Research praxis

Every research-practice note the newsletter has run, newest first. Each is a way of reading, writing or judging evidence that outlives the edition it appeared in. Several have a script in tools/; the entry says which.

Generated by scripts/build_companions.py from archive/; edit the source, not this file.

How to Read a Deletion Rate

From The Research Rundown: July 2026, 2026-07-08.

A deletion figure with no denominator tells you almost nothing, and yet it is always the number that leads the headline.

When you see “47 lakh electors removed,” run four checks before you react.

Denominator. Forty-seven lakh against Bihar’s 7.89 crore roll is 5.95 per cent. The absolute figure is built to alarm; the ratio is what you actually reason with.

Baseline. Rolls decay on their own through death and migration at a few per cent a year, so a genuine decadal revision should remove a fair slice simply by catching up. The real test is whether removals exceed that plausible churn and which way the errors point.

Draft or final. A name missing from a draft roll can be restored during the claims window, so a draft deletion number is a ceiling, not a verdict. Delhi’s draft lands on 5 August, and its final roll on 7 October; the gap between the two will tell you more than either one alone. Quoting the draft as if it were final is the most common way this statistic is abused.

Error borne by whom. Every verification system trades wrongful removals against wrongful retentions. The politics lies entirely in that trade-off, specifically in whether the citizen or the state has to prove.

Do this once, and no headline deletion count will carry an argument past you on its own again.

What a Poverty Line Actually Is

*From [The Research Rundown

May 2026](/SignalStack/archive/2026-05-08-the-research-rundown-may-2026.html), 2026-05-08.*

Every development researcher uses a poverty line. Most of us inherited the one our institution adopted without thinking carefully about what it encodes.

A poverty line is a consumption or income threshold below which a person or household is classified as poor. The most commonly cited global lines are the World Bank’s $2.15/day (extreme poverty), $3.00/day (low-income country standard), and $4.20/day (lower-middle-income country standard). India’s Tendulkar poverty line - the last official national benchmark - sits at roughly Rs 1,622 per person per month in rural areas at 2011-12 prices. Adjusted for inflation, this falls well below any of the international lines. As Utsa Patnaik and Jean Dreze have argued, it was calibrated below caloric adequacy thresholds from the start.

Three things to track whenever you read poverty data:

The reference period. India’s HCES 2022-23 is the first comparable consumption survey in fifteen years. Any poverty estimate prior to its release was extrapolated from 2011-12 data using modelling. The assumptions embedded in that extrapolation - which consumption components grew, at what rate - were rarely stated and are rarely audited.

The consumption basket. The HCES 2022-23 uses MMRP (Modified Mixed Recall Period), which captures food and non-food consumption differently from the URP method used in previous rounds. Switching from URP to MMRP reduces measured poverty, not because welfare improved, but because the survey methodology changed. Cross-period comparisons without explicit adjustment are comparisons of different instruments.

The counterfactual. If the poverty rate fell from 27 per cent in 2011-12 to 5 per cent in 2022-23 at the $3.00 line, what drove it? The World Bank attributes the fall to real consumption growth. Researchers, including Dipa Sinha and Himanshu, argue that a significant portion reflects the methodological shift. Clean causal attribution is unavailable given the data breaks; studies that claim it are overstating their evidence.

For evaluation practitioners: if your programme targets households below a particular poverty line and tracks outcomes against it, your results are only as good as your line. A programme that “lifted” 40 per cent of beneficiaries above the $2.15 line may have moved them to Rs 65 per day, which in Delhi today buys approximately half a meal.

On participation and power in research design

From The Research Rundown — March 2026 Edition, 2026-03-18.

The pressure to “do participatory research” has produced a great deal of research that is called participatory but functions extractively. Community members are consulted at the design stage, their input is recorded, and then an external research team goes away to write the report. The community never sees it.

True participatory research requires communities to have meaningful control over research questions, data interpretation, and the use of findings. The institutional barriers are real: funders set timelines, ethics boards set protocols, journals set citation norms — all of which tilt toward the researcher as the expert. Acknowledging this tension openly in research design documents, rather than burying it in a methodology footnote, is a small but important act of intellectual honesty.

Practical tip for your next proposal: Add a section explicitly titled Limitations of Participation in this Study and describe, candidly, where your design falls short of genuine community control. It will strengthen your ethics application and make your findings more credible. It will also distinguish your work from the large volume of research that claims participation without examining what it actually means.

