Datasets you can actually have today

1. 1The file has 37,009 rows and 757 observations. Every station-year is repeated about fifty times, identical apart from the object id. Run a test on the file as downloaded and your sample size is wrong by a factor of fifty, so everything comes out significant.
2. 22007 is missing from the dataset entirely.
3. 3The number of campaigns behind each annual mean ranges from 1 to 12. A mean from a single visit is not comparable with a mean from twelve, so filter before you compare.
4. 4The current year is present but incomplete, with a single campaign behind it.
5. 5The station roster grows over time, from 19 stations in 1995 to around 30 now. Comparing early years with late years compares different sets of places unless you restrict to the stations present in both.
6. 6One value sits at exactly 1000, which looks like the highest value the lab reports (a ceiling), not a real measurement.
7. 7Zeros and values of 0.8 recur, which suggests a detection limit rather than genuinely no bacteria.
8. 8The only lake bathing station, Pâquis, reads zero in 24 of its 28 years. The lake itself is clean. The pollution is in the rivers flowing into it (the tributaries), so a study of the lake shore has nothing to analyse.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• How has mean annual E. coli at Geneva's river monitoring stations changed between 1995 and 2024?
• A comparison of E. coli between stations on predominantly agricultural catchments and predominantly urban ones.
• Do stations downstream of a wastewater treatment plant differ from stations upstream of one?

• E. coli, CFU/ml (annual mean)
• Sanitary class, five ordered categories from Très bon to Mauvais
• Number of sampling campaigns behind that year's mean
• Station name and code, watercourse, year
• Swiss coordinates, so it maps

1. 1Missing values are recorded as −99, not as blanks, and nothing in the file says so. Twenty parameters carry them. Average silica without removing them and you get a large negative concentration, which is impossible, and a spreadsheet will calculate it without any warning.
2. 2Silica is −99 in 856 of 950 rows, so that column is 90% missing rather than 90% present. Check how much of a column is real before you build a question on it.
3. 31973, 1989 and 1990 are absent from the series entirely.
4. 4The network grew from 6 stations in 1969 to about 30 now, so early and late years describe very different sets of places. Any before-and-after comparison needs restricting to the stations present in both.
5. 5Station names are not unique. Twenty-three different stations are called "Embouchure", because every river has a mouth. Group by CODEMESURE, never by name, or you will silently merge twenty rivers into one.
6. 6There are no unit columns here, unlike the physico-chemistry file. The units are in the documentation, and you have to go and read it.
7. 7Every value is the year's 90th percentile, not its mean: nine samples in ten were at or below it, so it reports a station's bad days. The documentation in the zip says so for every parameter. Describe your variable that way, and don't average it with a file that gives means.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• How has phosphate (each year's 90th percentile) at Geneva's river monitoring stations changed between 1969 and 2025?
• A comparison of nitrate between stations on predominantly agricultural catchments and predominantly urban ones.
• Which parameter most often downgrades a station's grade, and has that changed over fifty years?

• Phosphate and total phosphorus, nitrate, nitrite, ammonium, total nitrogen
• Dissolved oxygen and saturation, biochemical oxygen demand, dissolved organic carbon
• Conductivity, pH, alkalinity
• Chloride, sulphate, sodium, potassium, calcium, magnesium, silica
• DECLASSANT: which parameter dragged that station-year's grade down

1. 1The same −99 convention as the major-elements file, and worse. 21% of mercury values are −99. The mean of the column as it downloads is −20.555; with those values removed it is 0.403. A fiftyfold error, in the wrong direction, producing a negative concentration that cannot exist.
2. 2Gadolinium and vanadium are −99 in about two thirds of rows, so those columns are mostly absent rather than mostly present.
3. 3Copper is the downgrading parameter in 255 of 683 station-years, which makes it the obvious variable and also the one everybody else in your class will pick.
4. 4With thirty metal columns, it is tempting to test pair after pair until one comes out significant. Decide which metal your question is about before you open the file, and say in your method that you did.
5. 5Station names are not unique. Group by CODEMESURE.
6. 6These are small numbers with real detection limits, so a value of exactly zero is likelier to mean "below the limit" than "none present".

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• A comparison of copper concentrations between stations on urban catchments and rural ones.
• How has zinc at Geneva's river monitoring stations changed between 1995 and 2025?
• Which metal most often determines a station's grade, and does that differ between river types?

• Copper, zinc, lead, cadmium, chromium, nickel, mercury
• About twenty more, from aluminium and iron to platinum and uranium
• DECLASSANT: which metal downgraded that station-year, most often copper
• Number of samples behind each annual value

1. 1It stops in 2017. It is the biggest and most detailed file in the family, so it looks like the main one, but it is the one that was discontinued. The major-elements file carries most of the same parameters through to 2025. Work out which one your question needs before you commit a day to this.
2. 2Missing values are marked differently here from the rest of the family: small negative numbers such as −0.005 or −0.05, rather than −99. Same publisher, same subject, two conventions, and no warning in either file.
3. 3Discharge is present in only 502 of 836 rows, so a question that normalises by flow loses 40% of the data before it starts.
4. 4About 120 columns, half of them unit columns. So many columns is a real hazard: it makes it very easy to test pair after pair until one is significant.
5. 5Station names are not unique. Group by CODEMESURE.
6. 6The overlap with the major-elements file is large but not exact. If you use both, make the parameter names match and check the units, because only one of the two carries unit columns.

• A comparison of dissolved oxygen before and after a named wastewater treatment plant opened.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• How did nitrate at Geneva's river monitoring stations change between 1962 and 2017?
• Which parameters improved over the period, and which did not improve at all?

• Discharge, temperature, conductivity, pH, oxygen and saturation
• Nitrogen and phosphorus in several forms, BOD and dissolved organic carbon
• About thirty metals
• A unit column beside almost every parameter, and quality classes for the regulated ones

1. 1There is no bulk download. Each station has its own Export button, so the river is five exports. The page warns that more than a year or two at once may fail; twenty years at a time worked when this was checked.
2. 2The export needs the page, not the link. The button submits a form, so a copied link or a script gets a page asking for your screen size instead of the file. Click Export on the station page itself.
3. 3Campaigns, not years. The secondary network is sampled every four or five years: 2003 (Calèves), 2006 (Nyon Lac), 2009, 2013, 2016, 2020 and 2025. Chéserex and Moulin Velliet were sampled once, in 2009, so they can show where along the river, never when.
4. 42009 at Nyon Lac is three samples. February, May and July. The same year's figures elsewhere are monthly, so a 2009 average for Nyon Lac is not the same kind of number.
5. 5The sampling method changed. In 2003 (Calèves), 2006 (Nyon Lac) and 2009 (Embouchure) the samples were 24-hour composites, `Durée [j]` = 1. Every later sample is a single grab, `Durée [j]` = 0. A composite smooths out peaks a grab can catch, so part of any before-and-after is the method.
6. 6The criterion moves. The ammonium threshold is 0.2 or 0.4 depending on water temperature, and the nitrite threshold is 0.02, 0.05 or 0.1 depending on chloride. Judge failures with the `QR` column (value divided by criterion; above 1 fails), not against one fixed number.
7. 7Text in the number column. Twelve values read `<LOQ` (below the limit of quantification) and one reads `nd`, which turns the column into text in a spreadsheet. Decide what to substitute and say so in your method; half the limit is the usual choice, and the limit itself is only recorded from 2016.
8. 8The station summary table uses a different statistic. It shows the 90th percentile of a year, except where fewer than twelve samples were taken, when it shows the maximum instead and adds a `w`. That is why 2020 at Nyon Lac looks about three times worse than its neighbours on the summary page.
9. 9It is an old-format .xls with French headings, and dates are written 04.12.2025, day first.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• How do nitrite and ammonium change between Chéserex and Moulin Velliet, and what does that suggest about a point source between them?
• Has the share of samples at Calèves that fail the nitrite criterion changed between 2003 and 2025?
• Is nitrate at Calèves higher in autumn than in spring, and has that difference changed since 2003?

