Research

Control Questions for Psionics Research Using Existing Data

2026-09-10

A useful control question resembles the real research task in relevant ways but removes or changes the relationship the proposed method claims to detect. Using existing datasets, controls can probe metadata clues, generic response tendencies or analytical artifacts. Choose controls from a causal explanation of the possible error, not from a desire to decorate the study with a control condition.

What exactly could make the result look better than it is?

Before designing a control, name the ordinary failure mode. Perhaps a filename reveals the answer, a common response succeeds because one category dominates, or an analysis pipeline produces a pattern even when the nominated relationship is absent. Each problem needs a different check.

Within Andrei Ursachi's exploratory applied-psionics framework, a control does not test whether an impression felt meaningful. It asks whether the translated research claim adds information beyond specific alternatives. The task remains computational and uses existing records rather than new participant or physical experiments.

Lipsitch and colleagues describe negative controls as tools for detecting certain forms of confounding and error under stated assumptions. Their framework motivates careful control selection; it does not imply that every null comparison is a valid negative control.

Sources: Lipsitch, Tchetgen Tchetgen and Cohen (2010), Negative controls.

A hypothetical archive with answer-revealing labels

Imagine an existing collection of simulation summaries labeled stable_case and unstable_case in their filenames. A reviewer is asked to nominate which summary contains an instability. Even if the numerical arrays are hidden, the question is already compromised by the visible metadata.

A metadata-only baseline can show how much the labels reveal. A revised dataset can replace those labels with neutral identifiers while retaining the scientific description needed for the actual question. The revision is a new evaluation condition, not proof that the original result was blind.

Another control might retain the document layout but use a target variable that the proposed mechanism should not predict. Its validity depends on whether that target is genuinely unrelated under the hypothesis and whether other shared causes could still produce an association.

Match nuisance structure without preserving the answer

A shuffled outcome is not automatically a good control. Random shuffling can destroy temporal dependence, group structure or sampling patterns, making the control far easier than the real problem. Preserve the nuisance structure relevant to the suspected artifact.

For example, if a proposed pattern concerns daily variation, a control that scrambles all timestamps may remove the very acquisition rhythm being investigated. A defensible comparison may require a preselected shift or within-group permutation whose assumptions are explicitly justified.

Use the simplest control that targets the concern and record what it cannot rule out. A successful metadata check does not rule out remembered literature. A failed temporal control does not tell you which causal mechanism produced the original association.

  • Identify the suspected artifact.
  • State which structure the control preserves.
  • State which relationship it removes.
  • Explain why the expected control result follows.
  • Record remaining alternative explanations.

Keep control design away from the final answer

Choose controls and scoring rules before examining the evaluation outcomes. Repeatedly redesigning a control until the preferred hypothesis looks distinctive turns the control into another tuning parameter. Preserve all attempted versions and label that work as development.

The scikit-learn guidance on data leakage emphasizes keeping test information out of model fitting and preprocessing. The same practical boundary matters here: selecting the control using the answer can contaminate the comparison before the final statistical test begins.

An existing public dataset may still require an exposure statement. If the reviewer or AI already knows its outcomes, do not describe a control as restoring ignorance. At best, it may isolate a narrower source of information or reveal a specific weakness.

Sources: scikit-learn, Common pitfalls and recommended practices.

Use control design as the first feasibility decision

For a consulting brief, describe the intended claim and the most plausible ordinary explanation for an apparent success. If your team cannot name a discriminating comparison in the available data, that may be the primary problem to solve.

A scoped review can identify whether an existing archive supports a fair test, propose a control and explain the remaining ambiguity. The output can be a recommendation to proceed, narrow the claim or stop because the evidence cannot separate the alternatives.

The practical takeaway is that a control should earn its place by ruling out a named error. A large suite of irrelevant controls is less persuasive than one well-chosen comparison whose assumptions are clear.

Questions this raises

Do controls require new participant studies?

No. The comparisons described here use existing datasets, metadata, archived predictions and computational baselines.

Does passing a negative control prove the proposed mechanism?

No. It may weaken one alternative explanation under the control's assumptions. Other alternatives and independent evaluation still matter.

Sources and their limits

Prepared with AI assistance. The linked sources support the specified technical points; they do not validate applied psionics as a whole or guarantee a result for a client.

Read the editorial and evidence standard.

Continue reading

Explore Scientific Oracle consultingfor a scoped review of an existing-data research decision. Start with a non-confidential outline of the question, available evidence and the decision it needs to inform.