Inference From Incomplete Digital Datasets in Competition-Law Enforcement

Introduction

Inference from incomplete digital datasets refers to the use of statistical, algorithmic, or economic techniques to draw conclusions about market conduct when the available digital evidence is partial, fragmented, biased, noisy, or generated under changing technological conditions.

This issue is increasingly important in competition-law enforcement because modern markets generate enormous quantities of data—transaction records, search histories, platform logs, API calls, prices, rankings, advertising auctions, clickstreams, algorithmic decisions, and communications. Yet enforcement authorities rarely obtain a complete dataset. Important observations may be unavailable because of data deletion, privacy restrictions, interoperability barriers, technical incompatibility, encrypted communications, changing algorithms, missing historical records, or the fact that relevant activity occurred outside the investigated platform.

The central legal problem is therefore:

How far can a competition authority infer anti-competitive conduct from incomplete digital evidence without converting statistical probability into proof of infringement?

The answer depends on the type of infringement, the applicable evidentiary standard, the quality of the inference, corroborating evidence, and whether the undertaking has a meaningful opportunity to challenge the methodology.

1. Meaning of Inference From Incomplete Digital Datasets

A digital dataset may be incomplete in several different ways:

A. Missing observations

Some transactions or users are absent.

Example:

• a platform supplies only 70% of historical transaction records;
• deleted accounts cannot be reconstructed;
• certain geographic regions are missing.

B. Missing variables

The dataset contains prices but not costs, or transactions but not consumer characteristics.

This is particularly important in:

• predatory-pricing cases;
• discrimination cases;
• personalized pricing;
• merger analysis;
• exclusionary conduct.

C. Selection bias

The available observations may systematically differ from the missing observations.

For example, a platform may provide data concerning active users but not users who abandoned the platform.

D. Censoring

The authority observes only activity above or below a particular threshold.

For example, an advertising platform may reveal winning bids but not losing bids.

E. Temporal incompleteness

Historical algorithmic decisions may be unavailable even though the current algorithm can be inspected.

This creates a serious problem because the algorithm operating during the alleged infringement may differ from the algorithm operating during the investigation.

F. Cross-platform incompleteness

The authority may possess data from one platform but not competing platforms.

A finding of market power based solely on the investigated platform's internal data may therefore be misleading.

2. Why Digital Markets Create Special Evidentiary Problems

Traditional competition cases often rely upon documents, contracts, emails, testimony and observable prices.

Digital markets generate a different evidentiary environment.

A platform's conduct may instead be reflected in:

• ranking changes;
• recommendation outputs;
• API requests;
• software releases;
• model parameters;
• bidding histories;
• A/B tests;
• user-level transactions;
• algorithmic logs;
• internal dashboards;
• machine-learning training datasets.

The enforcement authority may therefore have to reconstruct conduct indirectly.

For example:

Incomplete observations → statistical inference → economic model → inference concerning conduct → legal conclusion

The danger is that each additional inferential step creates another opportunity for error.

3. The Fundamental Legal Principle

The existence of a statistically plausible explanation does not automatically establish an infringement.

Competition authorities must distinguish between:

Evidence

Facts directly established by reliable evidence.

Inference

A conclusion reasonably drawn from established facts.

Hypothesis

A possible explanation that remains capable of alternative explanations.

This distinction becomes critical where:

Incomplete data + sophisticated modelling ≠ conclusive proof of anticompetitive conduct.

An authority should therefore demonstrate:

1. what data actually exist;
2. what data are missing;
3. why the missing information does not materially undermine the inference;
4. what assumptions have been made;
5. whether alternative explanations were tested;
6. whether the results remain robust under different assumptions.

4. Role of Econometric Inference

Incomplete digital datasets can nevertheless be extremely valuable.

Authorities may employ:

• regression analysis;
• difference-in-differences;
• event studies;
• instrumental variables;
• propensity-score methods;
• synthetic controls;
• counterfactual simulations;
• cluster analysis;
• structural modelling;
• natural experiments;
• machine-learning classification.

For example:

If prices systematically increase after a dominant platform restricts rival access, an authority may investigate whether the relationship survives controls for demand, costs, seasonality and other market factors.

