Are administrative data always government data?

No. Businesses, charities and other organizations also generate routine operational records.

Can administrative data replace a survey?

Sometimes they can supply some or all required information, but suitability depends on coverage, concepts, accuracy and the question. Combining sources may be more appropriate.

Does a missing record mean the event did not occur?

Not necessarily. It can mean no event, no contact with the source organization, delayed reporting, failed linkage or another recording limitation.

AI Glossary / Statistics and probability / Beginner

Administrative data

Definition

Administrative data are records created as organizations carry out routine operations, deliver services or administer programs. Examples include registrations, transactions, payroll records and service-use records. They are collected primarily to support those activities and may later be reused for statistics, research or machine learning.

What it means

The defining feature is the original operational purpose, rather than a particular file format or storage technology. Administrative data can come from public bodies or private organizations and can describe people, businesses, events or transactions.

These records may provide detailed information without asking people to complete another survey. However, they reflect the rules and processes of the organization that created them. A record of service use describes contact with that service; it does not necessarily describe everyone who needed it. Statistical usefulness therefore depends on understanding the recording process.

How it works

Start by identifying the source system, its operational purpose and the population it covers. Determine what each row represents and how identifiers, timestamps and status fields are maintained. A row may represent a person, an application, a payment or an updated version of an earlier record.

Before analysis, map operational fields to the statistical concepts required. Resolve duplicate records, define the relevant period and handle cancellations or revisions according to documented rules. Link sources only when appropriate, evaluating missed and incorrect matches as potential sources of error.

Assess coverage, missingness, measurement consistency and changes in administrative procedures. Eligibility rules or software updates can change counts even when the underlying phenomenon is stable. Maintain metadata and a traceable processing history. Appropriate access controls and disclosure protection remain necessary because detailed operational records can contain sensitive information even after obvious identifiers are removed.

Three practical examples

1. A payroll provider summarizes recorded payments to support an employment indicator, checking which employers and workers are included in its system.

2. An education analyst studies enrollment records while distinguishing enrollment events, individual learners and changes in program registration.

3. A service organization uses historical support records to train a routing model, examining whether recorded outcomes reflect genuine resolution or merely administrative closure.

Think of it this way

Think of administrative data as the paper trail left by everyday work. The trail can reveal a great deal about what the organization did, but it follows the organization's procedures. People who never entered the process may leave no trace, and a change in paperwork can change the trail without changing the underlying need.

Worked example

Consider an original monthly service extract containing 1,200 rows. A documented check identifies 100 exact duplicate records caused by an export issue, leaving 1,100 distinct recorded visits. These visits belong to 800 unique clients. The average is therefore 1,100/800 = 1.375 recorded visits per client.

Suppose a separate, compatible reference list contains 1,000 eligible clients for the same period and all 800 recorded clients are verified as members of that list. The observed proportion with a recorded visit is 800/1,000 = 80%.

The 1,200 raw rows are not 1,200 clients, and the 1,100 visits are not 1,100 unique people. Furthermore, the 200 eligible clients without a recorded visit cannot automatically be classified as having no need. They may not have sought service, may have used another provider, or may have missing records. The arithmetic describes the defined records and denominator; it does not resolve those explanations.

Illustrative case study

Fictional case study: Serviceledger uses appointment records to estimate demand for a community program. A sudden increase appears after a new booking system is introduced. Analysts discover that rescheduled appointments now create additional records instead of updating the original record.

They reconstruct a consistent event definition and document the break in the recording process. They also compare the administrative coverage with an independent population source before discussing unmet demand. Their published summary separates recorded activity from estimates requiring additional assumptions.

This fictional case illustrates why operational metadata matters. It does not claim that administrative records are inherently unreliable or that a survey would automatically measure the same concept more accurately.

Common misconceptions and mistakes

A large administrative dataset is not automatically representative of the population of interest. Complete coverage of a program's users is different from complete coverage of everyone eligible for or needing the program. Detailed records can still contain systematic omissions and classification errors.

Administrative does not mean publicly available, anonymous or collected solely by government. Nor does the absence of survey sampling eliminate all uncertainty. Processing errors, incomplete coverage, changing definitions and linkage mistakes can materially affect conclusions even when every available record is used.

Business use

Administrative data can reduce repeated collection effort, improve timeliness and support analysis of activity across periods. For AI projects, they can supply operational context and outcomes that would be expensive to collect separately.

Before using them, establish the intended decision and verify that the recorded variables correspond to it. An administrative completion code may reflect a workflow rule rather than customer success, and using it as a training label can teach a model the wrong objective. Monitor changes in source systems and preserve a versioned data dictionary. Evaluate performance across covered groups while acknowledging groups absent from the records. Match the level of access and detail to the approved analytical purpose.

Check your understanding

A system contains 500 transactions from 350 customers. Which count should describe the number of customers represented?

Answer: 350, assuming identifiers correctly distinguish unique customers. Five hundred describes transactions, a different unit of observation.

Related terms

Administrative data collection
Record linkage
Coverage error
Metadata
register
data provenance
statistical unit and secondary data

Learn next

trace one administrative field from its original operational entry through cleaning to a published statistic or model feature. Record its definition, possible missingness, update rules and limitations, then test how changes in those rules would affect the result.

Sources

Primary sources: Australian Bureau of Statistics, Statistical terms and concepts glossary, Administrative data: https://www.abs.gov.au/statistics/understanding-statistics/statistical-terms-and-concepts/statistical-terms-and-concepts-glossary

Australian statistical data integration guidance, Data management: https://statisticaldataintegration.abs.gov.au/topics/data-management

Australian statistical data integration guidance, Quality assurance: https://statisticaldataintegration.abs.gov.au/topics/data-management/data-quality/quality-assurance

Previous: Adjusting or controlling for a variable Next: Administrative data collection