AP Statistics Unit 3: Collecting Data — Free Review Games.

A census attempts to contact every individual in the population.

In an SRS, every possible sample of size n is equally likely.

Observational studies measure outcomes without intervening.

Experiments deliberately impose treatments to study cause and effect.

Self-selected samples overrepresent people with strong feelings about the topic.

Stratification divides into similar groups and samples from every stratum.

In cluster sampling, you randomly pick clusters and survey everyone in them.

Confounders are related to both variables, making it hard to determine causation.

The placebo effect is a psychological response to receiving a sham treatment.

Double-blind prevents both subject and evaluator bias.

Random assignment balances known and unknown confounders across groups.

Undercoverage means certain segments of the population have no chance of being sampled.

Only randomized experiments with proper controls can establish cause and effect.

Response bias results from misleading questions or subjects' desire to give socially acceptable answers.

Replication uses many subjects so results are not due to chance alone.

In systematic random sampling, a random starting point is chosen among the first k individuals, then every kth person thereafter is selected. This differs from stratified sampling (choice B), where individuals are drawn from within each subgroup, and cluster sampling (choice C), where entire groups are included wholesale. Convenience sampling (choice D) is not a probability method at all.

A convenience sample consists of individuals who are easy for the researcher to access, which often introduces bias because the readily available group may not represent the broader population. Choices B, C, and D describe probability-based sampling methods — simple random sampling, stratified sampling, and systematic sampling, respectively — all of which reduce bias through randomization.

Experimental units are the entities that receive treatments — they may be people, animals, plants, or objects. The outcomes measured on these units are called response variables (choice B), the conditions being compared are the treatments (choice C), and the people running the study are investigators, not experimental units.

A control group receives no treatment (or a placebo) and serves as a reference point. Without a control group, it is impossible to determine whether changes in the treatment group result from the treatment itself or from other factors such as the passage of time or the Hawthorne effect. Choice C is incorrect because no experimental design eliminates all variability — the goal is to reduce and account for it.

In a single-blind study, only the participants are kept unaware of their treatment assignment, while researchers retain that knowledge. This design reduces participant-side response bias but does not control for researcher bias (such as treating patients differently based on group assignment). When both participants and researchers are unaware, the study is double-blind (choice B), which provides stronger protection against bias.

A lurking variable is not part of the analysis but can influence the relationship between the explanatory and response variables, potentially creating a spurious or misleading association. This is distinct from the response variable (choice B) and the explanatory or treatment variable (choice C). When a lurking variable is unaccounted for in a study, it becomes a confounding variable that can distort conclusions.

A treatment is a specific experimental condition — for example, a particular drug dose, a teaching method, or a fertilizer type. The term applies broadly beyond medicine, making choice B too narrow. The measurement taken at the end is the response variable (choice C), and the process of dividing participants is random assignment (choice D), which is a method for allocating treatments, not the treatment itself.

In cluster sampling, the population is divided into groups (clusters), some clusters are randomly selected, and every individual within those clusters is included. Here, homeroom classes serve as the clusters. In stratified random sampling (choice B), individuals would be randomly drawn from within each group — entire groups are never included wholesale. The key distinction is that cluster sampling selects groups, while stratified sampling samples from groups.

Non-response bias occurs when those who respond differ meaningfully from those who do not. Highly engaged or strongly opinionated readers may be more likely to return the survey, making the 175 respondents unrepresentative of all 2,000 subscribers. This is distinct from voluntary response bias (choice C), which arises when individuals actively seek out a survey rather than being randomly selected first. Here, subscribers were selected first and then chose whether to respond.

Multistage sampling applies random selection at multiple successive levels (districts, then schools, then students). This differs from simple cluster sampling (choice C), where all members of selected clusters are automatically included without further sampling within them. In stratified sampling (choice B), the population is divided into subgroups and individuals are randomly selected from each subgroup — the subgroups themselves are not sampled.

In a matched pairs design, subjects with similar characteristics (such as age or baseline fitness) are paired, and then random assignment determines which member of each pair receives which treatment. This reduces variability attributable to the matching variable, improving sensitivity. Choice B lacks the essential element of random order assignment, which would introduce carryover or order effects. Choice C describes a completely randomized design with no pairing.

A completely randomized design uses random assignment as its only organizational structure — there is no blocking or pre-grouping. Choice B describes a randomized block design, which accounts for a known source of variability. Choice C describes a crossover or within-subjects design. A completely randomized design is simpler to implement but less efficient when a blocking variable is known to strongly influence the response.

Blocking controls for a known source of variability so it does not obscure the treatment effect. By comparing fertilizer to no-fertilizer within plots of similar soil type, the researcher removes soil-type variability from the experimental error, making differences between treatments easier to detect. Blocking does not equalize soil composition (choice B) — it accounts for existing differences by grouping similar plots together before randomization.

The question uses emotionally loaded language ('reckless' and 'critical') that frames one position as the reasonable answer, pushing respondents toward a negative view of the council's decision. This is response bias caused by question wording. Non-response bias (choice B) and undercoverage (choice C) relate to who participates in the survey, not to how the question itself is phrased.

Practical constraints make a census impractical for large, continuous production runs. In destructive testing scenarios — such as testing battery life by running them until they fail — a census would destroy the entire inventory. Choice B is false: a census, if feasible, would have less sampling error than a sample. Choice D is false because sampling does not eliminate measurement error; it introduces sampling variability that a census avoids.

In a factorial experiment, the total number of treatments equals the product of the number of levels for each factor. With 3 caffeine levels and 2 time-of-day levels, there are \(3 \times 2 = 6\) distinct treatment combinations (e.g., 0 mg morning, 0 mg afternoon, 100 mg morning, 100 mg afternoon, 200 mg morning, 200 mg afternoon). Adding the levels (choice B) is a common error that fails to account for all possible combinations of the two factors.

The most fundamental flaw is that the convenience sample is drawn from a location and time that systematically excludes most city residents. Weekday morning hardware shoppers are likely to be retired individuals, contractors, or homeowners — not a cross-section of the city. Increasing the sample size (choice B) would not fix this bias; a larger convenience sample would remain equally unrepresentative.

Observational studies identify associations but cannot establish causation because subjects self-select into groups, making it impossible to rule out confounding variables. Coffee drinkers may differ from non-drinkers in diet, exercise habits, alcohol consumption, or other lifestyle factors that could explain the difference. Choice B incorrectly claims causation from observational data. Choice D is wrong — observational studies can and do detect genuine associations; the limitation is in causal inference, not detection.

Though conceptually analogous, blocking and stratification serve different purposes. In a designed experiment, blocking groups experimental units by a known variable before randomizing treatments within each block, thereby reducing variability that would otherwise obscure the treatment effect. In survey sampling, stratification divides the population into subgroups to ensure each is adequately represented in the sample. Choice B is incorrect — while related in spirit, the procedures operate in different contexts and with different statistical goals.

This is a classic confounding scenario. Population illness burden drives both the number of hospitals (more hospitals are built where illness is prevalent) and the death rate (sicker populations have higher mortality). Without accounting for this confounding variable, the association between hospitals and deaths is misleading and does not support a causal claim. Choice B misidentifies the problem as response bias, which involves misreporting of data rather than the influence of an unmeasured third variable.

The most critical flaw is the absence of random assignment. People who actively seek out an energy drink likely differ in ways related to the outcome — they may be more motivated, health-conscious, or already accustomed to higher caffeine intake — creating systematic confounding. Without random assignment, observed productivity differences cannot be attributed to the drink itself. While blinding (choice B) is a genuine concern, it is secondary because the groups are not comparable to begin with.

Simpson's paradox occurs when a trend in aggregated data disappears or reverses when the data are divided into subgroups. Here, Procedure A was likely used more frequently on low-risk patients, who have higher baseline success rates regardless of procedure. This imbalance inflates Procedure A's aggregate rate. When controlling for risk level, Procedure B is superior in every subgroup. This illustrates why analyzing data only at the aggregate level can be deeply misleading when a confounding variable is unaccounted for.

Random assignment (not random sampling) is what justifies causal inference in an experiment — it ensures that treatment groups are comparable, so observed differences can be attributed to the treatment. However, scope of generalizability depends on how subjects were recruited. Since these are volunteers, not a random sample, results may not extend to all chronic back pain patients. Choice C is a key error: random assignment is sufficient for causation, even without random sampling. Choice B overclaims by extending results globally from a non-random sample.

When a variable is known to strongly predict the response, blocking on that variable before random assignment makes the design more sensitive by reducing within-block variability. Comparisons of treatment versus control are then made among students with similar parental education backgrounds, isolating the intervention's effect more cleanly. Choice B overstates the guarantee of random assignment — in small samples especially, random assignment may not perfectly balance all characteristics, and blocking is more reliable when a strong predictor is known. Choice C is an observational study and cannot establish causation.

