The Geneva rivers study, and the file it came from
The ski piste study shows what a fieldwork investigation looks like. This one is its opposite number: no site visit, no apparatus, no morning in a field. One published dataset, thirty years long, and everything that had to be done to it before it could answer anything.
It is here because the secondary route fails in different places from the fieldwork route, and reading about those places is not the same as watching them happen. Every number below came out of the real file. So did every problem with it.
Rain falls on a city, runs off roofs and roads into sewers, and when the sewers are overloaded some of what is in them reaches a river. Those rivers drain into Lake Geneva, which supplies drinking water and carries most of the region’s bathing. Geneva has spent thirty years and a great deal of money separating foul water from surface water and reshaping its watercourses. The canton has also been counting the bacteria that whole time, which means the question of whether any of it worked can be answered rather than asserted.
The dataset, and the protocol
On this route the method section is a set of instructions for getting the same file back. This table is the model: everything in it was on the screen at the moment of download, and none of it could have been reconstructed a month later.
It is about a megabyte. Opening it is the fastest way to understand everything below.
What the file actually was
A downloaded file arrives looking authoritative. This one has a cantonal publisher, a licence and an update date, and it is not what its row count says it is.
Duplicating a value fifty times does not move an average, so every mean in the file is fine. The sample size is destroyed. A statistical test run on the download has an n fifty times too large, which makes almost anything come out significant, and the p-value it reports looks entirely normal. It is the one error on this route that a reader cannot catch from the finished report, which is exactly why the method has to say you checked.
Thirty years, plotted
Once the duplicates are gone and the thin years are filtered out, this is what is left. Note what the chart does not do: it does not draw a line across 2007, because there is no 2007.
The rule, fixed before looking
Each value in the file is an annual mean, and the file records how many sampling visits each mean rests on. Some sit on twelve visits and some on one. A mean from a single visit is not comparable with a mean from twelve, so the rule was: keep station-years with at least ten campaigns, and drop the incomplete current year.
That removed 111 of the 757 observations and left 646. The rule was written down before the two periods were compared, which is the only thing separating a defensible selection from a convenient one. There is nothing in the finished report that would let a reader tell the difference, so the discipline has to come from you.
Note also what the filter is not doing: it is not removing values that looked wrong. It is removing values that were measured differently.
What it found
Comparing the first six years with the last six, restricted to the 49 stations that appear in both, so the two figures describe the same places rather than different ones.
The unrestricted comparison gives 40.0 against 18.2 and means of 89.1 against 27.9, close enough to the paired figures that the improvement is clearly not an artefact of the monitoring network growing. That is a sensitivity test, and it passed. Running it was worth doing even though it confirmed what was already there: an unpaired comparison that happens to be right is still one you cannot defend.
Of those 49 stations, 32 improved, 15 got worse and 2 were unchanged. One rose from 49 to 141. Another, named in the file as sitting downstream of a French treatment plant, rose from 33 to 58. “Geneva’s rivers got cleaner” is true of the median and false at nearly a third of the places it describes, and deciding which of those two sentences to lead with is the whole of Criterion E (Analysis and conclusion).
Eight things wrong with it
None of these were planted. They are what an afternoon with one published cantonal dataset turned up, and each one is taught somewhere in the guide.
- 1The file publishes 37,009 rows and holds 757 observations. Every station-year is repeated about fifty times, identical apart from an internal id. step 4 →
- 22007 is missing from the dataset entirely, with no note to say so. step 4 →
- 3The number of sampling campaigns behind each annual mean ranges from 1 to 12, so the values are not comparable with each other untouched. step 4 →
- 4The current year is present and incomplete, sitting on a single campaign, which makes it look like a collapse. step 5 →
- 5One value sits at exactly 1000, which reads as a reporting ceiling rather than a measurement. step 5 →
- 6Zeros and values of 0.8 recur, which suggests a detection limit rather than an absence of bacteria. step 5 →
- 7The monitoring network grew from about 19 stations to about 30, so early and late years describe partly different places. step 6 →
- 8The only lake bathing station reads zero in 24 of its 28 years, so the receiving water has nothing to say and the study had to move to the tributaries. step 1 →
Three strategies, and what is wrong with each
Criterion B (Strategy) needs one real strategy and one explained tension. Three were checked here, and the useful thing is that none of them is the obvious winner.
Who: The French State, the Rhône-Alpes region, the Canton of Geneva, the Haute-Savoie department, the Agence de l'eau Rhône-Méditerranée-Corse and basin users, with operational lead shared between the Communauté de communes du Genevois and the canton.
How it bears on the data: Names improvement of water quality through assainissement among its stated objectives, and covers 160 km² and nine watercourses. Three of them, the Aire, the Drize and the Laire, carry monitoring stations in this dataset. The tightest fit of the three for a question about bacteria, and it can be aligned with the years either side of it.
The tension: The instrument exists because the goals needed reconciling across a border: Geneva legislates and pays for its own network, while part of what arrives comes from French territory it cannot legislate for. Several of the stations that got worse between the two periods are cross-border ones.
See the sourceWho: The canton sets the PGEE, which decides sector by sector whether the system is combined or separate. The secondary network is owned by the communes: more than 1,300 km of foul and surface-water sewers and 28 pumping stations. Since 1 January 2015 the FIA mutualises the cost across all of them.