• Nitrate, nitrite, ammonium, total nitrogen, total phosphorus and dissolved reactive phosphorus, all in mg per litre as N or P
• Dissolved organic carbon, chloride, pH and water temperature
• The Swiss quality criterion for each measurement, and QR: the value divided by that criterion, so anything above 1 fails
• The limit of quantification (from 2016 only), station coordinates in the Swiss LV95 grid, and the sampling network (secondary or passive)

1. 1Results below the detection limit. Many results read <0.02 or <0.004: too little to measure, not zero. At Whitney Toll Bridge, where the Wye enters England, 39% of results are like this, and the limit itself changes over the years. Decide how to treat them before you look at the pattern, and say what you did.
2. 2Gaps. Whitney Toll Bridge has no samples from 2011 to 2018, and Ross drops to four a year from 2015 to 2018. Count samples per year before you compare years.
3. 3There are two River Wyes. Searching for Wye also finds the Derbyshire Wye at Buxton and Rowsley. The Herefordshire points have codes starting MD-500.
4. 4One year per file, and everything in it. Each download is every measurement of every substance at every point in Herefordshire for one year. Filter to your sampling points and to one determinand before anything else, then stack the years. For a long series, that is up to 26 files: plan it.
5. 5Monthly spot samples miss the blooms. Algal blooms happen in hot, low-flow weeks, and a sample once a month can miss them. Yearly averages look flat for that reason. Summer against winter, or low flow against high flow, shows more.
6. 6England only. The upper Wye is in Wales and is monitored by Natural Resources Wales, published separately. Volunteers in the Wye Alliance have taken over 50,000 samples since 2021, but their test kits are less precise: a good comparison, not a replacement.
7. 7Orthophosphate is not total phosphorus. It is the dissolved form algae can use straight away; total phosphorus includes what is bound to soil. Say which you used, and use one throughout.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Is the orthophosphate concentration in the River Lugg at Mordiford higher than in the River Wye at Wilton Bridge, 2021 to 2025?
• Is orthophosphate in the Wye higher below Hereford's sewage works (Carrots Pool) than above them (Victoria Bridge)?
• How does orthophosphate in the Wye at Ross-on-Wye vary between summer and winter?

• samplingPoint.notation and samplingPoint.prefLabel: the point's code and name, such as MD-50050, R LUGG AT MORDIFORD BRIDGE
• phenomenonTime: when the sample was taken
• determinand.prefLabel: what was measured. Orthophosphate, reactive as P is code 0180
• result and unit: the value, in milligrams per litre for phosphate. Some results start with <
• samplingPurpose: why the sample was taken; most are statutory monitoring

1. 1Search for the Gulf of America. The EPA renamed the Gulf in 2025, so the page, the file headings and the citation all say Gulf of America. Searching the EPA site for "Gulf of Mexico" misses it. Say in your method that the two names mean the same sea.
2. 2Six lines of notes first. Each CSV opens with a title, the data source, a note and the units, then a blank line, then the headings. A spreadsheet reads them as data: delete them, or tell the import to start at row 7.
3. 3Null is not zero. 1989 and 2016 read null: only half the area was mapped. Leave them out. 1988 is real: 15 sq mi, a drought year.
4. 4One cruise a year. The area is what one survey in late July found. It says nothing about how long the dead zone lasted or how big it was in June or August, and the page says so in its Limitations tab.
5. 5Weather breaks the pattern. 2019 was forecast to be one of the largest on record, but Hurricane Barry mixed oxygen down days before the survey. 2017 is the record. Look at these years rather than deleting them.
6. 6Square miles against square kilometres. The file is in square miles, the Task Force goal is 5,000 km² (about 1,930 sq mi). Convert once and state it.
7. 7Exhibit 2 is one week in 2021. Its 94 stations are enough for a question about where the oxygen is lowest, not about change over time. You work out each station's distance from the river mouth yourself, from its latitude and longitude, and that calculation is part of your method.
8. 8It stops in 2021. Later summers are only in NOAA's yearly announcements, each a source in its own right: 2022 (3,275 sq mi), 2023 (3,058), 2024 (6,705) and 2025 (4,402). Add them only if you cite each one, and note that the EPA's 2017 figure (8,772) and NOAA's (8,776) differ slightly.

• Was the Gulf dead zone larger in the summers after the Hypoxia Task Force's 2001 Action Plan (2002 to 2021) than before it (1985 to 2001)?

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• On the July 2021 survey, is the oxygen in the bottom water related to a station's distance from the mouth of the Mississippi?

• Year and Square miles in Exhibit 1: the area of bottom water below 2 mg/L of oxygen that July
• Latitude, Longitude and Dissolved oxygen concentration (mg/L) in Exhibit 2: one row per station, measured near the seabed, 25 to 31 July 2021
• The page's own text gives the thresholds that matter for marine life: bottom fish start to leave below about 3 mg/L, crustacean larvae die below 2.5, and fish avoid or die below 1

1. 1Check the phosphate units. The PO4 heading says µgP/L, but the values are in mg/L: 0.034 is 34 µg per litre. Multiply by 1,000 and say so in your method.
2. 2Semicolons and Latin-1. Open it as UTF-8 and "µg/L" in every heading turns into garbage characters. Tell your spreadsheet the file is semicolon-separated and Latin-1 (or Western European).
3. 3NaN means missing. Phytoplankton is missing in 9 months and Daphnia in 31. Leave those months out, never read NaN as 0, and count how many months each year really has before you average it.
4. 4The sampling depth changed. Phytoplankton was sampled in the top 10 m until 2001 and the top 18 m after. A step in biomass around 2001 could be the method, not the lake. The thermometer was replaced by a probe in 1998.
5. 5Months are not independent. Every year rises and falls with the seasons, so a correlation across all 444 months is mostly summer against winter. Compare the same month across years, or annual means, and say which.
6. 6Different amounts of water. Phytoplankton is per litre of the upper layer; Daphnia is per square metre of lake surface, down to 50 m. They are not the same kind of number, so compare how they change, not their sizes.
7. 7A1 to A5 are only names here. The file does not say which algae are in each group. That is in Anneville et al. (2018), Hydrobiologia 824: 121-141. Read it before you build a question on one group.
8. 8It stops in 2010. Later years exist but need an account and a data request from the observatory, which will not arrive in a lesson.

• Is annual mean phytoplankton biomass related to annual mean dissolved phosphate in Lake Geneva, 1974 to 2010?

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Is the number of Daphnia related to phytoplankton biomass in the same month in Lake Geneva, 1974 to 2010?
• Is the biomass of the summer phytoplankton group related to summer water temperature across the years 1974 to 2010?

• Year, Month
• PhytoTot: total phytoplankton biomass, µg per litre, in the top 10 m of water (top 18 m after 2001)
• A1 to A5: the biomass of five groups of phytoplankton that grow at different times of year, defined in Anneville et al. (2018)
• Daphnia: water fleas per square metre of lake surface, counted in a net pulled up from 50 m
• PO4: dissolved phosphate. The heading says µgP/L; the values are mg/L
• WaterTemp in °C and SolarRad in joules per cm²

1. 1It is not UTF-8. The file is Latin-1, so a spreadsheet that assumes UTF-8 turns Côte d'Ivoire and Türkiye into nonsense or refuses to open it. Choose the encoding on import. Everything else in the file is plain ASCII, which is why the fault is easy to miss.
2. 2Most of the numbers were never measured. Of 6,305 country values for total water withdrawal, 499 are marked A, official. The rest are estimated, imputed or borrowed. India has 48 years and not one of them is official.
3. 3Imputed means drawn as a straight line. Between two real values the gaps are filled with a straight line, and after the last one the value is copied forward: India reads exactly 761 every year from 2010 to 2022. About 23% of the withdrawal values equal the year before. A flat line here usually means nobody measured, not that nothing changed.
4. 4Nineteen of the areas are not countries. World, Europe, Sub-Saharan Africa, Least Developed Countries and fifteen more sit in the same column, so a sum or an average across areas counts places twice. Filter on the numeric code, not the name: Central African Republic and South Africa look like regions to a text search.
5. 5Percentages above 100 are real. Water stress is withdrawal as a share of renewable water, and 17 countries were over 100 in 2022, Kuwait at 3,850. Countries that pump fossil groundwater or desalinate seawater take more than rain replaces. One Kuwait makes an average across countries meaningless, so use the median.
6. 6Coverage grows over time. Total withdrawal has values for 2 countries in 1965, 54 in 1980 and 182 from 2015, so a total or an average drawn across the whole period is mostly measuring how many countries joined.
7. 7The unit sits inside the variable name, `[10^9 m3/year]`, which is cubic kilometres. The header also repeats itself: two columns called aquastatElement and two called timePointYears. Both year columns are identical; delete one.
8. 8The last year is 2022. Before you use a country's most recent years, check that they are not an earlier value copied forward.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• How far did Australia's agricultural water withdrawal fall during the Millennium Drought, and how quickly did it recover?
• Which countries withdraw more freshwater than they renew each year, and what do they have in common?
• Did Spain's total water withdrawal fall after 2015, using only the values marked official?

