Three districts, six spellings

Group this survey by district and you get six districts. There are three. One enumerator team wrote Nord-Ouest four different ways, and because the variants sort apart, the district with the worst water access is split into pieces small enough that none of them reaches the top of the table.

Nobody will tell you this happened. The table will look fine.

This lab runs locally in RStudio rather than in a hosted notebook. Colab can run R, but the runtime has to be switched by hand — and the course spent lesson 1 on a project that reopens a year later, which is not a thing a hosted session teaches you.

The file

wash-household-survey-2024.v1.csv — 2,403 household interviews across three districts and eighteen communities, covering water source, collection time, quantity, point-of-collection testing, sanitation and hygiene. Synthetic.

Set up first

An RStudio project with renv initialised, as lesson 1 describes, and the data in its own folder. No setwd(), no absolute path. You will run this twice — once now and once at the end, from a clean session — and the second run is the one that proves it.

The task

Place every household on the three JMP service ladders — drinking water, sanitation, hygiene — and report coverage by district.

Five things stand in the way.

The district names. Normalise before you group, with fct_collapse() or a recode, not with a chain of ifelse(). Then check: fct_count() should show three levels and no NA. If the fourth variant reappears next round spelled a fifth way, your code should fail visibly rather than quietly adding a district.

The ladder is not the source. An improved source more than 30 minutes round trip is limited service, not basic. Roughly a quarter of these households sit on limited service for that reason alone, and a classification built on water_source alone misses every one of them. Use case_when(), and put the 30-minute test before the source test so the ordering is explicit.

Shared is not basic either. An improved sanitation facility shared between households is limited service. shared_sanitation is the column; forgetting it inflates basic sanitation.

Two denominators, not one. Water quality was tested on roughly a third of households. Access indicators use the whole sample; quality indicators use the tested subsample. Report each against its own denominator and say which is which in the column name — this is the habit the course keeps returning to.

Units that lie. Eleven records hold collection time in hours rather than minutes, and fourteen hold litres for the whole household rather than per person. Both look plausible alone. Find them against household_size and the distribution, decide what to do, and record the decision.

What to hand in

An R script, sourced from a clean session, producing:

• a coverage table by district and ladder, with district as an ordered factor running worst to best rather than alphabetically
• the share of households below the Sphere minimum of 15 litres per person per day
• the water-quality table, on its own denominator, with the tested share stated
• a cleaning log: every row you dropped or corrected, and why

Then wrap the cleaning in a single function taking the file path and returning the cleaned tibble, so the next round is one call.

Check your numbers

The dataset notes state what a correct analysis finds:

	Expected
Below Sphere 15 L/person/day	about 13%
Round trip over 30 minutes	about 39%
Open defecation	about 13%
Basic hygiene service	about 34%

If you get six districts, the normalisation did not run before the grouping. If basic sanitation looks high, you have not applied shared_sanitation.

The questions to answer in prose

Three sentences each.

1. Normalising the district names moves one district’s ranking. Which one, in which direction, and what would have been decided differently if the table had gone out unnormalised?

2. About 530 households report treating their water but have no chlorine measurement. Dropping them is tempting and wrong. Say what those households have in common and what dropping them would do to the treated-water figure.

3. You reported water quality on a smaller denominator than access. A reviewer asks why the two coverage figures cannot be compared directly. Answer them.

How to know you are done

renv::restore() on another machine, one source(), and the same tables. Your cleaning function should run unchanged if the next round adds a fourth district — and should tell you it did, rather than silently producing a fourth row.

What this lab is not

It is not the JMP methodology. The ladder definitions are given to you here; how they were arrived at and what they leave out belongs with the WASH content itself. This lab is about the R: factors that hold their order, case_when() that reads in the order it evaluates, and cleaning that survives being run by someone else.