Skip to main content

01 — Introduction to statistics and data types

Learning goal: Explain why statistics matters; distinguish descriptive from inferential work and population–sample–unit; recognize brief sampling techniques and selection bias; classify measurement scales; and weigh primary vs secondary data — starting from cases that raise questions before concepts.

Lecture slides (Typst, problem-based):
- Open presentation · Download PDF
Previous: 00 — Course description & syllabus (Indonesian) · Next: measures of central tendency (mean, median, mode, quartiles).
Indonesian twin: 01 — Pengantar statistika dan jenis data.


Use left/right arrows or Space to change slides. Full screen · Download PDF.

Each topic in this session follows case → questions → concept. The notes below match the slide order without repeating the same example lists in many places.

Block 1 — Why study statistics

Case. A campus portal runs the headline: “Survey: 78% of students support fully online classes.” In the article body, an online questionnaire was filled by 43 students from a single class, with no random draw. Leadership is asked to consider this for next semester’s policy.

Questions arise almost automatically: does 78% describe all students, or only those 43 people? What is missing from the headline so that decisions go wrong easily? If you sit in a policy meeting, what do you ask before agreeing or rejecting?

Concept. The problem is not a shortage of numbers, but numbers without sampling context. Headlines cut away uncertainty; readers hear false certainty. Statistics trains three habits: ask who was measured and how many; separate a sample summary from a population claim; write the limits of a conclusion honestly.

Those habits show up in many other decisions — each with its own trap for reading numbers.

Weather forecast. “70% chance of rain” is a probability statement under similar conditions, not a guarantee that today will be wet. One sunny day does not “refute” 70%; deciding to carry an umbrella is weighing risk.

Discount “off the regular price.” A percent-off claim only means something if the baseline is clear: regular according to which store, when, and compared with market price or a phantom price. Without a comparator, “50% off” is easy to misuse.

Lab results and health risk. “Above the normal range” or “95% effective” needs definitions of the variable, the reference population, and sample size. An adverse-event proportion without a denominator and criteria pushes people toward excess fear or excess calm.

YouTube thumbnail A/B test. After 20 views, thumbnail A leads by 3 clicks — that can be sampling luck. Changing the default thumbnail needs adequate n (impressions), a clear metric (clicks? CTR? watch time?), and awareness that early viewers may not represent the whole channel audience.

Server latency and error rate. Averages hide a slow tail; an error percent without request volume misleads. You need a distribution (p95/p99), a time window, and a threshold before acting.

The pattern is the same across all five: limited data → a decision must be made → risk of misreading. The portal news case at the start of the block is the door into that habit of asking.

Block 2 — Descriptive versus inferential

Case. A program chair receives a report: mean midterm score in Statistics class A = 72 (n = 40). A proposal follows: “The cohort is already fine; no remedial class needed.” Classes B and C have not been tested; the cohort totals about 180 students. Core question for this block: which claim is descriptive, and which is inferential?

Concept. Descriptive work summarizes the data you hold: in class A, the center of scores is about 72 — a valid claim for those forty people. Inferential work appears when 72 is used to talk about the whole cohort — a generalization to a population that needs sampling design and a probability model. Rejecting remediation mixes the two: a descriptive number is hung with an inferential decision without evidence from the other classes.

From here comes the pair parameter and statistic. A parameter describes a population and is usually unknown (\(\mu\), \(p\), \(\sigma^2\)). A statistic is computed from a sample (\(\bar{x}\), \(\hat{p}\), \(s^2\)). In the case above, \(\bar{x} = 72\) is a statistic for class A; the cohort mean is a parameter that is not yet known.

Block 3 — Population, sample, unit, and sampling techniques

Case. A fintech startup releases a press statement: “90% of Indonesian users are satisfied.” Method: an in-app survey of 200 active users in one city, one week after a cashback promo. What is counted are accounts that finished the survey. Core question: which is the population, which is the sample, and what is the observation unit?

Concept. A population is the set you want to generalize to; a sample is the part you observe; an observation unit is one entity that is measured (one data row). In that case, the target population of the claim is Indonesian users, but the source/frame population is only active users in one city who filled the survey. The sample is 200 respondents; the unit is an account that completed the survey — not all downloads, not all citizens.

How the sample is drawn determines who may be generalized to. Simple random sampling gives every unit in the frame equal chance of selection. Stratified sampling splits the frame into strata (for example program or city) then samples within strata so proportions are controlled. An in-app survey filled voluntarily after a promo is neither — that is convenience / self-selection. Cashback adds selection bias: people who feel rewarded are more likely to answer. Sample size n = 200 does not save a national claim if the frame and technique are narrow.

