A complete operating manual for data-driven decision making, synthesized from MGT 6381 (Managerial Statistics, Fall 2024, Yasin Alan, Owen Graduate School of Management, Vanderbilt University) — fourteen classes, three group homework assignments, five online quizzes, five practice problem sets, and the final exam preparation booklet. Approximately 35,000 words across a landing page, ten module sub-pages, and an eleven-page Master Playbook.

The course in one paragraph

Managerial Statistics is Owen's foundational quantitative course — officially named for the subject, unofficially named for the thing the subject is for: Data-Driven Decision Making. The course's central claim is that the bottleneck in modern business is no longer access to data — companies are data rich but information poor — but the manager's ability to partner with quants, ask the right questions, convert technical answers back into managerial actions, and avoid the classical traps that statisticians have spent two centuries cataloging. Across fourteen classes, the course builds a complete toolkit: descriptive statistics, data visualization, sampling theory, confidence intervals, hypothesis testing, A/B testing, simple and multiple regression with diagnostics, logistic regression, and machine learning — ending with data ethics and a clear-eyed treatment of when removing demographic variables from a model can make algorithmic discrimination worse. Every concept is taught with R as the working tool, with a worked managerial case (Commodore Cereals, WSES, Vungle, Compass Maritime, Caesars, QWE, OSHA, Home Credit), and with a recurring discipline: what have you really learned today, and what would you do differently as a manager because of it?

The three anchor phrases

If you remember only three sentences from this course, remember these. Every framework, every formula, every case study is a footnote to one of them.

1. "What you want is to imbue in your organization a decision-making culture that is based on data and information and not on the intuition of hunch." — Gary Loveman, former CEO of Caesars Entertainment, MIT-trained economist who built the modern data-driven casino. This is the managerial north star of the course. It is not a claim that intuition is worthless; it is a claim about the default. When the default in your organization is gut feel, every meeting compresses to the loudest voice. When the default is data, the loudest voice has to bring a number — and any number can be tested, challenged, or contradicted.
2. "If you torture the data long enough, it will confess to anything." — Ronald Coase, Nobel laureate in economics. This is the epistemological warning. Statistics will give you an answer for almost any question; the discipline is in noticing when you have stopped asking the question and started fishing for the answer you wanted. The corollary, delivered by Yasin Alan on day one: never pressure your quants with comments like "see if you can find some evidence in the data to support my idea." The whole edifice of hypothesis testing — pre-registering H₀ before you look at the data, choosing α before computing p-values, separating training data from validation data — is institutional armor against this temptation.
3. "Big data does not eliminate the need for statistical sampling. Indeed, big data may lead to biased samples, conclusions, and decisions." — the course's running counterweight to data-maximalism. Tim Harford's Financial Times essay (assigned reading in Class 1) lays out the four articles of faith of the big-data movement — that big data delivers uncanny accuracy, that N=All eliminates sampling, that correlation displaces causation, and that the numbers speak for themselves — and shows that all four collapse under careful examination. The Literary Digest 1936 poll (n=2.4 million, very precisely wrong); Google Flu Trends (overstated flu by a factor of two as soon as the world's news coverage changed); Boston's Street Bump app (a beautiful pothole map that systematically favored young, affluent neighborhoods with smartphones); and Target's pregnancy predictor (whose famous "hit" rate was inflated by ignoring the false positives) are the four canonical examples. Larger populations do not require larger samples. What samples need is to be representative.

The Owen six-step Data-Driven Decision Making framework

Class 1 introduces the course's organizing operational metaphor: every decision a manager makes with quantitative help passes through six steps. Steps 2-5 are where the technical work lives; steps 1 and 6 are where the managerial work lives, and where the course argues that an MBA's value-add is concentrated.

1. Recognize the problem or question. "Are our marketing expenditures at the optimal level? Perhaps we are spending too much on TV advertising and too little on digital." This step does not have a formula. It is the manager's job. If you skip it — if you start with a method instead of a question — you will get answers to questions nobody is asking.
2. Frame the question as a testable hypothesis. Convert "are we spending too much on TV?" into "the population effect of TV advertising on incremental sales is less than [threshold]" — so that data can speak to it.
3. Choose the right test or model. A confidence interval? A hypothesis test? A simple regression? A logistic regression with interaction effects? An A/B test with random assignment? Modules 5-10 of this masterclass are a decision tree for this step.
4. Estimate the test or model. This is where R lives. The course's working tool is lm(), glm(), t.test(), prop.test(), pnorm(), pt(), qnorm(), qt(), vif(), stepAIC(), predict() — a vocabulary you will fluently command by Class 14.
5. Interpret the output statistically. Was the coefficient significant? Was the R² high enough? Did the residuals show patterns? Did the validation MAPE beat the training MAPE by a suspicious margin (overfitting)?
6. Present and act on the results. "Our analyses indicate that the digital marketing channel has a high return on investment while our TV ads have negligible impact on sales. We recommend reallocating the marketing budget as follows…" This is where the manager re-enters. The presentation cannot be the regression output — it must be the decision, framed in the language of business, with the statistical caveat (margin of error, confidence interval, prediction interval, validation MAPE, false-positive rate) translated into managerial risk.

A recurring pattern in the course: when a case goes wrong, it almost always went wrong at step 1 or step 6, not at steps 2-5. The Bet Performer ship was bid at $125.83M, sold at $133M, and a sister ship sold for $27M nine months later in the post-crash market — the regression was fine; the manager missed the structural break in shipping markets in 2008.

The course arc — fourteen classes, ten modules

The ten modules of this masterclass group the fourteen classes into the natural learning arcs: the four foundations (descriptive statistics, visualization, sampling, CLT) build the vocabulary; the inferential pair (confidence intervals, hypothesis testing) introduce the bridge from sample to population; experimentation (A/B testing) shows how the bridge gets built in modern companies; regression (simple, multiple, validation — Classes 8-11) becomes the workhorse for understanding and predicting; and the closing trilogy (logistic regression, machine learning, data ethics — Classes 12-14) extends regression to binary outcomes, scales it to high dimensions, and finally subjects the whole enterprise to ethical audit.