Expert Witness Practice Area

Platform Data & Measurement Expert Witness

Platform data disputes turn on whether the numbers a platform reports measure what a party claims they measure. Jonathan Hersh, PhD is an economist who evaluates platform metrics, telemetry, logs, and attribution methodology in litigation — assessing sampling, definitional changes, and inference from proxies, and explaining what the data can and cannot establish.

Request a Consult All Practice Areas

What these cases turn on

What does this metric actually count, and has its definition changed over the period at issue?
Is the underlying data a census or a sample, and how was it weighted?
How much of the reported figure is measured versus modeled or imputed?
Does the attribution methodology support the causal claim being made from it?
Are the gaps and anomalies in the log data material to the conclusion?

Analysis

Platform-reported numbers get treated in litigation as if they were direct observations. They rarely are. Most are the output of a measurement pipeline with a definition, a sampling design, an attribution model, and a revision history — and disputes frequently turn on a property of that pipeline rather than on the conduct itself.

The recurring problems

A metric's definition changed mid-period, so a before-and-after comparison is measuring two different things.
A figure presented as counted is partly modeled, and the model's assumptions are doing the work.
Attribution logic assigns outcomes to a channel by rule, then the rule's output is cited as causal evidence.
Sampling coverage differs systematically across the groups being compared.
Log gaps are treated as zeros rather than as missing data.

Why measurement is my area

Much of my published research is measurement under difficult conditions: estimating economic well-being from satellite imagery where survey data does not exist, and detecting war destruction from space using machine learning. That work is entirely about how far you can trust an inferred quantity, how to validate a proxy against ground truth, and how to characterize uncertainty honestly. Platform metrics present the same problem with better data and higher stakes.

Methods applied

Reconstruction of metric definitions and audit of definitional changes over time
Assessment of sampling design, weighting, and coverage in platform-reported data
Separation of measured quantities from modeled or imputed ones
Validation of proxy measures against ground truth where independent data exists
Evaluation of attribution models and the causal claims drawn from them

Peer-reviewed research grounding this work

Published, citable research in this area — the verifiable basis for the analysis above.

World Bank Economic Review · 2022

Poverty from Space: Using High Resolution Satellite Imagery for Estimating Economic Well-being

Uses high-resolution imagery to estimate consumption and poverty with robust out-of-sample performance.

Proceedings of the National Academy of Sciences · 2021

Monitoring war destruction from space using machine learning

Develops machine-learning methods to detect conflict-related infrastructure destruction at scale from satellite imagery.

Information Technology for Development · 2021

Open data for algorithms: mapping poverty in Belize using open satellite derived features and machine learning

Shows that open satellite feature sets can improve poverty prediction performance and reduce cost barriers for policy analytics.

Management Science · 2022

How APIs Create Growth by Inverting the Firm

Documents growth gains from API adoption and platform openness, while quantifying associated governance and security tradeoffs.

Frequently asked questions

What kinds of measurement disputes do you work on?

Matters where a platform-reported figure is contested — engagement or usage metrics, advertising delivery and attribution, telemetry and server logs, and any dispute where a party's damages or liability theory rests on a number the platform itself produced.

Can platform metrics be used to prove causation?

Sometimes, but not on their own. Attribution models assign outcomes to channels by rule, and the rule's output is a bookkeeping result rather than a causal estimate. Establishing causation requires a design that identifies it — an experiment, a natural experiment, or a defensible counterfactual.

Do you work with raw log data?

Yes, and I prefer it. Platform-reported summaries have already had definitional and attribution decisions baked into them; raw logs let those decisions be examined rather than assumed. Documentation of business rules matters as much as the data itself.