Measures of Center

What does "typical" performance actually mean?

Every Monday, the sales leadership team at Northwind Analytics, a mid-size business intelligence software company, receives a slide titled "Average Deal Size." Last quarter the number was $67,500. The VP (vice president) of Sales used it to argue that the team was closing "enterprise-grade" deals. The CFO (chief financial officer) used the same number to justify a higher quota for the next quarter. Two weeks later, a regional manager pointed out that eleven of twelve reps closed deals between $15,000 and $45,000, and one whale contract at $500,000 pulled the average up. Quota planning built on the mean would have demoralized most of the team and missed the real story: typical deals clustered near $29,000.

That argument is not a personality conflict. It is a measure of center problem. A measure of center is a single number that summarizes where a batch of observations tends to fall. It answers the executive question, "What is normal here?" But "normal" can mean the arithmetic average, the middle rank, or the most common value. Each definition is legitimate. Each leads to different decisions about quotas, staffing, inventory, and investor messaging. When managers treat one summary statistic as the whole truth, they allocate resources against a fiction.

Unit 1 taught you to classify data before you analyze it. From Types of Data and Measurement Scales, you know that most business KPIs (key performance indicators, the metrics leaders track) are ratio-scale numbers: revenue, units sold, handle time in minutes. Ratio data allows all four measures of center in principle. The harder question is which one matches the decision. This lesson builds that judgment. You will compute the mean, median, and mode; extend the mean to weighted averages when subgroups differ in size; and learn why pooled averages can reverse direction when customer mix shifts, a trap called Simpson's paradox.

The mean: the arithmetic center of gravity

The mean (also called the arithmetic average) is the sum of all values divided by the count of values. If you have n observations labeled x₁, x₂, …, xₙ, the sample mean is:

Mean (x̄) = (x₁ + x₂ + … + xₙ) / n = Σx / n

The Greek letter Σ (sigma) means "sum." The bar over x (x̄, read "x-bar") is standard notation for a sample mean.

The mean treats every observation equally. A $500,000 deal counts exactly five hundred times as much as a $1,000 deal in the numerator. That property makes the mean the right tool when you need a total to scale linearly: if every customer spent the mean amount, total revenue would match actual revenue. Finance teams use means constantly when they multiply "average price × volume" to forecast revenue, or when they compute average cost per unit for a P&L (profit and loss statement, the report of revenue minus expenses over a period).

The mean also has a serious weakness in skewed business data. Skew means the distribution has a long tail on one side. Revenue per customer, deal size, and CEO compensation are almost always right-skewed: many modest values and a few very large ones. In right-skewed data, the mean is pulled toward the tail. It can sit above what most customers or reps actually experience. Reporting only the mean to a board without context invites overinvestment in chasing whales or underinvestment in the core business.

When the mean is appropriate, three conditions usually hold. First, the data are roughly symmetric, or the tail values are economically representative rather than one-off anomalies you would exclude by policy. Second, you care about totals and linear scaling (inventory, payroll, aggregate revenue). Third, you report the mean alongside n (sample size) and ideally alongside a second center measure so readers can see whether skew is distorting the story.

For a symmetric batch like daily call volumes of 98, 100, 102, 100, 100 calls, the mean is 100 and every day feels "about 100." For Northwind's deal sizes with one whale, the mean is not what a typical rep experiences. That gap between mean and median is your first diagnostic for skew.

The median: the middle rank

The median is the value that splits the sorted data so half the observations fall at or below it and half at or above it. For an odd count, it is the middle value. For an even count, it is the average of the two middle values.

To find the median, always sort first. Order statistics are not optional. A common beginner error is to pick the middle number from an unsorted list.

The median ignores how far extreme values sit from the center. It cares only about rank. A $500,000 deal and a $50,000 deal occupy the same position in the ranking logic: they are simply "high." That robustness makes the median the default center for compensation surveys, housing prices, customer spend, and any metric where a few giants should not dictate the story of the typical case.

Compensation committees rarely benchmark to mean pay because a handful of executive packages would inflate the average. They benchmark to median and 75th percentile pay (covered in Lesson 3). Operations leaders staffing a call center use median handle time when a few marathon calls distort the mean. The median answers: "What does the person (or day, or deal) in the middle experience?"

