have written about the work by Anderson and Wiedenbeck supporting this argument. I have also written about the terrific work by Pirolli exploring different ways to teach recursion, which fed into the work by Anderson.

In the discussion on the earlier post, Shriram correctly pointed out that there are more modern ways to teach recursion, which might make it better to teach before iteration. Other respondents to that post point out the newer forms of iteration which are much simpler. Anderson and Wiedenbeck’s work was in the 1980’s. That sounds great — I would hope that we can do better than what we did 30 years ago. I do not know of studies that show that the new ways work better or differently than the ways of the 1980’s, and I would love to see them.

By default, I do not assume that more modern ways are necessarily better. Lots of scientists do explore new directions that turn out to be cul-de-sacs in light of later evidence (e.g., there was a lot of research in learning styles before the weight of evidence suggested that they didn’t exist). I certainly hope and believe that we are coming up with better ways to teach and better theories to explain what’s going on. I have every reason to expect that the modern ways of teaching recursion are better, and that the FOR EACH loop in Python and Java works differently than the iteration forms that Anderson and Wiedenbeck studied.

The problem for me is how to talk about it.  I wrote that earlier blog post thinking about teachers.  If I’m talking to teachers, should I put in all these caveats and talk about the possibilities that haven’t yet been tested with evidence? Teachers aren’t researchers. In order to do their jobs, they don’t need to know the research methods and the probabilistic state of the evidence base. They want to know the best practices as supported by the evidence and theory. The best evidence-based recommendation I know is to teach iteration before recursion.

But had I thought about the fact that other researchers would be reading the blog, I would have inserted some caveats.  I mean to always be implicitly saying to the researchers, “I’m open to being proven wrong about this,” but maybe I need to be more explicit about making statements about falsifiability. Certainly, my statement would have been a bit less forceful about iteration before recursion if I’d thought about a broader audience.

Making Predictions before Live Coding

I’m not consistent about how much evidence I require before I make a recommendation. For a while now, I have been using predictions before live coding demonstrations in my classes. It’s based on some strong evidence from Eric Mazur that I wrote about in 2011 (see blog post here). I recommend the practice often in my keynotes (see the video of me talking about predictions at EPFL from March 2018).

I really don’t have strong evidence that this practice works in CS classes. It should be a pretty simple experiment to test the theory that predictions before seeing program execution demonstrations helps with learning.

• Have a set of programs that you want students to learn from.
• The control group sees the program, then sees the execution.
• The experimental group sees the program, writes down a prediction about what the execution will be, then sees the execution.
• Afterwards, ask both groups about the programs and their execution.

I don’t know that anybody has done this experiment. We know that predictions work well in physics education, but we know that lots of things from physics education do not work in CS education. (See Briana Morrison’s dissertation.)

Teachers have to do lots of things for which we have no evidence. We don’t have enough research in CS Ed to guide all of our teaching practice. Robert Glaser once defined education as “Psychology Engineering,” and like all engineers, teachers have to do things for which we don’t have enough science. We make our best guess and take action.

So, I’m recommending a practice for which I don’t have evidence in CS education. Sometimes when I give the talk on prediction, I point out that we don’t have evidence from CS. But not always. I probably should. Maybe it’s enough that we have good evidence from physics, and I don’t have to get into the subtle differences between PER and CER for teachers. Researchers should know that this is yet another example of a great question to be addressed. But there are too few Computing Education Researchers, and none that I know are bored and looking for new experiments to run.

Code.org and UTeach CSP

Another example of the complexity of talking to teachers about research is reflected in a series of blog posts (and other social media) that came out at the end of last year about the AP CS Principles results.

• UTeach wrote a blog post in September about the excellent results that their students had on the AP CSP exam (see post here). They pointed out that their pass rate (83%) was much higher than the national average of 74%, and that advantage in pass rates was still there when the data were disaggregated by gender or ethnicity.
• There followed a lot of discussion (in blog posts, on Facebook, and via email) about what those results said about the UTeach curriculum. Should schools adopt the UTeach CSP curriculum based on these results?
• Hadi Partovi of Code.org responded with a blog post in October (see post here). He argued that exam scores were not a good basis for making curriculum decisions. Code.org’s pass rates were lower than UTeach’s (see their blog post on their scores), and that could likely be explained by Code.org’s focus on under-represented and low-SES student groups who might not perform as well on the AP CSP for a variety of reasons.
• Michael Marder of UTeach responded with two blog posts. One conducted an analysis suggesting that UTeach’s teacher professional development, support, and curriculum explained their difference from the national average (see post here), i.e., it wasn’t due to what students were served by UTeach. A second post tried to respond to Hadi directly to show that UTeach did particularly well with underrepresented groups (see post here).

I don’t see that anybody’s wrong here. We should be concerned that teachers and other education decision-makers may misinterpret the research results to say more than they do.

• The first result from UTeach says “UTeach’s CSP is very good.” More colloquially, UTeach doesn’t suck. There is snake oil out there. There are teaching methods that don’t actually work well for anyone (e.g., we could talk some more about learning styles) or only work for the most privileged students (e.g., lectures without active learning supports). How do you show that your curriculum (and PD and support) is providing value, across students in different demographic groups? Comparing to the national average (and disaggregated averages) is a reasonable way to do it.
• There are no results saying that UTeach is better than Code.org for anyone, or vice-versa. I know of no studies comparing any of the CSP curricula. I know of no data that would allow us to make these comparisons. They’re hard to do in a way that’s convincing. You’d want to have a bunch of CSP students and randomly assign them to either UTeach and Code.org, trying to make sure that all relevant variables (like percent of women and underrepresented groups) is the same in each. There are likely not enough students taking CSP yet to be able to do these studies.
• Code.org likely did well for their underrepresented students, and so did UTeach. It’s impossible to tell which did better. Marder is arguing that UTeach did well with underrepresented groups, and UTeach’s success was due to their interventions, not due to the students who took the test.  I believe that UTeach did well with underrepresented groups. Marder is using statistics on the existing data collected about their participants to make the argument about the intervention. He didn’t run any experiments. I don’t doubt his stats, but I’m not compelled either. In general, though, I’m not worried about that level of detail in the argument.

All of that said, teachers, principals, and school administrators have to make decisions. They’re engineers in the field. They don’t have enough science. They may use data like pass rates to make choices about which curricula to use. From my perspective, without a horse in the race or a dog in the fight, it’s not something I’m worried about. I’m much more concerned about the decision whether to offer CSP at all. I want schools to offer CS, and I want them to offer high-quality CS. Both UTeach and Code.org offer high-quality CS, so that choice isn’t really a problem. I worry about schools that choose to offer no CSP or no CS at all.

Researchers and teachers are solving different problems. There should be better communication. Researchers have to make explicit the things that teachers might be confused about, but they might not realize what the teachers are confused about. In computing education research and other interdisciplinary fields, researchers may have to explain to each other what assumptions they’re making, because their assumptions are different in different fields. Teachers may use research to make decisions because they have to make decisions. It’s better for them to use evidence than not to use evidence, but there’s a danger in using evidence to make invalid arguments — to say that the evidence implies more than it does.

I don’t have a solution to offer here. I can point out the problem and use my blog to explore the boundary.