A real, messy dataset — not a tidy textbook table — is where exam questions about sampling, cleaning and interpretation actually live. Knowing its structure in advance is half the battle.
7 min read3 worked examples
What the large data set is
Alongside the usual exam content, the specification includes a large data set — a real, pre-released set of measurements that you're expected to be familiar with before the exam. For this course it's a set of meteorological data: daily weather readings (such as daily mean temperature, total rainfall, and mean wind speed) recorded at several UK and overseas weather stations, for two separate years.
You won't need to memorise every number in it. What you're expected to know is its structure: what variables it records, roughly what units they're in, what a typical value looks like, and how the data is organised — so that when a question hands you a small extract, you can interpret it immediately rather than starting from scratch.
Why an exam uses a real dataset at all
Real data is messy in ways invented data never is: missing readings, rogue values from equipment faults, awkward units, or variables recorded to inconsistent precision. Working with the large data set in advance is practice at the skills a made-up, perfectly clean table can't test — spotting what's missing, deciding whether a value is plausible, and knowing which summary statistic is even appropriate for a given variable.
What the data set records
The meteorological data set covers daily readings from several UK weather stations and a small number of overseas stations, across two separate historical periods roughly thirty years apart — so that questions can compare a recent period against an older one. Each UK station's entry typically records:
Daily mean temperature (°C) — the average of the hourly readings across a 24-hour period.
Daily total rainfall (mm) — including melted snow and hail; very small amounts are recorded as "trace" rather than as a number.
Daily total sunshine (hours) — the number of hours of direct sunshine recorded that day.
Daily mean wind direction and windspeed — windspeed is given in knots and is also grouped into a category on the Beaufort scale.
Daily maximum gust (knots) — the single highest instantaneous windspeed recorded that day, together with the direction it blew from.
Daily maximum relative humidity (%) — values above roughly 95% are associated with misty or foggy conditions.
Daily mean cloud cover — measured in eighths of the sky covered ("oktas").
Daily mean visibility — the greatest distance an object can be seen in daylight, measured in tens of metres.
Daily mean pressure (hectopascals, hPa).
Overseas stations record a smaller subset of these variables — typically just mean temperature, total rainfall, mean pressure and mean windspeed — since not every instrument is available at every station.
Beaufort scale
Descriptive term
0
Calm
1–3
Light
4
Moderate
5
Fresh
Notation — knots
Windspeed in the data set is given in knots (nautical miles per hour), where $1$ knot $\approx 1.15$ mph. A "Fresh" reading of, say, 18 knots on the Beaufort scale corresponds to roughly 21 mph — worth being able to convert roughly if a question mixes units.
Working with an extract
Exam questions typically give you a small extract — a handful of rows and a few columns — rather than the whole dataset. Three skills come up repeatedly:
Missing data: some entries are recorded as "n/a" or left blank, usually because of an equipment fault or a station not recording that variable. These rows are often excluded from a calculation rather than treated as zero.
Units and precision: temperature is typically in °C, rainfall in mm, wind speed in knots — read the given units carefully rather than assuming, since a plausible-looking number can be wildly wrong in the wrong unit.
Sampling from the set: since the full dataset is large, exam extracts are themselves samples — the sampling methods from the previous lesson (systematic, stratified by location or month, and so on) apply directly to how an extract might have been chosen.
Strategy — treat "n/a" as a sampling problem, not an arithmetic one
A missing value isn't something to guess or interpolate in a hypothesis-testing or averages question — the usual approach is to calculate using only the values that are present, while being ready to comment that this could itself introduce bias if the missing values aren't random (for instance, if a sensor consistently fails during extreme weather).
Worked example 1
Reading an extract
An extract from the large data set gives the daily total rainfall (mm) at a UK weather station for five days in one month: $0.0,\ 2.4,\ \text{n/a},\ 0.0,\ 5.1$. Find the mean daily rainfall for the days with recorded data, and state how many days had no rain at all.
MethodExclude the missing value, then compute directly
1
One value is missing ("n/a"), so only four days have recorded data: $0.0, 2.4, 0.0, 5.1$.
Exclude, don't guess.
2
Mean $= \dfrac{0.0+2.4+0.0+5.1}{4} = \dfrac{7.5}{4} = 1.875\ \text{mm}$.
Sum the recorded values, divide by how many there are.
3
Two of the four recorded days ($0.0$ and $0.0$) had no rainfall at all.
Count the zero entries.
▪ solved
Examiner's eye
Divide by the number of recorded values ($4$), not the total number of days ($5$) — treating a missing value as if it counted (or as if it were zero) is one of the most common errors with this dataset.
Worked example 2
Comparing two years
The large data set records daily mean temperature at the same station for 1987 and for 2015. Explain one reason why simply comparing the mean temperature of the two years might not, by itself, give strong evidence of a genuine climate difference.
This tests interpretation, not calculation — the skill is spotting a limitation in a simple comparison.
MethodLook for a source of variation the comparison doesn't control for
1
A single year's mean temperature is affected by that year's particular weather patterns, which vary naturally from year to year regardless of any long-term trend.
Natural year-to-year variability.
2
Comparing just two individual years can't distinguish a genuine underlying shift from ordinary short-term fluctuation — two years is a very small sample of "years".
Small sample size, in a different sense.
3
A more reliable comparison would use data from several years around each date, or a formal hypothesis test, rather than two single snapshots.
Reduces the influence of any one unusual year.
▪ explained
The key insight
"The data shows a difference" and "the data proves a trend" are not the same claim. A descriptive comparison of two means can suggest a pattern, but confirming it's a genuine effect (rather than natural variation) is exactly the job of the hypothesis-testing topics later in the course.
Worked example 3
Spotting an implausible value · Challenge
A row in the large data set records a daily mean temperature of $47.2$°C at a UK weather station in October. Explain why this value is almost certainly an error, and suggest what should be done with it before calculating summary statistics.
MethodCompare against realistic bounds, then decide how to handle it
1
$47.2$°C is far beyond any daily mean temperature ever recorded in the UK, in any month — the UK's highest daily means sit well below $30$°C, even in summer.
Compare against known real-world bounds.
2
In October specifically, typical UK daily mean temperatures are far lower still, making this value even less plausible.
Context (the month) narrows the plausible range further.
3
This is very likely a recording or transcription error (for instance, a misplaced decimal point, or a sensor fault) rather than a genuine measurement.
A specific, reasoned explanation.
4
It should be treated as an outlier or an error and excluded from summary calculations, rather than included as if it were a real reading.
Don't let one clearly wrong value distort a mean.
▪ discussed
Look deeper — why context matters more than the number alone
Judging plausibility always needs outside knowledge of the variable and its typical range — a value that would be perfectly reasonable for one variable, month, or location can be impossible for another. This is precisely why familiarity with the large data set's typical values, built up in advance, matters for this style of question.
Common mistake
Don't include "n/a" or clearly erroneous entries as zero, or attempt to calculate with them as if they were valid numbers — both distort every summary statistic that follows. Always state explicitly that missing or implausible values have been excluded, and from how many values a calculation is actually based on.
Check yourself
An extract from the large data set gives daily maximum gust (knots) at a coastal station for six days: $18,\ 21,\ \text{n/a},\ 19,\ \text{n/a},\ 25$. Find the mean of the recorded values, and state how many values were excluded. (Hint: count only the numerical entries, both for the sum and for the divisor.)
Method
1
Recorded values: $18, 21, 19, 25$ — four values, two excluded.
2
Mean $= \dfrac{18+21+19+25}{4} = \dfrac{83}{4} = 20.75$ knots.