Earnings Management Then and Now: Revisiting Jones (1991)

Website Data, Replication, and Research Design

Introduction

Introduction

Who am I?


(c)RSM
Caspar David Peter
  • 🏢     Rotterdam School of Management
  • 👷     Associate Professor
  • 🏫     Doctorate (WHU)
  • 🏫     Dipl.-Ök. (RUB)
  • 👶     Bochum, Germany
  • 🏠     Rotterdam, Netherlands
  • 🍞🧈 Economic consequences of transparency

Introduction

Roadmap & Learning Goals

What are we up to today?

Part 1: Three papers
  • Method: Haans and Mertens (2024),
  • Application: Boulland et al. (2025),
  • Opportunity: Jones (1991)
Part 2: Research in Action: Replication & Brainstorming
  • Goal: Learn methodology, think critically, brainstorm novel research

Learning Goals

  • Understand & Learn about the usefulness of archival website data
  • Evaluate new disclosure measures critically
  • Practice research design thinking
  • See how new data can revisit classic questions

PART 1: THREE PAPERS

Haans and Mertens (2024)

The Internet Never Forgets

Introducing the Wayback Machine

Wayback Machine screenshot
  • Archive of billions of web pages since 1996
  • Accessible via archive.org or waybackpy
  • Example use cases in accounting research
    • Corporate disclosures
    • Investor relations
    • CSR reporting
    • Executive communications

Key takeaways

  • Framework to systematically scrape organizational websites
  • Scraping websites over time to construct longitudinal data (tutorial)
  • Open-source code and database: CompuCrawl

Haans and Mertens (2024)

The Internet Never Forgets

Sample Selection & Coverage over Time

Critical Findings

  • ~20% of websites are not archived in any given year
  • Survivorship bias (Compustat backfills successful firms)
  • Website address changes are not tracked historically
  • Coverage improves over time (poor in 1990s, better 2010+)

Takeaway #1

Website data offers unprecedented longitudinal access, BUT requires careful attention to coverage gaps and data quality

Boulland et al. (2025)

Company Websites: A New Measure of Disclosure

Research Question & Motivation

  • Can company websites serve as a novel measure of corporate disclosure?
  • Why websites?
    • Increasingly important communication channel
    • Reflects voluntary disclosure choices
    • Accessible via Wayback Machine for historical analysis

Key Contributions

  • Proposes using company websites as a novel data source
  • Highlights the importance of online presence for corporate transparency
  • Validation of new measure against traditional disclosure metrics
  • Extends the scope to U.S. and E.U. private firms

Boulland et al. (2025)

Company Websites: A New Measure of Disclosure

Measurement

  • Primary measure: Website SIZE (bytes, page counts)
  • Content categories: Products, Investor Relations, Geography, HR
  • Data: U.S. public firms + U.S. private firms + French firms
  • Method: Wayback Machine + text analysis

Validation & Applications

Validation of Website Disclosure Measure: Validation


Applications: Private Equity Nonfinancial Disclosures in France

Devil’s Advocate

Takeaway #2

Website-based measures complement traditional disclosure metrics BUT validity questions remain - interpret with caution

Takeaways so far

Takeaways so far

Summary

  • Websites are a novel and valuable data source for accounting research
  • Wayback Machine allows for longitudinal website data collection
  • Website-based disclosure measures can be validated against traditional metrics
  • Applications span public and private firms across different countries
  • Open source tools and databases facilitate replication and extension
  • Next: Revisiting a classic question in accounting research

Back to the future

Jones (1991)

Research Question & Main Finding & Contributions

Research Question

Do firms that would benefit from import relief attempt to decrease earnings through earnings management during ITC investigations?

