Exploring One-Variable Data and Collecting Data

Follow the study from target to conclusion

A population is the full group named by the question; a sample is the observed subset. A parameter describes the population and is usually unknown. A statistic is computed from sample data, so it can change if a different sample is selected. The calculation alone does not determine whether the sample represents the population.

A useful investigative question has a purpose fixed before the data are analyzed, names an answerable variable, and can be matched to a feasible collection plan. Do not rewrite the question after seeing the result. In the synthetic map below, five labeled units come from a 200-unit population. Their mean measurement is 12.4, which is a statistic; calling it the population mean would erase sampling variability.

Classify by meaning, not appearance

A variable records one characteristic for each observational unit. Categorical values name groups, even when software stores them as numbers. Quantitative values are measured or counted amounts for which arithmetic has meaning. A parameter summarizes a population variable; a statistic summarizes the corresponding sample variable.

A discrete quantitative variable has countable possible values, such as a number of completed cycles. A continuous quantitative variable can take any value in an interval, such as elapsed time before rounding. Type and role are separate decisions: first ask which values are possible and what operations make sense, then ask how the variable functions in the investigation.

A proportion is incomplete without its denominator

A frequency table shows how many observations fall in each category. A relative-frequency table divides each count by a stated total, so its entries should sum to one apart from rounding. The total row catches omitted categories and inconsistent denominators.

The synthetic responses have 20 records: nine red, six blue, four green, and one missing. Red therefore represents 9/20, or 0.45. If a prompt excludes missing responses, the denominator becomes 19 and that choice must be announced. Multiple-response questions form a different structure because one unit can contribute to more than one category.

Match the graphic to the denominator

Bar charts leave gaps because categories are distinct labels. Histograms instead use touching numeric intervals. A categorical magnitude axis should begin at zero when bar length carries the comparison; truncation exaggerates differences. Pie charts require mutually exclusive parts of one whole and become difficult to read with many small slices.

For groups of unequal size, raw counts can point in a different direction than conditional percentages. Group X has 12 category-A cases out of 20, while Group Y has 18 out of 40. Y has more A cases, but X has the higher conditional share: 60% versus 45%. Segmented bars should therefore have equal total lengths and display within-group proportions.

Every display trades detail for structure

A dotplot shows every value and duplicate. A stemplot also preserves observations but needs a place-value key. A histogram groups values into contiguous, nonoverlapping bins; it shows shape efficiently but hides exact values inside a bin. A boxplot compresses further into quartiles and potential outliers.

Histogram appearance depends on bin origin and width. That flexibility can reveal or hide features, so boundaries must be visible and all bins must follow one convention. For the authored values, left-closed bins [0,5), [5,10), [10,15), and [15,20) contain 4, 6, 1, and 1 observations. No value may disappear at a boundary.

Shape tells where values extend, not where they point

A complete description names modality and symmetry or skew, locates clusters, gaps, or unusual values, and supports center and spread with numbers. One main peak is unimodal, two prominent peaks are bimodal, and roughly even frequencies with no prominent peak are approximately uniform. Right skew means the longer tail extends toward larger values; it does not mean most values are on the right.

The values 1, 2, 2, 3, 3, 4, 10 form one cluster from 1 through 4 and a separated high value at 10, producing right skew. Reflecting them about 5.5 gives 1, 7, 8, 8, 9, 9, 10; center reverses position and the long tail now extends left. The labels follow geometry, not whether the context sounds favorable.

Center and spread must be read as a pair

The mean balances signed deviations, while the median splits ordered positions. Percentiles locate ordered values; the range is maximum minus minimum, and $IQR=Q_3-Q_1$ measures the middle half. Standard deviation measures typical distance from the mean using squared deviations and the sample denominator $n-1$. Median and IQR resist a distant extreme better than mean, range, and standard deviation.

For 2, 4, 4, 6, 9, the mean and median are both 5. Squared deviations total 28, so the sample variance is 28/4=7 and the sample standard deviation is $\sqrt7\approx2.646$. The 1.5-IQR rule and values more than two standard deviations from the mean are two different ways to flag potential outliers; neither proves a value is erroneous.

