CS Ed Podcast 2 - Dan Garcia on test creation
Episode 2 of the CS Ed podcast had Dan Garcia talk about exam creation.
This wasn't a podcast about the value of exams - in class, high stakes or otherwise. In fact Dan says in the podcast it would be great to "get grades out of the equation. Grades are gonna be an impediment to learning." But recognizes that we have not say in this most of the time (and I'll add that though I agree grades can and do perform a function) so we should be creative in terms of assessment.
Dan hit on a lot of important points and there's no way I can weave them into a sensible narrative so I'll just hit a bunch of points and riff from there.
Dan started right out front saying that he's come around to being ok or maybe even liking multiple choice having referenced the ability top make "good distractors."
This is something that I think I'll never agree with. On the one hand, there is the idea of multiple guess but what really gets me is that multiple choice questions, by definition, are gotcha questions. You're supposed to have distractors to pull you away from the right path. To make a mistaken answer seem ok. I hate gotcha questions. Besides, halfway decent multiple choice questions are really hard to write.
To me, MC questions are really a sometimes necessary tool to make up for the fact that teachers are overworked and never have enough time to grade. This might make them a necessary evil but it doesn't make them good. We might have to resort to them if we're "teaching" a class of hundreds or perhaps thousands but we shouldn't have classes anywhere near that size.
Another problem with MC is that multiple choice tests seem to frequently test speed rather than knowledge. APCS-A is a good example of this. You have 90 minutes to answer 40 questions. That's a bit over 2 minutes per question. That's a race not an assessment.
Finally, I find multiple choice questions hard to write and time consuming to typeset and format. I guess if I were better at the distractors this might not be the case but a short answer version of a MC question is always easier for me to write and not much harder to grade.
Half a day. Twelve hours. That's how long Dan says it takes him to make a test. That sounds about right. There's no two ways about it. Creating a good exam is hard work. Creating a bad one is easy and quick but that's not a good answer.
Let's think about this for a minute. A college professor might teach one or two classes a semester. A college lecturer three, maybe 4. A high school teacher teaches 5 classes. Typically two or three different subjects. They basically get 40 minutes at work a day to prep and that includes EVERYTHING - lesson planning, grading, test creation, working with students. The whole kit and caboodle. Even with two preps rather than three that's an additional 24 hours just to create those two tests. More since you'll need multiple versions of each.
Of course we can mitigate this a it by reusing and modifying questions from old exams and sharing with colleagues but writing exams is a bear.
It's no wonder teachers turn to multiple choice test banks.
Really glad Dan spent time on this as teachers have to realize that it's not just them.
How long is too long?
I frequently struggle with test duration. It sounds like it's a common problem. Dan has his TAs test the tests hoping they can do it in one sixth the students time. He found, though that his slowest TAs might take one third of the allocated time.
When I started, I heard ratios from math teachers ranging from students take twice the time I would on my exams up to five times longer.
It's hard to get it right but it's important that tests are tests of knowledge and ability not speed. At Stuy where class periods were 43 minutes I'd try for tests that took the typical student 33 - 35 minutes. Just enough time to go over the exam. You might ace the exam, you might fail but time wouldn't be the issue.
This was a big one. Dan talks about students expectations - are tests cumulative or just since the last unit. Making sure test content is proportional to lesson content. Watching out for test morale - letting kids know hard questions are hard, test flow and much more. Not too much to say here other than if you haven't listen to the podcast.
One content point that I do want to drill down on is Dan's stating that we sometimes give too much backstory on questions. "If you want them to sort a list, make them sort a list! You don’t have to tell a story about the list…." Too often tests are races and even when they aren't a student has to first read the question, understand it, make sure there are no gotchas, formulate and write down a solution. Giving a complex backstory from which the students have to glean the real question in a timed high pressure environment is just too much. There are other places where you can ask a general question with a big back story and have students solve the poblem.
This made me think of technical interviews which are all backstory. The truth though is that most kids solve them by pattern matching. They prep by doing similar problems and look for key phrases. Have unlimited memory and need fast access - there's probably a hash table involved. Something have an easy solution but it's too slow - think recurison. Recursion is obvious but too slow - dynamic programming. It's a "have I seen this question before" more so than a real test of a potential employees ability.