Main Finding

Managers of domestic producers significantly decrease reported earnings during import relief investigations

  • Discretionary accruals are income-decreasing in investigation year (year 0)
  • No significant effect in surrounding years (-1, +1)

Contributions

Methodological Innovation
  • Develops firm-specific time-series models to estimate “normal” accruals
  • Foundation for the famous “Jones Model”
Theoretical Contributions
  • Unique incentive structure: First study examining income-decreasing earnings management
  • Demonstrates accounting numbers have economic consequences in regulatory settings (wealth transfers)
    • More on wealth transfers later ⇢ Accounting and Wealth Tranfer session by Patricia Breuer in February

Jones (1991)

Institutional Setting: Why This Is A Clean Setup

The ITC Import Relief Process

  • U.S. International Trade Commission investigates injury to domestic producers
  • Profitability is an explicit factor in injury determination
  • Tariff increases/quota reductions are granted if the industry is deemed injured

Perfect Alignment of Incentives

  • Managers want to appear injured ⇢ decrease earnings
  • Debtholders tolerate lower earnings (expect relief benefits)
  • No sophisticated monitoring by regulator

Critical Feature

The ITC does not verify any information in the audited financial statements or 10-Ks, nor do they make any adjustments to these data… The ITC does not attempt to adjust the financial data for accounting procedures used or for accrual decisions made by the firms’ managers

Jones (1991)

Institutional Setting: Sample Industries

Five Industries, Six Investigations (1980-1985)

Sample Characteristics

  • Firm names from ITC reports
  • 23 firms with complete data on Compustat
  • Broad product lines (material to consolidated financials)

Jones (1991)

Hypothesis Development

Main Hypothesis

Managers of domestic producers make accounting choices that reduce reported earnings during ITC investigation periods compared to non-investigation periods

Addressing Potential Concerns

Conflicting incentives?
  • Usually managers increase earnings (debt covenants, compensation)
  • But import relief benefit outweighs these concerns
  • Debtholders are willing to waive covenants; future earnings are expected to improve
Free-rider problem?
  • All industry firms benefit, but only some may manage earnings
  • Solution: Supplemental test using only petitioner firms (that bear investigation costs)

Jones (1991)

Empirical Results

The Jones Model

\[ \frac{\text{Total accruals}_{it}}{AT_{i,t-1}} = \frac{1}{AT_{i,t-1}} + \beta_1 \frac{\Delta REV_{it}}{AT_{i,t-1}} + \beta_2 \frac{PPE_{it}}{AT_{i,t-1}} + \epsilon_{it}\]

Main Results

Discretionary Accruals by Year
  • Year -1: Not significantly different from zero
  • Year 0 (investigation): Significantly income-decreasing
  • Year +1: Not significantly different from zero
Robustness
  • Portfolio tests control for cross-sectional correlation
  • Four-industry t-statistics range: -3.635 to -5.035
  • Petitioner subsample shows consistent results

Key Implication

Managers systematically manipulate accruals downward specifically during investigation periods, providing strong support for the earnings management hypothesis

Jones (1991)

Why This Design Is Pretty Darn Good

Clear Incentive Alignment

  • Regulatory rule explicitly uses accounting profitability

  • All contracting parties benefit from appearing injured

  • Regulator doesn’t adjust for earnings management

Powerful Research Design

  • Time-series control for firm-specific “normal” accruals

  • Clear prediction: decrease in investigation year only

  • Multiple industries, multiple years

  • Addresses free-rider problem with petitioner subsample

Key Implication

This setting provides a specific motive for earnings management that is not typically present in other contracting situations where all parties have incentives to monitor and adjust for manipulation

PART 2: RESEARCH IN ACTION

PART 2: RESEARCH IN ACTION

From Papers to Practice

What Well Do Today

  • Move from theory to execution
  • See my Jones (1991) interpretation
  • Explore rough website analysis
  • All code/data on GitHub (live demos)
  • Your job: Critique, spot problems, suggest improvements

Materials Available

  • 🔗 GitHub repo
    • Jones replication page
    • Website analysis page
    • R & Python scripts for data and analysis
You’ve had access—we’ll walk through key decisions

Disclaimer

This is a re-interpretation of Jones (1991), not a strict replication. Analyses are preliminary and unreviewed; treat all findings as exploratory. Results may change as data and methods are refined.

PHASE 1: THE REPLICATION JOURNEY

PHASE 1: THE REPLICATION JOURNEY

What I Set Out to Replicate

Jones (1991) Main Finding

  • Income-decreasing DA in Year 0
  • Clear institutional setting
  • Time-series firm-specific estimation

My Approach

  • Cross-sectional industry-specific (pooled)
  • Modern sample (1980-2020)
  • Key decision: Why cross-sectional?

