Computer Science · Big Data · Topic 3 of 5

Processing the data correctly doesn't guarantee the conclusion is fair.

This topic — named directly in the spec as "interpretation and representation (bias and display)" — is about reading data critically: who is actually represented in a dataset, what a chart does and doesn't show, and how a technically correct number can still mislead.

AS91908 · Big Data·Builds on Topic 2's pipeline

Photo: Luke Chesser / Unsplash

Executive summary

Goal: name specific types of bias, apply a four-question checklist to any statistic, and critique a display choice rather than just describing it.

Section	What it covers
Where bias enters big data	Four named types, each with a real example
Reading a claim critically	A four-question checklist for any statistic you're given
Choosing and critiquing a display	What each display type is good for, and what it risks hiding
Worked examples	Two real past-paper questions about collection problems and human impact

Where bias enters big data

Bias can be introduced at almost any stage of the pipeline from Topic 2, before any analysis happens at all.

Type of bias	What it means	Example
Selection / sampling bias	The people in the dataset aren't representative of the wider population	A mobility report built only from users who opted in to location tracking over-represents smartphone owners who chose to enable that setting
Opt-in bias	Only some users have the feature that generates the data turned on	If a setting is off by default, the dataset skews toward more tech-engaged users who actively turned it on
Measurement bias	The way data is collected systematically distorts it	Comparing "visits" using relative percentage change hides the true underlying numbers
Coverage gaps	Some areas or groups simply don't generate enough data to be measured	A low-population region may not meet a platform's minimum data threshold at all

Case file: a real coverage gap

Google's 2020 New Zealand Community Mobility Report is built entirely from users who had the Location History setting turned on — which is off by default. This means the underlying sample is not the whole population, but specifically the subset of people who chose to enable a particular privacy setting. This is a textbook example of opt-in / selection bias: it doesn't make the report useless, but it does mean any conclusion ("New Zealanders reduced retail visits by 91%") technically describes that opted-in group, not the whole country.

The regional breakdown for Manawatu-Wanganui in the same report also shows a category — "Transit stations" — explicitly marked with a footnote: "Not enough data for this date: Google needs a significant volume of data to generate an aggregated and anonymous view of trends." This is a direct illustration of the volume characteristic from Topic 1 causing a real gap in what can be reported, and it's worth noting in an answer that a sensible dataset owner withholds a statistic rather than reporting an unreliable one.

(Source: 91908-res-2021.pdf, Resources E–G.)

Reading a claim critically

When you're given a statistic drawn from big data, check:

1. Population — who was actually measured? (everyone, or an opted-in/self-selected subset?)
2. Baseline — what is the number being compared against, and is that baseline itself reasonable?
3. Aggregation level — is this a national figure hiding large regional differences?
4. Causation vs correlation — does the data actually show why something happened, or only that it happened alongside something else?

Choosing and critiquing a display

The same underlying data can be shown in very different ways, and the choice changes what a reader notices first.

Display	Good for	Risk
Line graph over time	Showing a trend and its speed (velocity)	Can exaggerate small changes if the y-axis is zoomed in
Percentage change vs baseline	Quick comparison across categories	Hides the actual underlying numbers (a 90% drop from a small baseline may be a smaller real change than a 20% drop from a huge baseline)
Regional breakdown table/map	Showing that a national average hides local variation	Can appear more precise than the data supports, especially in low-volume regions
Raw counts	Showing true scale	Can be harder to compare across groups of very different sizes

Past-paper practice

What key problems or issues relating to the 3 Vs can you see with the collection of a dataset like this? How might these problems or issues be addressed? (Adapted from 91908-exm-2021.pdf, Question Three (b).)

Reveal how to answer well

• Name a specific V (e.g. veracity, or volume) and tie it to a concrete collection issue — don't just say "the data might be biased"
• For the Google mobility example: veracity/representativeness is limited because only opted-in Location History users are counted; volume is limited in low-population regions, producing "not enough data" gaps
• Suggest a realistic fix: widening the sample where possible, being transparent about who is and isn't represented, or explicitly withholding statistics below a minimum data threshold (as Google does) rather than reporting an unreliable figure

A companion question style asks you to weigh the human impact directly:

With reference to the resources, how do the technical capabilities and limitations of big data affect people's lives? Give examples to support your answer. (91908-exm-2021.pdf, Question Three (c).)

This wants you to connect a technical limitation (e.g. a minimum data threshold suppressing a statistic) to a real consequence for a specific group (e.g. policymakers in a low-population region having less evidence available to guide a local decision than a city does).

Exit question

A news report says "Regional District X saw a 60% drop in foot traffic." Using the four-question checklist above, list two questions you'd need answered before trusting that claim.

Key terms

• selection / sampling bias
• opt-in bias
• baseline
• aggregation
• data threshold

1. 1
2. 2
3. 3
4. 4
5. 5

← Back to Topic 2: Generation, Processing & Formats

Next → Topic 4: Tools, Technologies & AI