• Total water withdrawal, in 10^9 m³ a year (cubic kilometres): India 761, Spain 29.0, Australia 16.7 in 2022
• The same split three ways: agricultural, industrial and municipal withdrawal, each also per person
• SDG 6.4.2 water stress: withdrawal as a percentage of renewable freshwater, and each sector's share of it
• Total renewable water resources, per country and per person
• A status letter on every value: A official, E estimated, I imputed, X taken from another international organisation

1. 1Choose Original File Format. The page's default download is a tab-delimited file holding only the first sheet, a list of authors (3.8 KB). The Excel original has all five sheets.
2. 2Three lakes in one file. Filter waterbody_name to Geneva first: Annecy and Bourget are mixed in.
3. 3The headings are on row 2, under a title row, and the first column is a row number with no heading.
4. 4One row per stage, under several names. Daphnia arrive as adult females, egg-carrying females and juveniles, named DAPHNIA SP until 1990 and HYALINA, LONGISPINA and GALEATA from 1984, as the counters began naming species. Add every row that starts DAPHNIA on a date to get all the water fleas.
5. 5No row means none counted. The file has no zeros: a date with no Daphnia row had none recorded. Put the zeros in before you average, or those dates vanish and the mean goes up.
6. 6Two nets. Some dates used a 212 µm net and others a 64 µm one, which also catches smaller animals. No date has Daphnia from both, but the mix changes over the years: keep zoop_mesh_um and say how you handled it.
7. 7Uneven sampling. 2001 has a single sampling date and 1981 has 37. Average within each year before comparing years, and say how many dates each year had.
8. 8Chemistry is its own sheet, one measurement per row. Total phosphorus (TP) is sampled at the surface about 18 times a year; temperature comes twice per date, surface and deep, so choose one.
9. 9Not the same unit as the monthly plankton card. Daphnia here are per litre of the top 50 m; the 1974 to 2010 file counts them per square metre of lake surface.

• Is the annual mean density of Daphnia related to annual mean total phosphorus in Lake Geneva, 1974 to 2015?

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Is the density of Daphnia in spring related to spring surface water temperature in Lake Geneva, 1974 to 2015?

• Zooplankton_Abundance sheet: waterbody_name, year_yyyy, month_mm, day_of_month_dd, taxa_name_orig_upper (the animal and its stage, in French) and zoop_value, individuals per litre
• zoop_mesh_um: the mesh of the net, 212 or 64 µm
• Water_Chemistry sheet: W_parameter (TP is total phosphorus in µg per litre; also temperature, secchi and chlorophyll_a), W_value, and W_zone (surface, or hypolimnion for the deep water)

1. 1The file publishes 24,759 rows and holds 503 observations. Every station-year is repeated, usually 50 times but sometimes 28, 32, 33 or 34, identical apart from an internal id. The duplication is not even consistent, so checking one group and assuming the rest match will still leave you wrong.
2. 2Most scores rest on one or two campaigns, not twelve. That is far less underlying sampling than the chemistry files, so a single year's score is a noisier number than it looks.
3. 3The index is calculated, not measured: a published formula turns a whole community into one value. Read what the formula gives most weight to before you base a conclusion on it.
4. 4Because the scores come already sorted into five bands, it is tempting to analyse the bands. They are ordinal, so a mean of them means nothing. Use the numeric index and keep the classes for description.
5. 5Station names are not unique. Group by CODEMESURE.

• A comparison of the diatom index between stations upstream and downstream of urban areas.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Do the diatom index and the chemistry agree about which stations are the worst?
• How has the diatom index at Geneva's river monitoring stations changed since 1998?

• DI-CH, the Swiss diatom index, running about 1.4 to 8.0
• A five-level quality class from Mauvaise to Très bonne
• Number of campaigns behind the score, usually one or two
• Station, watercourse, year and coordinates

1. 1The cleanest file in the family: 554 rows and 554 observations, no duplication at all. That is itself the lesson: two other files in this family are duplicated fiftyfold, and nothing on the outside tells them apart. Every file has to be checked on its own.
2. 2The series stops in 2022. It sits alongside files running to 2025, so a comparison across the family silently ends three years early unless you notice.
3. 3Scores rest on between 1 and 4 samples, and 46 of them rest on a single sample. Filter, or say why you did not.
4. 4The index has a ceiling of 20, so improvement at an already-good station is compressed. A station moving from 17 to 18 is not comparable with one moving from 4 to 5.
5. 5Station names are not unique. Group by CODEMESURE.

• A comparison of invertebrate index scores between renatured reaches and channelised ones.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• How has the biological index at Geneva's river monitoring stations changed between 1995 and 2022?
• Does the number of taxa tell the same story as the index score?

• The mean biological index, running 2.0 to 18.25
• The number of taxa recorded
• A five-level quality class from Mauvaise to Très bonne
• The number of samples behind the mean, between 1 and 4

1. 192 rows is the whole file. 55 stations, 26 watercourses, and 28 of those stations were surveyed exactly once. This is too small for a t-test on its own. Use it alongside one of the chemistry files, not as the basis of an investigation.
2. 2The file is not UTF-8. It is Latin-1, and opening it as UTF-8 turns "Très bonne" into "Très bonne" in every row of the class column. The first header also carries a byte-order mark, so the first column arrives named `OBJECTID` unless your reader strips it.
3. 3Semicolon-separated, like the rest of this family.
4. 4CLASSE is ordinal: five ordered bands, so a mean of them means nothing. Use SCORE_FINAL for arithmetic and keep the classes for description. Seven rows have no class at all.
5. 5The four component columns are called PARAMETRE_1 to PARAMETRE_4 and nothing in the file says what they measure. The documentation folder inside the zip does, and it is the difference between a variable and a column heading.
6. 6Fifteen distinct years span 2009 to 2025, so the gaps are real. A station's two visits may be a decade apart, which is a long time in a river and a short time in a fish population.

• Do renatured reaches carry higher fish index scores than channelised ones?

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Which watercourses have never scored above Moyenne, and what do they have in common?
• Has the fish index at the stations surveyed more than once changed between visits?

• SCORE_FINAL, the fish index, running 0 to 12
• CLASSE, five ordered bands from Médiocre to Très bonne
• PARAMETRE_1 to PARAMETRE_4, the components behind the score, named only in the documentation
• Station code and name, watercourse, survey date, Swiss coordinates

1. 1Twenty-two lines of header before the column names, carrying the reef's coordinates and its own maximum monthly mean. Those are not decoration: the threshold is what everything else is measured against.
2. 2The header dates and the table dates are in different orders. The header says `First Valid BAA Date: 1985 31 03`, which can only be 31 March, while the table below is `YYYY MM DD`. Two date formats in one file, twelve lines apart.
3. 3The alert level is derived, not measured. BAA comes from DHW, DHW comes from HotSpot, and HotSpot is the SST minus this reef's own maximum monthly mean. The observation is the temperature; everything else is arithmetic on it, and a conclusion built on the alert level is a conclusion about a formula.
4. 4The threshold is local, so raw temperature does not compare across reefs. Aden's maximum monthly mean is 30.48 °C. A reef at 26 °C may be in severe stress and Aden at 29 °C may be perfectly comfortable. Compare HotSpot or DHW, never SST.
5. 5One reef file is 15,231 daily rows and there are 214 of them, so a question about several reefs is three million rows. Pick your reefs before you start downloading.
6. 6The series runs to yesterday, so the current year is always partial and any annual figure for it is a part-year, which matters most in the months when bleaching actually happens.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• How many days a year has one reef spent above its bleaching threshold, and has that changed since 1985?
• Do two reefs at similar latitudes accumulate heat stress at the same rate?
• In the worst year on record for one reef, how long did the heat stress last?