Honest reporting separates target from frame, states the sampling technique, and refuses or limits a “national” claim when sampling does not support it. In meeting 01, sampling is only an introduction. Formal sampling theory (sampling distributions, CLT, and what follows) is in meeting 09.

Block 4 — Measurement scales

Case. A teaching assistant builds a leaderboard: (1) the mean of program codes (SI = 1, TI = 2, IF = 3) as the “dominant program”; (2) the mean of a 1–5 Likert satisfaction score to rank units. The results go to a meeting as “objective evidence.” Core question for Block 4: what measurement scale does each variable have, and which operations are legitimate?

Survey excerpt (illustration):

Resp. Program Code Satisfaction (1–5)
R01 SI 1 4
R02 TI 2 5
R03 IF 3 3
R04 SI 1 4
R05 TI 2 2
R06 IF 3 5

Assistant’s figures: mean code = 2.00 (“dominant program ≈ TI?”); mean satisfaction = 3.83 (used to rank units).

Concept. Mathematical operations only make sense when they match the measurement scale. Stevens’s framework has four levels.

Scale Core idea Typical operations In the case
Nominal Labels without order Frequency, mode Program code
Ordinal Order, distances unsure Median, percentiles Likert 1–5
Interval Meaningful distance, arbitrary zero Differences, mean (with caveats) Temperature °C
Ratio Distance + absolute zero Differences and ratios Duration, click counts

Averaging program codes is meaningless because nominal data have no distance. Likert items have order, but equal spacing between scores is not guaranteed — means are often reported in practice, yet that is a choice to be aware of, not automatically valid as “perfect distance.”

Short worked examples per scale:

  • Nominal (elective enrollment, n = 12): SI = 6, TI = 4, IF = 2 → mode SI, SI share = 50%. A “mean code” of 1.67 has no program meaning.
  • Ordinal (satisfaction Likert, n = 9): sorted \(2,3,3,4,4,4,5,5,5\) → median = 4; mean = 3.89 looks tidy for ranking, but level gaps need not be equal.
  • Interval (server °C): 22, 24, 26, 28, 30 → mean = 26°C; an 8°C difference is meaningful; “30°C is twice 15°C” is not a thermodynamic heat claim.
  • Ratio (load time in seconds): 1.2, 1.5, 2.0, 2.4, 3.0 → mean = 2.02 s; 2.4 / 1.2 = 2× is valid because 0 seconds is absolute zero.

A useful complement in data classes: categorical (nominal + ordinal) versus numeric (countable discrete or continuous on a range). Discrete/continuous complements scale; neither replaces the other. Choose scale first (which operations are allowed), then discrete/continuous for numeric variables.

Quick check. Pass/fail status → nominal. Satisfaction 1–5 → ordinal. Server temperature °C → interval. Page-load duration (seconds) → ratio. Items in a shopping cart → discrete ratio.

Why is °C interval and Kelvin ratio? Celsius zero is a convention (water’s freezing point), not absence of temperature — differences are meaningful, but “20°C is twice 10°C” is not. Kelvin zero is a physical absolute; thermodynamic temperature ratios (for example 200 K vs 100 K) are meaningful. Quick test: if 0 means absence of the measured quantity, the scale is a ratio candidate; if 0 is only an agreed reference point, it is usually interval.

Block 5 — Primary and secondary data

Case. A student downloads a student-grades dataset from Kaggle (country and curriculum unclear), then writes: “Factor X significantly affects UHB student performance,” with no data collection on campus. What was downloaded is a public CSV; what was claimed is a local population.

Common misconceptions. A title containing “students” is treated as automatically suitable for UHB; secondary data is treated as “unscientific”; a large row count and a small p-value are treated as enough to generalize to any campus.

Concept. Primary data are collected for your purpose (your own survey, A/B test, observation, sensors you configure): high control, higher time cost. Secondary data were collected by someone else and reused by you (Kaggle, national statistics, campus logs, official reports): fast, but variable definitions and bias must be audited. Secondary analysis is allowed; what is not allowed is equating the dataset’s source population with UHB students without local data.

Before secondary analysis, four questions are almost always relevant: for what purpose was the data originally collected; who enters and who exits the frame; how are variables defined and measured; is reuse allowed ethically and legally. A proportional conclusion is limited to that dataset’s source population — not automatically “UHB students.”

Common traps — and the next meeting

The same trap patterns as in the cases above: headlines without sample size and method; one class treated as the whole cohort; a national claim from a narrow frame or voluntary sample; means on codes or raw Likert scores without caveats; a foreign dataset used as local institutional evidence. Large data volume does not save generalization if units enter the dataset in a biased way.

Meeting 02 moves to measures of central tendency — mean, median, mode, quartiles — after measurement type is clear, so the summary of “where the center of the data sits” is not chosen at random.