Unit of Analysis: Are You Counting Calls, People or Accounts?

By Trellus Team · November 1, 2026

A call export can contain twelve rows and describe only four people. If several of those people work at the same company, it may describe fewer accounts. Before calculating a percentage, decide which of those things your report is about.

The unit of analysis is the entity you want to describe or draw conclusions about: a call attempt, a person, an account, a rep or another clearly defined unit. The unit of observation is what you collect or measure. They can be the same, but they don't have to be. Concordia University's research-methods textbook makes that distinction and ties the analysis unit to the research question.

For a descriptive sales report, start with a sentence: “We want to know what happened to each ___ during ___.” Fill in the entity and the period before choosing a chart. Otherwise, an available export can end up choosing the question for you.

What the rows record and what the report describes

Suppose your source records one activity for each outbound call attempt. A call-level report might ask what share of those attempts led to a conversation. Each attempt contributes once to that calculation.

A different question is whether each intended person had at least one conversation during the week. You can answer it from the same call records, but first you need to group them by person. The call is the observation you collected; the person is the entity the resulting report describes.

At account level, you would group the people under their intended accounts and apply an account-level outcome rule. You haven't acquired more data by moving between levels. You've changed what one result in the report represents.

Keep the words attached to the number. “Four conversation attempts” and “four people reached” sound similar in a hurried review. They can describe very different work. A title such as “People with at least one conversation this week” is more useful than a tile labelled only “success rate.”

Twelve fictional attempts, viewed three ways

Consider a fictional report covering October 19 through October 25, 2026, using UTC dates. It includes the twelve distinct outbound attempts below. Each attempt has a known outcome. A conversation means that the intended person took part in a live conversation; unanswered calls, voicemail and conversations only with a receptionist don't qualify for this example.

The table summarizes the underlying attempt records by person so that the repeated attempts are easy to see. A1, A2, B1 and C1 are invented identifiers, not real prospect records.

At attempt level: four of the twelve attempts led to conversations. The share is 4 ÷ 12, or about 33.3%. All three conversations with A1 count because they occurred on three different attempts.

At person level: two of the four people had at least one conversation: A1 and B1. The share is 2 ÷ 4, or 50%. A1 contributes one reached person, even though that person participated in three conversations.

At account level: two of the three attempted accounts had at least one qualifying conversation: A and B. The share is 2 ÷ 3, or about 66.7%. Account A contributes once, although it contains two attempted people and three conversation attempts.

None of those percentages corrects the others. Each answers its own question. The 66.7% account result doesn't mean that two thirds of calls became conversations, or that two thirds of people were reached. Nor does a conversation establish qualification, a booked meeting or a sale.

This account calculation covers the three accounts attempted in the example. It says nothing about target accounts outside that group. For the difference between attempted-account reach and coverage of an entire target list, use the account reach rate guide.

Keep repeated attempts separate from duplicate records

Calling A1 four times can produce four legitimate attempt records. If your question concerns attempts, don't remove three merely because the person is the same. Doing so would discard work that belongs in the report.

Two copies of one activity are a different problem. An exported row repeated by a join, or the same activity copied from two sources, doesn't establish a second call. Resolve that duplication using the team's source records and correction process before counting distinct attempts.

Then group the valid attempts for the question you chose. For the person-level example, collect the four attempts under A1 and ask whether any qualifies as a conversation. For the account-level example, combine A1 and A2 under A and apply the account rule once.

Keeping both the source rows and the grouped result makes the calculation explainable. Someone reviewing account A should be able to see why it qualifies without mistaking its three conversation attempts for three different reached accounts. You don't need to put names, phone numbers or call notes into another broadly shared spreadsheet to provide that explanation; use the internal references and access your team actually needs.

The grouping rule belongs beside the result

A grouping key is the reference used to recognize the same entity across records. A stable contact identifier is generally more useful for a person-level report than a display name: two people can share a name, and one person's name can be written several ways.

Be just as clear about accounts. Does an account mean the parent company, a subsidiary or a particular operating location? If two locations are separate accounts for the question you're asking, don't combine them merely because their names look related. If the question concerns the parent company, a location-level export needs an agreed mapping before it can answer it.

Don't guess where an unmapped person belongs to make the totals line up. Show the unmapped count and state whether those records are excluded from the account calculation. Keep them available for a person- or attempt-level question if the necessary information is present. The number of records you can use may differ by level.

Missing outcomes need their own treatment too. The worked example deliberately has complete classifications. In a real export, an empty field may leave you unable to decide whether an attempt qualifies. The call outcome completeness guide explains how to review that uncertainty rather than silently turning blanks into “no.”

Changing the level changes what carries weight

In the example, account A supplies six of the twelve attempts. It therefore carries half the weight in an attempt-level percentage. In the account-level calculation, A contributes one of three accounts, just like B and C.

Choose that weighting deliberately. If the question is about the outcome of a randomly selected attempt from this recorded group, the attempt-level calculation fits. If the question is how many attempted accounts had any conversation, each account needs one result. Neither is a general measure of how well every rep performed.

Comparisons also need matching definitions. One team's 50% person-level result cannot be compared directly with another team's 50% attempt-level result. Even at the same level, the lists, outcome rules and reporting periods can differ. A shared label doesn't settle those differences.

Dates can change group membership independently of the unit you choose. A booking-level report still needs to say whether bookings enter by the date they were arranged or the date the meeting was scheduled to occur. The booking-date cohort guide works through that separate decision. Choosing the entity doesn't finish the report definition.

These are descriptive calculations for the records included. They don't establish why a result happened or predict a future list. If you're making a statistical comparison, the relationships among repeated observations need additional attention; a large number of activity rows alone isn't a justification for a confidence claim.

Write a report definition someone else can follow

Before handing over a percentage, save the question, entity, period, grouping reference and outcome rule. Include how duplicates, missing outcomes and unmapped records were handled. Keep the counts beside the percentage so a reviewer can reconstruct it.

For the fictional person-level result, a complete note could read:

Question: which attempted people had at least one qualifying conversation? Unit of analysis: person. Source: twelve distinct call attempts from October 19, 2026 at 00:00 UTC up to, but not including, October 26 at 00:00 UTC. Group by the fictional contact identifier A1, A2, B1 or C1. A person qualifies if at least one attempt includes a live conversation with that intended person. All twelve outcomes are known, all four people are mapped, and no duplicate activity copies are included. Two of four attempted people qualify: 50%. This describes the four people attempted in this example, not the entire target list or a meeting outcome.

To make it an account-level report, change the question, grouping reference, outcome rule and denominator together. Don't just rename the chart. Someone reading the chart should know whether the percentage describes calls, people or accounts.