The resulting inference can become important evidence.

But correlation alone does not establish causation.

5. Counterfactual Reconstruction

One of the most important uses of incomplete data is counterfactual reconstruction.

The authority asks:

What would the market have looked like if the allegedly anti-competitive conduct had not occurred?

For example:

Observed market

Dominant platform imposes restrictive API terms → rival traffic falls → prices rise.

Counterfactual

Without the restriction → rival traffic would allegedly have remained higher → prices would allegedly have remained lower.

The difficulty is that the counterfactual itself cannot normally be directly observed.

Consequently, incomplete datasets make the counterfactual partly inferential.

6. Missing Data and Burden of Proof

A crucial issue is determining who bears the consequences of missing evidence.

Ordinarily, the enforcement authority cannot simply rely upon an evidentiary gap and assume that the missing information would support its case.

However, the evidentiary position may change where:

• the undertaking exclusively controls the relevant data;
• the undertaking has failed to preserve evidence;
• the undertaking provides incomplete information;
• the authority has used compulsory investigative powers;
• the undertaking's own conduct created the information asymmetry.

Thus, digital competition enforcement increasingly involves a tension between:

Authority's evidentiary burden

and

undertaking's control over the evidence necessary to test the authority's theory.

7. Six Important Case Laws

1. Intel Corp. v. Commission

Court: Court of Justice of the European Union
Principle: Economic evidence must be properly assessed in exclusionary-abuse cases.

The Intel litigation is particularly important for digital-data inference because it demonstrates the danger of relying excessively on presumptions or incomplete economic analysis when determining whether conduct is capable of foreclosing equally efficient competitors.

The CJEU required consideration of the circumstances of the conduct and, where appropriate, economic evidence such as:

• market coverage;
• duration;
• magnitude of the conduct;
• competitors' position;
• potential foreclosure effects.

Relevance to incomplete datasets

Where an authority has only partial pricing or customer data, it should avoid treating one observed indicator as conclusive.

Lesson: Incomplete economic datasets require contextual assessment rather than mechanical inference.

2. United Brands v. Commission

Case: United Brands Company and United Brands Continentaal BV v Commission

The case remains foundational for determining market power and abusive conduct.

The Court examined numerous factors rather than relying upon a single market indicator.

Relevance

In digital markets, incomplete datasets may prevent precise calculation of:

• market shares;
• price differences;
• customer switching;
• competitive constraints.

United Brands demonstrates the importance of considering multiple indicators together.

For example:

market share + barriers to entry + customer dependence + competitive constraints + economic evidence

may provide a stronger inference than any single incomplete dataset.

Lesson

Where one digital metric is incomplete, enforcement should triangulate it with other evidence.

3. Hoffmann-La Roche v. Commission

Case: Hoffmann-La Roche & Co. AG v Commission

The Court established important principles concerning dominance and loyalty-inducing arrangements.

The case is significant because competition-law conclusions may be derived from the overall structure and economic effect of conduct, rather than from a single direct piece of evidence.

Digital relevance

Suppose a dominant platform's contractual records are incomplete.

The authority may nevertheless examine:

• exclusivity structures;
• rebates;
• customer dependence;
• duration;
• switching behaviour;
• internal platform data.

Incomplete digital data may therefore contribute to a broader evidentiary picture.

Lesson

Digital enforcement can legitimately employ circumstantial evidence, but the cumulative evidence must establish the relevant legal proposition.

4. Microsoft Corp. v. Commission

Case: Microsoft Corp. v Commission

The EU Microsoft litigation is especially relevant to technology markets.

The case involved interoperability information, software architecture and technological restrictions.

Relevance to incomplete datasets

Technology cases frequently involve information that is:

• technically complex;
• controlled by the investigated undertaking;
• difficult for authorities to independently reproduce.

An authority may therefore need to reconstruct competitive effects using indirect technical and economic evidence.

The Microsoft litigation demonstrates the importance of analysing:

• technological interoperability;
• network effects;
• market structure;
• barriers to entry;
• competitor dependence.

Digital-data lesson

An enforcement authority should not assume that the absence of a complete technical dataset means that exclusionary effects cannot be established. But the authority must explain the evidentiary chain connecting the available technical evidence to the alleged competitive harm.

