Administrative Data

A guide to obtaining, using, and managing administrative data for research, covering access processes, data flow options, data use agreements, ethical considerations, data validation, and common challenges.

Key Takeaways

Administrative data refers to information collected, used, and stored primarily for operational purposes rather than research.
While IPA projects usually rely on primary data, administrative data can also be a valuable input resource for research and impact evaluation.
Accessing and using Administrative Data requires careful planning across four stages: information gathering, implementation, legal compliance, and data validation.

What is Administrative Data?

Administrative data is typically collected by government agencies and organizations for registration, transaction, and record-keeping purposes. Examples of administrative data can include credit card transactions, electronic medical records, insurance claims, educational records, arrest records, and mortality records.

Administrative Data vs. Survey Data

Some challenges are common to both survey and administrative data. However, administrative data also comes with its own strengths and limitations. The table below compares the two approaches across key dimensions.

Dimension	Survey data	Administrative data
Recall and social desirability bias	High risk	Low risk
Attrition	Higher risk	Lower risk
Costs	Higher cost	Lower cost
Logistics	Easier to manage	More complex
Time to access	Faster access	Slower access
Documentation quality	Usually better documented	Often less documented
Control over study population	High control	Limited control
Control over data processing	High control	Limited control
Treatment affects measurement	Lower risk	Higher risk
Incentives to misreport	Lower risk	Higher risk

Real-World Example: IPA’s Experience with Administrative Data

IPA’s research on using administrative data for Monitoring and Evaluation (2016) highlights several key advantages:

Cost-Effectiveness: Administrative data reduces or eliminates the need for additional monitoring activities or surveys.
Timely Response: Regular updates in management information systems enable faster analysis of key indicators.
Large Sample Size: Coverage of entire beneficiary populations provides robust sample sizes.
Improved Accuracy: Reduces social desirability bias and recall issues common in self-reported data.

This research emphasizes that data accuracy and reliability should take precedence over cost savings. Organizations must balance data quality, actionability, and resource allocation when incorporating administrative data into their research design.

Standard Processes for Accessing Administrative Data

Accessing administrative data involves four main steps:

Finding Administrative Data

Implementing partners and government agencies are valuable sources of administrative data. Common sources include:

Health Data: Regional/national health departments, hospitals, health insurance records
Financial Data: Banks, credit unions, credit reporting agencies
Education Data: Schools, ministries of education, standardized testing agencies

Formulating a Data Request

When requesting administrative data, researchers should:

Define the time frame, format, and structure needed: “Primary school records from January 2022 to December 2023 in CSV format”
List specific variables of interest: Student ID, school level, Teacher ID, attendance, test scores
Specify whether you need identified or de-identified data: Request de-identified data when possible to reduce ethical complexity
Avoid broad requests: Instead of “all student data,” specify exact variables, time frames needed, and frequency of updates

Implementing Data Flow Strategies

A well-planned data flow strategy ensures smooth integration of administrative data:

Gather Identifying Information: Determine what identifiers are available in the study sample such as national ID numbers, phone numbers, email addresses
Link Datasets: Use pre-existing identifiers to match study data with administrative data. For example, match student IDs with test records using national ID numbers
Choose a Matching Strategy:
- Exact matching: When identifiers are identical such as national ID numbers
- Probabilistic matching: When using combinations of name, date of birth, and location information
Determine Who Performs the Link: Clarify whether the data provider, researcher, or a third party will conduct the linkage and deindentification to maintain data security and privacy

Data Flow Options

The structure of a data flow determines who has access to personally identifiable information (PII) and who conducts the match between study records and administrative data. Choosing the right option depends on the sensitivity of the data, the data provider’s legal constraints, and the research team’s security infrastructure. The table below summarizes the most common configurations.

Who conducts the match	Who has access to identified data	Notes
Researcher	Researcher receives identified data from agency	Simplest, but requires strong data security on the researcher’s end
Researcher, on-site at the agency	Researcher brings encrypted finder file; leaves with de-identified file	Useful when the agency restricts data from leaving its premises
Researcher, on agency device	Researcher conducts match and analysis on an agency-monitored computer	Agency retains oversight of the process
Data agency	Researcher never receives identified data; agency strips identifiers before transfer	Researcher agrees never to re-link the finder file with the analysis file
Data agency, with third-party intermediary	No single party has access to both identifiers and outcomes	Research partner assigns study IDs; agency conducts match; researcher receives only de-identified results
Data agency runs researcher-written code	Researcher never accesses raw data	Researcher writes and tests code, sends it to agency, and receives only analytic results

Matching errors affect statistical power

Exact matching virtually eliminates false positives but increases false negatives, which attenuates impact estimates and reduces statistical power. Probabilistic matching reduces false negatives but introduces the risk of false positives, which can be especially harmful if match quality correlates with treatment or control status. Researchers should document their matching strategy and assess whether matching errors are likely to be unrelated to treatment assignment.

Match rates in practice

Match rates vary considerably across studies depending on the quality of identifiers available, the data source, and the matching strategy used. The table below summarizes match rates from published studies that used administrative data linkage.

Study	Records being matched	Match rate
Health Care Hotspotting	Hospital discharge records for enrolled participants	~95%
Oregon Health Insurance Experiment (Finkelstein et al., 2012)	Credit reports for lottery participants	68.5%
Effect of pre-trial detention on conviction (Dobbie, Goldin & Yang, 2018)	Tax data for defendants	73–81%
Impact of kindergarten class on life outcomes (Chetty et al., 2011)	Parents’ tax records for kindergarten students	86%

These benchmarks illustrate that even well-resourced studies rarely achieve perfect linkage. Researchers should account for expected match rates when designing studies and assessing statistical power.

Data Use Agreements

A Data Use Agreement (DUA) is a legal document that outlines the terms under which a data provider shares data with a research institution. DUAs are also referred to as Data Sharing Agreements (DSAs), Memoranda of Understanding (MOUs), or Non-Disclosure Agreements (NDAs), depending on the context and the parties involved.

A well-structured DUA protects both the data provider and the research team by clarifying responsibilities, limiting liability, and ensuring that data are used only for agreed-upon purposes. The table below describes the elements that a DUA should include.

Project description: Summary of the research purpose and scope
Authorized users and analysts: Names or roles of individuals permitted to access the data
Data security procedures: Encryption, storage, and access control requirements
Data to be shared: Variables, time range, and format of the dataset
Timeframe: Duration of the agreement and data access period
Data destruction: Procedures for deleting or returning data at the end of the project
Publication review: Whether the data provider has the right to review outputs before publication
Data publication: Conditions under which derived data or results may be made public

DUAs must be signed by an institutional representative

A DUA should be executed by an authorized institutional representative, not by an individual researcher or staff member. This protects the researcher from personal liability and ensures that the institution assumes responsibility for compliance.

Tips for negotiating a DUA

Negotiating a DUA can be a lengthy process. The following practices help move negotiations forward:

Understand legal constraints from the start: Some provisions that appear as “requirements” from the data provider may be preferences that can be modified. Understanding what is legally mandated versus negotiable saves time.
Build the relationship early: Establishing trust with the data provider before formal negotiations often makes the process smoother.
Frame the research in terms of the provider’s mission: If the research is relevant to the agency’s own goals, emphasize that relevance to increase their willingness to share.
Consider using an intermediary: In complex or sensitive cases, a trusted third party can facilitate negotiations between the research team and the data provider.

Ethical Considerations for Using Administrative Data in RCTs

Using administrative data in research requires adherence to ethical and legal standards. Key considerations include:

Institutional Review Board (IRB) Approval

Most research using administrative data qualifies as human subjects research and requires IRB approval. This includes ensuring:

Proper handling of personally identifiable information (PII)
Justification for using identified vs. de-identified data
Security measures for protecting sensitive information

Data Security and Compliance

Researchers must implement robust security measures, including:

Encryption for storing and transferring sensitive data
Access controls to limit data exposure
Compliance with legal regulations such as the General Data Protection Regulation (GDPR) or Health Insurance Portability and Accountability Act (HIPAA), if applicable

For further clarification on IRB-related issues, any project with concerns should email humansubjects@poverty-action.org.

Data Validation

Data validation is the process of verifying that administrative data received from a provider has the expected characteristics before it enters a project’s analysis pipeline. Unlike data cleaning, which modifies records to correct errors, data validation checks whether the data as delivered matches what was agreed upon and what the analysis assumes and needs.

It is equally important to be clear about what data validation is not:

Not data cleaning: Validation identifies problems; cleaning modifies data to correct them. These are separate workflows.
Not a guarantee of data quality: Validation checks whether the received data matches the agreed-upon specifications, but it does not confirm that the underlying records accurately reflect reality.
Not statistical validity testing: Validation focuses on individual records and variables, not at the level of estimates or inferential claims.

Administrative data validation focuses on two dimensions:

Correctness: The reported value reflects the true observed value, accounting for allowable measurement error.
Consistency: Each record captures the same underlying construct across deliveries and across rows.

Data validation is different from data cleaning and from high-frequency checks used in survey data collection. The goal is to identify and triage errors in the data as received, not to alter the raw data.

When to validate

Validate as soon as data arrive

The first validation should occur immediately upon receiving a new data delivery. Running checks promptly allows the research team to identify problems while there is still time to contact the data provider for corrections. Automating validation scripts reduces processing time and ensures checks are applied consistently across multiple deliveries.

Validate each new delivery on the raw file

Validation scripts should run on the most recent file as received, using the variable names and formats the provider uses. This approach ensures that any errors identified are attributable to the source data, not to transformations the research team has applied. Validate before converting files to project-specific formats (e.g., .dta, .RData, or a relational database).

Keep validation and cleaning as separate workflows

Validation and cleaning serve different purposes and should be maintained as distinct scripts or processes. Validation documents what errors exist in the raw data. Cleaning modifies the data for analysis. Mixing the two makes it harder to trace the origin of any given value in the final analysis file and complicates future deliveries.

Common sources of error in administrative data

Source	Examples
Poorly designed forms or systems	No input constraints on data type or format; ambiguous questions that respondents interpret differently
Data entry errors	Misspellings, keystroke errors, misunderstanding of field definitions (e.g., annual income entered as monthly)
Misreporting or reporting bias	Incentives to over- or under-report; outcomes reported by program staff who have an interest in results
Data extraction errors	Incorrect queries, truncated exports, or format changes introduced during extraction
Data pre-preparation by the provider	Aggregations, recodes, or transformations applied before delivery that are not documented

Challenges When Using Administrative Data

While administrative data offers many advantages, researchers often face challenges such as:

Differential Coverage

Treatment and control groups may appear differently in administrative records, leading to bias. Examples include:

Identifiers Obtained After Enrollment: In a financial literacy program evaluation, treatment group participants may be more willing to provide bank account numbers for linking with administrative data, creating selection bias.
Program-Generated Data: If the intervention encourages healthcare visits, treatment group participants will appear more frequently in health administrative records, inflating apparent impact.

Reporting Bias

Some administrative data relies on self-reporting, which can introduce inaccuracies. Examples include:

Incentives for Misreporting: Agencies or individuals may have reasons to over- or under-report data.
Human Error in Data Entry: Manual data entry can introduce inconsistencies.

Cost of Administrative Data

Administrative datasets vary in cost, depending on:

The number of records requested
File-years needed
Data preparation time required by the provider

Conclusion

Administrative data is a valuable tool for randomized evaluations, offering cost-effective, accurate, and comprehensive insights. Researchers must navigate ethical, legal, and logistical challenges to ensure data quality and validity. By following standardized processes for information gathering, implementing appropriate data flow structures, executing well-designed data use agreements, and validating data at each delivery, research teams can produce reliable evidence from administrative sources.

References

Chetty, R., Friedman, J. N., Hilger, N., Saez, E., Schanzenbach, D. W., & Yagan, D. (2011). How does your kindergarten classroom affect your earnings? Evidence from Project STAR. The Quarterly Journal of Economics, 126(4), 1593–1660. https://doi.org/10.1093/qje/qjr041

Dobbie, W., Goldin, J., & Yang, C. S. (2018). The effects of pretrial detention on conviction, future crime, and employment: Evidence from randomly assigned judges. American Economic Review, 108(2), 201–240. https://doi.org/10.1257/aer.20161503 # codespell:ignore aer

Finkelstein, A., Taubman, S., Wright, B., Bernstein, M., Gruber, J., Newhouse, J. P., Allen, H., & Baicker, K. (2012). The Oregon Health Insurance Experiment: Evidence from the first year. The Quarterly Journal of Economics, 127(3), 1057–1106. https://doi.org/10.1093/qje/qjs020