11.1 Twenty distinctions that protect marks
- Rows are not automatically people; the key follows the row definition.
- A repeated ID can be a copy, a correction or a legitimate repeated observation.
- Correct arithmetic can still use the wrong denominator.
- Size includes missing entries; count excludes them.
- A valid numeric range does not establish source accuracy.
- A missing value is not zero and not automatically not-applicable.
- MCAR/MAR/MNAR describe mechanisms; response counts alone do not identify them.
- Equal response rates do not guarantee an unbiased comparison.
- Recovering a verified observation differs from imputing an estimate.
- Filling cells does not increase the number of observed responses.
- An IQR flag is not a deletion instruction; a whisker is not a fence.
- Mean/median agreement and near-zero skewness do not prove normality.
- Sample and population-style SD use different divisors by design.
- A z-score needs a reference group; it is not automatically a percentile.
- A future min-max value may lie outside 0-1.
- A past event's result can be unavailable; a future schedule can already be known.
- Fit learns from training; transform applies the saved rule.
- Perfect training accuracy and tiny-test accuracy do not prove future reliability.
- Pooled rates are weighted averages; changed composition can reverse a comparison.
- Grouping and standardising do not automatically establish a causal effect.
11.2 Reporting templates
Descriptive summary: "Among [valid n] observed [units] from [population/period], [quantity] had [centre] and [spread] in [measurement units]. The distribution was [shape]. [Missing count or other specific limitation] limits [the intended conclusion]."
Cleaning decision: "I [action] [record/value] because [dictionary/source evidence]. This changes [rows/cells/values] by [count]. I retained [raw/provenance] and [uncertainty]. The alternative would [specific consequence]."
Transformation: "I fitted [parameters] using [training rows]. For [new value], [substitution] gives [output]. The same rule is reused for test/future records; [availability or interpretation limit]."
Comparison: "[Group A] had [successes/denominator and rate], compared with [Group B]. The difference is [percentage points or other unit]. [Composition/selection factor] may explain part of it, so this does not establish [causal claim]."
Simpson explanation: "Within [strata], [direction] holds, whereas the pooled direction is [opposite]. [Group A] places more weight on [lower/higher-rate stratum], with weights [values]. These answer different composition-specific questions; [remaining causal limit]."
Use these as structures, not scripts to paste unchanged. Marks depend on the evidence and quantities in the question you are given.
11.3 Precision notes and slide inconsistencies
The source material contains useful reminders that should not be overgeneralised. Use the precise statement in calculations and explanations.
- W2 PDF p. 15: the heading says an impossible value moves a mean by 34 years, while the figure labels show 66.0 and 95.9, a difference of 29.9. Do not memorise this inconsistent arithmetic; recompute from the observations supplied in an exam. The valid lesson is that extreme values can move the mean strongly.
- W2 PDF p. 17: the caption says eight scores, but the displayed mini-list contains nine. Count the actual observations before locating positions.
- W2 PDF p. 29: the n-1 correction concerns the expectation of sample variance under sampling assumptions, not every realised sample and not exact unbiasedness of SD.
- W2 PDF pp. 25,28: range cannot shrink as observations are added to a fixed list, but may stay constant. SD and IQR can change; the distinction is their different behaviour, not that they never move.
- W2 PDF pp. 45,49: interpreting z=1 as about the 84th percentile needs a normal model; the later slide correctly separates distance from rank.
- W3 PDF pp. 13,50-51: percentage-to-proportion conversion and 0-100 score conversion target different units. Read the dictionary.
- W5 PDF pp. 11,53: the reported 2.6 percentage-point gap uses unrounded fractions; subtracting the displayed one-decimal rates loses precision.
- Deck numbering: some W2 printed slide numbers repeat or differ from PDF page order; references in this guide use PDF page positions. The Lab 2 note mentioning slides 2-65 exceeds the 60-page uploaded deck; rely on the actual attached material and ask about any missing revision if needed.
11.4 Readiness check
Before calling a topic learned, solve a new example without notes, explain each denominator and unit, identify a tempting wrong answer, and say what evidence would change your conclusion. Then repeat after a delay. A correct calculation copied from a worked example is not the same as independent mastery.
You are ready for this pack's topics when you can: predict a join's row count; explain a misleading average; calculate SD and quartiles with conventions; distinguish flags from deletions; explain missingness within groups; reconstruct a sensitivity bound; transform a future vector from saved parameters; identify availability and test leakage; compute held-out error; and explain a Simpson reversal using weights.
For the actual paper, scan all questions, allocate time in proportion to marks, show substitutions, label denominators and units, and use the last portion of your time to check omissions and overclaims. If asked for a choice, commit to the supported choice and state the condition that makes it defensible.
11.5 How the rest of your earlier syllabus connects
Your earlier syllabus continues with visual communication (Week 6), probability and Bayes (7), inference and A/B design (8), regression (9), classification and fairness (10), clustering and PCA (11), and time series, text, dashboards and responsibility (12). The current guide supplies their foundations: meaningful measurements, data quality, honest evaluation and cautious interpretation. Detailed preparation for those later topics needs their actual lecture and lab materials.