Thinking with data as a manager

Data Analysis for Managers is the fourth course in the first term of the weekend executive MBA at the Lakeshore School of Business in Chicago. It is taught by Dr. Mei-Ling Zhou, a statistician who led analytics teams at a national bank and at an airline before she moved into teaching. She starts the first Friday session by putting a single slide on the screen. It shows a dashboard with forty colored tiles, each one green, amber or red, and she asks the five managers in the room a simple question: which decision would you make differently tomorrow because of this screen? Nobody can name one. That, she says, is the problem this course is meant to solve.

The cohort is the same group that has worked through accounting in MBA-501, economics in MBA-502 and leadership in MBA-503. Daniel Osei is the plant controller at Northfield Components, a maker of metal stampings and welded assemblies. Hana Kowalski runs operations for Lakeview Health, a not-for-profit group of clinics and a community hospital. Leo Brandt is co-founder and chief financial officer of Kestrel Analytics, a software company that sells inventory forecasting to retailers. Grace Mwangi is a category manager at Fairway Grocers, a regional supermarket chain. Owen Tremblay is finance manager at Prairie Ridge Energy, which runs gas processing plants and a growing fleet of wind and solar farms. All five have more data than they can use. None of them, they admit, is confident about what the numbers are telling them.

Dr. Zhou makes three promises about the course. It will not turn the managers into statisticians, but it will make them demanding customers of analysis, able to ask the questions that expose a weak argument. It will be built around decisions rather than techniques: every method is introduced because a manager needs it to make a better choice. And it will be honest about uncertainty. Most of the damage done by data in organizations, she argues, comes not from complicated mathematics but from simple errors made confidently: confusing a correlation with a cause, reading noise as a trend, ignoring how a sample was chosen, or treating a forecast as a fact.

Start with the decision, not the data

The first habit of a manager who thinks well with data is to start from the decision. Before asking for an analysis, write down what you will do differently depending on the answer. If no plausible result would change the action, the analysis is not worth commissioning, however interesting it sounds. Hana describes a recent request from her board for a report on patient satisfaction by clinic. When Dr. Zhou asks what the board would do with it, Hana realizes that the board has no lever: clinic staffing is set by a provincial funding formula, and the board does not choose which clinics to keep open. The useful question is narrower. Which of the things Lakeview does control, such as appointment reminders, booking hours and the triage process, are associated with complaints, and would changing them be worth the cost?

A good analytical question has four parts. It names the decision and the person who will make it. It states the options being compared. It specifies the measure of success, such as cost, revenue, wait time or risk. And it sets a deadline that reflects when the decision must be taken, because a perfect answer delivered after the decision is worth nothing. Dr. Zhou calls this the decision brief, and she asks each manager to write one on a page before any data is pulled.

Data are measurements made by someone, for some purpose

Every number in an organization was produced by a process: someone decided what to count, how to record it and when. Before trusting a figure, a manager should know how it was generated. Grace gives an example from Fairway. The chain's reports show that shrink, the value of stock lost to damage, spoilage and theft, is far lower at its newest stores. On closer inspection, the newer stores use a different handheld scanner that only records spoiled produce when a supervisor approves the write-off. The measured difference is partly a difference in measurement. Dr. Zhou calls these data-generating processes, and she says that understanding them is worth more than any technique taught later in the course.

Three questions help. First, what population does the data describe, and who is missing? Customers who answer a survey are not the same as all customers, and patients who complete their treatment are not the same as all patients who start it. Second, how was each value measured, and could the measurement have changed over time or between groups? A new definition of an active user, a switch in accounting systems or a change in who records the data can create apparent trends. Third, why was the data collected? Data gathered for one purpose, such as billing, often fits awkwardly when used for another, such as measuring quality of care.

Variation, signal and noise

Almost every business number varies from week to week for reasons nobody can identify. Sales at a store differ between two Tuesdays because of weather, a local event, a delayed delivery or nothing anyone could name. The manager's job is to separate signal, a change that reflects something real and lasting, from noise, the background variation that would occur anyway. The most common error is to react to noise: praising the store manager whose sales rose last week, criticizing the one whose sales fell, and then watching both numbers drift back toward the average. The statistician Francis Galton described this drift, regression to the mean, in the 1880s when studying the heights of parents and children, and it is everywhere in management.

