Exploring One-Variable Data and Collecting Data

An observational unit is the person or object measured. A variable records a characteristic of that unit. Categorical values place units into groups; quantitative values describe numerical amounts for which arithmetic has a useful meaning.

An individual is the object described by data, such as a student, household, or manufactured part. A variable records a characteristic. Categorical variables classify individuals, while quantitative variables measure numerical amounts for which arithmetic is meaningful. An identification number is not automatically quantitative just because it contains digits.

A population is the full group of interest; a sample is the observed subset. A parameter describes the population, while a statistic describes a sample. State the population and measurement method before calculating: the average for volunteers who complete an online survey may not estimate the intended population well.

Bar charts display category counts or proportions with distinct categories. Histograms group quantitative values into intervals, while dotplots and stemplots can preserve individual values in smaller datasets. A boxplot summarizes center and quartiles but can conceal clusters or gaps that a fuller display reveals.

Describe a quantitative distribution in context using shape, center, spread, and unusual features. When comparing groups, make direct comparisons rather than writing two disconnected descriptions. Include units and specify the measure of spread, such as IQR or standard deviation. A long right tail indicates right skew even if most observations lie toward the left.

The mean uses every value and is sensitive to extreme observations. The median is the middle ordered value and is resistant. Standard deviation describes typical distance from the mean through squared deviations; IQR is Q₃ − Q₁ and summarizes the middle half. Neither measure is the full range.

The 1.5-IQR rule flags potential outliers below Q₁ − 1.5IQR or above Q₃ + 1.5IQR. A flag is a prompt for investigation, not automatic permission to delete data. Adding a constant shifts center but not spread. Multiplying values by a constant scales spread by its absolute value.

A normal distribution is symmetric and bell shaped, specified by a mean and standard deviation. Standardizing maps a value to its relative location in standard-deviation units. Probabilities correspond to areas under the density curve, not the curve’s height at one point.

Use a normal model only when it is appropriate or explicitly given. The empirical rule approximates 68%, 95%, and 99.7% within one, two, and three standard deviations. A percentile describes the proportion at or below a value under the stated convention; a z-score is not itself a probability.

An observational study records existing conditions without assigning treatments. An experiment imposes treatments. Random assignment helps balance lurking variables across groups, supporting a causal comparison when implementation is sound. Random sampling and random assignment serve different purposes; one does not replace the other.

Control, replication, and randomization strengthen experiments. Blocking groups similar experimental units before random assignment can reduce variability. A matched-pairs design compares naturally paired observations or two conditions on the same unit with appropriate order control. Blinding can reduce expectation effects, while a placebo can help separate treatment effects from the experience of receiving a treatment.

A well-randomized experiment may support causation for the studied units but not broad population generalization if the subjects are a narrow convenience sample. A representative observational sample may support an association in a population but not establish causation. Confounding occurs when effects of explanatory variables cannot be separated.

An ethical design requires informed participation and appropriate treatment of subjects. Missing data, dropouts, and noncompliance can undermine an otherwise good assignment procedure. Report limitations directly and avoid claiming that statistical sophistication repairs a fundamentally biased design.

Begin by identifying the observational units, the variable, and the population you want to understand. A histogram groups quantitative observations into intervals, so its shape depends partly on bin width. A boxplot emphasizes quartiles, median, spread, and potential outliers but hides clusters and gaps. Do not infer a detailed distribution shape from a boxplot alone.

Compare distributions using context, shape, center, variability, and unusual observations. “Group A is better” is not a statistical description. A stronger statement says which measured quantity has the higher typical value and which group varies more. For a strongly skewed distribution, the median and interquartile range often describe typical values more robustly than the mean and standard deviation.

Random selection supports generalizing from a sample to its population. Random assignment supports a causal comparison by distributing potential confounding variables across treatment groups. These are different design decisions. A randomized experiment using volunteers can support a causal conclusion for the participants without establishing that its effect generalizes to all students.

In a blocked experiment, first create groups whose members are similar on a variable expected to affect the response, then randomly assign treatments within each block. Blocking reduces unwanted variability; it does not replace random assignment. A matched-pairs design compares related observations, such as two measurements on each participant, so the analysis should preserve those pairs.

Sampling variability is the chance fluctuation among statistics from repeated random samples. Bias is a tendency for a method to systematically miss the population value in a particular direction. Increasing a biased sample from 100 to 10,000 does not automatically correct the selection process. A large voluntary-response poll can still overrepresent people with especially strong opinions.

Identify the mechanism and likely direction when possible. If a survey about commuting time reaches only people arriving before 8 a.m., it may omit a different commuting population. However, do not invent a direction without information connecting the omitted people to the response. State what is missing and why representativeness is uncertain.

Before calculating, identify one observational unit and what is measured on it. A numerical label such as a postal code is categorical because arithmetic on its digits does not measure a quantitative attribute. A count and a measurement can both be quantitative, but their possible values differ. This choice determines sensible displays and summaries.

Describe a quantitative distribution in context using shape, center, spread, and unusual features. A mean is pulled toward extreme observations; a median depends on order. Standard deviation measures typical distance from the mean in the variable’s units, while the IQR describes the middle half. A right-skewed distribution often has mean greater than median, but do not infer an exact mean from an unlabeled sketch.

Compare distributions explicitly: name which has the larger center or spread, rather than writing two disconnected descriptions. Check axes and sample sizes. A histogram’s bar height may represent frequency, relative frequency, or density; unequal bin widths require particular care. A boxplot conceals some shape information, so it cannot establish every feature visible in raw data.

Random sampling supports generalization to the population represented by the sampling frame, subject to response and coverage limitations. Random assignment balances potential confounders across treatment groups in expectation and supports causal comparison. These are different random processes. An experiment using volunteers can support a causal claim for the studied setting without making the volunteers a random sample of everyone.

In a stratified sample, sample from each relevant subgroup. In a cluster sample, randomly select groups and observe units within those groups according to the design. The names are not interchangeable: stratification aims to represent subgroups, while clustering often improves practicality. A convenience sample with many responses remains vulnerable to selection bias.

Undercoverage, nonresponse, and response bias affect different parts of data collection. An online survey can miss people without access; selected people may not respond; wording can influence answers. Increasing sample size reduces random sampling variability but does not automatically repair these systematic problems.

An experiment needs clear treatments, an outcome, replication, and an assignment procedure. A control group provides a comparison; a placebo can help isolate effects of receiving treatment. Blinding reduces behavior or assessment changes caused by knowing treatment assignment. Explain who is blinded and why it matters.

Blocking groups units using a relevant pre-treatment characteristic, then randomizes treatments within blocks. A matched-pairs design can use the same person under two conditions or carefully matched units. Pairing changes the analysis because differences within pairs, rather than all individual observations treated as independent, become central.

Confounding occurs when effects cannot be separated. If every treated plant is in one greenhouse and every control plant in another, greenhouse and treatment are confounded. Many leaves measured on one plant do not create many independently treated plants. Identify the actual experimental unit before counting replication.