Adding $a$ to every value adds $a$ to measures of center and position but leaves range, IQR, and standard deviation unchanged. Multiplying every value by $b$ multiplies center and position by $b$ and multiplies spread by $|b|$. Units change with the transformation, so transformed summaries need a fresh contextual interpretation.

A modified boxplot is a summary, not a miniature histogram

For 1, 2, 3, 4, 5, 6, 7, 20 under the declared median-of-halves convention, the median is 4.5, Q1 is 2.5, and Q3 is 6.5. Thus IQR=4 and the upper fence is 12.5. The value 20 is plotted separately; the upper whisker stops at 7, the largest nonoutlier.

A boxplot efficiently compares medians, IQRs, and flagged points but hides modality and most individual values. In a roughly symmetric distribution, mean and median tend to be close. A long right tail usually pulls the mean above the median; a long left tail usually pulls it below. That relationship is a clue about shape, not a rule that reconstructs the hidden data.

Compare features, not isolated numbers

Group A has values 2, 4, 4, 5, 5, 6, 6, 8; Group B has 4, 5, 7, 7, 8, 9, 10, 14. Their medians are 5 and 7.5, so B tends to be higher. A complete comparison also notes B’s greater range and high endpoint rather than claiming every B value exceeds every A value.

Aligned dotplots retain shape while parallel boxplots emphasize center and spread. Raw-frequency histograms can mislead when sample sizes differ; use relative frequency or density. For relative position, $z=(x-\mu)/\sigma$ measures signed standard-deviation distance. For example, 74 in a distribution with mean 70 and SD 8 has $z=0.5$; compare z-scores rather than raw values when scales differ.

Selection and assignment grant different licenses

An investigative question identifies the variable or parameter, points toward a comparison or estimation method, and names the population and kind of conclusion sought. A census records every member of that population. A survey uses a standard set of questions and remains observational unless researchers impose treatments.

An observational study records naturally occurring conditions; an experiment assigns treatments to experimental units and then measures a response. Random selection supports population generalization, while random assignment balances other causes across treatments and supports cause-and-effect. A study can have either random mechanism, both, or neither.

A confounder must be associated with both the explanatory and response variables and offer a competing explanation. Merely naming a third variable is not enough. Prospective observation is still observational, and a retrospective design is not automatically invalid; judge how data were generated, measured, and linked to the question.

Random does not mean haphazard

A simple random sample of size n makes every possible n-unit subset equally likely. Saying only that each individual has an equal chance is not enough; a design could meet that condition while favoring certain pairs. Selection must operate on a complete, relevant frame.

Stratified sampling takes a random sample from each internally similar stratum, ensuring representation across all strata. Cluster sampling randomly chooses natural mixed groups and observes every unit in selected clusters. Systematic sampling chooses a random start and every kth unit; periodic ordering aligned with k can create bias.

Locate the mechanism before naming the bias

Undercoverage occurs when the frame leaves some target units out or gives them too little chance. Voluntary response lets units self-select. Nonresponse occurs after selection when chosen units do not provide usable data. Response bias changes recorded values through wording, memory, interviewer effects, sensitivity, or measurement.

Bias is a directional tendency over repetitions, not every difference between statistic and parameter. In the synthetic population, subgroup H is 40% with target rate 0.80 and subgroup L is 60% with rate 0.30, so the population parameter is 0.50. A frame omitting H targets 0.30; repeated large samples from that frame become precisely wrong, about 0.20 low.

Randomization balances causes that were not measured

An experimental unit is the smallest unit independently assigned to a treatment. A sound experiment compares at least two conditions, randomly assigns treatments, replicates across independent units, and directly controls known extraneous variation. A control group provides a reference; it may receive no active treatment, a standard treatment, or a placebo.

A placebo effect is the average response to a placebo minus the average response to no treatment. In a single-masked study, either participants or interacting researchers do not know assignments; in a double-masked study, neither group knows. Masking reduces expectation and measurement effects but does not replace random assignment.

Blocking groups similar units before random assignment; matched pairs form blocks of size two or use each unit under both conditions in randomized order. Repeated readings on one unit are not independent replication. Ethical design still requires consent where relevant, reasonable risk, privacy, and a defensible reason for every condition.