Regression to the mean has a direct implication for evaluating interventions. If Northfield sends its three worst-performing production lines for extra training after an unusually bad month, those lines are likely to improve the next month whether or not the training works, simply because an unusually bad month is unlikely to be repeated. Daniel recognizes the pattern from a consultant's report that claimed a 30 percent improvement in defect rates after a quality program aimed at the worst lines. Without a comparison group of similar lines that did not get the program, the claim cannot be separated from regression to the mean.

# Regression to the mean with no intervention at all
import random
random.seed(504)
stores = 40
true_level = [random.gauss(100, 6) for _ in range(stores)]      # each store's real weekly sales level
week1 = [t + random.gauss(0, 8) for t in true_level]               # noise on top of the real level
week2 = [t + random.gauss(0, 8) for t in true_level]

order = sorted(range(stores), key=lambda i: week1[i])
bottom, top = order[:8], order[-8:]
avg = lambda xs: sum(xs) / len(xs)
print(f"Bottom 8 stores: week 1 average {avg([week1[i] for i in bottom]):.1f}, week 2 average {avg([week2[i] for i in bottom]):.1f}")
print(f"Top 8 stores:    week 1 average {avg([week1[i] for i in top]):.1f}, week 2 average {avg([week2[i] for i in top]):.1f}")
print(f"All stores:      week 1 average {avg(week1):.1f}, week 2 average {avg(week2):.1f}")

Bottom 8 stores: week 1 average 83.5, week 2 average 97.5
Top 8 stores:    week 1 average 114.7, week 2 average 103.4
All stores:      week 1 average 98.6, week 2 average 100.4

The simulation gives every store a fixed underlying level of sales and adds random weekly noise. Nothing changes between the two weeks. Yet the bottom eight stores appear to improve and the top eight appear to decline, because stores that looked extreme in week one were partly unlucky or lucky. A manager who had sent the bottom stores on a training course in between would have seen an apparent success. The only defense is a comparison: what happened to similar stores that did not receive the intervention?

Questions to ask of any analysis

Dr. Zhou gives the cohort a short list of questions to ask whenever someone presents an analysis. Compared with what? A number on its own means little until it is set against a baseline, a target, a previous period or a comparison group. How big is the effect in units that matter? A change that is statistically detectable may be too small to justify action, and a large effect may be estimated so imprecisely that the true value could be zero. How uncertain is the estimate? Every estimate from a sample comes with a range of plausible values, and a single number presented without one invites overconfidence. What else could explain the result? Alternative explanations, especially selection, confounding and chance, should be ruled out before a causal story is accepted. And what would change our mind? A conclusion that no evidence could overturn is a belief, not an analysis.

Leo, whose company sells analytics to retailers, admits that his own sales team rarely asks these questions of the results they show clients. A typical slide reports that customers using Kestrel's forecasting tool reduced stockouts by 18 percent. Compared with what? Retailers who chose to buy the tool may already have been better managed than those who did not, and stockouts fell across the industry that year as supply chains recovered. The honest claim is narrower and less impressive, but it is the claim that will survive a skeptical buyer's analysis.

Data sources for managers

Many of the most useful numbers for a manager are published free by statistical agencies. The Census Bureau publishes data on population, business formation, retail sales and much more, including a regular survey that asks businesses whether they are using artificial intelligence; it has reported use rising but still concentrated in larger firms and a minority of businesses overall. The Bureau of Labor Statistics publishes employment, wages and prices, and the Federal Reserve Bank of St. Louis gathers thousands of series in its free FRED database. Using an official series as a benchmark, for example comparing Fairway's sales growth with regional grocery sales, is one of the simplest ways to answer the question compared with what.

Internal data has the advantage of being specific to the business and the disadvantage of reflecting the business's own blind spots. Transaction records show what customers bought, not what they wanted and could not find. Complaint logs record the people who complained, not the larger number who left quietly. Survey data show what people say, which may differ from what they do. A manager who combines internal records with external benchmarks and a small amount of carefully designed new data, such as a test or a short survey, usually learns more than one who searches ever larger internal databases.

The course ahead

The rest of the course follows the path of a typical analysis. Lesson 2 covers describing data well, because a good chart often answers the question before any test is needed. Lessons 3 to 6 deal with probability, sampling and testing, including the experiments that companies such as Leo's now run constantly. Lessons 7 to 9 turn to relationships between variables: correlation, causation and regression, the workhorse of business analysis. Lesson 10 is about forecasting, and lesson 11 about decision analysis, which brings uncertainty and value together in a single framework. The capstone in lesson 12 returns to Cedarline Outdoor, the apparel company the cohort analyzed in MBA-501, MBA-502 and MBA-503, and asks for a recommendation grounded in its data.