• SST_MIN and SST_MAX, the day's range in °C
• SSTA, how far the day sits from normal for that reef
• HotSpot, degrees above the reef's own maximum monthly mean
• DHW, degree heating weeks: heat stress accumulated over the previous twelve weeks
• BAA, the bleaching alert level from 0 to 4, which is derived from the two above

1. 1Records measure searching as well as wolves. Confirmed records fell from 6,376 in 2024 to 4,408 in 2025, while the OFB's population estimate rose from 1,013 to 1,082. A count of evidence is not a count of animals, and the OFB's own estimate comes from genetics, not from this file.
2. 2Keep only Retenu. Three records in ten are not a confirmed wolf: 7,206 rejected, 9,854 impossible to tell, 3,199 still being analysed. Every year has some still being analysed, so counts can change after you download.
3. 3The location is blurred on purpose. Every point sits at the centre of its commune, and nothing appears until a month after it is found. Count by commune or department, never by distance.
4. 4The evidence types are single letters, and the file does not say what they mean. From the genetics and the OFB's own list: V is a sighting (half the file), F droppings, T tracks, H a howl, U urine, P hair, S blood. C looks like the carcass of wild prey and D a dead wolf, whose yearly count tracks the culling. E is unexplained. Check against the OFB before you rely on any of them.
5. 545 confirmed records are duplicates: the same reference code, date and commune under two ids. Remove them by the r column, which is how 46,953 becomes 46,908.
6. 62026 is a part year, running to late August at the time of checking.
7. 7Livestock attacks are not in this file. Those are a separate state system; this one is evidence of the wolf itself. The attacks are published by canton and year, 2010 to 2022: they have their own card. Joining the two needs INSEE's 2021 commune list, since this file names the commune but not its canton.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• How has the number of French communes with confirmed wolf evidence changed between 2013 and 2024?
• When did confirmed wolf evidence become regular in the Ain, Doubs and Jura departments, and how fast did it grow after that?
• Is evidence reported by members of the public confirmed as wolf less often than evidence reported by OFB staff?

• date_indice: when the evidence was found
• d and c: department and commune. The map point is the centre of the commune, never the find itself
• f: the verdict. Retenu (confirmed wolf), Non retenu (not a wolf), Invérifiable (cannot tell), En cours d'expertise (still being analysed)
• ti: the type of evidence, as a single letter
• o: who reported it: OFB staff, national parks, hunters' federations, the forestry office, or a member of the public
• g_sp: the species a genetic test found, for evidence that was tested

1. 1It is not the Big Butterfly Count. That is the survey everyone has heard of, but it publishes only yearly headlines, such as 10.3 butterflies per 15-minute count in 2025. It is a good way into the issue. This file is the one you analyse.
2. 2The index is not a count. It is the logarithm (base 10) of a relative index, scaled so that each species' average over the whole series is 2. An index of 2.3 means about twice the average year, not 2.3 butterflies. To compare species, compare their change, never their index values.
3. 3Five countries in one file. Every species appears for the UK and again for England, Wales, Scotland and Northern Ireland. Filter to one COUNTRY first, or the UK is counted alongside its own parts.
4. 4The number of sites grew enormously. The Peacock was indexed from 37 sites in 1976 and 2,859 in 2023. Early years rest on far fewer places, so say how many sites stand behind your first years.
5. 5Species start in different years. 36 of the 59 start in 1976; others start as late as 2003 and 2009. Compare species over years they all share.
6. 6Habitat specialist and wider countryside are JNCC's groups, not yours. Take the lists from JNCC's UK Biodiversity Indicator C6 (26 specialists, 24 wider countryside species) and cite it. Sorting species yourself would decide your result.
7. 7Two versions. This file ends in 2023. The UKBMS official statistics spreadsheet covers 2025 but gives one trend per species, and its significance stars sit inside the numbers (-48***), which turns the cell into text.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Have UK habitat specialist butterflies changed in abundance between 1976 and 2023 more than wider countryside species?
• Is the change in abundance since 1976 different in England and in Scotland, for the species recorded in both?
• How has the abundance of one farmland butterfly, such as the Small Tortoiseshell, changed between 1990 and 2023?

• COMMON_NAME and SPECIES: the butterfly, in English and Latin
• YEAR
• COLLATED_INDEX: how abundant the species was that year, on a log scale where the species' own average is 2
• N_SITES: how many sites the index was calculated from
• YEAR_RANK: the year's rank for that species, 1 being its best year
• COUNTRY: UK, England, Wales, Scotland or Northern Ireland

1. 1Pollack is not pollock. Pollock (Alaska pollock) is the certified-sustainable fish in fish fingers, caught in the North Pacific and imported. Pollack, listed here as Pollack (Lythe), is the UK's own inshore fish, and scientists advised zero catch for it in 2025. Coley (saithe) is a third species. Check which one a source means before you use it.
2. 2Only five years per species. Table 2.2 covers 2020 to 2024. For a longer species series you have to copy the same table from earlier years' reports and stack them, checking that the species names and units have not changed.
3. 3Octopus has no row. The 2025 octopus bloom off Devon and Cornwall, the largest of four in 125 years, is real, but octopus sits inside Other Shellfish here. The MMO publishes octopus figures only as monthly summaries on its octopus bloom page. Use the bloom as a way into the issue, not as your dataset.
4. 4Thousands of tonnes, to one decimal place. Pollack falls from 1.4 to 0.8. The table cannot see a change smaller than 100 tonnes, which is a large share of a small fishery.
5. 5Landings are not catches, and not the stock. Fish thrown back, fish caught by anglers and fish landed abroad are not in the landings into UK ports. A fall in landings may mean fewer fish, a lower limit or fewer boats.
6. 6Value and quantity move differently. Value is in cash of each year and is not adjusted for inflation. Compare weights, or adjust the values and say how.
7. 7Do not download the underlying data file. The species-by-port-by-month version on the same page is 108 MB, and over 3 GB once opened. The summary tables in this spreadsheet are the usable form.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• How have UK landings of demersal, pelagic and shellfish species changed between 1938 and 2024?
• How have UK landings of pollack changed between 2020 and 2024, compared with the catch limit set for it?
• Which ports landed the most shellfish in 2024, and how much of it was crab and lobster?

• Species: 36 named species, grouped as demersal (bottom-living fish), pelagic (open-water fish) and shellfish
• Quantity: live weight landed, in thousand tonnes to one decimal place
• Value: first-sale value, in £ million
• Year: five years side by side, as columns

1. 1A catch is not a population. It depends on how many people fished, the rules, and the weather: in a dry summer rivers run low and salmon are hard to catch. Say that your data measures catch, and that the population is estimated separately in the annual salmon stock assessment.
2. 2The Wye has not fallen most. Its average catch for 2021 to 2025 is 23% of its 2008 to 2012 average; for Wales as a whole it is 16%, and for the Usk 8%. If your question assumes the Wye stands out, this table may disagree. That is a finding, not a failure.
3. 3Salmon are falling everywhere. Survival at sea has dropped across the North Atlantic, so a decline on one river does not show a cause on that river. Compare with other rivers before you blame anything local, including chicken farms.
4. 4Effort is only for the latest season. Table 37 gives days fished for 2025 alone. To compare catch per day across years, take Table 37 from earlier editions of the same publication.
5. 5Released fish are included. The figures count every salmon caught, kept or not, and on the Wye every salmon must now be released. A fish caught twice may be counted twice.
6. 6The Wye is listed under Wales, although much of it runs through England. Net catches (the Wye lave nets) are in a separate table, Table 32.
7. 7Do not let one year carry the trend. 2025's 60 fish is the lowest in the series. Use averages over several years, as in the comparison above, rather than first year against last.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• How has the declared salmon rod catch on the River Wye changed between 2008 and 2025, compared with the total for Wales?
• Has the salmon rod catch fallen more on the Wye than on other Welsh rivers, comparing 2021 to 2025 with 2008 to 2012?
• Is the share of multi-sea-winter salmon in the 2025 rod catch different on the Wye from other large rivers?

