A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
Surveys collect self-reported data on attitudes, behaviors, or experiences from large samples via questionnaires. Survey quality depends on clear wording, appropriate response scales, logical order, and piloting. Sampling strategy determines whether results generalize; response rate and representativeness affect validity. Surveys are cost-effective for descriptive and correlational research.
Critique published survey instruments for clarity, response bias, and relevance. Draft a brief survey and pilot it with colleagues, noting confusion or skip patterns. Compare online, paper, and in-person administration modes.
Your prerequisite on variable definition and operational measurement established that psychological constructs — anxiety, motivation, trust, satisfaction — must be operationalized: translated from abstract concepts into concrete, observable, measurable responses. Surveys are the most widely used operationalization vehicle in social science. Building a good survey means solving the operationalization problem at the item level, for every question on the instrument, while simultaneously managing the conditions under which responses are collected.
Every survey item is an attempt to extract a reliable signal about some internal state. The challenge is that the path from internal state to recorded response passes through several steps: the participant must interpret the question, retrieve relevant information from memory, form a judgment, and map that judgment onto the provided response options. Each step introduces potential distortion. Response biases — systematic tendencies to respond in ways unrelated to the true construct — are the primary threat. Acquiescence bias is the tendency to agree with statements regardless of content; it inflates positively worded items and can be partially controlled by including reverse-scored items. Social desirability bias is the tendency to present oneself favorably rather than accurately — particularly strong for sensitive topics like drug use, sexual behavior, income, and prejudiced attitudes. Both biases produce systematic error that mimics real variation in the construct, making them harder to detect than random error.
Question wording is the most controllable source of bias. Double-barreled questions ("How satisfied are you with the price and quality?") force a single response to two distinct questions and produce uninterpretable data — a respondent who loves the quality but hates the price cannot answer honestly. Leading questions ("Don't you agree that the policy was unfair?") embed an evaluative frame that pulls responses toward a predetermined answer. Loaded terms and abstract language trigger idiosyncratic interpretations: if one participant reads "frequently" as "more than once a week" and another reads it as "more than once a day," their responses are not measuring the same thing. Best-practice item writing uses specific, neutral, concrete language that a thoughtful stranger with no context would read in only one way.
Response scales shape the distribution and meaning of responses as much as question wording does. The number of scale points, the presence or absence of a neutral midpoint, and the verbal labels on endpoints all matter. A 5-point scale with a labeled neutral midpoint gives genuinely indifferent respondents a valid option; a forced-choice 4-point scale requires a lean in one direction — appropriate when you believe "neutral" is actually avoidance rather than genuine ambivalence. Order effects operate at both the item and survey levels: early items prime the cognitive context for later ones, and demographic questions at the beginning can activate identity-based response patterns that color substantive answers. Standard practice places sensitive items after rapport-building items and demographics at the end.
Sampling links instrument quality to research validity. A perfectly constructed survey administered to a non-representative sample produces internally valid but ungeneralizable findings. Probability sampling — where every unit in the target population has a known, nonzero chance of selection — is the basis for statistical generalizability. Simple random sampling gives equal probability to every unit; stratified sampling ensures adequate representation of key subgroups by sampling within strata separately; cluster sampling draws entire naturally occurring groups (schools, neighborhoods) when individual-level sampling is impractical. Non-probability samples (convenience, snowball) are common in practice but require explicit acknowledgment of generalizability limits. Response rate interacts with representativeness in a non-obvious way: a high response rate from a poorly defined sampling frame is less valuable than a moderate response rate from a probability sample of the actual target population. What matters is not how many people responded, but whether the people who responded are representative of the people you wanted to describe.