This poll suffers from voluntary response bias — only users who feel strongly about political advertising bother to tap and vote, making the sample self-selected and unrepresentative of all users. Additionally, social media users as a population differ from the broader adult population in age, political engagement, and internet access, creating undercoverage if conclusions are extended beyond platform users. Choice B incorrectly frames the issue as low participation (55,000 responded, which is high) and conflates it with response bias, which involves misreporting rather than self-selection.

Random assignment distributes potential confounding variables—both known and unknown—roughly equally across treatment groups, making the groups comparable before the treatment is applied. This is distinct from random selection, which addresses representativeness of the sample (choice A). Random assignment does not eliminate the placebo effect; blinding addresses that concern (choice B). Equal chance of selection describes random sampling, not random assignment (choice D).

Stratified random sampling involves dividing the population into subgroups (strata) based on a shared characteristic, then randomly sampling from each stratum. The strata here are the grade levels. Cluster sampling (choice A) involves dividing the population into groups and selecting entire groups, not random individuals from each group. Systematic sampling (choice B) involves selecting every \(k\)th individual from a list. Simple random sampling (choice D) gives every individual an equal chance of selection without dividing into subgroups first.

Undercoverage bias occurs when some members of the population have little or no chance of being included in the sample. People without landlines—who may tend to be younger, lower-income, or more mobile—are systematically excluded from the sampling frame. Nonresponse bias (choice A) occurs when individuals chosen for the survey decline to participate. Response bias (choice B) occurs when participants give inaccurate answers. Voluntary response bias (choice D) occurs when individuals self-select to participate in a survey rather than being chosen by the researcher.

A placebo controls for the placebo effect—the tendency for participants to report improvement simply because they believe they are receiving treatment. By comparing the active treatment group to a placebo group, researchers can isolate the actual pharmacological effect of the medication rather than the effect of expectation. The placebo group is not primarily a safety measure (choice A). Sample size is not the reason for including a placebo group (choice C). A dose comparison would require multiple active treatment groups, not a placebo (choice D).

The defining feature of an experiment is that the researcher actively imposes a treatment or condition on subjects and observes the effect, rather than simply recording what occurs naturally. Choice A is incorrect because experiments do not require random sampling of participants—many experiments use volunteers. Choice C is incorrect because experiments can occur in natural field settings and observational studies can occur in controlled environments. Choice D is false; both study types can involve either type of variable.

Nonresponse bias occurs when the individuals who respond differ systematically from those who do not. The 120 households that returned surveys may be more engaged, more opinionated, or have more time—making them unrepresentative of all 500 selected households and, in turn, the broader population. Sample size alone is not the fundamental problem (choice A); the quality and representativeness of responses matter more. Choice B is incorrect because the original selection was random. Voluntary response bias (choice D) refers to self-selected samples, not to nonresponse among a randomly chosen group.

The control group receives no treatment (or the standard/baseline condition) and provides a comparison point for evaluating the effect of the treatment. Without a control group, researchers cannot determine whether any observed differences in focus are due to the music or to other factors. The experimental group (choice A) receives the treatment being tested—here, the music. A placebo group (choice B) receives an inactive substitute designed to mimic the treatment; silence is simply the absence of treatment. A stratified group (choice D) is not a standard experimental role.

A sampling frame is the list or set of individuals from which a sample is actually selected. Here, the voter registration database is the source from which the researcher draws the sample. If the database does not include all members of the target population (e.g., eligible voters who have not yet registered), undercoverage bias may result. The parameter of interest (choice A) is a numerical characteristic of the population, not a list. The sampling method (choice C) describes how individuals are chosen from the frame—for example, SRS or stratified sampling. The population (choice D) is the broader group the researcher wishes to study, which may be larger than the frame.

Selecting a random starting point and then every \(k\)th unit is systematic random sampling. Each unit does have an equal probability of selection (since the start is random), so choice C is incorrect. The critical concern unique to systematic sampling is periodicity: if the production process has a cycle whose period matches the sampling interval—for example, if the machine is recalibrated or a shift change occurs every 100 bulbs—then the sample could be biased toward particular production conditions. This does not apply to stratified or cluster sampling (choices A and B), which describe fundamentally different designs.

A confounding variable is associated with both the explanatory variable and the response variable. Socioeconomic status (SES) is associated with fast food restaurant density (lower-income areas tend to have more fast food outlets and fewer grocery stores) and with childhood obesity rates (lower-income families may have reduced access to healthy food and safe recreational space). This means SES could explain the observed association between fast food density and obesity without any direct causal link. Total restaurant count (choice A) is less directly tied to obesity outcomes. School enrollment (choice C) and neighborhood size (choice D) are not strongly linked to both variables simultaneously.

In proportional stratified sampling, each stratum contributes to the sample in the same proportion as it appears in the population. The total population is \(400 + 300 + 200 + 100 = 1000\) students. Juniors make up \(\frac{200}{1000} = 0.20\), or 20%, of the population. The sample should therefore include \(100 \times 0.20 = 20\) juniors. This ensures that each grade level is represented proportionally, unlike simple random sampling, which could by chance over- or under-represent a grade level.

Undercoverage occurs when certain groups in the population are systematically less likely to appear in the sample. Adults who work full-time are unlikely to be at a public library on weekday mornings, so they will be largely excluded from this convenience sample. This is a serious problem because full-time workers may have different internet usage patterns than those who are available during weekday mornings. Retired adults (choice A) actually have more weekday flexibility and may be overrepresented. Regular library visitors (choice C) and people living nearby (choice D) would be overrepresented, not underrepresented.

Blinding the evaluating physicians prevents experimenter bias—if a physician knows a patient received the active drug, they may (even unconsciously) interpret ambiguous symptoms as improvement or interact differently with that patient. Since depression assessment often involves subjective clinical judgment, this bias could distort the results in favor of the treatment. Choice A describes an administrative error-prevention measure, not a statistical rationale. Choice B addresses dosing consistency, which is handled by the trial protocol, not blinding. Choice D addresses ethics and privacy, not the statistical purpose of double-blinding.

Blocking removes a known source of variability (prior skill level) from the unexplained error in the experiment, making it easier to detect the true effect of the new instructional approach. If students of widely varying skill levels were randomly mixed without blocking, the natural differences in scores could mask or inflate the treatment effect. Blocking does not equalize students' abilities (choice A)—it simply accounts for the differences that exist. Blocking is unrelated to the placebo effect (choice C), which is addressed through blinding. Representativeness (choice D) is a sampling concern, not a blocking rationale.

A prospective study recruits participants in the present, measures exposures or behaviors now, and then follows them forward in time to observe future outcomes. The researchers are not assigning any treatment—they are observing who naturally develops cardiovascular disease—making this an observational study, not an experiment (eliminating choices C and D). A retrospective study (choice A) looks backward in time, asking people who already have a disease to recall past behaviors. The key distinction here is the direction of time: forward (prospective) vs. backward (retrospective).

In cluster sampling, the population is divided into groups (clusters)—here, city blocks—and entire clusters are randomly selected. All members of the chosen clusters are included in the sample. This is different from stratified sampling (choice A), where only a random subset of members is selected from each stratum, and every stratum is represented. Systematic sampling (choice B) involves selecting every \(k\)th individual from a list. Multistage sampling (choice D) would involve an additional stage—for example, randomly selecting blocks and then randomly selecting a subset of households within each selected block, rather than including all households.

Response bias occurs when participants give inaccurate answers. In a face-to-face interview about a sensitive behavior like alcohol consumption, participants often underreport to avoid judgment—a social desirability effect. This effect is stronger when a human interviewer is present than in an anonymous questionnaire. Choice A describes undercoverage in who agrees to participate, not how they answer. Choice B describes nonresponse, which affects who participates, not the accuracy of responses given. Choice D describes voluntary response bias, which applies to self-selected samples, not to the format of a scheduled interview.

This is a classic example of Simpson's paradox: a trend that appears in aggregated data can reverse when the data are broken into subgroups. The explanation is that female applicants disproportionately apply to more competitive departments (with low acceptance rates for everyone). Even though women perform as well as or better than men within each department, their overall rate is lower because they are concentrated in harder-to-enter programs. The paradox arises from the confounding effect of department selectivity on the aggregate comparison. This is not a data collection or measurement problem (choices A, B, D)—it is a structural property of combining groups with different sizes and base rates.

Because physicians know who received the active drug and are conducting subjective clinical interviews, their expectations may unconsciously influence how they probe for symptoms, interpret ambiguous responses, or rate improvement. This experimenter bias directly threatens internal validity—differences between groups may reflect the physician's beliefs rather than the drug's effect. This is why double-blinding is the gold standard. Choice A is incorrect; 200 patients can be adequate depending on the expected effect size. Choice D is technically true but random assignment with 200 patients generally produces comparable groups; this is a minor concern relative to the active experimenter bias introduced by single-blinding. Choice C is an ethical concern, not a threat to internal validity.