• River, with the Environment Agency area or Wales it is listed under
• One column per year, 2008 to 2025: salmon caught by rod, released fish included
• Days fished and catch per licence day (Table 37): how hard anglers fished, for the latest season only
• Grilse and multi-sea-winter salmon (Table 38): how many returned after one winter at sea and how many after two or more

1. 1One year per file, and no year in it. Each export is one hunting year for every canton, and the CSV has no year column. Add one as you stack the files, or 34 files become one undated pile.
2. 2A missing canton is not a zero. Cantons with no report are left out of that year's table, not listed as 0. Between 2004 and 2025, 30 canton-years are missing, almost all from cantons with few or no boar (Nidwalden, Obwalden, Glarus, Uri, Schwyz). Treat them as no data and say so.
3. 3The names won't join. This file says Argovie, Genève, Berne and Soleure; the land-use file says Aargau, Geneva, Bern and Solothurn. Add a two-letter canton code to both before you match them.
4. 4Divide by area. Vaud pays more than Basel-Landschaft because it is six times bigger. Compare damage per km² of canton, or per hectare of farmland, never raw francs.
5. 5Francs paid are not damage done. The figure is compensation, and each canton sets its own rules: small claims are not paid, and a farmer without an electric fence may get nothing. Say that your variable is compensation.
6. 6Hunting years, not calendar years. In most cantons the hunting year runs from 1 April to 31 March, so "2018" ends in March 2019. The land-use survey uses the calendar year the photographs were taken.
7. 7The range is still spreading. Bern and Fribourg have plenty of arable land but little damage (CHF 12 and 56 per km² in 2014-18), because boar spread in from the north-west and are only now reaching the central plain. Part of your pattern is where the boar have got to, not only what the land is like.
8. 8Big swings between years. Damage was CHF 4.28 million in 2024 and 2.74 million in 2023. Good years for beech and oak nuts, and snowy winters, move boar numbers. Average several years before comparing cantons.
9. 9Geneva shoots no boar. In the Chasse table Geneva is always 0, because hunting is banned there. Its culls are under Tir spécial. Add Chasse and Tir spécial together for every canton, or Geneva looks empty.
10. 10Sum the columns. The shot and found-dead tables split into males, females, young and undetermined. Before 1992 everything is in undetermined, so a total needs all four columns.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Is wild boar damage per km² related to the share of arable land across the 26 Swiss cantons, 2014 to 2018?
• Is wild boar damage per km² more closely related to forest cover or to arable land across Swiss cantons?
• Has wild boar crop damage in Switzerland risen between 1992 and 2025, and do the years with high damage match the years with many boar shot?

• Area: the canton, in French (Argovie, Genève, Vaud)
• Dégats en francs: damage caused by wild boar in that hunting year, in Swiss francs
• On the Faune sauvage tab, for Sanglier: Chasse (shot by hunters), Tir spécial (special culls, including Geneva's wardens) and Gibier péri (found dead), each split into males, females and young from 1992
• Pair it with the land-use card above for each canton's arable land, farmland and forest

1. 1Most rows are zero. Every count lists all 38 species, so 67,148 of 76,228 rows read 0. That is real data (no birds of that species seen), not a gap. Average across counts, don't drop the zeros.
2. 2Count the counts, not the rows. One five-minute count is one Site, Transect, Station, Date, Time and Observer together. The number of counts varies from 40 to 179 per forest per season, so compare birds per count, never totals.
3. 3Before is winter, after is summer. Birds are easier to hear in summer, so every forest goes up after winter 2012, poisoned or not. Compare the treated forest with its untreated partner, not with its own winter.
4. 41080 kills rats and possums. Rats are on the chew cards too, and on the West Coast they were back on 48% of cards by 2014/15. You can't say the birds responded to possums alone. And watch the mice: on the East Coast treated site they rose to 82% of cards.
5. 5Two names, one bird. dunnock and hedge_sparrow are the same species, entered under both names. Add them together. finch and unknown are birds nobody could identify: leave them out and say so.
6. 6Which birds are native? Silvereye is the most counted bird in the file (3,308) and arrived from Australia by itself in the 1850s, so it is usually treated as native. On the West Coast most of the treated forest's late gain is silvereye. Decide before you count, and check your answer both ways.
7. 7Two forests per side. Hundreds of counts come from just two poisoned and two untreated forests, and most stations were counted twice on the same day. The counts are not independent forests. Say so in your evaluation.
8. 8Dates are day/month/two-digit year (7/12/12 is 7 December 2012), and the files end .txt but are comma-separated.

• Is the number of native birds per five-minute count higher in the 1080-treated forests than in the untreated forests, in the third summer after the drop?
• How did the share of chew cards bitten by possums change in treated and untreated forests between winter 2012 and summer 2014/15?
• Did bellbirds increase more in the treated forests than in the untreated forests after the 1080 drop?

• species, count: the bird and how many were seen or heard in the five minutes. Every count lists all 38 species, so most rows are 0
• Area, Site, treatment: East or West Coast, the forest, and whether it was poisoned (treatment) or not (control)
• Season: Winter 2012 (before), then Summer 2012/13, 2013/14 and 2014/15
• Transect, Station, Date, Time, Observer: which count it was; together they identify one five-minute count
• Sun, Wind, Rain, Temp, Noise: conditions during the count, which change how many birds you hear
• Chew cards: species (possum, rat, mouse), total_cards and cards_marked, so the share bitten is marked divided by total

1. 1Records are not population. There are 605 records for 2020 and 9 for 2023, because of when scientists published, not how many lionfish there were. Use where and when lionfish were first recorded, never the number of records in a year.
2. 2No record is not proof of absence. Presence-only data can show that lionfish reached a place, not that they are missing from one. For your "not recorded" places, choose coasts with plenty of divers and scientists, and say that is what you did.
3. 3Memories, not sightings. Records from Spain (2002), Turkey (2005) and northern Greece (2008) come from studies that asked dive centres and locals when they first saw lionfish. Check associatedReferences, and decide whether to keep them before you start.
4. 4Twenty years of nothing. The first record is Israel, 1991, then nothing solid until Lebanon in 2012. A lag like this is common in invasions, so don't draw a straight line from 1991.
5. 5Year-only dates. 1,178 of the 2,019 records give only a year. That is fine for first records by country, not for anything by month.
6. 6Use this copy. The same database on SEANOE is the older 2022 version: 4,015 records up to 2020, with commas as decimal points. The GBIF archive is updated to January 2026.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Is the February sea surface temperature higher at places where lionfish have been recorded than at places in the Mediterranean where they have not?
• Did each country's first lionfish record come when its February sea surface temperature had passed about 15 °C?
• How far north has the northernmost lionfish record in the Mediterranean moved each year since 2012?

• scientificName: the species. The file holds 234 exotic fish; keep Pterois miles
• eventDate: when it was seen. Year only in 1,178 rows, year and month or a full date in the rest
• country, locality, decimalLatitude, decimalLongitude: where
• individualCount: how many were seen, blank in 375 rows
• associatedReferences: the paper each record comes from. Read it for any record that looks wrong

1. 1The gear changed in 2024. All sharks caught ran at about 500-850 a year until 2022, then 948, 1,497 and 3,430 in 2023-25, as new locations, daily servicing and new drumlines came in. That is effort, not more sharks. Stop long-term trends at 2023, or compare months and species within the same years.
2. 2Catch is not the number of sharks. It depends on how much gear is in the water, where and how often it is checked. Say what your variable is: sharks caught.
3. 3Add up NumberCaught, don't count rows. 4,003 rows record more than one animal.
4. 4White sharks are rare. 214 in 30 years, almost all south of Bundaberg. Pool the years for monthly patterns rather than testing year by year.
5. 5Names with a star are a group, not a species. HAMMERHEAD SHARK * means a hammerhead nobody identified to species; there are also 81 UNKNOWN SHARK. Decide whether to include them.
6. 6The 2026 rows stop in May. A part year: leave it out of any yearly total.

• Do nets and drumlines differ in the share of their shark catch that is white sharks, in the four areas that set both gears, 1996 to 2023?

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Do white sharks and bull sharks peak in different months on the southern beaches, and does that match the water temperature each prefers?
• Are bull sharks caught in a higher share of months in the southern areas in the warmest years than in the coolest, 1996 to 2023?