Here is a comparison of key dimensions between the original Jones (1991) study and my replication approach:

Dimension Jones Original My Replication
Estimation approach Time-series (firm-specific) Cross-sectional (industry)
Data requirements Long firm history Industry observations
What it controls for Firm-specific effects Industry dynamics
Statistical power Lower (fewer obs/firm) Higher (pooled data)
Parameter interpretation Firm-specific “normal” Industry-specific “normal”

PHASE 1: THE REPLICATION JOURNEY

The Timing Problem I Had to Solve

The Problem

  • Fiscal years ≠ calendar years ≠ investigation dates
  • Jones found effects in “Year 0”
  • First-try: I found effects in Years -2 to -1… Why?

Example Timeline

Investigation announced: June 1985 (calendar)
Firm fiscal year ends: Dec 1984
Firm fiscal year ends: Dec 1985
Which "year" is Year 0?

My Solution

# See phd-lecture-analyses.R script ...
time = case_when(
    # Fiscal year ends before investigation decision - no adjustment needed
    fyear == event_year & month(datadate) < year_zero_month ~ time,
    # Fiscal year ends after investigation decision - shift time by +1
    TRUE ~ time + 1
  )

Why This Matters

  • Firms may start managing earnings before formal announcement
  • Anticipation of the investigation
  • Fiscal year overlap creates measurement issues

Key Insight

The timing “discrepancy” isn’t a bug—it’s a feature revealing that firms may manage earnings in anticipation of investigations

PHASE 1: THE REPLICATION JOURNEY

Finding and Linking the Investigations

The Problem

  • Several types/subjects of investigations - which to choose?
  • Jones uses domestic producers of broad product lines

Investigations

USITC Top 5 Subjects Treatment Cohorts Histogram

My Solution

  • I focused on “Safeguard” investigations Why?
  • Jones picks only “Escape Clause” cases - subset of Safeguards
  • “[…] only investigations pertaining to a broad product line […]”

Why This Matters

  • Investigation type affects incentives & sample
  • If the industry is small (part of consolidated sales)
    • less incentive to manage

Key Insight

Institutional knowledge is key here - not yet fully utilized! Knowledge of investigation types is crucial for sample selection and interpretation of results. There is very much room for improvement here.

PHASE 1: THE REPLICATION JOURNEY

Live Demo: Replication Results

What we’ll look at:

  1. Sample construction & coverage
  2. Key results table (compare to Jones Table 5)
  3. Discretionary accruals
    • First-difference approach
    • Event-study (DiD-style plot)
  4. What we could add to align with contemporary EM literature


⚒ Jones-esque analysis page

PHASE 1: THE REPLICATION JOURNEY

Replication Takeaways

What Worked

  • Jones findings are robust to modern methods ✔
  • Cross-sectional approach provides adequate power ✔
  • Timing issue resolved with fiscal year alignment ✔
  • Effects are economically and statistically significant ✔

Open Questions

⚠️ Would time-series give identical results?

  • Enough observations per firm to test ✔
  • Need to identify firms from reports ⏳ (outsource to 🤖?)

⚠️ The Earnings Management Literature has advanced

  • E.g. Dechow et al. (2010) or Breuer and Schütt (2023)
  • Issues with two-step estimations, see e.g. Chen et al. (2018) or Chen et al. (2023)

🤔 Why do effects appear in Year -1 to Year 0?

  • Accrual reversal in Year +1? Can we test that?

The Bigger Question

Replication is validating, but not publishable. What NEW insight can we add 35 years later?

PHASE 2: WEBSITE DATA EXPLORATION

PHASE 2: WEBSITE DATA EXPLORATION

Adding Website Data: The Hypothesis

In 1991…

  • One financial statement for all audiences
  • Limited tools for strategic communication
  • No firm websites existed
  • ITC reviews audited financials only

Today…

  • Multi-channel disclosure strategy
  • Rich web presence during investigations
  • Different audiences (investors, regulators, employees, customers)
  • Opportunity for coordinated manipulation

Research Question

Do firms strategically manage website disclosures during import relief investigations?