A lurking variable is an unmeasured variable that drives a spurious association between two other variables. Age causes both foot growth and reading development. As children get older, both shoe size and reading ability increase—the two variables are correlated because they share a common cause (age), not because one causes the other. This is a classic example of a confounding lurking variable. Choice A is incorrect; correlation is statistically valid here, but the causal interpretation is wrong. Choice C is incorrect; observational studies routinely reveal correlations—experiments test causal hypotheses. Choice D addresses external validity (generalizability), not the flawed causal inference from foot size to reading ability.

An interaction effect occurs when the effect of one factor on the response variable depends on the level of a second factor. Here, the benefit of practice problems over flashcards is present at 3 hours but not at 1 hour—meaning the effect of study method depends on study time. This is the definition of an interaction. A confounding effect (choice A) would mean study time is associated with study method in a way that distorts the comparison, not that the effect of method varies across time levels. Blocking (choice C) is a design technique for controlling variation, not a description of a pattern in results. A lurking variable (choice D) was not introduced here because the study used random assignment to all four treatment combinations.

The fundamental flaw is voluntary response bias. People who choose to click on a poll and submit a response are self-selected—they tend to have stronger opinions (positive or negative) than the general public. Residents who are indifferent or only mildly opinionated are unlikely to participate. This makes the results systematically unrepresentative regardless of the large sample size, so choice A is incorrect. Choice C is plausible as written but sidesteps the main problem. Choice D suggests a design improvement that would not address voluntary response bias, since the self-selection problem persists even if respondents were stratified after the fact.

Blocking by baseline insomnia severity is the most powerful and appropriate design when a variable is known to affect the response. By randomly assigning within each severity block, the researcher ensures both treatment groups contain comparable proportions of mild, moderate, and severe cases, while also being able to estimate and remove the variability due to severity from the error term. This increases the power to detect a treatment effect. Choice A is statistically valid but less efficient with only 80 participants, where random chance imbalance is more likely. Choice B restricts generalizability and wastes data. Choice D describes retrospective matching after data collection, which does not achieve the same control as prospective blocking and can introduce selection bias.

This is multistage sampling: random selection occurs at three successive levels (district, school, student). The key statistical disadvantage is clustering: students within the same school share teachers, curricula, and community resources, making their scores more homogeneous than scores from students randomly selected across the entire country. This within-cluster similarity reduces the effective sample size and yields less precise estimates than a true SRS of \(80 \times 4 \times 25 = 8000\) students. However, multistage sampling is vastly more logistically feasible. Choice A describes single-stage cluster sampling (selecting whole districts with no further subsampling). Choice B is incorrect; stratification is not occurring here. Choice D describes systematic sampling, which is not the design used.

Even a large, well-controlled observational study cannot prove causation. First, unmeasured confounders—variables not recorded in the study, such as personality traits, genetic predispositions, or chronic illness—may explain the association. Second, reverse causation is a serious concern: people experiencing depression may exercise less because of their condition, meaning the causal arrow may point from depression to inactivity rather than the reverse. Only a randomized experiment, where participants are assigned to exercise or not, can provide strong evidence for causation. Choice A is incorrect; large samples improve reliability, not distort significance. Choice C suggests a design improvement but does not identify why the causal claim is invalid from an observational study. Choice D raises a valid distinction between statistical and clinical significance, but does not address the word 'proves' or the causal language in the headline as directly as choice B.

An SRS requires that every possible sample of size \(n\) has an equal chance of being selected — a stronger condition than simply giving each individual an equal probability of inclusion. For example, in systematic sampling each individual may have equal probability of being chosen, but not every sample of size \(n\) is equally likely (since only samples consistent with the fixed interval can occur). Choice A describes a weaker requirement satisfied by several non-SRS methods. Choice B describes stratified sampling. Choice C describes systematic sampling.

A voluntary response sample consists of individuals who choose to participate on their own, typically those who feel most strongly about the issue. This produces biased results because respondents are self-selected — listeners who are especially opposed to or enthusiastic about the curfew are far more likely to call in. Choice A involves selecting every \(k\)-th individual from a list. Choice B divides the population into subgroups and samples from each. Choice D involves randomly selecting entire groups and surveying all their members.

In an experiment, the researcher actively imposes treatments on subjects by assigning them to groups, which is what permits causal conclusions to be drawn. In an observational study, the researcher merely observes without controlling assignments. Choice A is false — observational studies can be larger than experiments. Choice B is false — time frame varies independently of study type. Choice C describes how a sample is obtained, which is desirable in both study types and does not distinguish an experiment from an observational study.

The placebo controls for the placebo effect — the tendency for people to feel or report improvement simply because they believe they are receiving treatment. Without a placebo, any improvement in the treatment group could be partly due to this psychological effect rather than the medication itself. A placebo does not reduce costs (Choice B), affect the experimental group's dosage (Choice C), or substitute for randomization (Choice D), which serves the separate purpose of balancing known and unknown confounding variables across groups.

The control group receives no treatment (or a standard treatment) and provides a baseline against which the fertilizer's effect can be measured. Choice B (placebo group) typically refers to a group receiving an inert treatment designed to mimic the real one — most relevant in human studies where psychological expectation matters. Choice C (experimental units) refers to all individual subjects under study, not just one group. Choice D (blocking group) is not a standard term for this role in experimental design.

Systematic random sampling involves selecting every \(k\)-th unit from a list or sequence, often after a randomly chosen starting point. Here, every \(15\)th bottle is inspected. Choice B (SRS) would require randomly selecting individual bottles from all bottles produced. Choice C (stratified) would divide bottles into distinct groups and sample from each group separately. Choice D (cluster) would involve randomly selecting entire production runs or batches and inspecting all bottles within the selected batches.

Replication means applying each treatment to multiple experimental units so that natural variability among units can be estimated and random chance is less likely to account for any observed differences. Without replication — if only one unit receives each treatment — it is impossible to distinguish a true treatment effect from random variability. Choice B describes scientific reproducibility across labs, a separate concept. Choice C describes the matched pairs design, not replication. Choice D describes the use of a control group, which is a related but distinct design principle.

In cluster sampling, the population is divided into groups (clusters), clusters are randomly selected, and every member of each selected cluster is surveyed. Here, neighborhoods are the clusters, 3 are randomly chosen, and all residents in those neighborhoods participate. Choice A (multi-stage) would involve randomly selecting a subset of residents within each chosen neighborhood rather than surveying everyone. Choice B (stratified) would require randomly selecting individuals from every neighborhood. Choice D uses a fixed selection interval from an ordered list.

A confounding variable is associated with both the explanatory variable (book ownership) and the response variable (reading scores), making it difficult to isolate book ownership as the cause. Household income affects both: wealthier families can afford more books and also provide tutoring, better schools, and enrichment activities that boost reading scores. Choice A (TV watching) may correlate with reading time but is not as directly tied to book ownership itself. Choices B and C are not strongly or systematically linked to both variables simultaneously.

This is a convenience sample — whoever was easiest to reach was selected. People who go to a fitness center exercise far more than the typical adult, so the sample differs systematically from the general population, producing an overestimate. Choice A is incorrect — random order of exit from the fitness center does not correct for the fact that all respondents are fitness center patrons. Choice C is implausible and directionally wrong. Choice D is wrong because sample size alone cannot eliminate systematic bias; a biased sample of 60 is still biased.

A randomized block design groups experimental units into blocks based on a variable expected to affect the response (here, GPA level), then randomly assigns treatments within each block. This controls for the effect of GPA and reduces within-group variability, making it easier to detect differences among study strategies. Choice A (completely randomized) would assign all students to strategies without accounting for GPA differences. Choice B (matched pairs) pairs individual subjects rather than grouping them into blocks. Choice C is a sampling method, not an experimental design.

The question uses loaded, negatively framed language ('wasteful,' 'at the expense of American taxpayers') that steers respondents toward opposition — a form of response bias called a leading question or wording effect. This produces a systematic overestimate of opposition to foreign aid. Choice B (social desirability) describes people giving the answer they think is socially acceptable, which would likely favor supporting aid rather than opposing it. Choice C is incorrect — question wording substantially influences responses. Choice D is incorrect — the effect shifts the distribution in one direction rather than increasing spread.

With only a 21% response rate, nonresponse bias is the central concern. Residents who feel strongly enough to return the survey — perhaps those passionate about environmental issues or those strongly opposed — may not represent all residents. The 82% approval figure applies only to those who responded, and generalizing it to all 1,000 sampled residents (or to the entire city) is unjustified. Choice A is incorrect — 1,000 is a reasonable sample size. Choice B is a false generalization. Choice C is a trivial formatting issue with no bearing on statistical validity.