The median has limits. It is not designed for "multiply by count" forecasting. If every customer spent the median amount, total revenue would generally understate actual revenue in a right-skewed business because the mean exceeds the median and the largest customers contribute disproportionate dollars. Smart reporting pairs median (typical experience) with mean (aggregate economics) and states n.

When mean and median diverge by a wide margin, treat that gap as a signal to segment the data before making policy, not as noise to ignore.

The mode: the most common value

The mode is the most frequently occurring value in a dataset. A distribution can have one mode (unimodal), two modes (bimodal), or no clear mode if all values appear once.

The mode shines with nominal and ordinal categories, where mean and median are meaningless. The most common support ticket category, the most frequent product return reason, the peak day-of-week for web traffic: these are mode questions. For continuous dollar amounts, exact modes are rare unless you bin data (group into ranges like "$20k–$25k") or round.

Bimodality often reveals hidden segments. If ticket resolution times cluster at 1 hour and 48 hours, the blended mean might suggest a 12-hour service level is achievable with one process. The mode structure tells you that you actually run two workflows: simple tickets and escalations. Lesson 3 develops shape further; here, note that mode thinking pushes you to ask whether "one average customer" exists at all.

For strictly numerical deal sizes with no repeats, purists say there is no unique mode. In management practice, you may report a modal bin ("most deals landed in the $25k–$35k band") after grouping. Always label that choice explicitly so readers know you binned.

Weighted mean: when subgroups differ in size

A weighted mean averages values where each observation counts in proportion to its weight. If value xᵢ appears with weight wᵢ:

Weighted mean = (w₁x₁ + w₂x₂ + … + wₙxₙ) / (w₁ + w₂ + … + wₙ) = Σ(wᵢxᵢ) / Σwᵢ

The ordinary mean is a special case where every weight equals 1.

Business data is almost never a flat list of equal-importance numbers. Conversion rate across channels must weight by traffic. Gross margin across product lines must weight by revenue. Portfolio return blends asset returns by allocation weights. Using a simple average of subgroup rates without weights is one of the most expensive mistakes in growth analytics.

Simpson's paradox (also called the Yule–Simpson effect) occurs when subgroup trends go one direction, but the pooled average goes the opposite direction because the mix of subgroups changed. The paradox is not magic. It is a reminder that a blended rate is a weighted mean, and weights moved.

From From Business Questions to Data Questions (Unit 1), you learned to decompose business questions before touching data. Weighted means are the arithmetic expression of that discipline: if channels, regions, or customer tiers differ in size, you must carry the weights forward or your "overall improvement" can be a mix illusion.

Trimmed mean: reducing outlier pull without hiding whales

A trimmed mean drops a fixed percentage of the smallest and largest values, then averages what remains. A 10% trimmed mean on 20 observations removes the two lowest and two highest, then averages the middle 16.

Compensation surveys and some macroeconomic series use trimmed means to limit outlier distortion while still reflecting more of the distribution than the median alone. Policy choice matters: trimming 10% of deal sizes for quota setting differs ethically from trimming 10% of employee performance scores. Always disclose the trim rule. A whale deal may be rare but real; excluding it from investor revenue guidance without disclosure crosses into misleading reporting.

For Northwind, a 10% trimmed mean on 12 deals removes one low and one high (conventional rule: trim equal counts from each tail). After removing $15k and $500k, the trimmed mean summarizes the core mid-market band. Use it as a sensitivity check, not a substitute for transparent median and mean reporting.

Choosing and reporting center statistics in management packs

When you present to mixed audiences, structure the narrative in three layers. Layer 1 states the business question in plain language ("What should a new rep expect to close in the first year?"). Layer 2 names the statistic you chose and why ("Median deal size because the distribution is right-skewed with occasional enterprise whales"). Layer 3 shows the number with n, time window, and segment definition ("Median $29k, n=12 Q3 enterprise deals, North America only").

Avoid the phrase "average" without clarification. In conversation, "average" might mean mean, median, or even "typical in my experience." In writing, use mean or median explicitly. If you must use "average" for a general audience, define it once in parentheses.

Cross-functional conflict often comes from each function preferring a different row in this table. Finance needs mean totals. Sales leadership needs median quotas. Product needs modal feature requests. A literate analytics lead publishes all three in one slide with roles labeled, rather than letting each function screenshot its preferred number from different reports.

Link back to Types of Data and Measurement Scales (Unit 1): never compute a mean on nominal data (region codes) or treat ordinal survey labels as precise intervals without caution. Center statistic choice starts with measurement scale, then skew, then decision.