• Year, Month: when
• Area, BeachName: where, in 12 areas from Cairns (about 17°S) to the Gold Coast (28°S)
• Gear: Net or Drum (a baited hook on a buoy)
• SpeciesGroup: SHARK, TURTLE, MAMMAL or OTHER. CommonName and ScientificName: the species
• NumberCaught: how many. Alive/Deceased and Fate: whether the animal was released, died or was killed

1. 1Don't read the Pivot sheet. It is a ready-made summary that opens filtered to one small fleet (UK-Turks and Caicos). Work from the Data sheet.
2. 2Choose the stock. Shortfin mako has three stocks in the file (ATN, ATS, MED) and the 2022 ban applies to ATN, the North Atlantic, only. Filter on Stock before you add anything up.
3. 3Add the dead discards to the landings. When keeping a species is banned, a fleet's catch can move from L to DD rather than stop: Spain's North Atlantic mako is 870 t landed in 2020 and 585 t discarded dead in 2021. Reading L alone mistakes a change of label for a change in fishing. Sharks released alive are not recorded here.
4. 4More countries reported over time. Two flags reported North Atlantic mako in 1975 and 19 by 2020, and in the 2000s over 90,000 t of sharks were recorded only as a group (SHX, SKH). Some of the rise before 2000 is better reporting, not more fishing: compare fleets that reported throughout, or start later.
5. 5Tonnes caught is not the number of sharks alive. It depends on how much fishing was done: most of these makos are caught on longlines set for swordfish (Spain's biggest mako fleet is coded EU.ESP-ES-SWO). A country with a big longline fleet catches more, whatever its wealth.
6. 65,960 rows are zero. They are reports of no catch, not missing data. Keep them for counting who reported; leave them out of averages.
7. 7Estimated rows. CatchSource marks 1,113 shark rows as Estimated: ICCAT's scientists filled a gap, often by carrying the last reported value forward. Decide whether to keep them and say so.
8. 82024 is the latest year and may still be revised. Download the current version and cite its date.

• Across the countries whose fleets catch North Atlantic shortfin mako, is the tonnage caught related to the country's GDP per person, 1990 to 2021?
• After the ban on keeping North Atlantic shortfin mako in 2022, did the total dead catch (landed plus dead discards) fall below ICCAT's 250 t limit, and what share of it was thrown back dead?
• Did porbeagle landings fall more in the North-east Atlantic, after the EU's zero catch limit in 2010, than in the North-west Atlantic over the same years?

• Species, ScieName, SpeciesGrp: what was caught. Sharks are the groups 4-Sharks (major) and 5-Sharks (other). SMA is shortfin mako, BSH blue shark, POR porbeagle
• Stock: which population. ATN is the North Atlantic, ATS the South Atlantic, MED the Mediterranean (porbeagle is split four ways)
• YearC: the year
• FlagName, FleetCode: which country's boats, in words rather than a standard code. EU members appear separately, e.g. EU-España
• GearGrp: LL is longline, the gear that catches most of these sharks
• CatchTypeCode: L for landed, DD for dead discards (thrown back dead)
• Qty_t: the catch in tonnes

1. 1Warm months are also busy months. More people swim and surf in summer, and this file has no count of people in the water. A link between bites and temperature may be a link between bites and beach use: say so, and think about how you could separate them (the same months in warmer and cooler years, for instance).
2. 2The nets are seasonal. In New South Wales they are in the water from September to April, the warmer half of the year.
3. 3A month with no bite has no row. Count bites per state and month yourself, and put in the zeros, or you will only ever compare months that had a bite. Most months have none: 528 months off NSW from 1982 to 2025 and 211 unprovoked bites.
4. 4Decide on provoked bites before you look. 423 incidents were provoked (spearfishing, handling a caught shark, feeding). Their timing follows the people, not the sharks, so most questions use unprovoked bites only.
5. 5Tidy the text first. State has one "Qld" beside 378 "QLD", and Site.category and Victim.injury have stray capitals and a trailing space. Make them consistent or your counts split in two.
6. 6The shark is often a judgement. Most species were identified from the bite marks and where it happened, not seen, and 66 incidents name no species or "unknown". Wobbegong bites (216) are small and nearly half provoked: decide whether they belong in your question.
7. 7The file's own water temperature is nearly empty. Water.temperature.°C is filled in for only 93 incidents, so use NOAA's sea temperature instead and say why.
8. 8The record grew with the coast's population. Recorded bites rise decade by decade as more people use the sea and records improve. Compare the same months across years, or the share of bites in each month, rather than raw totals across decades. 2026 stops in June.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Are unprovoked shark bites in New South Wales more frequent in the calendar months when the sea off Sydney is warmer, 1982 to 2025?
• Is the sea off each state warmer, on average, in the months when white sharks are blamed for unprovoked bites than in the months when bull and tiger sharks are?
• Were there more unprovoked bites off New South Wales in warmer summers (December to February) than in cooler ones, 1982 to 2025?

• Incident.year, Incident.month: when
• State, Location, Latitude, Longitude: where
• Provoked/unprovoked: whether the person touched, fed, speared or caught the shark first
• Shark.common.name, Shark.identification.method: which shark, and how anyone knows
• Victim.activity, Victim.injury: swimming, boarding, diving and so on; injured, uninjured or fatal
• Site.category: coastal, estuary/harbour, river, island open ocean or ocean/pelagic

1. 1Blank cells mean "same as above". It is a pivot-table export: region, department, arrondissement, canton and id appear only on the first row of each group. Fill them down before you sort or filter.
2. 2The last row is a grand total. Total Résultat (34,995 reports) sits at the bottom. Delete it, or every sum doubles.
3. 3Three rows have no place. Donnée géographique non renseignée covers 23 reports with no canton.
4. 4One canton can take several rows in one year. Fifty cantons are cut by an arrondissement boundary. Add the rows up by department, canton and year: 1,862 rows become 1,687 canton-years.
5. 5A missing canton-year means no paid claim, not missing data. Only canton-years with at least one compensated report are listed, so fill the gaps with zero when you join.
6. 6It counts paid claims, not attacks. Only cases where compensation was decided and the wolf was not ruled out are listed, and the compensation rules changed in 2019.
7. 7Today's cantons for every year. The boundaries are those drawn in 2015, even for 2010 to 2014, and ten codes are towns rather than real cantons (0599 is the whole of Gap).
8. 8It stops in 2022. The note promises 2023 after consolidation; in September 2026 later years exist only as PDFs.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Across French cantons, is the number of compensated wolf attacks on livestock related to the number of confirmed wolf records, 2013 to 2022?
• How did the number of sheep killed or injured for each compensated report change between 2010 and 2022?

• INSEE_DEP and INSEE_CAN: department and canton number. Together they make the INSEE canton code: 04 + 12 = 0412
• Annee: the year
• Compter - numero_constat: the number of compensated attack reports
• Somme - retenues_ovin: sheep killed or injured. Beside it, cattle, goats, dogs, horses and other animals
• INSEE_REG, INSEE_ARR, ID: region, arrondissement and the map id of the canton piece

1. 1You download every Swiss snail and mussel. 344,823 rows and 72 columns: filter to the quagga before opening it in a spreadsheet.
2. 2Two mussels with similar names. 2,888 rows are zebra mussel; the quagga is Dreissena rostriformis bugensis.
3. 3Not only Switzerland. 212 of the 585 quagga rows are in Germany, Austria or France, and stateProvince mixes canton codes with names like Baden-Württemberg.
4. 4No lake column. The lake is not named in any Swiss row, and every point is a 5 km square. A canton is not a lake.
5. 5A record is not an arrival. The earliest Swiss row is December 2016, while RTS reports Lake Geneva colonised since 2015.
6. 6Rows pile up. 277 of the 373 Swiss rows share a date and grid point with another. Do not read the number of rows as how many mussels there are.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• In which order were quagga mussels first recorded in each Swiss canton, 2016 to 2025?
• How many Swiss cantons had a quagga mussel record in each year from 2016 to 2025?