A matched pairs design is a special form of blocking in which each subject serves as their own control by receiving both treatments. Because each person's baseline reaction time is controlled for, individual differences are removed as a source of variability, increasing the power of the study. The random ordering of conditions prevents order effects (e.g., fatigue or practice). Choice A would randomly assign different volunteers to the caffeinated or decaffeinated condition, with each person receiving only one treatment. Choice B uses blocks of multiple subjects, not individual subjects as their own controls. Choice C mischaracterizes the study.

A lurking variable is one not included in the analysis that explains an observed association. Cities with older or sicker populations naturally require more hospitals and also experience higher death rates — the lurking variable (population health and age structure) drives both. This is a classic example of confusing correlation with causation. Choice A correctly notes that experiments support causal claims, but it does not identify the specific flaw in the journalist's argument. Choice C is wrong — \(r = 0.91\) is a very strong correlation. Choice D is a separate statistical consideration and not the core error.

With only a 22% response rate, nonresponse bias is the central concern. Families who returned the survey likely differ from those who did not — they may be more engaged with the school, have stronger opinions about technology programs, or have had notably positive or negative experiences. Reporting 88% satisfaction as representative of all sampled families (or the entire district) is unjustified. Choice A is incorrect — 800 is a reasonable sample size. Choice C is a false generalization. Choice D may not be feasible and would not necessarily prevent nonresponse bias even in a census if some families decline to participate.

Multi-stage sampling involves multiple rounds of random selection. Here, schools are randomly selected first (stage 1), then students within each school are randomly selected (stage 2). This differs from cluster sampling (Choice A), in which every member of each selected cluster is surveyed rather than a random subset. Choice B (stratified) would divide students into meaningful groups (strata) and sample from all of them. Choice D uses a fixed interval from an ordered list rather than separate random selections at each stage.

Stratified random sampling divides the population into subgroups (strata) and samples from each, ensuring that each department is represented even if one department is much smaller than others. It also tends to reduce overall sampling variability by creating more homogeneous groups within each stratum. A simple random sample might by chance severely underrepresent a small department. Choice A is false — stratification adds logistical complexity compared to SRS. Choice C describes proportional stratified sampling, a specific type not implied by this design. Choice D is incorrect — stratified sampling takes a random sample from each stratum, not a census.

Blocking on a known source of variability keeps that source constant within each block, reducing within-group noise and improving the experiment's power to detect differences due to sleep duration. If age groups were mixed without blocking, age-related variability in reaction time would inflate the error term, making it harder to detect sleep effects. Choice A is incorrect — blocking accounts for age's influence but does not remove it from reality. Choice B is wrong — participants were not randomly assigned to age groups, so no causal claim about age is warranted. Choice C is wrong — random assignment within blocks is essential to the design and cannot be eliminated.

Random assignment justifies causal conclusions — the app (not some other variable) caused the observed reduction in anxiety among participants. However, because the sample consists of volunteers from a single university campus, it likely does not represent all adults or even all college students, which limits external validity. Choice A overgeneralizes to all adults without justification. Choice C incorrectly conflates random assignment (which supports causation) with random sampling (which supports generalizability) — these are separate design features. Choice D confuses statistical significance with clinical efficacy or universal applicability.

External validity refers to how well study results generalize to other settings, populations, and conditions. A tightly controlled laboratory environment may produce results that do not replicate in actual schools where distractions, varying materials, instructor differences, and student motivation differ considerably. The colleague is correctly questioning whether lab-based findings will hold in the real world. Choice A is incorrect — the question states random assignment was used. Choice B is incorrect — comparing two active instructional methods without a no-treatment group is a valid design for this research question. Choice D confuses statistical power with validity.

An interaction effect occurs when the effect of one factor on the response variable depends on the level of another factor. Here, the benefit of the low-carbohydrate diet appears only when exercise intensity is high — the diet's effect changes depending on exercise level. This is the definition of an interaction. Choice A (main effect of exercise only) would imply exercise affects weight loss consistently across diet types without diet mattering differently at each exercise level. Choice B (main effect of diet only) would imply diet affects weight loss consistently across exercise levels. Choice D is contradicted by the described pattern, which shows a clear conditional relationship.

Blinding physicians prevents observer bias (also called experimenter bias): a physician who knows a patient is receiving the new drug and expects it to be superior may unconsciously rate that patient's pain as lower, artificially inflating the apparent treatment benefit. This is why outcome assessors must also be blinded, not just patients. Choice A is incorrect — group size equality comes from randomization, not blinding. Choice B is incorrect — the placebo effect refers to changes in patients' own responses due to belief; blinding physicians addresses how evaluators measure outcomes. Choice D is incorrect — blinding and randomization serve distinct, complementary purposes and neither replaces the other.

Without a control group receiving the existing curriculum, the researcher cannot determine whether the \(9\)-point improvement resulted from the new curriculum or from natural student growth and maturation over the school year — a threat to internal validity called the history or maturation effect. Students typically improve in reading each year regardless of instructional approach, so the gain may have occurred without any curriculum change. A concurrent control group would allow the researcher to separate curriculum effects from natural growth. Choice B is an empirical claim requiring context — it is not the identified flaw. Choice C describes a measurement preference, not a validity threat. Choice D misapplies the replication principle.

Main effects are assessed by comparing marginal means. Since \(A_1\) and \(A_2\) both have marginal means of \(30\), there is no main effect of Factor A. Since \(B_1\) and \(B_2\) both have marginal means of \(30\), there is no main effect of Factor B. However, the cell means reveal a striking interaction: when Factor A is at \(A_1\), Factor B at \(B_2\) yields a higher response (\(40\) vs. \(20\)); when Factor A is at \(A_2\), Factor B at \(B_1\) yields a higher response (\(40\) vs. \(20\)). The direction of Factor B's effect completely reverses depending on the level of Factor A — a classic crossover interaction. Choices A, B, and D all incorrectly claim main effects where marginal means are identical.

An SRS requires that every possible sample of size \(n\) — not just every individual — has an equal chance of selection. Choice B describes a weaker condition that could be satisfied by other methods (e.g., systematic sampling) that do not give all samples equal probability. Choices C and D describe stratified and systematic sampling, respectively.

Voluntary response sampling occurs when individuals choose whether to participate, typically attracting people with strong opinions. This makes the sample unrepresentative of the broader population. The other choices all involve some form of active random selection by the researcher rather than self-selection by respondents.

A placebo is a treatment with no active ingredient, designed to look identical to the real treatment. It allows researchers to separate the actual effect of the medication from the psychological effect of believing one is being treated (the placebo effect). A control variable is held constant, and a block is a group of similar experimental units — neither is an inert treatment.

Systematic random sampling involves selecting every \(k\)th individual from an ordered list after a random start. Here \(k = 15\). Cluster sampling randomly selects entire groups. Stratified sampling divides the population into subgroups and samples from each. Convenience sampling requires no structured selection process.

Stratified random sampling divides the population into non-overlapping subgroups (strata) based on a relevant characteristic, then draws a random sample from each stratum. This differs from cluster sampling, where entire groups are randomly selected rather than individuals from within each group.

A control group receives either no treatment or a standard treatment, allowing researchers to attribute any difference in outcomes to the experimental treatment rather than to other factors. Without a baseline for comparison, it is impossible to determine whether observed changes are due to the treatment. The control group does not eliminate the placebo effect — that is addressed through blinding.

Convenience sampling selects individuals who are easiest to reach rather than using a random mechanism. Shoppers at a mall entrance are not representative of the general population — they are available and willing, which introduces selection bias. The other methods all involve formal random selection procedures.

A confounding variable is one that is related to both the explanatory and response variables and therefore provides an alternative explanation for an observed association. 'Lurking variable' is sometimes used synonymously, but in AP Statistics, a confounding variable specifically refers to one whose effect cannot be separated from the explanatory variable's effect. A blocking variable is intentionally accounted for in the design.

Undercoverage occurs when parts of the population are systematically excluded from the sampling frame. Adults who rely solely on mobile phones — a group that tends to skew younger — are completely absent from a landline-only frame, making the sample unrepresentative. While nonresponse is also a real concern with telephone surveys (choice C), the most fundamental structural flaw is undercoverage of an entire segment of the population.

The phrase 'reckless plan' is loaded language that steers respondents toward opposition, making this a leading question. This creates response bias — a systematic tendency for responses to differ from the true opinion due to the question's wording. Undercoverage and nonresponse are related to who participates, not to how the question itself influences answers.

Cluster sampling involves randomly selecting entire groups (clusters) and then including all or most members of those groups. Here the clusters are school districts. This differs from stratified sampling, where the population is divided into subgroups and a random sample is drawn from within each subgroup — in cluster sampling, only selected clusters are sampled, not all of them.

