Comparing and explaining student cases

This is a workshop, not a lecture. Three seller profiles are placed on the table, and the task is to compare them without flattening their differences. The profiles are teaching fiction written for this course; they are not real students, shops or transactions. That matters because the lesson is about method, not about discovering which real business is better. By the end, each group must produce a comparison table, a set of labelled claims, two alternative explanations and one bounded conclusion. The quality of the conclusion depends on the discipline of the steps before it.

The first temptation is to compare headlines. One seller reports monthly revenue, another reports sales per livestream, a third reports weekly orders. The numbers look commensurable because they are all money or counts, but they answer different questions. Scope alignment fixes this: who is the party, which market, which product, over what period, measured in value or units, and whether returns are deducted. A comparison that skips these six choices will produce a conclusion that cannot be checked. The §1.3 contrast between ecommerce and traditional commerce is useful here: online versus offline is one dimension, not a verdict about efficiency.

The second step is the four-dimension table. Parties: is the seller a company, an individual or a platform, and is the buyer a consumer or a business? Ownership: who holds inventory, who sets price, who owns the customer relationship and the data? Channel: how does the buyer discover the offer, and where does the transaction complete? Occasion: is the purchase planned or impulsive, and what payment and delivery context surrounds it? The four dimensions are not a causal model. They are a way to keep unlike cases from collapsing into one another, and to see which difference is likely to matter.

Then attach evidence. Case A has paper sales books, supplier receipts and the owner's account; Case B has a platform backend export; Case C has chat logs and market receipts. These are not equally strong for every claim. A backend export is strong for orders and returns inside that store, but silent about why customers bought. An owner's account is strong for decisions and weak for prevalence. A chat log is strong for what was said and weak for whether the customer would have bought elsewhere. The course uses a simple grading: first-hand records, operator-published data with a stated scope, direct observation or interview, media relay, and finally model or inference, which is not evidence of fact.

The third step separates fact, inference and judgement. A fact is verifiable within a stated scope: B's backend shows 1,020 orders in March. An inference explains: B sells more because the host is persuasive. A judgement says what should be: A should close the shop. The three can appear in one paragraph, but they must be labelled, because each is tested differently. Facts are checked against records; inferences against alternative explanations and data; judgements against values and trade-offs. When a judgement is written as a fact, discussion becomes a contest of preferences.

The fourth step is the alternative explanation. B's volume could come from lower prices, a returning customer base, platform traffic support, seasonal demand for home textiles, or host skill. Each has a test: average order value, share of new versus returning customers, traffic source, the same host selling a different category, and the promotion calendar before and after the session. If two explanations predict the same observation, the case is underdetermined and the honest answer is that the evidence cannot decide. This is not a failure of the analysis; it is the analysis working.

Five errors recur. Single-cause attribution turns one plausible reason into the reason. Survivorship bias studies only the sellers still trading. Level mixing compares a platform's growth with a shop's revenue. Scope drift changes the denominator midway, for example comparing A's monthly revenue with B's best session. Precision bias treats a decimal as proof. Each error can be corrected by a question: what else could produce this, who is missing from the sample, at what level is the claim, is the denominator constant, and what is the margin of error?

A bounded conclusion names its limits. A useful template is: for this type of seller, in this market, selling this product through this channel, during this period, the evidence supports this claim; it does not support that larger claim. For example: for a four-person team selling home textiles on one platform in 2026, three two-hour sessions a week covered ad spend and returns; this does not show that livestreaming outperforms physical shops in Dhaka. The second clause is not decoration. It prevents a local observation from becoming a universal law.

The workshop runs in five moves: read the profiles, align scope, fill the four dimensions, label claims and write alternatives, then draft the conclusion. Groups of four divide roles - one keeps the scope, one fills the table, one checks evidence, one writes the conclusion - and then exchange sheets. Peer review asks only two questions: which claim is unlabelled, and which conclusion is unbounded? The teacher does not supply the answer; the teacher asks for the missing test.

The assessment rewards process, not rhetoric. Scope alignment carries two points, the four dimensions two, fact-inference separation two, alternatives with tests two, and the bounded conclusion two. A group that writes a modest, well-bounded answer scores higher than one that writes a confident, unbounded one. This is also the habit needed outside the classroom: most claims about ecommerce are local observations travelling without their scope. The workshop trains the reader to ask where a number came from before asking what it means.