Must interviews be face to face?

No; telephone interviewing is also possible.

Does an interview survey require open-ended questions?

No; it can use structured response options.

AI Glossary / Statistics and probability / Advanced

Interview survey

Definition

An interview survey collects answers through verbal interaction with an interviewer, who records the responses. It describes how a questionnaire is administered, not how respondents are selected. Interviews may occur in person or by telephone, and the recording process may use paper or software.

What it means

The interviewer presents questions, manages the sequence, and records what the respondent says. This can help people navigate a questionnaire, but it also introduces opportunities for inconsistent explanations, prompting, and recording mistakes.
An interview survey is not automatically a representative survey. Interviewing whoever is easiest to reach remains a selection decision with consequences. Likewise, a structured interview survey intended to estimate proportions differs from an exploratory conversation intended to understand experiences in depth.

How it works

Define the population and sampling approach separately from the interview mode. Prepare a questionnaire with clear wording, reference periods, response options, and instructions for questions that should be skipped. Test it with potential respondents before full collection.
Train interviewers to deliver questions consistently, use agreed neutral clarification procedures, and distinguish a respondent's answer from an interpretation. Record contact outcomes and incomplete interviews so the final dataset does not conceal who was not represented.
Evaluate both the questionnaire and its administration. Cognitive interviews can reveal how people understand questions; examination of interviewer-respondent behavior can reveal inconsistent delivery or answers that do not fit the options. Analysis should address the sampling design and relevant missing responses rather than treating every collected answer as equally representative by default.

Three practical examples

A company uses telephone interviews to ask sampled customers about a recent support interaction. Interviewers apply the same reference period and response categories.
A research team interviews employees about AI-tool usage during the preceding week. It defines whether use includes embedded features as well as standalone assistants.
A service organization offers interviewer assistance for a questionnaire that some respondents find difficult to navigate. It evaluates whether the assisted mode changes how questions are understood or answered.

Think of it this way

The interviewer is part of the measurement process, like the person positioning a measuring instrument. Consistent handling helps, but neither careful handling nor a sophisticated instrument fixes an unsuitable selection of people to measure.

Worked example

Consider an invented survey selecting two hundred eligible employees: one hundred from Department A and one hundred from Department B. Both departments have equal numbers of employees in the target population. Assume the initial sample was selected appropriately within each department.
The final contact records show 140 completed interviews, twenty partial interviews, twenty refusals, and twenty people not reached. These categories sum to two hundred. The simple completed-interview proportion is 140/200 = 70 percent. This is a simplified disposition summary, not a claim to implement every element of a formal response-rate standard.
Suppose all one hundred selected employees in Department A complete interviews, but only forty of the one hundred selected in Department B do. The interview dataset therefore contains 100/140, or about 71.43 percent, Department A respondents, even though the two departments are equally large in the target population.
Now examine a particular yes-or-no question about weekly AI-tool use. Among the completed interviews, ninety Department A respondents and thirty-six Department B respondents provide an answer. Fourteen completed interviews therefore lack this particular item: ten in A and four in B. Completing an interview does not guarantee that every question has an answer.
Seventy-two of the ninety valid Department A answers are yes, giving 80 percent. Twelve of the thirty-six valid Department B answers are yes, giving one-third, or about 33.33 percent.
Pooling the valid answers gives eighty-four yes responses out of 126, or about 66.67 percent. Dividing eighty-four by all 140 completed interviews would instead give 60 percent, but that would treat the fourteen missing item responses as if they were no. The denominator must match the reported quantity.
For a purely illustrative department adjustment, give each department half the population weight. The result is 0.5 times 0.8 plus 0.5 times one-third = about 56.67 percent. This differs by ten percentage points from the unweighted proportion among valid answers.
That adjustment is not automatically a valid population estimate. It assumes, among other things, that the observed within-department answers adequately represent those missing within each department. If Department B's nonrespondents use the tool differently, matching the department proportions does not remove that bias.
Suppose each completed interview takes twenty minutes. The 140 completed interviews alone require 2,800 minutes, or forty-six hours and forty minutes. Contact attempts, partial interviews, preparation, and quality checks require additional time. Treating completed-interview duration as the whole collection budget would understate the work.
Finally, imagine an interviewer rephrases weekly use as regular use. A respondent who used a tool once during the specified week might answer differently. Neither the department adjustment nor the correctly calculated denominator repairs that change in the measured question.

Illustrative case study

Fictional company Alder Research prepares interviews about employees' use of AI features. A pilot reveals that some respondents count embedded writing suggestions while others count only standalone chat tools.
The team clarifies the question, retests it, and gives interviewers a consistent explanation. It keeps records of unanswered items and unsuccessful contacts, then reports who the results describe.

Common misconceptions and mistakes

Interview mode does not determine whether the sample is random or representative.
A high completion proportion is useful information but does not by itself establish the absence of nonresponse bias.
An interviewer should not silently convert uncertainty or refusal into a substantive answer.
More detailed probing is not always better if it changes the meaning or comparability of a standardized question.
Weighting cannot automatically repair misunderstood questions, inaccurate recording, or every form of missingness.

Business use

Choose interviews when their assistance and interaction suit the population and information needed. Budget for recruitment, interviewer preparation, supervision, and analysis as well as conversation time. Report the sampling method, interview mode, fieldwork period, participation outcomes, item denominators, and any weighting so the findings can be interpreted in context.

Check your understanding

How many completed interviews lacked the AI-use item?

Answer: Fourteen.

What was the unweighted yes proportion among valid answers?

Answer: About 66.67 percent.

Does department weighting prove the missing respondents are represented correctly?

Answer: No.

Related terms

Interviewer error
Questionnaire
sampling frame
unit nonresponse
item nonresponse
cognitive interviewing
survey weighting
mode effects

Learn next

Study questionnaire pretesting, neutral interviewing, response-disposition reporting, and the assumptions behind nonresponse adjustments.

Sources

NIST statistical glossary: https://www.itl.nist.gov/div898/handbook/glossary.htm . U.S. Census Bureau, Questionnaire Testing and Evaluation Methods: https://www.census.gov/about/policies/quality/standards/appendixa2.html ; Measures and Indicators of Nonsampling Error: https://www.census.gov/about/policies/quality/standards/standardd3.html . Survey figures and company scenario are original illustrations.

Previous: Intervenability Next: Interviewer bias