This is an observational study — students were not randomly assigned to sports participation. Students who choose to participate in sports may differ systematically from those who do not in ways that also affect GPA (e.g., motivation, family support, socioeconomic status). These confounding variables prevent a causal conclusion. Only a randomized experiment can control for such differences.

A double-blind design conceals treatment assignment from both participants and those evaluating outcomes. This prevents the placebo effect from influencing patient-reported pain and prevents evaluator bias from affecting measurements. A single-blind design blinds only one party (usually the participant). The term 'placebo-controlled' refers to the presence of a placebo, not to blinding.

Stratified sampling improves precision by ensuring representation from each subgroup (stratum). If the strata have different distributions of the response variable, sampling from each stratum separately reduces the variability of estimates compared to an SRS of the same total size. Choice B is a property of SRS, not stratified sampling. Choice D describes systematic sampling.

A matched pairs design groups experimental units into pairs that are similar on variables that could affect the response (here, soil quality and sun exposure), then randomly assigns one unit in each pair to each treatment. This reduces variability and makes it easier to detect a true treatment effect. A completely randomized design assigns all units to treatments independently without pairing.

A prospective study recruits participants before outcomes occur and follows them forward in time to observe who develops the condition of interest. A retrospective study looks backward at existing records. A cross-sectional study measures exposure and outcome at the same point in time. Since participants are randomly selected but not randomly assigned to diets, this is observational, not experimental.

Causation cannot be inferred from observational data because lurking variables may be responsible for the observed association. People who carry reusable bottles may generally be more health-conscious, leading them both to use bottles and to drink more water — the bottle itself may not be the cause. Only a randomized experiment (randomly assigning some people to use reusable bottles) could support a causal claim.

Blocking controls for a known source of variability by grouping experimental units that are similar (same elevation) and then randomly assigning treatments within each block. This allows the researcher to separately estimate the effects of elevation and irrigation method, reducing confounding. Choice A (completely randomized design) does not control for elevation. Choice B reduces external validity by limiting scope. Choice D does not randomly assign different treatments within pairs.

Recall bias is a major threat in retrospective studies: individuals who experienced a serious health event (heart attack) may be more motivated to search for reasons and therefore recall or report their past diet differently than healthy controls. This systematically distorts the association between diet and heart attack risk. While randomized experiments are stronger for causal inference, observational studies are not invalid — they simply require careful interpretation.

The critic is questioning whether the results can be generalized to the broader student population — this is a question of external validity. Because only motivated volunteers participated, the effect seen may not apply to typical students who would not seek out a study-skills workshop. Internal validity (whether the workshop caused the improvement within this sample) is actually well-supported by random assignment, which controls for differences between groups.

Systematic sampling can introduce periodicity bias when the sampling interval (\(k = 20\)) happens to coincide with a natural cycle in the ordered list. If every 20th record falls disproportionately in peak or off-peak purchase periods, the sample will not accurately represent all customer types. This is a known limitation of systematic sampling and does not affect SRS, which has no fixed interval.

When evaluators know which treatment a participant received, their assessments can be unconsciously influenced — a form of observer or measurement bias. This is why double-blinding (concealing treatment from both participants and evaluators) is the gold standard in clinical trials. The study is single-blind at best: patients may not know their treatment, but the clinicians do. Random assignment (choice B) addresses group equivalence, not evaluator bias.

This is an observational study susceptible to confounding. Wealth is a lurking variable: wealthier cities invest in walkable infrastructure and also have residents with higher incomes, better diets, and greater healthcare access — all of which independently reduce obesity. Without randomly assigning walkability to cities (which is impossible), it is impossible to disentangle walkability's effect from the effects of these confounders.

This is an example of regression to the mean as a confound: if last year's scores were unusually low (perhaps due to disruptions, illness, or other temporary factors), scores would likely rise the following year regardless of any intervention. Without a control group that did not receive the new curriculum, it is impossible to separate the effect of the curriculum from natural score recovery. This is a classic threat to the internal validity of before-after studies.

Because communities volunteered rather than being randomly assigned, there is self-selection bias: communities willing to adopt a public health campaign may already have populations more motivated or prepared to change behavior. This pre-existing difference — not the campaign itself — could explain the greater smoking reduction. Randomization at the community level (choice D) would help, but only if assignment were truly random rather than based on willingness to participate.

A simple random sample requires that every possible sample of size n has an equal probability of being selected — not merely that each individual has an equal chance. Choice B describes a weaker condition satisfied by other methods such as systematic sampling, where each individual is equally likely to be chosen but not every sample of size n is equally likely. Choice C describes systematic random sampling, and Choice D describes stratified random sampling.

Voluntary response bias occurs when individuals self-select into a survey, rather than being randomly chosen. People who feel strongly — particularly those who oppose the ban — are far more likely to take action and text in. This produces a sample that overrepresents those with strong opinions. Nonresponse bias (Choice C) differs: it occurs when randomly selected individuals fail to respond. Here, no one was randomly selected at all — anyone may participate.

A control group receives no active treatment (or receives a placebo) and serves as a baseline comparison. Without it, there is no way to know whether changes in the treatment group result from the treatment itself or from other factors such as the passage of time, the Hawthorne effect, or natural recovery. The control group does not eliminate the placebo effect — it accounts for it by allowing researchers to measure how much improvement occurs even without the active treatment.

In a single-blind experiment, the participants do not know which treatment they are receiving. This prevents them from changing their reported outcomes or behavior based on expectations. When both the participants and the researchers interacting with them are unaware of treatment assignment, the design is called double-blind. Double-blinding provides stronger protection against bias because it also prevents researchers from unconsciously treating participants differently or interpreting outcomes in a biased way.

A convenience sample consists of individuals who are easily accessible to the researcher rather than randomly selected from the population. People near the journalist's office building are likely not representative of the entire city — they may share similar occupations, commuting patterns, and neighborhoods. This makes convenience samples prone to bias. A cluster sample (Choice D) would involve randomly selecting groups and surveying all their members, which is not what occurred here.

A confounding variable is one that is associated with both the explanatory variable (coffee consumption) and the response variable (heart disease), making it difficult to determine whether the explanatory variable itself causes the response. For example, smokers tend to drink more coffee and also have higher heart disease risk — so smoking could create an apparent relationship between coffee and heart disease even if no true causal link exists. Choice A describes an explanatory variable and Choice B describes the response variable.

A census collects data from every individual in the entire population of interest — here, every farm in the United States. While a census provides complete and accurate information, it is typically expensive, time-consuming, and sometimes logistically impossible for large populations, which is why most large-scale studies use probability sampling methods instead. None of the other choices apply because this is not a sampling procedure at all.

In cluster sampling, the population is divided into groups called clusters. A random sample of clusters is selected, and all members within the selected clusters are measured. Here, schools are the clusters. This differs critically from stratified sampling, where individuals are randomly selected from within each group (stratum) — not all members of selected groups. Cluster sampling is practical and cost-effective when a population is naturally divided into accessible groups, but it can produce higher variability if clusters differ substantially from one another.

Undercoverage bias occurs when part of the population has no chance of being selected because they are absent from the sampling frame. Using only landline numbers excludes the roughly 50 to 60 percent of US households that rely solely on cell phones — a group that skews younger, lower income, and may differ in health behaviors. This is a fundamental limitation of the sampling frame itself. Nonresponse bias (Choice A) is also a concern but addresses what happens after someone is contacted, not who can be reached at all.

The placebo effect occurs when participants experience genuine, measurable improvements in their condition simply because they believe they are receiving an active treatment. This is a real psychological and sometimes physiological phenomenon, not merely biased reporting. Including a placebo control group is essential in clinical trials precisely to quantify this effect and separate it from the true impact of the medication. Double-blinding (Choice C) is a design feature used to reduce bias, not an explanation for why the placebo group improved.

Response bias due to question wording occurs when the phrasing of a question systematically leads respondents toward a particular answer. The first version uses loaded terms like 'wasteful' and frames the policy as harming Americans, pushing respondents to oppose it. The second version uses neutral and positive framing. Good survey design requires neutral, balanced language that does not prime respondents toward any answer. This is distinct from sampling variability (Choice D), which refers to random fluctuation between samples, not systematic distortion caused by wording.

In an observational study, the researcher does not control or randomly assign the explanatory variable (sleep duration). People who sleep fewer than six hours may differ in many other ways — they may have more stressful jobs, poorer diets, less exercise, or higher BMI — all of which are associated with diabetes risk. These confounders could fully or partially explain the observed association. Only a randomized experiment, where sleep duration is experimentally assigned and other variables are controlled, would support a causal conclusion. Choice C is incorrect because 800 is a reasonable sample size for detecting such associations.

