Path
Data Scientist for NGOs
For the data scientist working where the constraint is not the model but the decision — small samples, targeting rules people can contest, and predictions that get acted on by a person visiting a household.
Competencies
Problem framing
AdvancedTurn a programme question into an analysis with a decision attached, and recognise when a model is the wrong instrument.
Composite indicators
AdvancedBuild a score from several indicators, state its assumptions, and test whether its conclusion survives dropping an input.
Fairness in targeting
AdvancedReport who a rule includes and who it excludes, disaggregated, and separate what the indicators produced from what was chosen.
Explainable outputs
AdvancedProduce a ranking or flag a caseworker can act on and a district authority can contest, without a model nobody can interrogate.
Before you start
- The Data Analyst path, or its equivalent — you can already clean a real export and defend every change you made to it.
- A tolerance for being told the model is not the deliverable. Most of the value in this role sits upstream of it.
Continues fromData AnalystThis path assumes that one rather than repeating its courses.
The route through
Stage 1
Ground the numbers
Know what the target variable counts before you fit anything to it, because a model inherits every ambiguity in its label.
Stage 2Choose one
Pick the language your team uses
Go deep in one of Python or R. Python is the usual choice where a model has to be deployed; R is the better answer where the work is analysis handed to statisticians.
These are alternatives, not a sequence. Take the one your team already uses — the second is far cheaper to add once you have the first.
Stage 3
Make the data trustworthy
Turn a raw export into a table you would defend line by line, and hand over a cleaning log that answers the auditor's question before it is asked.
Stage 4
Put the files together
Assemble several exports into one analysis table you would defend column by column, with every join proved, every grain stated and every denominator sourced.
Stage 5
Check it like an auditor
Assess your own data before a donor does — five dimensions with measures, a recount against the source, and a report where every finding has an owner and a date.
Stage 6
Define what you report
Write indicator definitions two analysts compute the same way, defend the denominator, and set a baseline and target that survive a mid-term review.
Stage 7
Put an interval on it
Analyse a cluster survey the way its design requires — weights from the frame, a measured design effect, and an interval that says what the sample can and cannot settle.
Stage 8
Know where it came from
Read a routine reporting system as the database it is, pull an extract you can point at months later, and answer what was counted for any figure it produces.
Stage 9
Apply it to nutrition
Take the methods into one sector — WHO growth standards, SMART plausibility, the IPC phases and the Sphere performance thresholds, on a survey and a treatment register.
Stage 10
Apply it to public health
Rates with person-time denominators, treatment cascades, coverage three ways, and an outbreak line list turned into a curve, attack rates and a case fatality you can defend.
Stage 11
Apply it to WASH
The JMP service ladders and the Sphere minimums on a household survey, then a monitoring register with repeat visits that turns one functionality rate into three and bounds each against the rounds nobody drove.
Stage 12
Apply it to food security
FCS, HHS, rCSI and the livelihood coping module built from raw components and found to disagree by a factor of seven, then a price series whose seasonality dwarfs the programme effect and the evidence table that reports both.
Stage 13
Apply it to protection
The analyses you must decline to publish, alongside the ones that matter — a consent-gated referral pathway, a nineteen-point equity gap located at a specific gate, and a caseload that explains two other tables.
Stage 14
Apply it to education
Gross against net enrolment on a projected denominator, two attendance numbers twenty-six points apart, a cohort through promotion and repetition, and two assessment rounds whose instruments differ.
Stage 15
Say how sure you are
An interval on every proportion, the right test for a comparison, an effect size beside every p-value, and the count of comparisons that turns two striking schools back into noise.
Stage 16
Model more than one thing at once
A coefficient is a comparison — which one, between which units, adjusted for what. An odds ratio your reader will misread, a covariate that removes 42% of the effect, and a model that explains three per cent and settles a targeting decision.
Stage 17
Say what caused it
A seven-point gain that is entirely the school year, a comparison group imbalanced on every characteristic measured, and the minimum detectable effect that decided the answer before any data existed.
Stage 18
Put it in front of them
The mark the comparison implies, an interval that stops a ranking, a palette that already means something to this audience, and a figure generated from the dataset so the chart and the sentence cannot drift.
Stage 19
Make it rerunnable
Raw data read-only and code the only thing edited, an environment pinned so a colleague's laptop gives your numbers, checks that stop the pipeline rather than producing a plausible wrong one, and a handover a successor can act on.
Stage 20
Put it in front of the decision
A page that answers one question rather than twenty, the definition panel that stops the monthly argument, and the report whose findings, limitations and recommendation survive being read separately.
Recommended projects
How the skill is assessed
Can build a composite score and demonstrate that its selection is stable under dropping any single component.
EvidenceA stability check reported with the ranking, naming what moves.
Can report the composition of a targeted caseload against the surveyed population, disaggregated.
EvidenceA composition table with the over- and under-represented groups named in the report, not only computed.
Can state where a decision boundary is not supported by the data.
EvidenceA boundary analysis showing the households either side of the cut and what separates them.
The hard part of this job is rarely the method. It is that the output is a list of named households, and somebody will act on it.
That constraint rules a lot out. A model whose ranking cannot be explained cannot be contested, and a targeting rule that cannot be contested will be — by a district authority reading the caseload as a statement about who deserves assistance. The food security project here ends by reporting that displaced households are 19% of the surveyed population and 33% of the selected caseload, and that no displacement term appears anywhere in the score. Both sentences have to be in the report.
The other habit this path builds is admitting where a boundary is arbitrary. In that same project the forty households either side of the cut span less than a tenth of a point, adjacent pairs differ by thousandths, and nothing in the data distinguishes the last household included from the first excluded. Saying so is more useful than a precision the data does not have.
What this path still needs
Three courses are published: Data Analysis Foundations, Python for Programme Data and R and the Tidyverse for Programme Data. The modelling content sits in Statistics and Modelling on the programme roadmap — applied statistics, regression, impact evaluation — and none of it is written yet. See the programme.