5. Google Shopping

Case: Google and Alphabet v Commission / Google Shopping litigation

This case is highly relevant to modern digital enforcement because the alleged conduct concerned:

• search results;
• ranking;
• traffic;
• visibility;
• click data;
• competing comparison-shopping services.

The General Court considered evidence concerning the evolution of traffic to competing services and the impact of Google's practices.

Relevance to incomplete digital datasets

Digital ranking systems generate enormous quantities of data, but authorities may still lack:

• complete user histories;
• counterfactual rankings;
• all competing platforms' data;
• historical algorithmic versions.

Therefore, the authority may have to combine:

traffic data + ranking evidence + algorithmic information + market structure + qualitative evidence.

Lesson

Digital competition cases demonstrate how incomplete platform data can still support a broader theory of harm, provided the authority establishes a sufficiently reliable evidentiary connection between the conduct and competitive effects.

6. Qualcomm

Case: Qualcomm Inc. v Commission

The Qualcomm litigation concerned exclusionary payments and their potential effects on competition.

The General Court emphasised the importance of carefully examining economic evidence and causation when assessing whether the conduct was capable of producing exclusionary effects.

Relevance

A digital investigation may reveal that a dominant undertaking's conduct coincides with:

• rival exit;
• declining rival sales;
• reduced access;
• changes in prices.

But coincidence does not necessarily establish causation.

Lesson

Authorities should test whether the observed competitive harm can plausibly be explained by factors other than the investigated conduct.

7. CK Telecoms UK Investments v European Commission

Case: CK Telecoms UK Investments Ltd v Commission

This merger case is particularly important for evidence-based assessment of competitive harm.

The General Court scrutinised the Commission's approach to assessing whether a transaction would significantly impede effective competition.

Relevance to incomplete datasets

Merger authorities frequently operate with incomplete information because the transaction has not yet occurred.

They must therefore construct predictions concerning:

• future competition;
• innovation;
• pricing;
• product quality;
• strategic interaction.

This is inherently inferential.

Digital-market relevance

In technology markets, the problem becomes more significant because:

• innovation cycles are short;
• products are differentiated;
• user data are difficult to value;
• future technological developments are uncertain.

Lesson

Predictive competition analysis must be supported by sufficiently reliable evidence and cannot rest on speculative modelling alone.

8. Inference in Algorithmic Collusion Cases

Incomplete datasets create a particularly difficult problem in suspected algorithmic collusion.

Suppose authorities observe:

• highly parallel prices;
• synchronized algorithmic changes;
• similar price responses;
• stable margins.

But they do not possess:

• complete algorithmic code;
• all communications;
• complete transaction histories;
• every competitor's pricing data.

The authority must distinguish:

Conscious coordination

Human decision-makers communicate or intentionally coordinate.

Tacit coordination

Firms independently recognise mutual interdependence.

Algorithmic parallelism

Algorithms independently respond to similar market signals.

Algorithmic facilitation

Algorithms make coordination easier or more stable.

These categories have very different legal implications.

9. Inference From Platform Ranking Data

Suppose a dominant search platform changes its ranking algorithm.

Available data show:

These numbers may suggest foreclosure.

But the authority must investigate whether the decline resulted from:

• consumer preferences;
• improved rival products;
• seasonal changes;
• SEO changes;
• independent algorithmic optimisation;
• market entry;
• changes in advertising;
• technical problems.

Thus:

Observed correlation → possible theory of harm

but not automatically:

Observed correlation → infringement.

10. Missing Data Caused by the Undertaking

A particularly difficult situation arises where the undertaking itself controls the missing evidence.

For example:

• historical logs were deleted;
• algorithm versions were overwritten;
• internal datasets were not preserved;
• APIs do not retain historical queries;
• relevant communications occurred through ephemeral systems.

This can create an evidentiary asymmetry.

Competition authorities may therefore place considerable importance on:

• document-preservation duties;
• forensic imaging;
• audit trails;
• disclosure obligations;
• data-production orders;
• reproducibility requirements.

Nevertheless, the absence of evidence should not automatically be transformed into affirmative proof of infringement.