Possible Mechanisms

  • Emphasize “injury” narrative (layoffs, plant closings, foreign competition)
  • De-emphasize expansion/investment language
  • Adjust messaging by section (IR vs Products vs Careers)
  • Coordinate with accruals management strategy

PHASE 2: WEBSITE DATA EXPLORATION

What I Actually Did (Quick & Dirty Version)

Data & Method

Data source: Cleaned website text from Haans and Mertens (2024)

Method: Keyword frequency analysis

Analysis: Event study around investigation dates

Spoiler: There’s sometimes something there, but it’s messy1

Red Flags 🚩

  • Wayback coverage varies wildly

  • Keywords might be too crude/noisy

  • Don’t know who controls website content

  • Website timing issues remain

Important Context

This is NOT a finished analysis. Most likely it will change until the lecture takes place! This is a “should I keep going?” check.

PHASE 2: WEBSITE DATA EXPLORATION

Live Demo: Website Analysis

What we’ll look at:

  1. Data collection approach - already done by Haans and Mertens (2024)
  2. Keyword frequency trends around investigations
  3. Apply same design as for revisiting Jones (1991)
  4. Preliminary correlation with discretionary accruals

⚒ Website analysis page

PHASE 2: WEBSITE DATA EXPLORATION

Website Analysis Takeaways

What I Found

📊 There IS variation in website language around investigations

📈 “Injury” keywords increase in Years -1 to 0

☣️ The signal is weak and noisy: Very sensitive to sample restrictions and keyword choices

But Is It Real?

🤷 Not clear if it’s strategic or mechanical

  • Are firms crafting a narrative?
  • Or just reporting investigation news?

🎯 Attribution problem

  • Can’t distinguish CFO strategy from marketing communications

⏰ Timing ambiguity

  • Don’t know when updates happen
  • Snapshots may miss key moments

Critical Assessment

There’s a signal. Maybe. But I have five serious concerns that could kill this project.

PHASE 2: WEBSITE DATA EXPLORATION

What Keeps Me Up at Night

Concern #1: Measurement Validity

Are keywords capturing “strategic injury narrative” or just “we’re in an investigation”?

Concern #2: Wayback Coverage & Timiming

Only later firm-years & events have snapshots available & when do websites get updated? Quarterly? Ad hoc?

Concern #3: Attribution Problem

Websites controlled by whom? CFO? Marketing? Compliance?

Concern #4: Spurious Correlation

Is any website-(accruals-)investigation correlation real or just noise?

Concern #5: Using different EM measures & addressing 2-stage critique?

Which other EM measures exist and fit the purpose?

Reality Check

Any ONE of these could sink the project. Are they fixable, or fatal?

PHASE 3: YOUR TURN

Break into groups of 2-3. Choose ONE prompt (10 minutes discussion):

PHASE 3: YOUR TURN

Structured Critique

PROMPT A: The Concerns Triage

Pick two of my five concerns. For each:

  1. How serious is this? (Scale 1-10, where 10 = “paper-killer”)
  2. Is it fixable? If so, how? If not, why not?
  3. What would addressing it require?
    • More/better data?
    • Different methods?
    • New theory?
    • Something unfeasible?

PROMPT B: Design Improvements

You have 3 more months and access to better NLP tools.

What ONE thing would you change/add to make this stronger?

Consider:

  • Different text analysis approach?
  • Additional data sources?
  • Alternative research design?
  • New validation strategy?

Be specific: Not “better analysis” but “sentiment analysis using FinBERT on earnings call transcripts to validate website tone”

PHASE 3: YOUR TURN

Structured Critique (continued)

PROMPT C: Alternative Outcomes

Instead of keyword frequency, what ELSE could you measure on websites?

Possibilities:

  • Sentiment by section: Different tone in IR vs Products vs Careers?
  • Changes in emphasis: Which topics get more prominent placement?
  • Visual content: Images, infographics, charts—do they change?

Pick one and explain:

  • Why it might be better than keyword counts
  • What it would reveal that keywords miss
  • What new problems it would create

PROMPT D: The Pivot

Assume one of the concerns kills the website idea.

What OTHER data source could you use to test “multi-channel disclosure strategy during investigations”?

