What is a data dictionary, and why should funder and partner share one?

In short: a data dictionary writes down what each reported number means, where it comes from and how it may be added up. When the funder and every funded partner use the same one, a number from one partner means the same thing as a number from another.

The reporting agreement lists the metrics. The dictionary gives each a definition someone else can apply without guessing, so "placed in a job" cannot mean 90 days in one report and six months in the next.

Build each entry in four layers. Start with the metric and its plain definition. Link it to a dimension of impact, which says what kind of evidence it is. Add a standard code if one fits. Then say how it rolls up across partners.

In a fictional impact fund, the portfolio manager and investee develop the theory of change and agree core metrics during the onboarding call. Existing documents inform that conversation; a consented transcript or confirmed notes supports drafts that both sides review. The approved dictionary then supplies the definitions for recurring collection and donor or LP reporting.

Auroville International USA channels more than $1M a year to about 80 project partners, who report every quarter. It defines terms such as a "complete quarterly report" and a "clean-water beneficiary" once, so five different water projects count the same person the same way.

What does every dictionary entry need?

In short: eight things: a plain definition, the collection point, the data type, the unit, the disaggregation, the source, a missing-value rule and an owner. If a colleague can apply the entry to real records and get your count, it is complete.

The definition says who is counted, what qualifies and when. Collection point and source name the form the value comes from. Data type and unit stop one partner sending a percentage while another sends a headcount.

The missing-value rule matters most. A person with no follow-up answer is unknown, not "not placed". Keep unknown separate from no, and report how many are unknown. The owner is the person who approves a change to the definition.

How do the Five Dimensions of Impact fit in?

In short: use them as a guide to what kind of evidence each metric is, not as a form to fill for every row. They show which questions your list already answers and which it leaves open.

The Five Dimensions of Impact, maintained by Impact Frontiers, ask five questions of any outcome. Tag each dictionary entry with the one or two it answers.

"How much" covers scale, depth and duration, so one headcount cannot answer it alone. If no metric touches Contribution or Risk, write that down as a known gap rather than adding metrics nobody will collect.

Do you need IRIS+ codes in your dictionary?

In short: no. A standard code is optional. Add one when a funder asks for it or when you want to compare with others who use the same code, and only when your definition truly matches the standard's.

IRIS+, from the Global Impact Investing Network, is a catalog of metric definitions with codes. An IRIS+-aligned dictionary for a job training program uses its metrics for counting clients.

A code is a label, not a definition you can skip. Read the standard's own wording before you map to it. If your "placed in a job" uses a 90-day window and a standard does not, record the difference beside the code.

How do you build a data dictionary from your agreement?

In short: take each metric in the agreement, write its entry in the four layers, then test it on a handful of real records. Stop when every agreed metric has an entry; do not add metrics the agreement does not need.

Test with awkward records: a person who enrolled twice, a missing follow-up, a late answer.

Here is a starting dictionary for the fictional workforce fund, drawn from its agreement. Copy the columns into a spreadsheet and replace the rows with your own.

If the agreement lives in a table, an AI tool can draft these rows. Treat the draft as a list of questions to settle.

Below is the metrics table from our reporting agreement between [FUNDER] and [FUNDED PARTNER].

Draft one data dictionary row per metric. For each row give:
metric name, plain definition (who is counted, what qualifies, when),
collection point, data type, unit, disaggregation, source,
missing-value rule, owner, Five Dimensions of Impact tag (one or two of
What, Who, How much, Contribution, Risk), and roll-up rule across partners.

Rules:
- Use only what the agreement says. Where it is silent, write "ASK: [question]".
- Do not suggest an IRIS+ code unless I ask. If I ask, quote the code and its
  name and say where our definition differs.
- Missing answers are "unknown", never zero or "no".
- Return a table, then a short list of the ASK questions.

[PASTE AGREEMENT TABLE]

How do you share the dictionary and keep it current?

In short: the funder sends the same dictionary to every funded partner before the first report, and the team keeps a short change log whenever a definition moves. Shared definitions are what make adding up possible.

For the funder, the dictionary is the ruler for every report that arrives. It is how you check each partner report against the agreement, and it decides which numbers you may roll up and benchmark. In the fictional fund, Partner C reported placements within six months, so its count cannot be added to the 90-day counts from A, B and D until C confirms.

For the funded partner, the dictionary tells you what each funder means. When several funders use different words for the same count, map each ask to one entry as shown in map every funder's ask to one evidence base.

Definitions change. Keep a change log as a team practice: the date, old and new wording, who approved it and which reports are affected. A 30-day and a 90-day follow-up are different metrics, so give the new one its own row rather than overwriting the old.