Northwind Analytics closed 12 new enterprise contracts in Q3. Amounts are in $000s (thousands of dollars). The CRO (chief revenue officer) wants one slide for the board. The sales ops analyst must choose center statistics deliberately.

Part A: Setup and fact pattern

Assumptions: All deals are closed-won ACV (annual contract value, one year of subscription revenue). No multi-year prepayments. n = 12. One deal (#12) is a strategic whale in a new vertical.

Total contract value matters for the board; typical rep experience matters for quota and hiring.

Part B: Mean calculation

Step 1: Sum all deal sizes.

15 + 18 + 20 + 22 + 25 + 28 + 30 + 32 + 35 + 40 + 45 + 500 = 810 ($000s)

Check by grouping: (15+45) + (18+40) + (20+35) + (22+32) + (25+30) + (28+500) = 60 + 58 + 55 + 54 + 55 + 528 = 810 ✓

Step 2: Divide by n.

Mean = 810 / 12 = 67.5 ($000s) = $67,500

Interpretation: If every deal were exactly the mean size, total Q3 ACV would still be $810,000. The mean is correct for aggregate revenue math. It is misleading as a label for "typical rep performance" because one observation contributes 500/810 ≈ 61.7% of the sum.

Part C: Median and mode

Step 1: Sort ascending (already sorted in the table).

Step 2: With n = 12 (even), the median is the average of the 6th and 7th values.

6th value = 28, 7th value = 30

Median = (28 + 30) / 2 = 29 ($000s) = $29,000

Mode: No value repeats. Modal bin (width $5k): the band $30k–$35k contains two deals (30 and 32 if binning 30–34 and 35–39); with $5k bins starting at 15, the $25k–$29k band holds 25 and 28, the $30k–$34k holds 30 and 32. The richest bin depends on bin rules. Report: "No single mode; most mass between $20k and $45k except the whale."

Reconciliation check: Median ($29k) << Mean ($67.5k). Gap confirms right skew. ✓

Part D: Managerial read

Board question: "If we exclude the whale, what is run-rate?" Eleven-deal sum = 810 − 500 = 310. Mean of remaining 11 = 310/11 ≈ $28,182, close to the median. Quota at $67k would misalign incentives.

Northwind runs self-serve trials on mobile and desktop. The growth team reports "overall conversion up!" while each channel's conversion fell. The CMO (chief marketing officer) needs the weighted math before reallocating ad spend.

Part A: Month 1 (baseline)

Step 1: Converted trials per channel.

Mobile: 8,000 × 0.040 = 320

Desktop: 2,000 × 0.060 = 120

Step 2: Pooled conversions.

Total converted = 320 + 120 = 440

Step 3: Weighted mean conversion.

Blended rate = 440 / 10,000 = 0.044 = 4.4%

Formula check: (8,000×0.040 + 2,000×0.060) / 10,000 = (320 + 120)/10,000 = 4.4% ✓

Month 1 story: mobile-heavy mix (80% of trials) pulls blended rate below desktop's 6%.

Part B: Month 2 (mix shift)

Traffic shifts to 50/50 after a mobile app campaign. Each channel's conversion worsens (product friction on mobile, desktop promo ended):

Mobile converted: 5,000 × 0.035 = 175

Desktop converted: 5,000 × 0.055 = 275

Total converted = 175 + 275 = 450

Blended rate = 450 / 10,000 = 0.045 = 4.5%

Paradox: Blended conversion rose from 4.4% to 4.5% while mobile fell (4.0% → 3.5%) and desktop fell (6.0% → 5.5%). ✓

Cause: Weights shifted toward desktop, the higher-yield channel, even at lower absolute yield than Month 1 desktop.

Part C: Wrong analysis (unweighted trap)

Simple average of channel rates in Month 2: (3.5% + 5.5%) / 2 = 4.5%. Accidentally correct here because weights are equal. In Month 1, that wrong method gives (4.0 + 6.0)/2 = 5.0%, which is not the true 4.4% blended rate. Never average rates without weights.

Part D: Managerial read

Operator decision: Segment dashboards by channel always. Overall conversion is a mix index, not a pure performance index.

Investor / board narrative: "Conversion improved 0.1 points" is false progress if both channels degraded. Report channel rates, trial mix, and weighted total.

Action: Fix mobile friction (3.5%) before celebrating blended 4.5%.