11. Privacy and Data-Protection Constraints

Digital enforcement can also be affected by privacy rules.

An authority may technically possess access to a dataset but be unable to use every variable because of:

• data-protection requirements;
• anonymisation;
• purpose limitation;
• confidentiality;
• trade secrets.

This may force authorities to work with:

• aggregated data;
• pseudonymised information;
• sampling;
• differential-privacy techniques;
• synthetic datasets.

The resulting inference must account for information lost through aggregation.

12. Data Imputation and Competition Enforcement

Authorities may use imputation to estimate missing observations.

For example:

15% of transaction records are missing.

An economist might estimate those transactions using observed characteristics.

But the authority should disclose:

1. the imputation method;
2. assumptions;
3. confidence intervals;
4. sensitivity analysis;
5. alternative specifications.

If the infringement finding changes dramatically when the missing observations are estimated differently, the conclusion may be insufficiently robust.

13. Machine Learning and Black-Box Inference

Machine-learning systems create an additional problem.

A model may identify a strong statistical association between:

• platform behaviour;
• user characteristics;
• competitor outcomes.

But a competition authority must still explain why the model's result supports the legal theory.

A black-box prediction such as:

"The model predicts exclusion with 94% probability"

does not by itself establish:

"The undertaking intentionally or unlawfully excluded competitors."

Therefore, explainability and evidentiary traceability become critical.

14. Confidence Intervals and Evidentiary Reliability

Competition authorities should distinguish between:

Point estimate

and

uncertainty surrounding the estimate.

For example:

Estimated foreclosure effect = 12%.

If the confidence interval is:

12% ± 10%,

the authority cannot treat the result as equivalent to a precisely established 12% effect.

Digital enforcement should therefore consider:

• statistical significance;
• confidence intervals;
• model specification;
• sampling error;
• measurement error;
• missing-data assumptions.

15. Robustness Testing

A strong enforcement inference should survive reasonable alternative assumptions.

For example:

Model A

Effect = 15%

Model B

Effect = 13%

Model C

Effect = 14%

The conclusion is relatively robust.

But:

Model A

Effect = 20%

Model B

Effect = 3%

Model C

Effect = –2%

The evidence is considerably weaker.

Thus, authorities should conduct sensitivity analysis rather than relying on a single preferred model.

16. Circumstantial Evidence

Incomplete digital datasets often make circumstantial evidence particularly important.

A competition authority may combine:

• suspicious communications;
• algorithmic changes;
• pricing patterns;
• internal documents;
• market outcomes;
• customer complaints;
• competitor responses.

Each item may be insufficient individually.

Together, however, they may establish a coherent evidentiary narrative.

The legal strength comes from convergence of independent evidence, not simply from the sophistication of the statistical model.

17. Risks of False Positives

Inference from incomplete datasets can generate false positives.

A false positive occurs where lawful competitive conduct is incorrectly classified as anticompetitive.

Examples include:

• similar prices caused by similar costs;
• parallel algorithmic responses caused by common demand shocks;
• reduced rival traffic caused by superior product quality;
• high switching costs caused by genuine consumer preferences;
• algorithmic price increases caused by legitimate scarcity.

This is particularly dangerous in innovation-intensive markets because aggressive competition can itself generate unusual data patterns.

18. Risks of False Negatives

The opposite problem is also serious.

A weak-data requirement can allow sophisticated digital misconduct to escape enforcement.

For example, a dominant platform may structure its systems so that:

• relevant data are fragmented;
• algorithms change frequently;
• historical logs disappear;
• critical decisions are automated;
• third-party data cannot be obtained.

If enforcement requires perfect data, the most technologically sophisticated forms of anticompetitive conduct could become practically unenforceable.

Therefore, competition law must balance:

evidentiary reliability

against

practical enforceability.

19. Standard of Proof

The precise standard varies by jurisdiction and type of proceeding.

Nevertheless, a useful general principle is:

The greater the inferential distance between the available evidence and the alleged infringement, the greater the need for corroboration, methodological transparency and robustness.

An authority should therefore establish:

Step 1 — Data provenance

Where did the data originate?

Step 2 — Completeness

What proportion is missing?

Step 3 — Reliability

Can the data be independently verified?