Replication in experimental design means applying each treatment to multiple experimental units (here, multiple plots). With only one plot per treatment, any difference in yield could easily be due to random variation in that particular plot — its specific soil conditions, drainage, or sun exposure — rather than the wheat variety itself. Using multiple plots per treatment allows researchers to estimate the natural variability in yield and determine whether the observed difference is larger than what chance alone would produce. Choice A describes scientific reproducibility, which is a different concept.

Identical twins share the same DNA, so pairing them controls for genetic differences that might otherwise affect memory performance. By randomly assigning one twin in each pair to caffeine and the other to decaf, the researcher ensures that any systematic difference in memory scores within pairs is more likely due to caffeine rather than underlying genetic variation. A completely randomized design could by chance place twins with naturally stronger memories disproportionately in one treatment group. Random assignment is still used (Choice D is incorrect) — it occurs within each pair, not between pairs.

Nonresponse bias occurs when individuals who respond to a survey differ systematically from those who do not. Voters who feel strongly about a tax increase — particularly those who oppose it — may be more motivated to return the survey, while indifferent voters may discard it. This creates a sample skewed toward strong opinions. There is no universal response rate threshold that invalidates results (Choice B is incorrect); what matters is whether respondents are representative. Choice C describes undercoverage, a different bias introduced by the sampling frame, not by nonresponse.

Question order effects are a form of response bias in which earlier questions prime or anchor respondents, influencing how they interpret and answer subsequent questions. When specific complaints are raised first, respondents are primed to think critically, which lowers overall satisfaction ratings. When overall satisfaction is asked first, respondents give an impressionistic positive rating before complaints are mentioned. This is why well-designed surveys carefully consider question ordering, sometimes randomizing or counterbalancing the sequence. Sampling variability (Choice D) refers to random fluctuation between samples, not systematic differences caused by design.

Blocking groups experimental units that are similar with respect to a variable that is likely related to the response (here, age affects baseline blood pressure and cardiovascular health). By assigning treatments randomly within each block, the comparison between treatment and control occurs among people similar in age. This reduces the variability that would otherwise make it harder to detect the true treatment effect. Drawing separate conclusions for each age group (Choice D) is a secondary possibility, but improving precision across the entire study is the primary purpose of blocking.

The researcher's design is stratified random sampling: the population is divided into strata (schools), and a random sample is drawn from each stratum, ensuring all schools are represented. The colleague's design is cluster sampling: entire groups (schools) are randomly selected and all members surveyed. Stratified sampling typically produces more precise estimates when the strata are internally similar with respect to the variable of interest, because it guarantees representation from every group. Cluster sampling is more convenient but can be less precise if clusters differ from each other in screen time habits.

Undercoverage bias at the sampling frame level occurs when a segment of the target population has no chance of selection. Adults who do not use social media tend to be older, less digitally engaged, and may have very different smartphone usage patterns — possibly far less screen time or fewer apps. Because this group is systematically excluded and differs meaningfully from social media users, estimates of 'US adults' smartphone habits based solely on social media users will be biased. Choice A (multiple accounts) causes unequal probability of selection but does not exclude entire segments, so it is a less fundamental flaw.

Seasonal temperature is the classic lurking variable in this scenario. During warmer months, people both purchase more ice cream and spend more time swimming — directly increasing the opportunity for drowning accidents. Ice cream sales and drowning rates are both responses to hot weather and are not causally related to each other. This illustrates a fundamental principle: correlation does not imply causation. A lurking variable can create a spurious association between two variables that have no direct causal link. Population size (Choice B) might explain city-level differences but does not explain month-to-month variation within a region.

Random assignment of treatments supports a causal conclusion — it controls for confounding variables, so the observed difference in muscle gains is attributable to the supplement. However, because participants were recruited from a specific population (college athletes who volunteered), the results can only be generalized to similar individuals, not all college students or the general public. This reflects the distinction between internal validity (supported by random assignment) and external validity (limited by the volunteer, athlete-specific sample). Choice C incorrectly denies causation — random assignment is what enables it, regardless of whether participants know they are in a study.

The critical flaw is that treatment assignment is based on when members signed up, not on random assignment. January sign-ups may be driven by New Year's resolutions and may differ from February sign-ups in baseline motivation, fitness level, and commitment — all of which could affect fitness gains independently of the program. This is a form of selection bias. Without random assignment, the groups are not comparable, so any difference in outcomes cannot be confidently attributed to the HIIT program. Double-blinding (Choice B) would be ideal but is a secondary concern — the non-random assignment is the more fundamental threat.

This is a classic example of Simpson's paradox, where a trend present in aggregated data reverses when data is broken into subgroups. If female applicants disproportionately apply to highly selective colleges (such as law or medicine) that have low acceptance rates for all applicants, the overall female acceptance rate will be lower — even if women succeed at equal or higher rates within each college. The college applied to acts as a lurking variable associated with both the applicant's gender and the probability of acceptance. Analyzing only aggregate data without accounting for this lurking variable produces a misleading picture of admissions bias.

For a variable to be a confounder, it must be associated with both the explanatory variable (pet ownership) and the response variable (blood pressure). Physical activity meets both criteria: people who are more active are more likely to own pets that require exercise (such as dogs), and exercise independently lowers blood pressure. This means physical activity could fully or partially explain the observed difference in blood pressure between pet owners and non-owners without pet ownership causing any direct benefit. Choice C describes reverse causation — a different validity threat — and Choice D describes how to improve the study design, not how physical activity acts as a confounder in the current study.

In multi-stage (cluster-based) sampling, the final sampled units are geographically clustered. Households within the same census tract tend to share similar income levels, racial and ethnic composition, housing type, and political views — they are more alike than a set of households drawn at random from across the country. This within-cluster similarity means each additional household from the same tract adds less new information than an independently selected household would. The result is that the effective sample size is smaller than the nominal sample size, producing wider confidence intervals than a true SRS of the same count. Despite this, multi-stage sampling is far more cost-efficient for national surveys and does produce valid (though less precise) probability-based estimates.

Stratified random sampling divides the population into non-overlapping subgroups called strata (here, grade levels) and then takes a random sample from each stratum. Cluster sampling would involve randomly selecting entire grade levels and surveying everyone in them. Simple random sampling would not guarantee representation from each grade level.

A census collects data from every member of the population of interest. Since the principal is measuring all 1,200 enrolled students — not a subset — this is a census. A sample only includes a subset of the population.

Voluntary response bias occurs when individuals choose whether to participate, and those who feel most strongly about an issue are more likely to respond. Callers who oppose increased parking fees are disproportionately motivated to call in. This is different from nonresponse bias, which occurs when selected individuals fail to respond to a survey they were asked to complete.

The defining feature of an experiment is that the researcher actively assigns treatments to subjects. In an observational study, the researcher simply measures variables as they naturally occur without intervening. Both study types can identify associations, but only well-designed experiments support causal conclusions.

A control group receiving a placebo provides a baseline response against which the treatment group can be compared. Without a control group, researchers cannot tell how much improvement is due to the drug versus natural recovery, the placebo effect, or other factors. The control group does not by itself create a double-blind design.

The placebo effect occurs when subjects experience real changes in symptoms simply because they believe they are receiving treatment, even when the treatment is inert. The Hawthorne effect refers to behavior changes caused by the awareness of being observed, which is a related but distinct phenomenon.

Systematic sampling involves selecting every \(k\)th individual from a list or sequence after a random starting point. Here, every 20th item is selected, making this a systematic sample. Convenience sampling would involve selecting items that are easiest to access, not every \(k\)th item.

Replication means applying each treatment to multiple subjects (or repeating the experiment). This allows researchers to distinguish true treatment effects from random chance variation. Without replication, it is impossible to know whether an observed difference reflects the treatment or just natural individual-to-individual variability. Replication does not eliminate the need for a control group.

Cluster sampling divides the population into groups (clusters) and randomly selects entire clusters to study. Here, schools are the clusters, and every student in the selected schools is surveyed. Stratified random sampling would require selecting a random subset of students from every school, not all students from only some schools.

Nonresponse bias occurs when those who respond differ systematically from those who do not. Alumni who feel strongly (either positively or negatively) may be more motivated to return the survey than those with neutral opinions. Although the original 1,000 were randomly selected, only an \(18\%\) response rate raises serious concerns about whether the 180 respondents represent the full group.

A double-blind design keeps both the participants and the researchers evaluating outcomes unaware of who received which treatment. This prevents both the placebo effect (from subjects) and evaluator bias (from researchers unconsciously rating treated patients more favorably). A single-blind design would only hide the assignment from participants.

In a matched pairs design, each subject is measured under both conditions (before and after), so each person serves as their own control. This controls for individual differences in baseline focus. A completely randomized design would randomly assign different subjects to either the music condition or a no-music condition.