How it bears on the data: The most direct of the three. Separating foul water from surface water is the mechanism by which sewage stops reaching a river when it rains, which is exactly what an E. coli series is measuring.
The tension: Who pays, and on what basis. The FIA is funded by a one-off connection charge plus two annual ones: the canton and the communes pay on the impermeable public road surface connected to the network, and property owners pay on their drinking-water consumption. Mutualising across every commune means the bill and the benefit do not land in the same place. Note that I could not find documented opposition to it, so a student choosing this one has to find sourced positions rather than assert the disagreement.
See the sourceWho: The Grand Conseil, which in April 1997 added seven articles on renaturation to the cantonal water law of 5 July 1961 and created the fund. It is fed mainly by the hydraulic royalties paid by SIG and the Société des Forces Motrices de Chancy-Pougny, by pumping taxes, by federal subsidies and by donations, at around 11.8 million francs a year with a floor of 6 million.
How it bears on the data: It aims at river morphology and flood behaviour rather than at bacteria, so the link to an E. coli question has to be argued rather than assumed. Step 3 permits exactly this: the strategy need not be what you measured, provided the connection is stated and credible, and doing it well is worth more than picking the obvious one.
The tension: The best-documented of the three. Farmers in the sector judged the project's land footprint oversized, one description calling it science fiction, and argued that agricultural realities had not been taken into account and that farmland is the farmer's principal working tool and inheritance. Part of the land at stake sits in the surfaces d'assolement, the protected cantonal quota of the best arable land. An economic and a cultural claim against an environmental one, which is a tension along three of the five lines the criterion names.
See the sourceNotice the shape of that problem. The strategy with the tightest link to the data has no documented opposition attached to it. The one with the best-evidenced disagreement is aiming at river shape rather than at bacteria. Whichever you choose, you are arguing for it, and step 3 says in as many words that a strategy need not be the thing you measured provided the connection is stated and credible.
The sources
All four are real and public. Follow each link to the original; the summary is mine.
Every bacteriological measurement made on Geneva's watercourses since 1995, station by station and year by year, counting Escherichia coli as an indicator of faecal pollution. The catalogue record gives the licence, the update date and the download links.
This is the whole investigation. One click, no account, and the conditions of use arrive in the zip alongside the data.
Open the catalogue recordAn agreement between the French State, the Rhône-Alpes region, the Canton of Geneva, the Haute-Savoie department, the Rhône-Méditerranée-Corse water agency and basin users, covering 160 km² and nine watercourses. Improving water quality through assainissement is among its stated objectives.
Three of its nine watercourses carry monitoring stations in the dataset, which is what lets a strategy and a set of numbers actually meet.
Read the contract recordThe plan that decides, sector by sector, whether the sewer system is combined or separate. The secondary network it governs is owned by the communes: more than 1,300 km of sewers and 28 pumping stations, with the costs mutualised since 2015.
Separating foul water from surface water is the mechanism by which sewage stops reaching a river when it rains, which is exactly what this dataset measures.
Read about the PGEESeven articles added to the cantonal water law in April 1997 created a fund of around 11.8 million francs a year, drawn largely from hydraulic royalties. The renaturation of the Aire is its most visible project.
It aims at river shape rather than at bacteria, so a study using it has to argue the connection rather than assume it. In exchange it comes with the best-documented disagreement of the three.
Read about the renaturation programme- 1.2 Systems, Rain, sewer, river and lake as one system with flows and stores, and a strategy that acts on one arrow in it.
- 1.3 Sustainability, Thirty years of investment against a third of the stations getting worse, which is the sustainability question with numbers attached.
- 4.2 / 4.4 Water quality and pollution, Faecal indicator organisms, point and diffuse sources, and why an indicator species stands in for a hazard.
- 8.1 / 8.2 Human populations and urban systems, Impermeable surfaces, combined sewer overflows, and who pays for the network underneath a city.
- The lake bathing station reads zero almost every year while the rivers feeding it do not. What does that tell you, and what does it fail to tell you?
- A third of the stations got worse while the overall median improved. Which of those two findings would you lead with, and why?
- The renaturation programme has the clearest disagreement attached to it but the weakest link to bacteria. Is it a defensible choice for Criterion B?
- The sampling is fortnightly and scheduled, while contamination arrives in storm peaks. How much does that undermine a thirty-year comparison?
- Several worsening stations sit downstream of French infrastructure. What can a cantonal strategy actually do about that?
In your IA: notice how much of this happened before any analysis. The protocol, the shape check and the selection rule are three quarters of the work and all of Criterion C (Method), and the statistics at the end took about ten minutes.
In class: the file is a megabyte and the eight problems are all findable in an afternoon. Handing a class the download and asking how many rows it really contains is a better lesson about data than any amount of explaining.
The Geneva rivers investigation used on the secondary-data route of this guide is the author’s own analysis of a published cantonal dataset. Every figure quoted comes from the Canton of Geneva’s open data, downloaded on 10 September 2026. The duplication, the missing year and the varying campaign counts described above are as found on that date and may since have been corrected. The selection rules, the analysis and the conclusions are the author’s and not the canton’s. It is written to show how a secondary investigation is built, not to report findings about water quality in Geneva.