River
Y CombinatorBacked by Y Combinator
Log InStart Free

Free template

Program Evaluation Plan Template

An evaluation plan, findings report and recommendation note, plus sheets that count every participant and pair each finding with what else could explain it.

Edit with AIDownload Word and CSV

Free download · No account needed

What you get

Pack of 6 files: 3 documents, 3 spreadsheets

Program Evaluation Plan Template
  • Evaluation PlanDocument
  • Findings ReportDocument
  • Recommendation NoteDocument
  • Outcome AnalysisSpreadsheet
  • Participation and Dose AnalysisSpreadsheet
  • Limitation RegisterSpreadsheet

How to write a program evaluation plan

A program evaluation plan sets out what you want to learn about a program and how you will find out, before anyone looks at the results. A good one names the decision the evaluation will inform and the few questions it must answer. It states the design and what it cannot prove, which measures you will take and when, how you will count everyone who enrolled, and the limitations you expect.

Write and date it before you see any outcome data. A plan written after the results can be bent to fit them, even without anyone meaning to.

A workshop leader points to handwritten planning notes on a flip chart while seated participants listen
Agree on the questions and the limits as a team, before the data arrives. Photo: Centre for Ageing Better on Unsplash

Start the evaluation plan with a decision

An evaluation earns its cost when someone will act on the answer: a board deciding whether to expand a course, or a funder weighing a renewal. The University of Kansas Community Tool Box advises that you begin by asking what you actually need to learn and what decisions the information will help people make1. Put that decision, its owner and its deadline on the plan's first page.

Then narrow "Does it work?" into questions this year's data can answer. CDC's 2024 evaluation framework says the number of evaluation questions should be limited to those that can be answered in a timely manner, with the resources available2.

Broad aims made answerable, for a six-session diabetes course

Question typeToo broadAnswerable
Reach"Are we reaching the right people?""What share of enrollees have the recent diagnosis the course is designed for?"
Delivery"Is the course running well?""Did every cohort get all six sessions, on schedule, with the planned content?"
Outcome"Does the course work?""How much do completers' knowledge-check scores rise between session one and session six?"
Mechanism"Do the support calls help?""Do people who get more calls gain more than people who started at the same score?"

Choose a design that fits the claim

Decide early whether you are measuring change or testing a cause. US federal rules for AmeriCorps grantees draw the line: performance measurement tracks how participants change but is not intended to establish a causal relationship3 between the program and the outcome. Evaluation, as the same rules define it, compares outcomes with what would have happened without the program.

Evaluators call that what-if the counterfactual, and a single group can't supply it. The What Works Clearinghouse, part of the US Department of Education, is blunt about this. Under its standards, a study with no comparison group cannot illustrate a counterfactual4, even with repeated tests before and after.

Four common designs and what each can show

DesignWhat it can showWatch for
Post-test onlyWhere participants ended upNo starting point, so no measure of change
One group, before and afterHow much participants changedRetesting, time or regression to the mean can mimic a gain
Before and after, with a comparison groupWhether participants changed more than similar people who didn't take partGroups that started at different levels
Random assignment, such as a lottery for placesWhether the program caused the differencePeople leaving one group more than the other

The Community Tool Box notes that not every evaluation requires an experimental design1. Choose the strongest design your question, ethics, data, timeline and resources allow, and say plainly what it can't show.

Plan each measure and count everyone

For each question, write down the measure, the tool that collects it, when and by whom. Then set dates for the analysis and for the decision itself. If your outcomes aren't defined yet, draft a logic model first.

Ask for records with one row per person, not summary counts. A tally of "people who improved" can't be split later by attendance or starting score, and both splits matter below.

Plan to count people at three points: enrolled, completed and measured. People who leave early may differ from those who stay. In this page's template, 117 of 331 enrollees (35.3%) never reached the post-test, so every result describes only the other 64.7%.

Protect the people in your data

Decide who can see names, where records are kept and for how long, and get participants' consent. If you use health records, or a university partner plans to publish, check privacy and research ethics rules before collecting anything.

Write the analysis plan before the data

Before anyone opens the outcome data, write down the comparisons you will run, how you will split the results and what pattern would count as a finding. If the plan changes later, the American Evaluation Association's principles ask evaluators to record and promptly communicate any changes to the original evaluation plans5, with the reasons.

Name each limitation as a mechanism, not a disclaimer. "Results may vary" commits to nothing. "Staff call whoever seems to be struggling, so the high-call group starts lower" is a claim you can check.

That limitation matters because of regression to the mean. Adrian Barnett and colleagues, in the International Journal of Epidemiology, call it a statistical phenomenon that can make natural variation in repeated data look like real change6. Unusually low scores tend to be followed by scores closer to the average. So people picked for extra help because they scored low tend to improve on retest, even if the help does nothing.

The plan therefore needs a baseline check: compare the starting scores of the groups that got the most and least help. The What Works Clearinghouse says a study whose groups start more than 0.25 standard deviations apart does not demonstrate equivalence4, even after statistical adjustment.

ExampleA pattern checked against the baselineAn illustrative example written for this guide, from the template's made-up clinic.