• scientificName: the species. Quagga mussel is Dreissena rostriformis bugensis
• eventDate and year: when it was recorded
• countryCode and stateProvince: the country, and the canton as a two-letter code (VD, GE)
• decimalLatitude, decimalLongitude: where, on a 5 km grid

1. 1Four lines of notes sit above the header. The header is the fifth line and begins Country Name, and there is an empty column after the last year.
2. 2Aggregates sit among the countries: 47 rows such as World, Euro area and Low income. The zip's Metadata_Country file leaves Region blank for every one of them, so use it to take them out.
3. 3Years run across. One column per year: turn them into one column before you match the file to anything with a year column.
4. 4Values jump when the database changes. Switzerland reads 24.8% in 2013 and 9.7% in 2015: a change in what was counted, not land losing its protection. Check any country that jumps before you rely on its year.
5. 5Protected is not managed. The figure counts land designated as protected or conserved, whether or not anyone enforces it.

• Across the countries with at least 1 million hectares of tree cover, does protecting more land in 2015 go with losing less tree cover from 2016 to 2025?

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Across countries, is the share of land that is protected related to GDP per person?

• Country Name, and Country Code, the three-letter ISO code
• 1960 to 2025: one column per year, empty before 2013

1. 1Swiss waters only. The cantons report their own fishers; the French catch on the south shore is not in it. Say that your variable is the Swiss catch.
2. 21911 to 1920 read 0. Those years are missing, not empty nets. Leave them out.
3. 3Use the professional series for anything long. Recreational catches start in 1994, so the total jumps there when they are added in.
4. 4The file does not say what it is. It is named after the species alone (Corégones, 1904-2024.csv), with no lake, fishing or unit in it. Rename it at once.
5. 5A catch is not a population. It depends on how many fishers there are, how many nets they set and the rules they fish under. The official sheet records the nets, but this table does not.
6. 6Use a browser. The table is built in the page, so it cannot be fetched by other programs.

• Is the professional whitefish catch in Lake Geneva's Swiss waters related to the lake's total phosphorus, 1974 to 2015?
• Is the professional perch catch related to total phosphorus in Lake Geneva, 1974 to 2015?

• Année and one column named after the species: Corégones (whitefish) or Perche (perch)
• In the page, before you download: the fishing (Rendement de la pêche professionnelle, de loisir or totale), Poids (tonnes) or Nombre (fish), and the species

1. 1A species that was not seen has no row. 599 of the 5,253 transects in the northern Channel Islands have no purple urchin row at all. List every transect surveyed, from all of its rows, and give it 0 before you average, or every mean comes out too high.
2. 2Group by site_status, never by MPA_Name. An outside site carries the name of the reserve it is paired with: SMI_CUYLER is listed under Harris Point SMR, but it is a reference site.
3. 3Two halves of one place can sit either side of a boundary. SCI_SCORPION_W is inside the Scorpion reserve and SCI_SCORPION_E is outside it. Treat every site code as its own site, as PISCO does.
4. 4A conservation area is not a reserve. SMCA sites (Painted Cave, west Anacapa) allow some fishing, so they fit neither side of an inside-or-outside comparison.
5. 5The site table has many rows per site, one for each year and method, and some are blank. Filter method to SBTL_SWATH_PISCO and drop the rows with a blank site_status before you look anything up. Five island sites (SCI_VALLEY_CEN, _E and _W, SCI_YELLOWBANKS_CEN and _E) have no status at all.
6. 62013 is thin. Only 19 island sites were surveyed that year, against 34 to 69 in every other year from 2008 to 2024.
7. 7Giant kelp comes in several rows per transect, one for each number of stipes (the size column), so add them up. Urchins are counted only over 2.5 cm across, and kelp only over 1 m tall.
8. 8The package is updated. This card counted version 1.11 (January 2026). A later version can add years and move every count.

• In the northern Channel Islands, did purple urchin density differ between kelp forest sites inside and outside no-take reserves before (2010 to 2012) and after (2014 to 2024) sea star wasting disease removed the sunflower star?
• Are red urchins, which are fished, denser at kelp forest sites inside the Channel Islands reserves than at sites outside them?

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Across the northern Channel Islands' kelp forest sites, is purple urchin density related to the density of giant kelp?

• site: the site code, such as SCI_GULL_ISLE_E, which is also the key to the site table
• year, month, day, zone and transect: together, one transect on one day
• classcode and count: the species and how many were seen. STRPURAD is purple urchin, MESFRAAD red urchin, MACPYRAD giant kelp and PYCHEL sunflower star; the taxon table in the same package lists the rest
• In the site table, site_status (mpa or reference) and mpa_type (SMR, a state marine reserve where all fishing is banned, or SMCA, which allows some)

1. 1Not one EU member state is in this file as a country. They report through a single "European Union (28)" row, because the Union ratified the treaty in its own right. Germany, France, Italy, Spain and Poland are simply absent, and nothing says so: you search for Germany, find nothing, and think you made a mistake. No other problem in this file limits your questions more: it cuts the developed countries you can compare from fifty down to eighteen.
2. 2Removing the aggregates takes two rules, not one. Only the European Union row has a blank code. The other seven, the six continents and World, carry codes beginning OWID_, so the obvious "delete the rows with no code" leaves seven aggregates sitting among your countries and your country count reads 176 instead of 169. Unlike the CO2 file, no real country here carries an OWID_ code, so the two-part rule is safe.
3. 3Consumption can be negative, and 211 country rows are. UNEP defines it as production plus imports minus exports, so a country running down a stockpile posts a negative figure: Argentina is −528 ODP tonnes of carbon tetrachloride in 1990. It is a real accounting entry rather than an error, and either way it will distort a mean.
4. 4The zeros are filled in, not measured. Our World in Data's companion series on these same 5,642 rows is labelled "zero-filled", and 297 country rows read zero for every one of the six chemicals. Zero here mixes "none consumed" with "nothing reported", and the difference is the whole story of compliance.
5. 5The last year is a cliff, not a collapse. 2022 carries 14 countries. 2021 carries 168. Plot a global total to 2022 and it shows a huge fall that is really 92% of the reporting countries missing.
6. 6Whether a party is an Article 5 party is not in the file, and it is the one column a question about the treaty's fairness needs. You have to bring it: the Handbook for the Montreal Protocol defines the two groups, and the Ozone Secretariat's list of parties says which group each country is in. Check it country by country: seven former Soviet states, including Azerbaijan and Kazakhstan, are on the developed side.
7. 7ODP tonnes are weighted by how much damage a substance does, not by its mass. A tonne of halon counts for many times a tonne of HCFC, so these numbers cannot be added to anything measured in real tonnes, and part of any trend you find comes from the weighting.
8. 8Eight aggregates share the entity column with the 169 countries: the continents, the EU-28 and World.
9. 9The download has one column per chemical family and no total. If your question is about ozone-depleting substances as a whole, you are making that total yourself, and the sibling chart that already has it is a different file.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• In 1996, the deadline for developed parties, how much of its own peak was each party still using?
• In one country, did HCFC consumption rise as CFC consumption fell, and by how much of the gap?
• Compare the phase-out paths of two countries at different income levels against the dates the treaty set for each.

• Chlorofluorocarbons (CFCs), the original problem
• Hydrochlorofluorocarbons (HCFCs), the substitute that became the next problem
• Halons, carbon tetrachloride, methyl chloroform, methyl bromide
• Entity, code and year

1. 1Four lines of preamble, and the site's name sits on its own row. Point a spreadsheet at the file and your header is "Data supplied by UK-AIR on 29/6/2024". The real header is row five.
2. 2134 columns for 44 pollutants, because every pollutant is followed by its own status column and its own unit column. Select a pollutant by its position and you may pick up a status column instead.
3. 3Column names carry HTML. Particulates arrive as `PM<sub>10</sub> particulate matter`, which no lookup you type by hand will ever match.
4. 4Missing hours are blank, and there are hundreds. 120 hours of nitrogen dioxide are missing in 2015, 474 in 2019 and 430 in 2024. An annual mean is a mean of the hours the instrument was working, and the number of those changes year to year.
5. 52024 has 8,784 rows rather than 8,760. It is a leap year, and anything that assumes a fixed row count breaks on it quietly.
6. 6The status letter matters: R is ratified, P and P* are provisional. Recent years arrive provisional and the numbers can still change.
7. 7There is unusually heavy confounding here, and ignoring it is the main way to lose marks. The zone began in April 2019, the pandemic emptied the road in 2020, the zone expanded twice afterwards, and a congestion charge has been operating on the same street since 2003. A before-and-after across 2020 compares two very different situations.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Did nitrogen dioxide at Marylebone Road fall faster in the five years after April 2019 than in the five years before it?
• How does the weekday and weekend difference in nitrogen dioxide at one site compare before and after the zone began?
• Which pollutant at one site fell most between 2015 and 2024, and which barely moved?

