What a tracer study actually requires
Following the same trainees six and twelve months later: stable identifiers, attrition you can explain, and modules that survive contact with reality.
Training completion rate · against target
1 November 2026 · 4 min read
A tracer study sounds simple. You trained 400 electricians last year. Find them, ask what they earn now, report an employment rate. Funders like the design because it measures what actually mattered. The difficulty is not the questions. It is finding the same person twice, a year apart, and proving you did.
Identifiers come first
Everything depends on the identifier. The trainee in the baseline is a row in one spreadsheet. The trainee in the endline is a different row in a different spreadsheet, and somewhere a person decides they match. Decide the identifier at baseline, not at analysis. A national ID is best if you are allowed to store it. In practice the workable combination is a programme ID plus full name plus date of birth plus phone number, stored as separate fields so a misspelt name does not break the match.
Write the matching rule down before the first follow-up call. Two trainees with the same name in one district is common, so name alone is never the rule. If your rule is an exact programme ID match, say so in the methods section, and report how many cases it failed on.
People move, numbers change
Between waves a share of any cohort changes phone numbers, and some move districts for work, which is often the very outcome you are measuring. An electrician who took a job two provinces away is a success story who no longer answers the number in your baseline. Plan for it at baseline, not in panic at follow-up.
- Collect two contact numbers, plus the name and number of someone who will always know where the trainee is.
- Ask at every contact whether the number is still theirs, and update the record. Never overwrite the old one.
- Budget a tracing window, two to three weeks of calls per wave, as real work with a named owner.
Attrition you can explain
Attrition is not failure. Unexplained attrition is. A study that finds 78 percent of a cohort, and can say the missing 22 percent are mostly young men who migrated for work, keeps its credibility. A study that reports a 92 percent employment rate while quietly losing the hardest people to find does not, and the number flatters the programme for exactly one audit.
Report the contact rate next to the result. Compare respondents and non-respondents on everything baseline captured: sex, age, district, baseline earnings. If the people you lost look like the people you found, say so. If they do not, that belongs in the limitations, not in a footnote.
Baseline and endline must pair
An employment module that asks 'are you working' at baseline and 'have you worked in the last 30 days' at endline cannot measure change, because the two answers are not comparable. Write the endline question first, then write the baseline question that will sit beside it in the analysis. Same recall window, same categories, same order of options, and income bands defined the same way, adjusted for nothing until analysis.
Keep the endline short. Every extra module is a per-minute cost paid on hundreds of calls, and fatigue shows up as flat answers in the last third of a long interview.
Consent, and what you keep
You will hold contact details for a year or more, which makes you a data custodian whether the design mentions it or not. Get consent for recontact at baseline, as a separate yes, not buried in the interview consent script. Store contact details against the case, never in the interview data, so the analysis file can be shared without phone numbers in it. And re-consent at the start of every call, because a trainee who agreed to a baseline in 2025 did not agree to a recording in 2027.
What breaks, and what saves it
| What happens in the field | What it does to your data | What saves it |
|---|---|---|
| Phone lost, that was the only number | The trainee is unreachable | A second contact collected at baseline |
| Number changed, the SIM was reused | A stranger answers, counted as non-response | Name and programme ID verified before every interview |
| Moved district for work | A success story recorded as a drop-out | Current district captured every wave, migration asked directly |
The tooling for this is unglamorous. Each trainee is a case, not a row: a record with an identifier, a call log, a status, and waves of interviews attached to it. When the baseline, the six-month call and the twelve-month call hang off one case, matching stops being a spreadsheet exercise and attrition becomes a query instead of a weekend of work.
A tracer study is one of the few designs where the data work is the study. Getting the identifier right at baseline costs an hour. Getting it wrong costs the follow-up.
Read next
1 November 2026 · 3 min read
When a form builder stops being enough
KoboToolbox and ODK are excellent at collection. The honest signals you have outgrown them, and what changes when reporting stops being a spreadsheet job.
1 November 2026 · 3 min read
From nine days to an afternoon
The monthly donor report as an assembly job, what changes when the indicator, the dashboard and the report read one dataset, and what still needs a person.
1 November 2026 · 4 min read
Offline collection: what breaks
Flat batteries, ninety seconds of signal, duplicate households, stuck photos, a form update mid-flight. What a real system does about each one.
See it with your own data
Everything this blog describes, the collection, the indicator, the trail from a figure back to a submission, is what the platform is built to do.
No card, no countdown, and your data stays yours.