Theory of change

A theory of change is an explicit, testable causal chain from what you do to the outcome you want, with the assumptions named. It originated in program evaluation and international development, and it is the most rigorous way to state what an OD program is actually claiming.

Evidence: Not a model — an evaluation method. Widely required by funders and evaluators, and there is limited direct evidence that producing one improves program outcomes. Its defensible claim is narrower and does not need a trial: an intervention whose assumptions are written down in advance can be falsified, and one whose assumptions are not written down will be evaluated retrospectively against whatever happened. The discipline is the deliverable.

Use when: launching any program you will later be asked to justify; when stakeholders disagree about what a program is for; when you want the program to be falsifiable.

Do not use when: the work is genuinely exploratory and premature specification would foreclose learning. Even then, state the hypothesis.

The chain

Inputs → Activities → Outputs → Outcomes → Impact

• Inputs: money, people, time, tooling.
• Activities: what you do — workshops, coaching, redesign, tooling changes.
• Outputs: countable products of activities — sessions delivered, people trained, process documented. Not evidence of success.
• Outcomes: changes in behavior, capability, or condition. Short, medium, long term.
• Impact: the ultimate goal — patient outcomes, delivery performance, retention, mission achievement.

The discipline that matters: between each pair of boxes there is an assumption, and the assumption is where programs fail. Write the assumptions down. They are the falsifiable content.

See https://github.com/joelparkerhenderson/outputs-vs-outcomes.

Building one backwards

Always work right to left.

1. State the impact you want, in measurable terms.
2. Ask: what outcomes must occur for that impact to follow? What must be true about behavior?
3. Ask: what outputs would produce those outcomes?
4. Ask: what activities produce those outputs?
5. Ask: what inputs do the activities require?
6. For every arrow, write the assumption, and rate your confidence in it.
7. For every low-confidence assumption, decide how you would test it early and cheaply.

Worked example: psychological safety program

Assumptions, each stated explicitly:

• Leaders can change behavior with coaching alone. (Low confidence — test.)
• Reported safety translates to actual speaking-up behavior. (Medium.)
• Higher reporting reflects better reporting, not more incidents. (Critical: this metric moves the wrong way if misread. Pre-agree the interpretation with leadership before the data arrives.)
• Earlier surfacing reduces harm rather than merely relocating it. (Medium.)

The third assumption is the kind that sinks programs. A rising reporting rate is success, and will be presented by someone as failure. Agree in advance.

Examples by audience

• Health care: mandatory in most grant and improvement funding, and routinely produced as compliance paperwork rather than used. Its real value is forcing agreement on whether a rising metric is good or bad before it rises.

• Software: a developer experience program whose theory of change stops at outputs ("we shipped a new CI system") cannot answer whether it helped. Chain it to lead time, change failure rate, and retention. See dora-metrics.

• Executive: ask for the theory of change behind any major program before approving it, and specifically for the assumptions list. A program that cannot name its assumptions has not been thought through.

Limitations

• Can become an elaborate document that is written once and never revisited.
• Implies more linearity than complex systems allow; in the Complex domain treat it as a hypothesis to be revised, not a plan. See cynefin-framework.
• Vulnerable to reverse-engineering: writing the theory to justify an intervention already chosen.

See also

• measurement-and-metrics
• evidence-based-practice
• https://github.com/joelparkerhenderson/objectives-and-key-results