Applying Microlearning Principles to Development Programs

From The Research Rundown - August 2025 Edition (erm… in September), 2025-09-01.

The standard approach is to design comprehensive training workshops that cover all aspects of a topic. A microlearning approach: Break learning into specific, actionable units delivered when people need them

Why this matters: Hermann Ebbinghaus’s research proves humans forget rapidly without reinforcement. Most development programs offer one-off training, expecting lasting behavioural change—no wonder impact evaluations often disappoint.

Practical tip: Before designing any intervention, ask: “What specific decision do we want people to createifferently?” Then design the minimum effective content to influence that specific decision.

Example: Instead of a day-long climate adaptation workshop, create brief modules answering: “What should I plant if rains are late?” “How do I interpret weather app data?” “What are three water-saving techniques I can use tomorrow?”

Start by separating three different kinds of participation

*From [The Research Rundown

June 2025](/SignalStack/archive/2025-06-26-the-research-rundown-june-2025-edition.html), 2025-06-26.*

When planning a study, I find it useful to distinguish consultation, collaboration and control.

Consultation: researchers retain the design and ask participants for views.

Collaboration: researchers and participants make some substantive decisions together.

Community control: a community organisation or participant group holds decision-making authority over major parts of the research.

None of these is automatically the correct model.

A national labour-force survey cannot plausibly hand questionnaire design to every respondent. A village-level action-research project that claims to be community-led should probably give participants considerably more authority than a one-off focus group.

The choice should follow the purpose of the research.

The important thing is to describe the arrangement accurately. Calling consultation “co-production” creates the appearance of shared authority without demonstrating it.

One question usually exposes the difference:

Could participants change something the research team did not want changed?

If the answer is no, their role may still be useful. It probably is not control.

A finding is not an inference. An inference is not a recommendation.

*From [The Research Rundown

April 2025](/SignalStack/archive/2025-04-28-the-research-rundown-april-2025-edition.html), 2025-04-28.*

Evaluation reports often move too quickly across these three stages.

A finding might be:

Forty-two per cent of surveyed women reported difficulty reaching the health facility.

The inference could be:

Travel time or transport availability may be constraining service use.

The recommendation might be:

Introduce transport support for antenatal visits.

Each step requires additional evidence.

The survey establishes the reported difficulty. It may support an inference about access, depending on the instrument and sample. It does not automatically establish that transport support is the right intervention, how it should be designed or whether it is cost-effective.

Recommendations become stronger when the report shows the bridge between evidence and action.

For each major recommendation, I find it useful to ask four questions:

1. Which finding supports this?
2. What causal or institutional assumption connects the finding to the recommendation?
3. What evidence supports that assumption?
4. What would make the recommendation wrong?

The fourth question is often the most useful.

Information mapping

*From [The Research Rundown

June 2024](/SignalStack/archive/2024-06-05-the-research-rundown-june-edition.html), 2024-06-05.*

Evaluation reports often fail for an embarrassingly simple reason. They make the reader work too hard.

Information mapping starts from the assumption that readers need structure before detail. The technique is useful for long evaluations, technical reports, policy briefs and documents written for several audiences.

Start with the decision the report has to support. Build the structure around that.

Keep findings together with the evidence that supports them. Separate description from interpretation. Use headings that tell readers what a section contains. Put methodological caveats where they affect interpretation instead of burying all uncertainty in an annex.

And keep the summary.

A good executive summary is not a ceremonial page placed in front of a report. It should tell a reader what was evaluated, what the evidence permits you to say, what remains uncertain and what decisions follow.

One test I use: give the summary to someone who has never seen the project. If they cannot explain the main findings five minutes later, the problem may be the report rather than the reader.

Robert Horn’s work on knowledge mapping is useful here. His 2001 paper examines large visual maps for complex public problems, where causation, disagreement and institutional responsibilities cannot easily be represented as a tidy linear argument.

Robert E. Horn. 2001. Knowledge Mapping for Complex Social Messes.

https://faculty.washington.edu/farkas/TC510-Fall2011/Horn-SocialMesses.pdf

For a later discussion of his “info-mural” approach:

Robert E. Horn. 2021. “Art + Science + Policy: Info-Murals Help Make Sense of Wicked Problems.” Cadmus Journal 4(5).

https://cadmusjournal.org/article/volume-4/issue-5/art-science-policy