• Nitrogen dioxide, nitric oxide and NOx, µg/m³, hourly
• PM10 and PM2.5, ozone, carbon monoxide, sulphur dioxide
• At this site, about thirty hydrocarbons and six black-carbon channels as well
• A status column and a unit column after every single pollutant

1. 1The header is on the second row. The first line is a title, `Land-Ocean: Global Means`, so a reader that assumes row one is the header treats the column names as data.
2. 2Missing values are three asterisks. Nine cells read `***`, all of them the current year's months and the two overlapping-year columns for 1880. Any tool that converts the columns to numbers will either fail or quietly leave them as text.
3. 3Monthly and summary columns sit in the same row. J-D, D-N and the four seasons are averages of the months beside them, so if you reshape the file into one row per month without removing them, you count every year twice.
4. 4J-D and D-N are not the same year. One is January to December, the other December to November, and they differ by a tenth of a degree in places. Choose one and say which.
5. 5An anomaly is a difference from the 1951 to 1980 average, not a temperature. Every number in this file is relative to a baseline, and that baseline is not the one the 1.5 °C target uses.
6. 6The last row is a part-year. 2026 has values to August and asterisks after, so an annual figure for the current year does not exist and a chart drawn to it ends on a month rather than a year.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• How does the rate of warming since 1980 compare with the rate before it?
• Which season has warmed fastest, and does that hold across the whole record?
• How many of the ten warmest years in this file are in the last fifteen years of it?

• Twelve monthly anomalies, in degrees Celsius against the 1951 to 1980 mean
• J-D and D-N, two different ways of averaging a year
• DJF, MAM, JJA and SON, the four seasons
• Nothing else. It is a small file that says one thing very well

1. 1Missing months are −99999. Brest has 312 of them in 2,622 rows, 12% of the record, and there is no header to warn you. Averaging the column as it downloads produces a number thousands of millimetres wrong, in the wrong direction, and it looks like a plausible sea level.
2. 2There is no header row. Four columns separated by semicolons, and the only way to know what they are is the documentation page. A reader who guesses will guess wrong about the fourth.
3. 3The date is a decimal year. 1807.0417 is mid-January 1807. Nothing will parse that as a date for you, and rounding it to 1807 puts twelve months on top of each other.
4. 4The datum is local, and this is the trap that matters. Values are millimetres above a Revised Local Reference set roughly 7,000 mm below that station's own mean sea level. Brest reading 6,900 and another station reading 6,900 are not the same height above anything. Only the change within a single station means something; the difference between two stations means nothing at all.
5. 5A tide gauge measures the sea against the land, and the land moves. Scandinavian stations show sea level falling because the land is still rising after the ice sheets left. That is not an error in the data, it is what the instrument measures, and any conclusion about the ocean has to say so.
6. 6348 of Brest's values carry an attention flag in the fourth column. They are not wrong, they are marked, and saying whether you kept them is part of the method.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• How much has mean sea level risen at one named station, and has the rate changed over its record?
• Do two stations on the same coast agree, and if not, what could make them differ?
• How does the trend at a station in a subsiding city compare with one on stable ground?

• Mean sea level for the month, in millimetres
• The date, as a decimal year: 1807.0417 is mid-January 1807
• A missing-data indicator, and a flag marking values the compilers want you to look at
• The station list adds latitude, longitude, name and country for all 1,620

1. 1The file is anomalies, not temperatures. Each value is how far that year sat above or below the 1961 to 1990 average. For the mean annual temperature in °C, add the average printed beside the Download link on the page: 25.8 °C. For a trend or a correlation you do not need to: adding the same number to every year moves the line without changing its slope.
2. 2Two averages, and they do not quite agree. Tick 91-20 on the page and every anomaly drops by exactly 0.25 °C, but the printed average rises from 25.8 to 26.0 °C, only 0.2. Both averages are rounded to one decimal place, so the two routes to °C give years 0.05 °C apart. Choose one, say which, and never mix them.
3. 3No headings, and the first column is two dates run together. `199301199312` is January 1993 to December 1993, so the year is the first four characters: in Excel, =LEFT(A1,4). The columns are separated by spaces: use Data › Text to Columns, Delimited, Space, and tick "Treat consecutive delimiters as one".
4. 4Seven points, 220 km apart. The series averages the seven 2° grid points of NOAA's ERSST v5 that lie in or near the Marine Park. It describes the reef region, not any one reef, and NOAA says ERSST suits long-term, basin-wide studies best. For a single reef, use Coral Reef Watch.
5. 5Ships, buoys and floats, and fewer of them early on. ERSST is built from measurements at sea and filled in statistically where there are none, so the early 1900s rest on far fewer observations than recent decades.
6. 6The numbers can change. The Bureau recalculates the series whenever NOAA updates ERSST: in March 2020 every value from 2008 onwards changed when missing buoy reports were restored. Record the date you downloaded and keep your copy.
7. 7Each year is one point, and it trends. Almost anything that has changed steadily since 1993 will correlate with it. Correlate the change from one year to the next as well as the values.
8. 8Old links to the graph are dead. Links of the form bom.gov.au/climate/change/#tabs=Tracker now open a general climate change page without the graph. Use the page this card links to.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• Has the sea over the Great Barrier Reef warmed faster or slower than the globe between 1993 and 2023?
• Has the Great Barrier Reef warmed as fast as the Coral Sea as a whole between 1993 and 2023?
• Has the tropical Great Barrier Reef warmed faster or slower than the temperate Tasman Sea between 1993 and 2023?

• First column: the twelve months the value covers, start and end run together, so `199301199312` is January to December 1993
• Second column: the sea surface temperature anomaly in °C against 1961 to 1990, to two decimal places
• On the page, beside Download: the 1961 to 1990 average, 25.8 °C, which turns an anomaly into a temperature
• Change Region on the page for the Coral Sea, the Tasman Sea and five other Australian regions

1. 1Three comment lines before the header. Title, units and base period come first, each starting with #, so a spreadsheet that assumes row one is the header reads the title as a column name.
2. 2The date is one number. 185001 is January 1850, and plotted as a number the series jumps 89 units between every December and January. Split it into year and month (divide by 100) before you draw anything.
3. 3The warmest month is a February. February 2024 (+3.95) tops the file, and the top eight are all winter months. An anomaly measures how far a month was from its own normal, and European winters swing further (standard deviation 1.31 against 0.75 for summers). "Hottest" and "most unusual" are different questions; say which you are asking.
4. 4Monthly values jump around; yearly ones do not. Since 1990 Europe's monthly values vary about three times as much as the global series. Take annual or seasonal means before you fit a trend, or your trend will just show the weather.
5. 5Three baselines, three zeros. This file uses 1901 to 2000. NASA's GISTEMP uses 1951 to 1980, and the Paris limit uses 1850 to 1900. On this file Europe's 1850 to 1900 average is −0.34, so warming since pre-industrial is the anomaly plus 0.34.
6. 6The end year is part of the web address. The download URL ends 1850-2026. Ask for a year that has not started and you get an HTML error page saved with a .csv name, not a shorter file. In 2027, edit the year in the URL.
7. 72026 has eight months, January to August. An annual mean for 2026 is a mean of eight months, and it will not be the same number once winter is in.

Background for step 1, not a research question: none of these sets what people do against what happens to nature.

• How much faster has Europe's land warmed since 1975 than the globe as a whole?
• Have European summers or winters warmed faster since 1975, and which season varies more from year to year?
• How many of Europe's ten warmest complete years came after 2010?

• The month, written as one six-digit number: 185001 is January 1850
• The departure from the 1901 to 2000 average for that calendar month, in °C: from −4.60 to +3.95
• Nothing else, which is what makes it quick to pair with the global series from the same page