Question: among 214 people who finished a six-session diabetes course, do those who got more support calls gain more on a 20-point knowledge check?

Pattern: average gains of 3.1, 4.4 and 6.0 points for people who got 0-1, 2-3 and 4-5 calls.

Check: the 4-5 call group started at 10.1 and the 0-1 call group at 14.2, a gap of 4.1 points. By the post-test the gap was 1.2 points, so 70.7% of it had closed.

  • Not equivalent. With a standard deviation of about 3.5 points, the 4.1-point gap is about 1.2 standard deviations, more than four times the 0.25 limit.
  • Two other explanations. Regression to the mean could close part of the gap with no help at all. And the low-call group, starting at 14.2 of 20, had less room to grow.
  • What the report says. The pattern and the baseline gap in the same paragraph, with no claim that the calls caused the gain.

Pair each finding with its limits

When results come in, put each finding and its limitation in the same paragraph, not in a section at the back. The American Evaluation Association asks evaluators to make clear the limitations of the evaluation approach, findings, and recommendations5. A reader who finds an explanation you left out has reason to doubt the rest of the report.

Keep the recommendation separate from the causal claim. A cheap step, such as calling the people who now get one call or none, can go ahead on the pattern alone. Testing whether the calls work takes a design change. Next cycle, give a random half of each new cohort the full call schedule and the rest a lighter one.

Before anyone analyzes the data

  • The plan names the decision it informs, who makes it and by when
  • The design is named, along with what it cannot show
  • Every measure has a tool, a time and a person who collects it
  • Enrolled, completed and measured will all be counted and reported
  • Any help that staff assign by judgment is flagged for a baseline check
  • The plan is dated, and any later change will be logged with its reason

For participant surveys, the survey results report tool writes up tables with the count behind each figure. The nonprofit impact report template carries results into your yearly report to funders. This page's template has a made-up clinic's evaluation filled in, and Edit with AI opens it in River to build yours.

How to use this template

Use the files as they are, or open them in River. Every file comes filled in with a made-up clinic's evaluation of its diabetes course, so you can see a finished plan, findings and limitation register before you start yours. With Edit with AI, River asks which decision the evaluation must inform and whether any comparison group exists. It also asks for your outcome data, one row per participant.

It writes and dates the Evaluation Plan before analyzing any results, then builds the completion funnel and checks that no count is larger than the one before it. Ready-made prompts then split results by how much help people got, check the baseline gap, count early leavers and pair each finding with what else could explain it.

  1. Download it or open it in River

    Get the three Word documents and three CSV sheets free, or open the template in River to evaluate your own program.

  2. Write the plan first

    Name the decision, the questions, the design and its limits, then date the Evaluation Plan before anyone opens results.

  3. Count, split, then check

    Build the funnel from enrolled to measured, split gains by how much help people got, then compare their starting scores.

  4. Report findings with their limits

    Give each finding a row in the Limitation Register, then write the Findings Report and the Recommendation Note.

What's in the template

The pattern, and the check that keeps it honestGain by dose band, the baseline gap between the extremes, and every finding paired with a quantified alternative explanation.
Program Evaluation Plan Template
RiverOutcome Analysis

Outcome Analysis

Illustrative, for a fictional organisation called Cedar Line Community Health. One cohort of a six-session diabetes education course, one knowledge check administered before session one and again after session six.

Completion funnelCountShare
Enrolled331100.0%
Completed all six sessions21464.7%
Paired pre/post knowledge-check score214100% of completers
Knowledge-check score, of 20 pointsValue
Pre-test mean11.9
Post-test mean16.5
Gain4.6, 23.0% of the scale
Direction of change, among the 214 measuredCountShare
Improved17280.4%
No change3014.0%
Declined125.6%

4.6 points is the number a report would lead with. It is also an average of three very different bands, and the next tab is where the range comes from and what it does to the finding.

RiverDose Analysis

Participation and Dose Analysis

Support calls between sessions are not randomly assigned. Staff informally prioritise whoever seems to be struggling most.

Dose bandnPrePostGainShare
0-1 calls reached5214.217.33.124.3%
2-3 calls reached8812.016.44.441.1%
4-5 calls reached7410.116.16.034.6%
All measured completers21411.916.54.6100.0%
The check: baseline gap vs. post-test gapValue
Baseline gap, 0-1 calls minus 4-5 calls4.1 points
Post-test gap, same two bands1.2 points
Gap closed by post-test2.9 points, 70.7% of the baseline gap
Exclusion accountingCountShare of enrolled
Enrolled331100.0%
Completed and measured, above21464.7%
Early leavers, never reached a post-test11735.3%

The 4.1-point baseline gap is about 1.17 standard deviations on an estimated 3.5-point scale spread, more than four times the 0.25 standard deviation gap above which the What Works Clearinghouse says a study does not demonstrate equivalence. A gap that size closing by 70.7% at post-test is consistent with regression to the mean, which can close part of it with no programme effect at all. The dose-response pattern on the previous table and this baseline check are the same finding read two ways, and a report that shows only the first one is not lying. It is choosing not to look.

