What the record shows.
A defined contrast, estimated at the appropriate unit with an account of its uncertainty.
For example: a difference in per-fly bout duration over matched recording windows under a stated scoring rule.
OUR APPROACH
FROM QUESTION TO UNDERSTANDINGMake uncertainty
useful.
A question. Competing explanations.
A comparison that can separate them.
A useful design starts with an uncertainty we can name. We define what would change our interpretation, choose measurements that could reveal it, and preserve the analytical decisions behind the result.
Explore the research cycle01 / THE WORKING METHOD
From question
to evidence.
Description characterizes a measured pattern or difference. Prediction asks how well a rule anticipates outcomes in new data. Causal explanation asks what would change if a proposed cause were altered. We specify which task the study can perform.
Choose a step to explore the method.
01 / FRAME
What, exactly, are we trying to estimate?
Specify the population, conditions, outcome, and time window. Decide whether the aim is to describe a pattern, predict an outcome, or test a causal explanation.
Decisions to make
- Name the outcome and the contrast between conditions.
- Identify which units supply independent evidence for that contrast.
- Define a meaningful effect and the main analysis before examining the contrast.
EXAMPLE WORKING OUTPUT
A question brief
For the sleep example: compare each fly’s median bout duration between conditions in matched, valid recording windows. Specify how flies without an observed bout enter the analysis.
02 / EXAMINE
What do the records actually contain?
Trace the acquired signal to the derived measurement. A scored state depends on detector sensitivity, sampling resolution, and the rule used to classify it.
Decisions to make
- Preserve source versions, units, acquisition settings, and subject identifiers.
- Distinguish missing observations from inactivity and document exclusions.
- Map treatment assignment, repeated observations, and shared batches.
EXAMPLE WORKING OUTPUT
A source and decision register
For the sleep example: verify how movement was detected, the inactivity threshold, clock alignment, and missing intervals. A gap in recording cannot establish an uninterrupted bout.
03 / TEST
How large is the difference, and how uncertain?
Estimate the planned contrast using the structure of the study. Then examine the assumptions that could materially change its magnitude or interpretation.
Decisions to make
- Report effect sizes and intervals at the appropriate independent unit.
- Account for repeated observations, batches, and relevant model assumptions.
- Use justified sensitivity checks; label analyses prompted by the results.
EXAMPLE WORKING OUTPUT
An estimate and sensitivity summary
For the sleep example: estimate the between-condition difference without counting every bout as an independent replicate. Check a justified scoring alternative and report whether the estimate changes.
04 / REFINE
Which alternative explanation remains?
State the conclusion the design supports. Choose the next measurement or comparison for its ability to distinguish the strongest remaining explanations.
Decisions to make
- Separate a measured association from a proposed causal mechanism.
- Identify which uncertainty limits the next decision.
- Choose measurement validation, independent evidence, or a targeted experiment.
EXAMPLE WORKING OUTPUT
An interpretation and next-step note
For the sleep example: a bout difference could reflect state regulation, movement detection, or unequal coverage. An independent behavioral measure can address a different uncertainty from another analysis of the same signal.
The cycle uses an illustrative sleep-analysis question. Read the reasoning behind the design.
02 / A QUESTION, MADE CONCRETE
What does a total
leave out?
Consider an illustrative sleep question: could two records contain a similar amount of rest, yet organize that rest differently?
ILLUSTRATIVE
QUESTION BRIEF
QUESTION BRIEF
How is rest distributed within a defined recording window?
Start with the recorded signal and a justified classification rule. Compare records on the same time basis, then examine what duration, continuity, and timing each preserve.
Pattern lab / Synthetic records
Same total.
Different story.
Two 24-hour records. Eight hours of rest in each. Change the arrangement of Record B, then scan both records to see what a total leaves out.
04:00
Half-hour steps. Every arrangement stays within this day.
One day. Two records.
30 MIN / BIN- A · rest
- B · rest
- Activity
Each row has 48 half-hour bins. Tall colored marks show rest; short gray marks show activity. Record A has one eight-hour rest period from 04:00 to 12:00. Record B has 4 rest periods totaling eight hours, begins at 04:00, and has a longest period of 2 h. The outlined bin is 06:00–06:30: A is rest; B is activity. The summary and interval tables below provide the chart values.
06:00–06:30
Move the slider to inspect the outlined bin. Arrow keys move one half hour.
| Measurement | A Fixed | B Yours |
|---|---|---|
| Total rest | 8 h | 8 h |
| Rest periods | 1 | 4 |
| Longest period | 8 h | 2 h |
| First rest | 04:00 | 04:00 |
What changes?
Both contain 8 h of rest. B divides that time into 4 periods, with a longest period of 2 h. The starting time is the same.
