By AtramUpdated 2026-11-20
Atram measured the season with automated phone surveys after three rain events, a comparison group and two statistical tests. This note covers the Kenya long rains 2026 data page and says what its numbers cannot say. Method version 1 · February to June 2026.
We publish the method so you can check it. If you find an error or a better method, tell us. We change the note and say what changed.
Surveys followed three of the six rain events with warnings
One microfinance lender in Kenya enrolled about 10,000 of its clients to receive rain warnings by text message. Warnings went out for six of the season's ten rain events. After three of them, automated voice calls (IVR) asked a sample of clients what they did and what they lost.
Calls and survey completers per surveyed event
| Event 5 | Event 8 | Event 9 | Season | |
|---|---|---|---|---|
| Dates, 2026 | 18 to 22 March | 22 April to 3 May | 8 to 16 May | February to June |
| Calls attempted | 7,229 | 9,510 | 2,989 | 19,728 |
| Survey completers | 607 | 772, of whom 376 in the comparison group | 166 | 1,545 |
Atram, season data, Kenya long rains 2026. Cells under 30 respondents, and totals that would reveal them, are not shown.
A completer is a client who answered the questions that define a measure, not a client the survey system tagged as complete. Atram collected the surveys. The data describes the lender's clients, so no branch, client or loan appears on any page.
Six steps turn the raw data into the published figure
- Count reach per event. A client counts as reached when they received at least one warning in the event window. One client appears in several events, so unique clients are never summed across events.
- Survey after the event. Call a sample of warned clients after events 5, 8 and 9. Also call a comparison group, almost all of it in event 8; event 9's is too small to report.
- Compute the preparation rate. Divide the completers who took one or more preparatory steps after the warning by all completers.
- Build the comparison group. It joins clients pre-selected in the lender's customer files and warned clients who did not recall the warning.
- Convert losses to dollars. Clients report losses in Kenyan shilling bands. Each band becomes the value in the table below, and we exclude "not sure". A value is not always the midpoint of its band: the band 1,000 to 5,000 takes 2,500, and its midpoint is 3,000. The total converts at one rate, the rate of Atram's season analysis: 133 shillings to the dollar.
- Test the differences. Preparation, warned clients against the comparison group in event 8, uses a Fisher exact test. Losses, prepared against unprepared small businesses in event 8, use a Mann-Whitney U test.
The loss bands and the value we use for each
| Band reported, Kenyan shillings | Value used |
|---|---|
| Nothing | 0 |
| 1,000 to 5,000 | 2,500 |
| 5,000 to 20,000 | 12,500 |
| 20,000 to 50,000 | 35,000 |
| Over 50,000 | 75,000 |
| Not sure | Excluded |
The two tests give these results. In the event 8 test, warned clients prepared at 74.2%, against 53.5% in the comparison group. The test finds the gap unlikely to be chance, p below 0.001. Small businesses that prepared reported losing USD 80, against USD 115 for those that did not, p=0.04.
The survey also asked clients who prepared whether they would have prepared without the warning. Across the season, 57.6% said no. This is a self-reported answer about what clients would have done, so read it with care.
The numbers show what clients did, not what the lender's portfolio did
| The numbers can say | The numbers cannot say |
|---|---|
| In event 8, warned clients reported preparing more often than one comparison group | That the warnings lowered loan defaults or arrears. Loan-level data was not available . The exact size of the gap: the comparison group is not a clean randomised control, as step 4 shows |
| Small businesses that prepared reported smaller losses than those that did not | That the warnings cut losses. The loss comparison sets prepared against unprepared clients, not warned against unwarned ones |
| What clients reported losing during the event, in Kenyan shilling bands | What it cost them to recover and trade again. The survey did not ask . An audited loss: nobody checked a shop or a field |
| What one lender's surveyed clients did in one season | What would happen in Kenya as a whole. We have 1,545 completed surveys among about 10,000 enrolled clients, not the country |
| What event 9's surveyed clients did as a whole | What each segment did in event 9. With 166 completers, its findings by segment are directional only |
| The recall rates of events 8 and 9, side by side | How event 5's recall rate compares with them. Event 5 asked yes or no; events 8 and 9 added "not sure" |
| How many clients the warnings reached in events 6, 7 and 10 | What those clients did. Events 6, 7 and 10 had no survey |
This note leaves out the avoided-loss estimate on purpose
Atram's season analysis also estimates the losses the warnings avoided across the season, with two methods. We do not publish that estimate. It is a model built on self-reported answers, and it counts clients' losses, not the lender's. We also leave out the cost of the service, the lender's portfolio figures and any figure below the whole surveyed population.
Anyone can contest this method by email
Write to hello@atram.ai. We answer, and if you are right we change the note and say what changed.