Consider:

  • Press releases: More structured, official, time-stamped
  • Conference calls: Q&A reveals management framing
  • 10-Ks/10-Qs: Traditional disclosures, but more frequent
  • Something else?

Evaluate: Would your alternative be better or worse than websites? Why?

PHASE 3: YOUR TURN

Discussion

10 minutes: Share your critiques/designs with the class

WRAPPING UP

WRAPPING UP

Key Lessons from Today

What You Saw

  1. Replication is valuable
    • Validates methods
    • Builds intuition
    • Creates foundation for extensions
  2. “Quick checks” before major investments
    • Test signal existence first
    • Identify deal-breaker problems early
    • Iterate, don’t perfect
  3. All research has serious concerns
    • It’s about managing them, not eliminating them
    • Transparency > false confidence

What I Hope You Practice

  1. Combining classic papers/theories with new data
    • What has changed since publication?
    • Does new data add construct validity?
    • Are incentives/institutions still aligned?
  2. Thinking in stages
    • Check ⇢ Refine ⇢ Invest
    • Each stage has decision points
    • Knowing when to stop is a skill
  3. Making thinking visible
    • To yourself (document decisions)
    • To others (transparent process)
    • To reviewers (justify choices)

WRAPPING UP

Access Everything

GitHub Repository

🔗 Main repo

🔗 Jones-esque analyses

🔗 Website analysis

🔗 Scripts

What You’ll Find

  • Complete replication code
  • Website scraping scripts
  • Preliminary text analysis
Everything is documented—you can reproduce or extend this work

For Your Own Research

Think of this repo structure as a template. Clear organization, documented decisions, reproducible code—these habits will save you (and your coauthors) countless hours.

Thank you for your attention and good luck with your research!

APPENDIX SLIDES

References

Boulland, R., Bourveau, T., Breuer, M., 2025. Company websites: A new measure of disclosure. Journal of Accounting Research.
Breuer, M., Schütt, H.H., 2023. Accounting for uncertainty: An application of bayesian methods to accruals models. Review of Accounting Studies 28, 726–768.
Chen, B., Chen, W., Hribar, P., Melessa, S., 2023. Measuring dual effects in accounting research: Limitations of absolute residuals and the role of quantile regression. https://doi.org/10.2139/ssrn.4572590
Chen, W., Hribar, P., Melessa, S., 2018. Incorrect inferences when using residuals as dependent variables. Journal of Accounting Research 56, 751–796.
Dechow, P., Ge, W., Schrand, C., 2010. Understanding earnings quality: A review of the proxies, their determinants and their consequences. Journal of accounting and economics 50, 344–401.
Haans, R.F., Mertens, M.J., 2024. The internet never forgets: A four-step scraping tutorial, codebase, and database for longitudinal organizational website data. Organizational Research Methods 10944281241284941.
Jones, J.J., 1991. Earnings management during import relief investigations. Journal of accounting research 29, 193–228.

Appendix

Haans and Mertens (2024)

Web Scraping Steps

Sampling steps

Back

Haans and Mertens (2024)

Coverage

Coverage

Back

Haans and Mertens (2024)

Coverage Over Time

Coverage Over Time

Back

Boulland et al. (2025)

Validation of Website Disclosure Measure

Validation of Website Disclosure Measure

Back

Boulland et al. (2025)

Information Asymmetry and Website Disclosure

Information Asymmetry and Website Disclosure

Back

Boulland et al. (2025)

Application: Private Equity Deals

Application: Private Equity Deals

Back

Boulland et al. (2025)

Application: Nonfinancial Disclosures in France

Application: Nonfinancial Disclosures in France

Back

Boulland et al. (2025)

Devil’s Advocate

Are these correlations validating or concerning?

Interpretation A (Optimistic):

✅ Low correlations = websites capture different information

✅ Multi-stakeholder (not just investors)

✅ Real-time vs. quarterly

✅ Voluntary vs. mandated

Interpretation B (Pessimistic):

😕 Low correlations = measurement noise or invalidity

😕 Size ≠ quality (could be marketing fluff)

😕 Unaudited, unstandardized

😕 Designer decisions, not CFO strategy

Which is it? Back