Why the sequence mattersRead the rest intervals
| Start | End | Duration |
|---|---|---|
| 04:00 | 12:00 | 8 h |
| Start | End | Duration |
|---|---|---|
| 04:00 | 06:00 | 2 h |
| 07:00 | 09:00 | 2 h |
| 10:00 | 12:00 | 2 h |
| 13:00 | 15:00 | 2 h |
Each interval includes its start and ends just before its end time. All remaining bins are classified as activity. A rest period is a continuous run of rest bins in this example.
Synthetic example. Eight hours are fixed by construction. Adding periods changes their duration and the span they occupy. This illustrates what summaries preserve; the half-hour bins are not a biological sleep-scoring rule.
| View | What it asks | Decision to make explicit |
|---|---|---|
| Total duration | How much observed time meets the rest definition? | Valid coverage, missing time, and the denominator. |
| Bout structure | How is that time divided into episodes? | Resolution, scoring threshold, interruptions, and truncated bouts. |
| Timing | When do episodes and transitions occur? | Clock alignment, light schedule, and matched comparison windows. |
| Inference | What differs between conditions, and by how much? | Independent units, repeated observations, and uncertainty. |
Bout count and mean duration are related through the total; they are not independent confirmations of an effect. A measured difference becomes useful when its size, precision, and dependence on scoring choices are clear.
See the current sleep project03 / THE STRENGTH OF A CLAIM
What would change
the explanation?
A signal can support several accounts of what happened. We write those accounts down, specify where their predictions differ, and choose the next observation because it addresses that difference.
Read about evidence across scalesWhat might explain it.
A proposed process that generates predictions beyond the pattern used to suggest it.
Altered state regulation is one possibility. Detector sensitivity, missing time, and batch differences may also explain the observed pattern.
What would distinguish them.
A new comparison that addresses the specific ambiguity left by the first analysis.
Validate the movement measure to test a detection explanation; use an appropriate intervention to test a causal biological hypothesis.
04 / AN EXPLANATION, PUT TO A TEST
Recovery is one question.
The next response is another.
Suppose two groups have similar measured baselines after different exposure histories. A later challenge lets us ask whether the histories still matter. Keeping an unchallenged group for each history makes the comparison interpretable.
| Prior history | No later challenge | Later challenge | Challenge contrast |
|---|---|---|---|
| Control history | 10 | 6 | 6 − 10 = −4 |
| Prior exposure | 10 | 8 | 8 − 10 = −2 |
THE CONTRAST OF INTEREST
−2 − (−4) = +2 units
The challenge-associated change differs by history. In a real study, its uncertainty must be estimated from independent experimental units, with batch and repeated measurements handled explicitly.
WHAT THE DESIGN STILL NEEDS
This pattern would support a history-dependent response. It would leave several explanations open: persistent regulation, residual injury, altered development, or selection of survivors.
Exposure-removal checks, survival records, matched developmental stages, and a targeted intervention answer different parts of that problem. A mechanistic claim needs the relevant additional test.
See the stress-recovery design05 / A RESEARCH RECORD OTHERS CAN FOLLOW
Leave a clear
trail.
A research record should let another reader reconstruct the calculation and question its assumptions. These are the elements our releases are intended to preserve.
The source
A versioned source inventory, data dictionary, acquisition settings, units, subject and batch identifiers, and conditions of use. A reader should know which records can answer the question.
The decisions
An eligibility log, outcome definitions, missing-data rules, and reasons for exclusions. Distinguish choices specified before analysis from changes made after inspecting the data.
The analysis
Executable steps from source records to the values in each figure, with environment versions, model assumptions, and justified sensitivity checks. Retain an export of the plotted estimates.
The interpretation
Effect estimates, uncertainty, checks against alternative explanations, and the scope of the comparison. End with the next measurement that would change the conclusion.
A useful release pairs results with methods, definitions, and data-use conditions. Each project page identifies its stage and available materials.
Read the reproducibility noteA FEW TERMS, IN PLAIN LANGUAGE
Precision starts
with definitions.
Experimental and observational units
An experimental unit receives a treatment independently; observational units are what we measure. Many images, cells, or time bins from the same treated unit do not create additional treatment replicates. The design and dependence structure determine how uncertainty is estimated. Read the example and source.
Operational definition
The explicit rule connecting a concept to a measurement: detector, resolution, threshold, and classification. A rule can be applied consistently while missing part of the biological phenomenon. Its suitability needs evidence in the relevant context. See the sleep measurement example.
Sensitivity analysis
Re-estimate the same target under justified alternative assumptions or processing choices. Report which choices change the effect and why; selecting the most favorable version defeats the purpose. Several analyses of the same data are not independent replications.
Reproducibility and validity
Computational reproducibility means obtaining consistent results with the same data and analytical steps. Replication uses new data. Neither alone establishes that a measure captures the intended construct or that a proposed mechanism caused the result. Read the definitions and sources.
Research partnerships
Define the question.
Test the explanation.
Bring a measurement, method, or alternative explanation that can strengthen a defined study.
Discuss a collaboration