The other thing it made me think is that while I very much agree with Dan, it flies in the face of what teachers have been forced to deal with in the last couple of decades. It's all about the word problem where the kid must sleuth out the question. So many high stakes exams are as much reading exams as they are content exams. I'm not even going to get into how this affects non English speakers but let's just say this is a real issue.
Tests on computers and other test taking formats
Towards the end, Dan talks about having an on-computer part of an exam. I like having the students do live on computer exams but they also have their issues. I'm not so concerned about cheating. I give them the resources they can use and truth be told, the end results don't differ much from when I've used paper exams or other assessments. I am always concerned about a computer breaking but fortunately that hasn't really been a problem.
What has been a problem however is the fact that speed can become an issue. Some kids know how to type others don't. This can be a HUGE advantage in a CS0 or CS1 class. When you're thinking about what letter to type you're not thinking about the problem and in general beginners are very slow. I always tell my second year students to look at what they can do as an overnight assignment. They realize that what's now overnight was a large semester project back in their first year. This all means that you really have to be careful about the length of a computer based test.
Another thing Dan mentioned was giving group tests. Students would work in teams. He noted that in the group, The collective score was always higher than the top individual score. I get the idea and it's probably true trend wise but can't be universally true if your tests are actually compatible. Given a fair test, surely some students can ace it in which case the group can only equal the top scorer.
This group test reminded me of cooperative learning which was all the rage back in the day. Of course cooperative learning is no longer "it" but while it never was the silver bullet it is notable that it had some good ideas and it sounds like Dan is implementing some of them albeit without the same name.
Thoughts on grading
I'll disagree here with Dan's contention that we should all use Gradescope. I've come out against autograders before so I'll just summarize here. I'm not entirely against autograders. They're great for rudimentary answers and can also provide a level of instant student feedback. The flip side is that in order to get to know your students you have to look at their work. You might not learn much from multiple choice answers but you will if you look at code they've written out or other long form answers. I get that you can't do this with huge classes but again, I'll say we shouldn't have huge classes. To use an autograder as a necessity is one thing but to say it's better rather than just a time save is another. Now, you can use autograding tools effectively to cull and sort and draw your attention to things to look at by hand and that's a plus but I haven't seen too much of that in current tools.
Rather than an autograder, I try to use testing frameworks on short answers and to provide instant feedback. Specifically I use doctest for C++ and unittest for Python. They provide students with instant answer feedback while at the same time introducing them to practical software engineering sensibilities.
Some of my memorable tests
I thought here I'd share a few exams I've given that I found memorable.
First was a five question long answer exam. I gave it to my kids telling them "Answer three of the five questions. You can select any three but when grading, I will only grade quesitons 1, 2 and 4." Some students had some real issues with this. Led to an interesting discussion as to about choice and consequences. I didn't do this in order to mess with the kids. It was back in the day when you had to send exams to the central copy room to be copied and needed about a weeks lead time. I didn't know the direction the class would take so wrote questions 3,4, and 5 to hedge my bets. By the time we got to the test, question 4 made sense while 3 and 5 would have been really super hard.
A second test, well quiz I gave once was a single question that was something like this:
take the next 20 minutes to share with me your thoughts on cellular automata. Consider why we studied it. Was it interesting? Why, why not? What did you learn?
Some students absolutely loved it. Others loathed it. The freedom gave some kids license to give me super creative interesting answers. Some really needed more constraints and had tremendous difficulty.
Finally, I once gave a semi-gag test. The instructions on the front said that you had to proceed in order. You WERE NOT TO look ahead. It was a mix of real CS questions and goof questions like:
- sprint up to the blackboard, draw a smiley on the board, then return to your seat and go on to the next question.
- Stand up, do 10 jumping jacks, sit and proceed to the next question
- Stand up and raise your right hand. When you see someone else with their hand raised, walk to them and give them a high five. Then return to your seat nad proceed to the next question.
The last page had an answer key and instructed the students to grade themselves before handing it in.
Overall the class enjoyed the exam.
What was interesting, was that after class a few student appoached me either in person or via email saying that they felt really bad but they looked at the end of the exam when it started so saw the answers. They felt they cheated and wanted to let me know and understood if they got zeros (which of course, I didn't give them).
Wow. That was long but I said before that there was a lot to unpack.
I left a bunch out of this and it's already a really long post. I might write more on test creation and administration at some other point but the takeaway from this is listen to the podcast if you haven't already.