Six Unethical-Practice Risk Patterns
Section 3.4 teaches ethics through recurring failure modes. A common feature is the false appearance of trustworthiness: the numbers may look factual while timing, scale, definitions, selection, transformation or hidden bias changes the meaning.
1. Timing
Selecting or omitting observations based on timing can manufacture a false impression.
Recognition cue: the distortion depends on when observations are included or excluded.
Repair: use a representative/appropriate time window and disclose the timing context.
2. Misleading visualizations
Real numbers can still deceive through manipulated axes, omitted points, irrelevant comparisons or misuse of visual conventions.
Recognition cue: axis, scale, size, color, omitted points, graphic framing.
Exam trap: “the values are numerically correct” is not enough. Ethical presentation also requires honest context.
3. Unclear definitions / invalid comparisons
Two accurate numbers can form an invalid comparison when they describe different populations, denominators or definitions. Ethical presentation makes meaning explicit and compares like with like.
The chapter also warns that repeated/exhaustive statistical searching can create apparently significant results by chance.
Recognition cue: similar labels hide different populations or definitions.
4. Bias
Bias can enter collection, selection, sampling, analytical method, context/culture and presentation. Justice creates a positive duty to notice bias where data-driven decisions affect people.
Five bias patterns
| Pattern | What goes wrong |
|---|---|
| Data collection for a pre-defined result | Collection is shaped to reach the desired conclusion. |
| Biased use of data collected | Data may be sound, but analysts cherry-pick/manipulate it to confirm a preferred approach. |
| Hunch and search | Only confirming evidence is pursued; alternatives are ignored. |
| Biased sampling methodology | Sample selection systematically distorts who/what is represented. |
| Context and culture | Assumptions embedded in one cultural/contextual frame distort interpretation. |
Not every business selection criterion is automatically unethical. The risk depends on purpose, stakeholders, likely outcomes and potential harm. High-impact uses deserve stronger transparency, accountability and bias precautions.
5. Transforming and integrating data
Transformation changes data as it moves. Ethical risk rises when the organization cannot establish:
- origin and provenance;
- lineage;
- stable meaning and reliable Metadata;
- Data Quality;
- ownership and protection requirements;
- auditable remediation/change history.
A dataset loading successfully proves only technical movement — not preserved meaning or trustworthiness.
Why remediation history matters
Even well-intended remediation can alter data improperly or illegally. Formal, auditable change control preserves evidence of what changed and why; it does not by itself guarantee that the new value is correct.
6. Obfuscation / redaction
Masking, removing identifiers or aggregating data can reduce exposure, but downstream combination may re-identify people.
Techniques discussed include: - aggregation; - data marking/classification; - data masking; - sensitivity analysis and governed access.
Core caution: masking is a layer, not a magic shield. Ethical protection still requires purpose, context, Governance and consideration of downstream combination/use.
Fast classifier
| If the distortion is mainly… | Think… |
|---|---|
| observation window | Timing |
| graphic scale/framing | Misleading visualization |
| mismatched meaning/population | Definitions / invalid comparison |
| systematic slant toward a result/group | Bias |
| unknown provenance/meaning/change history | Transformation/integration |
| residual re-identification after masking | Obfuscation/redaction |
Source anchors: Chapter 2 §3.4.1–3.4.6, pp. 59–63.