Harbor Logistics recorded weekly parcel volumes (thousands of packages) for six distribution centers last month:

Tasks:

1. For Center B, compute the mean and median weekly volume. Which better describes "typical" week? Explain in two sentences.
2. Harbor reports company-wide average weekly volume per center as the unweighted mean of the three center means. Compute that number. Is it equal to the weighted mean of all 12 weekly observations? Show both calculations.
3. Center B's week 3 spike (50) was a one-time retail event. Compute a 25% trimmed mean for Center B (one lowest, one highest removed). Compare to the median.

Solution

1. Center B mean and median

Sorted Center B: 8, 9, 10, 50

Mean = (8 + 9 + 10 + 50) / 4 = 77 / 4 = 19.25 (000s) = 19,250 packages

Median = (9 + 10) / 2 = 9.5 (000s) = 9,500 packages

The median better describes a typical week because the mean is pulled upward by the 50k anomaly week. Operations should staff for roughly 9,500–10,000 packages, not 19,250.

2. Unweighted vs weighted company average

Center A mean = (10+12+11+13)/4 = 46/4 = 11.5

Center B mean = 19.25 (above)

Center C mean = (20+22+21+23)/4 = 86/4 = 21.5

Unweighted mean of center means = (11.5 + 19.25 + 21.5) / 3 = 52.25 / 3 ≈ 17.42 (000s)

Weighted mean of all 12 weeks:

Sum = 46 + 77 + 86 = 209 (000s)

Weighted mean = 209 / 12 ≈ 17.42 (000s)

Check: In this balanced design (4 weeks per center), unweighted mean of group means equals overall mean. ✓ If weeks per center differed, they would diverge.

3. Trimmed mean for Center B

Remove lowest (8) and highest (50). Remaining: 9, 10

Trimmed mean = (9 + 10) / 2 = 9.5 (000s), equal to the median here.

Task: A CS (customer success) team tracks time to first response (minutes) for 10 tickets: 5, 6, 6, 7, 7, 8, 9, 10, 12, 120.

1. Compute mean and median.
2. The SLA (service level agreement, a promised performance standard) says "average response under 15 minutes." Does the team meet SLA on mean? Should the SLA use mean? Explain.
3. Compute the 10% trimmed mean (remove one min and one max).

Solution

1. Sorted list given.

Sum = 5+6+6+7+7+8+9+10+12+120 = 190

Mean = 190 / 10 = 19 minutes

Median = (7 + 8) / 2 = 7.5 minutes

2. Mean = 19 exceeds 15, so the team fails a mean-based SLA. Yet nine of ten tickets were under 15 minutes. Mean is the wrong SLA center for skewed response times because one 120-minute outlier dominates. Median or 90th percentile (Lesson 2) aligns with customer experience.

3. Remove 5 and 120. Remaining sum = 190 − 125 = 65, count = 8

Trimmed mean = 65 / 8 = 8.125 minutes ≈ 8.1 minutes, closer to median than raw mean. ✓

Lesson 1 without Lesson 2 produces false confidence. Example recap using Northwind Q3 deals ($000s): mean 67.5, median 29, range 485, SD among the eleven non-whale deals ≈ 9.8 (tight core). A slide showing only mean invites overquota; a slide showing only median understates dollar volume. The executive pack needs:

Spread among core deals is low; spread including whale is enormous. Segment enterprise hunters vs core reps before applying one SD to quota variance.

Key takeaways

• The mean balances totals; the median balances ranks. Report both for skewed business metrics.
• Weighted means are mandatory when combining rates or averages across unequal-sized groups.
• Simpson's paradox: pooled improvement can hide subgroup decline when mix shifts.
• Always state n; center statistics without context mislead boards and teams.
• Choose the center statistic to match the decision: forecast totals (mean), typical experience (median), common category (mode).

1. Pull one skewed KPI from your company (deal size, ticket value, session length). Compute mean and median; note the gap and what decision each would imply.
2. Find a blended rate (conversion, margin, on-time %) that could hide Simpson's paradox if segment mix shifts. List the segment rates you would monitor separately.
3. Continue to Lesson 2: Measures of Dispersion.

Unit bridge: Lessons 1–5 share Northwind examples so you can trace one business from mean deal size through dispersion, latency percentiles, honest charts, and dashboard layout. Carry that continuity into your own company: pick one KPI and walk it through all five lenses this week.