A lurking (or confounding) variable is one that is associated with both the explanatory and response variable, creating the appearance of a causal relationship when none may exist. Denser, lower-income urban areas may have both more fast food restaurants and higher cardiovascular disease rates due to diet, access to healthcare, and other factors. While an experiment would be ideal, the primary flaw stated here is the failure to account for confounding.

Convenience sampling selects individuals who are easy to reach, but this produces a sample that may not represent the target population. People who attend a gym at 6 AM likely have stronger exercise habits than average Americans, so conclusions drawn from this group cannot be generalized. The sample size alone does not determine representativeness — a well-designed random sample of 50 could be more useful than this larger convenience sample.

In a randomized block design, subjects are first grouped into blocks based on a variable expected to affect the response (here, grade level), and then randomly assigned to treatments within each block. This reduces variability and allows the treatment effect to be estimated more precisely. Stratified sampling is a technique for data collection, not an experimental design.

Undercoverage occurs when some members of the population have no chance of being selected because they are not included in the sampling frame. Households without landlines cannot be reached by this method. Since landline ownership correlates with age and income, the sample systematically excludes certain demographic groups. Note that nonresponse bias (choice D) is also possible but requires the household to be in the frame first; undercoverage is the more fundamental issue here.

Response bias can result from question wording that leads respondents toward a particular answer. Survey 1 introduces a premise — rising violent crime — that frames the question in a way likely to increase support for police funding. This is called a leading question. Question wording effects are well-documented in survey research and can produce meaningfully different results even from equivalent random samples.

Simpson's paradox occurs when a trend present in aggregated data disappears or reverses when the data are disaggregated into subgroups. Here, Hospital A may treat a higher proportion of mild cases (which have lower mortality regardless of hospital quality), making its overall rate look better. Once case severity is controlled for, Hospital B performs better in both categories. The confounding variable is case-mix severity.

A confounding variable is associated with both the explanatory variable (carrying a lighter) and the response variable (lung cancer). People who smoke are much more likely to carry lighters and to develop lung cancer. The apparent relationship between lighters and cancer is not causal — it is explained by the lurking variable of smoking. Statistical significance alone (choice D) does not establish causation, especially in observational studies.

When subjects (or institutions) self-select into treatment conditions, the groups may differ in important ways before the treatment begins. Schools motivated enough to volunteer for a new curriculum may differ systematically from those that did not, making a fair comparison impossible. This is a lack of random assignment — the most fundamental requirement for drawing causal conclusions. Increasing the sample size (choice B) would not resolve this confounding.

When a population contains clearly defined subgroups that are internally homogeneous but differ from each other (high-income vs. low-income neighborhoods), stratified sampling reduces sampling variability compared to simple random sampling. By sampling within each stratum, the researcher captures the full range of the variable of interest. Cluster sampling (choice B) often increases variability because individuals within a cluster (neighborhood) tend to be similar to each other.

External validity (generalizability) refers to whether findings from a study can be applied to other populations, settings, or times. Internal validity — achieved here through random assignment — only ensures that the observed effect is attributable to the treatment within this study. College students in psychology classes are a convenience sample and may not represent working adults, older populations, or those with different sleep routines. External validity is a concern for all research designs, not just observational studies.

An interaction effect (also called effect modification) occurs when the effect of one variable depends on the level of another variable. Here, the advantage of active learning depends on school type: large in urban schools, reversed in rural schools. This means the two factors do not operate independently. Concluding that active learning is superior overall (choice D) would ignore this interaction and could lead to harmful policy decisions in rural contexts.

A fundamental trade-off of cluster sampling is the design effect: because individuals within the same cluster (county or ZIP code) tend to share demographic, economic, and political characteristics, the effective sample size is smaller than the nominal sample size. This inflates sampling variability relative to a true simple random sample of the same \(n\). Despite this limitation, multi-stage designs are used in practice because they are far less costly than a true national simple random sample.

Attrition bias (differential dropout) is a serious threat to internal validity in longitudinal studies. When dropout is related to treatment assignment and outcome — here, sicker or less tolerant participants in the drug group left early — analyzing only completers creates groups that are no longer comparable. The drug group completers may be those who tolerated the drug best, making the \(35\%\) improvement rate an overestimate of effectiveness in the full enrolled population. A complete-case analysis violates the intent-to-treat principle. Choice C raises a valid concern about unblinding, but the more direct and primary threat given the \(40\%\) differential dropout rate is attrition bias.

The defining feature of an experiment is that the researcher actively imposes treatments by randomly assigning subjects to conditions. This random assignment allows researchers to make causal conclusions. Observational studies simply observe subjects without intervening, so any association found cannot be attributed to cause and effect. Sample size, number of variables, and setting are not what define the difference.

An SRS requires that every possible sample of size \(n\) has an equal probability of being selected — this is a stronger condition than simply saying each individual has equal probability. Choice A describes equal inclusion probability, which is a consequence of an SRS but does not fully define it. Systematic sampling (choice C) and stratified sampling (choice D) are distinct methods that do not necessarily give every sample of size \(n\) equal probability.

Cluster sampling divides the population into groups (clusters), randomly selects some clusters, and then includes all members of the selected clusters. Here, neighborhoods are the clusters. In stratified sampling (choice A), you sample from every group, not just selected ones. Systematic sampling (choice B) selects every \(k\)th individual from a list. Convenience sampling (choice D) uses whoever is easiest to reach, without a random mechanism.

Nonresponse bias occurs when individuals who do not respond differ systematically from those who do. Here, the 1,600 households that did not return surveys may have different opinions or characteristics than the 400 who responded. Response bias (choice A) refers to inaccurate answers from those who do respond. Undercoverage (choice B) means some members of the population had no chance of being included. Voluntary response (choice D) occurs when people choose themselves, rather than being chosen.

The control group serves as a baseline for comparison and does not receive the active treatment. It may or may not receive a placebo. The placebo group (choice A) is a specific type of control group that receives an inactive treatment, making it a subset of the broader concept of a control group. Blocking (choice C) is a design technique to reduce variability. Replication (choice D) refers to repeating the experiment or applying each treatment to multiple subjects.

A voluntary response sample consists of people who choose to participate on their own. Such samples are biased because people with strong opinions — especially negative ones — are more likely to respond, making the results unrepresentative. None of the random sampling methods (choices A, B, or D) involve self-selection; they all use a defined random mechanism to determine who is included.

A confounding variable (also called a lurking variable) is related to both the explanatory and response variables, making it difficult to determine whether the explanatory variable truly causes changes in the response. For example, if wealthier people both exercise more and eat better, wealth confounds a study of diet and health. Blocking variables (choice C) are used in experimental design to control for known sources of variability, not to describe bias in observational studies.

Stratified random sampling divides the population into non-overlapping groups (strata) based on a shared characteristic, then draws a random sample from each stratum. Here, grade level defines the strata and the researcher samples from all four. This differs from cluster sampling (choice B), where only some groups are selected and all members in those groups are included. An SRS (choice A) would not guarantee representation from all grade levels.

The phrase 'most experts agree' is a leading cue that pressures respondents to agree, producing response bias — inaccurate answers caused by how the question is worded or asked. This inflates the apparent support for restrictions regardless of respondents' true views. Nonresponse bias (choice A) and voluntary response bias (choice D) relate to who participates, not how questions are worded. Undercoverage (choice B) relates to who has a chance to be included in the sample.

Blocking is used in experiments to control for known sources of variability by grouping similar experimental units together (into blocks) before randomizing treatments within each block. This reduces the chance that the variable (age, here) confounds the results. Stratified sampling (choice A) is a sampling technique, not an experimental design technique. Matched-pairs (choice D) specifically pairs two similar subjects, whereas blocking is a broader approach applied here to three groups of more than two.

Even though all students were emailed, nonresponse bias is introduced because the 800 who responded likely differ from the 9,200 who did not. Students with strong feelings (positive or negative) are more likely to take time to respond. The 72% figure may not represent the opinions of all 10,000 students. A sample of 800 from 10,000 (choice C) can actually be quite adequate if representative — sample size alone does not determine bias.

In a single-blind experiment, one party — typically the subject — does not know whether they are receiving the treatment or control. Here, subjects do not know which drink they received, but the experimenter does, making this single-blind. A double-blind design (choice A) would require that neither the subjects nor the people measuring responses know the assignments, which reduces experimenter bias. Since subjects are randomly assigned to conditions, this is an experiment, not an observational study (choice C).

Replication means applying each treatment to multiple experimental units so that results can be attributed to the treatment rather than chance variation among individuals. With only one worker per method, a single unusually skilled or unskilled worker could dominate the results. Blocking (choice B) would involve grouping workers by a known characteristic (like experience level) before randomizing. Control (choice D) refers to having a baseline group for comparison.

