(Note: this post should probably be titled "Quantitative Methods of Curricula Planning" but I thought the current title would draw more interest–though they would both lose out to "These Weird Approaches To Lesson Planning Will Leave You Speechless")
Suppose you were tasked with teaching a course about a field of study. There would be, of course, several topics that you are expected to cover by the course end date; how would you decide the order in which to teach them?
Most people would say that the topics should build on one another, with monotonically increasing levels of difficulty. Further, no topic should be brought up that requires comprehension of another topic yet unlearned.
Planning the syllabus under these constraints would, perhaps, come naturally to skilled and empathetic lecturers. But,
- not all lecturers are skilled and empathetic
- even satisfying all of these constraints, there are objectively superior and inferior lesson plans
- there are some subjects for which these constraints cannot be satisfied (statistics)
For these reasons, having a suite of quantitative methods for choosing the best order of topics in teaching a field of study would be valuable to pedagogy (not to mention providing challenging problems for me to focus on instead of writing).
I started thinking about this topic as I began to plan writing my book about learning introductory statistics with R. There are, of course, myriad other very good books on this very topic, so I figured that one way I can stand out is to organize the topics in a way that best facilitates mastering the material. I thought that this would be especially appreciated with a field of study that is notoriously scary and difficult to the uninitiated (like statistics is.)
Anyone, anywhere, teaching introductory statistics will be expected to touch on the common topics: measures of central tendency, measures of dispersion, probability, the central limit theorem, sampling theory, etc… I know how everyone else have arranged the topics, but what's the best way?
It might seem strange, but answering that question was probably the hardest thing about putting together this book and in all of my (admittedly limited) experience designing statistics curricula.
Let's speak of graph theory
To explore optimal paths through the topics, we can represent the subject of statistics as a big graph, or network. Each topic would be a node and there would be directed edges indicating when knowledge of a particular topic is a prerequisite to understanding another. Specifically, if there is a edge connecting topic "a" to topic "b", topic "b" requires an understanding of "a"–like long division requires knowledge of subtraction.
This is what a topic network of an excerpt of introductory stats topics might look like.
In graph theory, this is known as a directed acyclic graph (DAG). DAGs have the property that there exists at least one ordering of nodes such that no node in the ordering is connected to ("pointing to") a node earlier in the ordering. This is called a topological sort. For most DAGs, there are a number of different orderings that satisfy the ‘dependency’ constraints.
Now that I have your attention, let's now speak of monads
To get a list of all of them, I wrote a small library and set of algorithms in Haskell. You can view it here but the "meat" of the algorithm is in the following snippet that recursively adds all nodes with no children (topics that have no topics that depend on them) to a list of possible alternatives and removes the childless nodes. This is repeated until there are no nodes left to remove. A potential snag is that the function only takes one path but each function call may generate multiple alternate paths. However, if we view the output of the "gatherAllChildless" function as a non-deterministic computation, we can exploit the fact that the path of nodes is a monad and have the function recursively call itself inside of a monadic bind.
allTopoSorts :: Graph -> [Path]
allTopoSorts g = if null result
then error "no topo sort exists"
else map fst result
where result = gatherAllChildless (, g)
gatherAllChildless :: (Path, Graph) -> [(Path, Graph)]
gatherAllChildless (ns, g)
| null (nodes g) = [(ns, g)]
| otherwise = (map (\n -> ((n:ns), (removeNodeByName (name n) g)))
childlessList) >>= gatherAllChildless
where childlessList = getChildless g
This has a sub-quadratic time complexity (< O(n^2))… not too bad.
There are 26 possible orderings of the topics that satisfy these “knowledge dependencies” including:
probability -> central tendency -> measures of dispersion -> sampling theory -> sampling distributions -> probability distributions -> central limit theorem -> statistical inference -> NHST central tendency -> probability -> measures of dispersion -> probability distributions -> sampling theory -> sampling distributions -> central limit theorem -> statistical inference -> NHST
There are a few of the ordering that intuitively seem like poor choices. Taking the first one, for example: it might be strange to start a book on statistics with probability when readers may want to get starting with univariate analysis right away. Looking at the second one, it seems strange to stick "probability" in between "central tendency" and "measures of dispersion", even though it can technically be done, because most people expect highly related topics to be positioned next to each other.
One way of cutting down on the list is to label each topic node with a difficulty level, and choose the ordering which causes the fewest backwards jumps in difficulty level. This should represent the path that has the most gentle level-of-difficulty slope.
Given the algorithms from lines 67 to 78 of TopoSort.hs and the following (subjective) difficulty mapping:
"central tendency": "1" "measures of dispersion": "2" "sampling theory": "3" "sampling distributions": "3" "central limit theorem": "5" "probability": "4" "probability distributions": "3" "statistical inference": "5" "NHST": "5"
the “optimal” ordering is:
central tendency -> measures of dispersion -> sampling theory -> probability -> sampling distributions -> probability distributions -> central limit theorem -> statistical inference -> NHST
Yay! This is pretty close to the ordering I chose.
The most truly difficult thing about sorting this out is that the statistics topic network diagram is not a DAG. This means that there is no ordering possible that doesn’t appeal to topics yet unlearned. For example, explaining why sample standard deviation divides by n-1 instead of n requires appealing to sampling theory, which requires a good foundation in measures of dispersion to understand. There are a few more of these cyclical relationships in the field.
All of these instances require some hand-waving on the part of the writer or lecturer ("don't worry about why we divide by 'n-1', we’ll get to that later") and adds to the learner's perceived difficulty of grasping the field.
The best way to reconcile these circular knowledge dependencies is to introduce weight to the edges that represent the extent to which a topic requires knowledge of another. Then, a cycle detection algorithm can be run on the graph. Once all the cycles are detected, the edges in the cycles with the lowest weight can be systematically removed until there are no more cycles and the graph is a DAG. At that point, the specialized topo sort from above may be used. I plan on implementing this when I have more time :)
It's my hope that these and other qualitative methods for planning curricula can be applied to other legendarily confusing fields of study. These methods can even be applied to entire undergraduate course catalogues and major requirements to guide students over 4+ years of undergraduate study.share this: