Births and deaths are mentioned in my paper for the 2022 Joint Statistical Meetings in Washington, DC:

Application of Efficient Sampling with Prediction for Skewed Data:

https://www.researchgate.net/publication/362370770_Application_of_Efficient_Sampling_with_Prediction_for_Skewed_Data

In section 4.1, they are discussed in terms of the US Environmental Protection Agency (EPA) Toxic Release Inventory.

Further, mergers would be easy to describe as a sum of x-values mapped to a single y-value, for each data item in each case, though this would require more programming in practice, and here, more knowledge of the data than I had.

The following thread is on the kind of applications where this maintenence is used:

Do you know an example of Official Statistics produced from repeated establishment surveys? | ResearchGate:

https://www.researchgate.net/post/Do_you_know_an_example_of_Official_Statistics_produced_from_repeated_establishment_surveys

This is an excellent and very timely question. Tracking births, deaths, and mergers in establishment survey frames is one of the central organizational and methodological challenges in economic statistics. As you noted, it affects ratio analysis, totals, revisions, and ultimately the reproducibility of historical results.

Drawing from both international best practices (UNECE, Eurostat) and U.S. statistical experience (BLS, Census Bureau, EIA), I would highlight a framework that integrates longitudinal architecture, robust versioning, and practical operational tools.

1. Longitudinal Business Register as the Core Framework

Instead of treating survey frames as periodic snapshots, design them as a longitudinal business register (LBR)—a living system where every establishment has a lifecycle.

• Persistent Identifiers: Assign each unit a permanent, immutable ID (e.g., EIN, plant ID). This allows longitudinal tracking regardless of ownership or legal changes.
• Event Logging: Treat frame events as “vital statistics”: Births: New entities added, detected via tax files, incorporation records, or industry-specific permits. Deaths: Units marked inactive (not deleted), preserving history for potential reactivation. Mergers/Acquisitions: Map predecessor units to the surviving unit, with metadata on event type (absorption, consolidation, split). This ensures pre- and post-merger ratios and totals remain comparable.

This mirrors systems like the U.S. Longitudinal Business Database (LBD), which underpins Census Bureau products such as Business Dynamics Statistics.

2. Version Control and Revisions

Borrowing from data engineering, treat the survey frame like a version-controlled repository.

• Temporal Tables: Store every data item with valid_from / valid_to timestamps. When data are revised, insert new records instead of overwriting.
• Snapshots for Publication: Archive frozen versions of the frame and data at each release point (e.g., monthly or annually). This ensures historical reproducibility even as the live register evolves.
• Change Logs: Document the source and reason for updates (respondent correction, administrative adjustment, merger reconciliation).

This approach resolves the “archiving problem” you described, ensuring that past outputs can be exactly reconstructed.

3. Integration with Sampling and Estimation

Frame management must be tightly linked to survey methodology.

• Metadata Flags: Label each data point as observed, administrative, imputed, or modeled. This transparency helps distinguish genuine frame changes from estimation artifacts.
• Handling Births/Deaths in Samples: For births, consider temporary oversampling to stabilize estimates in early years. For deaths, adjust imputation models to avoid overestimation. For mergers, carry forward the combined weights of predecessors into the new entity.
• Benchmarking and Backcasting: Use universe data (e.g., Unemployment Insurance records, IRS filings) to realign sample-based estimates annually and apply “wedge-back” adjustments to prior periods.

4. Practical Tracking and Update Mechanisms

• Automated Detection: Use administrative data feeds (tax, social security, incorporation filings) to identify births and deaths in near real time. For sectoral data (like energy), integrate domain-specific sources (e.g., EIA forms, permits, utility registrations).
• Record Linkage: Apply deterministic (ID-based) and probabilistic (name/address-based) matching to ensure continuity across sources.
• Profiling Large Units: For multi-establishment enterprises, conduct manual “profiling” using corporate reports to ensure the frame reflects actual economic activity.

5. Quality Assurance and Documentation

• Develop quality indicators (coverage, timeliness, accuracy) for the register itself.
• Maintain a metadata repository documenting continuity rules, revision handling, and the impact of frame changes on published indicators.
• Foster inter-agency collaboration—for example, the BLS uses the Quarterly Census of Employment and Wages (QCEW) as a foundation for many programs, while the Census Bureau supplements with administrative tax and payroll data.

Examples of Survey Frames

• U.S. Bureau of Labor Statistics (BLS): Current Employment Statistics (CES): Uses UI-based frames with modeled birth/death adjustments. Job Openings and Labor Turnover Survey (JOLTS): Relies on QCEW as its frame.
• U.S. Census Bureau: Economic Census / Annual Survey of Manufactures (ASM): Frames refreshed using the Business Register and LBD. Business Dynamics Statistics (BDS): Focused explicitly on entry/exit dynamics.
• Energy Information Administration (EIA): Establishment-level energy surveys, where unit-level changes (e.g., a plant ceasing coal use) illustrate the “partial death” problem you described.

Adnan - Was the above AI generated? What parts would you agree with here? Do you have any experience in this area or reasoning or discussion you would like to share in your own words as to why some part of this is particularly important?

James R Knaub The answer I provided reflects my own understanding of survey frame management, but I did make use of AI tools to help me structure and sharpen the response, mainly to ensure it was concise, well-organized, and comprehensive.

To be clear: the underlying concepts I agree with and endorse. In particular, I strongly support:

• The primacy of a longitudinal design with immutable keys: This is the non-negotiable foundation for tracking births, deaths, and mergers accurately over time. It is the only way to solve the ratio comparison problem you described.
• The concept of "versioning" the entire frame and dataset: Your point about the irreproducibility of historical results is crucial. Treating the data repository with temporal validity stamps is the definitive solution, moving beyond simple "overwrite" archiving.
• The tight integration of frame management with estimation methodology: The frame cannot be a separate administrative exercise. It must directly inform sampling weights, imputation routines, and benchmarking processes, especially for the volatile segments (births, deaths) you work with.

These aren't theoretical concepts; they are the operational pillars of modern statistical agencies. My response was intended to map these established principles to your specific concern about energy data and economic statistics.

I appreciate you asking for clarification.

Above I mentioned births, deaths, and mergers with respect to ratio model predictions (which are estimates for a random variable using regression, not forecasts). These are by item though the sample is by unit which may report for multiple items/attributes. In section 4.1 of https://www.researchgate.net/publication/362370770_Application_of_Efficient_Sampling_with_Prediction_for_Skewed_Data , I did not have the knowledge of the database and subject matter or resources to look at this thoroughly, as it was just in support of showing an example of using methodology I developed for Official Energy Statistics in other settings, as it can be very useful especially when there is an occasional census and frequent sampling of the same data items. There it was from the US EPA Toxic Release Inventory. I noted the following thread where this situation would occur:

https://www.researchgate.net/post/Do_you_know_an_example_of_Official_Statistics_produced_from_repeated_establishment_surveys  

Does anyone have an example to share on that thread or this one regarding such use of ratio models, multiple regression, or quadratic linear regression, etc?  On this thread, that would include using a ratio, etc, when the frame changes between predictor data collection and the matching sample variable of interest.