Step 4 — Methodology

How were missing observations treated?

Step 5 — Causation

Why does the evidence indicate that the undertaking's conduct caused the competitive effect?

Step 6 — Alternatives

Were lawful alternative explanations considered?

Step 7 — Robustness

Does the conclusion survive reasonable methodological variations?

Step 8 — Legal characterisation

Do the established facts satisfy the elements of the relevant competition-law infringement?

20. Relevance to Merger Enforcement

Incomplete datasets are particularly significant in merger cases.

A merger authority may not know:

• future innovation;
• future entry;
• future consumer preferences;
• future technological developments.

It must therefore forecast.

This is particularly challenging with:

• AI firms;
• cloud platforms;
• foundation models;
• semiconductor ecosystems;
• data-driven advertising;
• digital marketplaces.

Authorities should distinguish between:

evidence-based prediction

and

speculative prediction.

21. Relevance to Dominance Cases

For Article 102 TFEU or equivalent national provisions, incomplete digital data can affect:

• market definition;
• dominance;
• foreclosure;
• exploitative abuse;
• discrimination;
• refusal to supply;
• tying;
• self-preferencing.

For example, a platform's market share may be known while its actual competitive constraints are not.

Authorities should therefore combine quantitative market evidence with:

• entry barriers;
• network effects;
• switching costs;
• multi-homing;
• access to data;
• ecosystem effects.

22. Relevance to India

In India, the issue is especially relevant to enforcement under the Competition Act, 2002, particularly in digital-platform investigations.

The Competition Commission of India may encounter:

• platform-generated datasets;
• algorithmic pricing;
• ranking systems;
• digital advertising records;
• API data;
• app-store information;
• consumer-level transaction data.

Incomplete evidence can affect investigations into:

• abuse of dominance;
• discriminatory conditions;
• denial of market access;
• leveraging;
• tying;
• self-preferencing;
• cartel-like algorithmic coordination.

The evidentiary challenge is to use sophisticated digital and economic evidence while preserving procedural fairness and avoiding conclusions based solely on opaque algorithmic predictions.

23. Relevance to UK Competition Law

Under the UK's competition regime, the Competition and Markets Authority similarly faces challenges involving:

• platform datasets;
• algorithmic evidence;
• digital markets;
• merger forecasting;
• online pricing;
• ranking systems;
• data access.

The UK's increasingly data-intensive enforcement environment makes methodological transparency particularly important.

An authority should be able to explain not merely what its model predicts, but why the prediction establishes the statutory elements of the alleged infringement.

24. Evidentiary Framework for Incomplete Digital Data

A useful enforcement framework is:

Incomplete Dataset

↓

Identify Missingness

↓

Determine Why Data Are Missing

↓

Test Selection Bias

↓

Construct Appropriate Counterfactual

↓

Estimate Competitive Effect

↓

Run Sensitivity Tests

↓

Compare Alternative Explanations

↓

Corroborate With Documentary/Technical Evidence

↓

Apply Legal Test

↓

Determine Infringement

This prevents the statistical model from becoming a substitute for legal reasoning.

25. Key Doctrinal Lessons From the Case Law

Conclusion

Inference from incomplete digital datasets is neither inherently impermissible nor inherently sufficient for competition-law enforcement.

The central question is the reliability of the inferential chain.

Digital enforcement inevitably operates with imperfect information. Requiring complete datasets in every case would make many technologically sophisticated infringements practically impossible to investigate. At the same time, allowing authorities to convert incomplete data and opaque algorithms into definitive conclusions would create substantial risks of false positives and procedural unfairness.

The appropriate approach is therefore triangulated, transparent and robustness-based enforcement:

Incomplete data + reliable methodology + corroborating evidence + tested counterfactual + consideration of alternative explanations = potentially persuasive enforcement evidence.

By contrast:

Incomplete data + untested assumptions + opaque model + speculative causal inference = weak basis for an infringement finding.

The emerging principle for digital competition law is consequently that data incompleteness affects the weight and reliability of evidence, rather than automatically determining its admissibility or value. The more important the legal conclusion, the more carefully the authority should expose the assumptions, uncertainties and inferential steps underlying it.