Observational studies cannot establish causation because confounding variables may be responsible for the observed association. Students who sleep more may also have lighter course loads, less financial stress, or better mental health — all of which could raise GPA independently of sleep. Self-reporting inaccuracy (choice A) is a valid concern but is secondary to the fundamental limitation of observational design. GPA is a quantitative variable (choice D), making that choice factually incorrect.

A longitudinal study follows the same subjects over time, collecting data at multiple points. This is also called a prospective study when subjects are followed forward in time. A cross-sectional study (choice B) collects data from different individuals at a single point in time — for example, comparing the diets of people of different ages simultaneously rather than tracking the same people over decades. Since no treatment is assigned, this is observational, not an experiment (choice A).

Multistage sampling involves selecting sampling units in stages, each using some form of random selection. Here, stage 1 randomly selects zip codes (clusters) and stage 2 randomly samples individuals within those zip codes. Pure cluster sampling (choice A) would survey all residents of the selected zip codes, not a random subset. Stratified sampling (choice B) would require sampling from all 50 zip codes, not just 10. Systematic sampling (choice D) uses fixed intervals on a list.

A large sample size reduces sampling variability but does not reduce bias from a flawed sampling method. Voluntary response samples systematically overrepresent people with strong opinions, often enthusiasm. The 85% figure reflects whoever chose to respond, not a representative sample of all viewers. This illustrates a key principle: precision (from a large \(n\)) and accuracy (freedom from bias) are independent qualities. A biased poll of 50,000 is worse than an unbiased poll of 1,000.

Blocking controls for a known source of variability (gender) by ensuring it is balanced across treatment conditions. If gender affects studying performance independently of music, blocking prevents gender from being a confounding variable. Without blocking, random assignment might by chance place most males in one group, distorting the comparison. Blocking does not increase sample size (choice A) and has nothing to do with blinding (choice D).

Without random assignment of students to note-taking method, self-selection creates potential confounding. Students who choose to write by hand may differ from laptop users in study habits, learning style, prior academic achievement, or course type — any of which could explain the difference in scores independently of note-taking method. This is the fundamental limitation of observational studies. Random assignment (as in a true experiment) is the only design feature that controls for all confounding variables, including those not yet measured.

Stratified sampling improves precision (reduces variance) when the strata are internally similar but differ from one another — a property measured by low within-stratum variance and high between-stratum variance. By ensuring representation from every stratum, the estimate of the population parameter is more stable across repeated samples compared to an SRS of equal size. However, stratified sampling does not eliminate bias (choice A) — both SRS and stratified sampling are unbiased estimators under proper random selection. Neither design enables causal conclusions (choice D), which requires an experiment.

When cluster sampling is used (here, selecting whole schools), individuals within a cluster share common influences, producing intra-cluster correlation. Standard formulas for standard error assume independent observations; when observations within a cluster are positively correlated, the effective sample size is smaller than the actual count of students, meaning the true standard error is larger than the formula suggests. This design effect must be accounted for with multilevel or cluster-robust methods. Random selection of schools (choice A) addresses external validity but does not create independence within schools.

The placebo effect — whereby patients improve simply from receiving treatment and believing it may help — is a well-documented psychological phenomenon. Double-blind, placebo-controlled trials are specifically designed to isolate the drug's effect above and beyond the placebo effect. A 40% improvement in the placebo group is not a design flaw; it is expected and accounted for. The drug is judged effective only if the treatment group improves significantly more than the placebo group. Removing placebo responders (choice D) would introduce bias by altering group composition after randomization.

Internal validity — whether the experimental treatment actually caused the observed effect — is strengthened by random assignment. External validity — whether results generalize to other people, settings, or times — depends on how subjects were recruited and selected, not on the \(p\)-value. A study with a highly significant result from a convenience sample of college undergraduates may have strong internal validity but weak external validity for the general adult population. A small \(p\)-value only tells us the result is unlikely under the null hypothesis; it says nothing about generalizability.

Calling households on weekdays between 9 AM and 5 PM systematically excludes people who work outside the home. Stay-at-home caregivers, retirees, and those who work from home are overrepresented. This is undercoverage — a segment of the target population (working adults) had no chance of being included. The direction of bias is uncertain but plausible: income, lifestyle, and food-purchasing behavior likely differ between those home during business hours and those at work. Choice D incorrectly claims random selection eliminates nonresponse bias — it does not.

An interaction effect occurs when the effect of one factor on the response variable depends on the level of a second factor. Here, lighting does not have a consistent directional effect on performance across all tasks — its effect reverses depending on whether the task is creative or analytical. A main effect (choice A) would mean lighting consistently improves or hinders performance regardless of task type. Confounding (choice B) would mean the factors are not independently controlled, but since subjects are randomly assigned to all four combinations, this is a properly designed factorial experiment.

A simple random sample gives every individual — and every group of individuals of the same size — an equal chance of being selected. Using a random number generator on a numbered list is the textbook implementation of this method. A stratified sample would require dividing students into subgroups (strata) first and sampling within each. A cluster sample would randomly select entire pre-formed groups. A systematic sample would pick every kth individual from the list.

Voluntary response bias occurs when individuals self-select into a sample, typically because they feel strongly about the topic. People who oppose the ordinance are far more motivated to call in than those who mildly support it, skewing results dramatically. Nonresponse bias applies when selected participants fail to respond — here, no one was randomly selected in the first place. Undercoverage bias occurs when part of the population has no chance of being sampled, which is a secondary concern but not the primary flaw in this design.

The control group provides a baseline — a reference point showing what happens under standard or no-treatment conditions. By comparing the treatment group to the control group, researchers can attribute differences to the treatment rather than to chance or outside factors. The control group does not eliminate the placebo effect; it actually helps measure it, since control subjects often receive a placebo. Random assignment is a design feature of the entire experiment, not a function of the control group specifically.

Stratified random sampling divides the population into non-overlapping subgroups called strata — here, the four grade levels — and then takes a random sample within each stratum. This ensures representation from every grade. Cluster sampling would involve randomly selecting entire grades and surveying everyone within the chosen grades. The fact that equal numbers (30) are drawn from each stratum is a choice by the researcher, not what defines stratified sampling.

This is a classic confounding variable problem. Wealthier or more walkable neighborhoods may simultaneously attract more coffee shops and have residents with healthier lifestyles, better healthcare access, and lower stress — all of which reduce cardiovascular disease. The presence of coffee shops is not necessarily causing better health outcomes; both may be caused by a third variable. Observational studies can legitimately measure correlation; the flaw is inferring causation from it. Sampling method is not relevant here — the issue is causal reasoning from an observational dataset.

Selecting every kth individual from a list or sequence is systematic sampling. The key concern is periodicity: if hospital arrivals follow a repeating pattern — for example, certain types of patients consistently arrive at the same times of day or days of the week — then every 15th patient may systematically over- or under-represent certain groups, violating the goal of representativeness. This is a known weakness of systematic sampling compared to simple random sampling. Cluster and stratified sampling involve different structures that do not apply here.

A double-blind experiment means that neither the subjects nor the researchers interacting with or evaluating them know who received which treatment. This controls for two sources of bias: the placebo effect (subjects who know they received treatment may report improvement regardless) and researcher bias (clinicians who know the treatment assignment may unconsciously rate outcomes differently). Because subjects are randomly assigned to treatment conditions, this is an experiment, not an observational study. A single-blind design would only keep one party — usually the subject — unaware of the treatment.

The primary purpose of blocking is to reduce variability — or experimental error — by grouping subjects who are similar with respect to a variable that affects the response (here, stress level). Within each block, subjects are comparable, so differences between treatment and control within a block are more likely due to the app itself. This makes the experiment more sensitive to detecting a real treatment effect. Blocking does not eliminate confounding entirely; it controls for one known source. Testing whether stress level modifies the treatment effect would require a factorial design or interaction analysis, which is a separate question. Blocking has nothing to do with blinding.

Undercoverage bias arises when part of the population has no chance — or a systematically lower chance — of being included. Adults who rely only on cell phones (disproportionately younger people), those who work outside the home on weekday afternoons, and those without phone access are all excluded from this sampling frame. If these groups hold systematically different opinions about education policy, the estimate will be biased in ways that the margin of error — which only reflects sampling variability, not bias — cannot correct. The 400 non-completers represent nonresponse bias, not voluntary response bias. A \(\pm 3\%\) margin of error is a standard and acceptable level for national polls.

The Hawthorne effect refers to the tendency for people to change their behavior — usually improving it — simply because they know they are being observed or are part of a study, regardless of the specific treatment. If this effect dominates, both groups would show substantial productivity gains relative to their baseline, but the difference between Program X and Program Y would be modest or negligible, since both groups are being observed equally. The best design fix is including a true no-treatment control group that is also observed — this allows researchers to estimate how much improvement is attributable to being studied versus to the actual wellness content. A matched-pairs design addresses between-subject variability, not observation effects.