What is CATI? The method and the decisions behind it

Computer-assisted telephone interviewing, or CATI, combines a spoken interview with a programmed questionnaire. The interviewer asks the questions and enters responses; the software presents the appropriate items, applies routing rules and stores the data. The Australian Bureau of Statistics’ account of survey methods also describes immediate data checks, call scheduling and supervision among its practical advantages.

That arrangement gives researchers considerable control over administration. It can stop an interviewer skipping a required question, route a respondent around an irrelevant section and retain information about unsuccessful contacts. Its value depends on the specification it implements. A routing mistake programmed into the instrument can be repeated just as consistently as the correct rule.

The sample exists before the interview

CATI describes how interviews are administered. A study still needs a target population, a sampling frame and a rule for selecting cases. Randomly generated telephone numbers, an organisation’s customer list and a recruited research panel offer different routes to respondents. Each has its own coverage limitations and selection process, which remain relevant after the interviews are complete.

The unit reached by a call may also differ from the unit the study intends to measure. A household number can reach several adults; a business number may reach a receptionist rather than the person qualified to answer. If the design requires selection within a household or organisation, the introduction and screening questions must implement that rule. Interviewing whoever happens to answer can change the selection process.

The sample-management record should make those decisions recoverable. Link each attempt to a sampled case, record its outcome and retain the final disposition. Explain how duplicate contact details and repeated attempts are handled. This is the information needed to account for the issued sample and calculate response rates with an appropriate denominator. The presence of CATI software alone says nothing about whether inclusion probabilities are known or whether a quota has been filled.

The questionnaire becomes an executable specification

Programming translates a questionnaire into decisions about what is read, what is displayed only to the interviewer, which answers are accepted and what happens next. A complete specification includes response categories, reference periods, rotations, eligibility rules and the treatment of uncertainty. It also identifies which instructions must be spoken and which are guidance for administering the question.

Suppose a service-use survey asks how many appointments a respondent attended in the previous month, then asks about the most recent one. Zero appointments should skip the follow-up. An inability to remember the number may require a different route, because the respondent could still describe the latest appointment. A refusal to answer is another state again. Collapsing all three into zero produces clean-looking data by discarding meaningful distinctions.

The codebook and routing specification should preserve those states, including questions that were never asked. Test the programmed instrument with deliberately chosen cases: someone who is ineligible, someone at the boundary of a numeric range, someone who corrects an earlier answer, and someone who declines a question that controls later routing. Compare the exported record with the expected result. A question can look right on screen while writing the wrong value to the dataset.

Data checks also need judgement. An impossible calendar date can justify a hard correction; an unusually large but possible value may justify a confirmation prompt. Treating every unusual answer as an error can encourage interviewers to obtain a more conventional answer. Researchers should decide in advance which conditions require correction, which allow confirmation, and how unresolved cases are recorded.

Standardisation includes rules for clarification

Reading the same words to everyone gives a study a common starting point. Respondents may still interpret those words differently. Definitions of a household member, paid work or a service visit can be clear to the research team and uncertain at the other end of the telephone. The interviewer’s response to a request for help becomes part of the measurement process.

Schober and Conrad’s 1997 experiment examined this problem using professional telephone interviewers and questions drawn from government surveys. Respondents answered from fictional situations, allowing the researchers to assess accuracy. Standardised and more flexible interviewing both worked well when the situations mapped clearly to the survey concepts. Allowing clarification improved accuracy for ambiguous situations, at the cost of longer interviews.

The laboratory design does not establish that unrestricted conversation improves every field survey. It does show why identical wording cannot guarantee identical understanding. For CATI practice, the useful implication is to develop clarification rules alongside the questions: approved definitions, neutral probes, and a clear boundary on what an interviewer may explain.

A respondent asking whether a student living away from home counts as a household member needs the study’s membership rule. Repeating the question may leave the uncertainty unresolved; an improvised definition may introduce a different rule. Cognitive testing can help identify these ambiguities before fieldwork. Interviewer briefings should then cover how the agreed definitions apply to difficult cases, as well as how to operate the software.

Quality control connects conversations to records

The US Behavioral Risk Factor Surveillance System offers a documented example of CATI operating within a wider quality process. The CDC’s 2024 overview describes interviewer training, supervision, monitoring and checks of disposition patterns by interviewer. These activities help identify problems that a completed questionnaire file cannot reveal on its own.

For a particular study, useful review questions include whether the introduction was delivered as approved, whether response options were read when required, and whether clarification remained within the protocol. Open answers need to be captured with enough fidelity for their intended analysis. A summary entered by an interviewer and a verbatim response are different data products; the specification should say which is required.

Patterns in the fieldwork record can direct further investigation. Unusually short interviews may indicate a routing problem, rushed administration or simply an eligible subgroup with fewer questions. Higher refusal rates for one interviewer may reflect the cases or shifts assigned to that person. Investigate those explanations before attributing the difference to performance.

When review leads to a correction, record what changed, which cases were affected and how the analytical file was treated. Some errors can be repaired from the available record. Others require recontact or leave an answer missing. An explicit correction history allows an analyst to understand the limits of the final data, including decisions that would otherwise disappear during cleaning.

Keeping a tracking series interpretable

Reusing a questionnaire is only one part of maintaining comparability. A tracker can change because its frame, respondent selection, contact protocol, weighting or mode changes. A software revision can also alter the reading of a scale or the route into a section. Keep these changes in a wave-by-wave methods record.

BRFSS provides a consequential example. In 2011 it incorporated cellphone data and introduced a new weighting method. The CDC advises against directly comparing estimates from before and after those changes. Continued use of telephone interviewing did not, on its own, preserve comparability.

Where continuity matters, plan an overlap or bridge study before replacing an established procedure. Running the old and new protocols during the same period, with comparable or randomly assigned samples, can help distinguish a method effect from temporal change. The design and sample size should be adequate for the outcomes whose trends need to be preserved.

Mode choice deserves equal attention at the start of a study. An interviewer may help administer a question that needs clarification, while a long list of response options can be difficult to retain by ear. Tasks requiring diagrams, documents or careful comparison may suit a visual format. The ABS guide to data collection methods treats the nature of the questions, respondent burden and practical constraints as reasons to choose a method deliberately.

Replacing the interviewer with a voice AI system introduces another set of administration decisions. Recognition, clarification and answer coding need evaluation against the intended instrument and population; we examine that evidence in Evaluating AI voice interviewers for survey research. For any CATI study, the commissioning question remains concrete: can the proposed sample, instrument and fieldwork procedure support the estimates the research needs?