RiverLimitation Register

Limitation Register

Written before the narrative. Every finding gets a row naming what else could produce it and, wherever the data allows, a number that bounds it.

FindingWhat it says
F-1Average gain across 214 measured completers: 4.6 of 20 points
F-2Gain rises with support calls reached: 3.1, 4.4, 6.0 points across three bands
F-3No completer, in any cohort, has a six-month clinical measure on record
LimitationConstrainsQuantified bound
L-1F-1, F-2No comparison group of any kind exists; a single group with repeated pre/post tests cannot illustrate a counterfactual under What Works Clearinghouse group design standards
L-2F-2Calls are staff-assigned, not random; baseline gap of 4.1 points (about 1.17 SD) between the extreme bands is over four times the WWC's 0.25 SD limit for demonstrating equivalence, and 70.7% of it closed by post-test, consistent with regression to the mean
L-3F-1, F-2117 of 331 enrollees, 35.3%, never reached a post-test and are absent from every finding above
L-4F-3Not an alternative explanation; this limitation is the finding itself. No source field for it exists in any cohort's data

A funder's own evaluator checks for exactly this table first. Four rows, each naming a specific mechanism and a number, is what lets the dose-response finding stay on the page instead of getting cut in the review that would otherwise catch it.

  • Evaluation PlanThe question, the design and what it can and cannot show, dated before any results are analyzed.
  • Findings ReportEach finding with its limitation in the same paragraph, including what the data cannot answer at all.
  • Recommendation NoteWhat to do next, kept apart from the causal claim, and how to test the finding properly next cycle.
  • Outcome Analysis sheetThe completion funnel from enrolled to measured, before and after scores, and how many improved or declined.
  • Participation and Dose Analysis sheetGains by number of support calls, the starting-score gap between groups, and every early leaver counted.
  • Limitation Register sheetEvery finding paired with what else could explain it and, where the data allows, a number that bounds it.

Common questions

What should a program evaluation plan include?

A program evaluation plan should name the decision the evaluation informs and the questions it will answer. It then describes the program briefly, states the design and what it cannot show, and lists each measure with its timing and source. It ends with how everyone enrolled is counted, the analysis plan, expected limitations, data protections and a timeline.

What is the difference between monitoring and evaluation?

Monitoring, also called performance measurement, tracks what a program delivers and how participants change over time. Evaluation goes further and asks how well and why the program works. An impact evaluation tests whether the program caused the change, by comparing participants with similar people who did not take part.

Can you evaluate a program without a comparison group?

Yes, but only to describe change and delivery, not to show the program caused the change. A single group measured before and after shows how much scores moved. It cannot rule out practice on the test, time passing or regression to the mean. Say so in the plan, and consider a lottery or staggered start next cycle.

When should you write a program evaluation plan?

Write the plan while the program is being designed if you can, and always before anyone analyzes outcome data. Planning early lets you collect starting scores you would otherwise miss. If the program is already running, write the plan now for the data still to come, and note any decision made after results were known.

Do we need an outside evaluator?

It depends on your funder. Under US federal rules, competitive and direct AmeriCorps grantees whose grants average $500,000 or more a year must arrange an independent evaluation7; smaller ones must run an internal one. Check your grant agreement, and consider an outside evaluator when a skeptical reader has to trust the result.

Is this program evaluation plan template free?

Yes. The three Word documents and three CSV sheets download free, with no signup. Each comes filled in with a worked example from a made-up health clinic, so you can see a finished evaluation before replacing it with your own program. Edit with AI opens the same files in a free River workspace.

What does Edit with AI do?

It opens the files in a free River workspace. River asks what decision the evaluation informs, requests outcome records with one row per person, and asks whether any comparison group exists. It writes and dates the Evaluation Plan before analyzing results, and ready-made prompts pair each finding with what else could explain it. Nothing goes to a funder unless you send it.

Sources

  1. Community Tool Box. “Chapter 36., Section 5. Developing an Evaluation Plan - Main Section.” ctb.ku.edu (opens in a new tab)

  2. CDC. “CDC Program Evaluation Framework, 2024.” cdc.gov (opens in a new tab)

  3. eCFR. “45 CFR 2522.700 — How does evaluation differ from performance measurement?” ecfr.gov (opens in a new tab)

  4. What Works Clearinghouse. “Designing Quasi-Experiments: Meeting What Works Clearinghouse Standards Without Random Assignment.” ies.ed.gov (opens in a new tab)

  5. American Evaluation Association. “Guiding Principles for Evaluators.” eval.org (opens in a new tab)

  6. International Journal of Epidemiology. “Regression to the mean: what it is and how to deal with it.” academic.oup.com (opens in a new tab)

  7. eCFR. “45 CFR 2522.710 — What are my evaluation requirements?” ecfr.gov (opens in a new tab)

Written by

Chandler Supple

Co-founder & CEO, River

Chandler has spent the last six years working in AI, building tools for millions of users and Fortune 100 companies. In 2025, he co-founded River to build a new kind of operating system for AI agents.

LinkedInGitHubX

Reviewed by

Brian Joseff, Co-founder, River

LinkedInX

Related tools and templates

More free templates