Chapter map

1. 1. Why several studies need more than pooling
2. 2. The shared-variable contract
3. 3. The factor-analysis foundation
4. 4. Common loading structure
5. 5. Study-specific loading structure
6. 6. A two-study covariance example
7. 7. Covariance, not identical observations
8. 8. Residual variances and scale
9. 9. Expectation Conditional-Maximisation
10. 10. Rotations, signs and identifiability
11. 11. Choosing common and specific ranks
12. 12. Held-out studies as a stronger test
13. 13. Interpreting loadings responsibly
14. 14. Comparing MSFA with nearby methods
15. 15. A student simulation
16. 16. Governance, careers and communication

Visit the How Mathematics Works hub for the wider series and its connections to education, reasoning, science, computing, design and careers.

---

1. Why several studies need more than pooling

This is where modelling choices begin to affect evidence. Independent studies can measure the same variables but contain different populations, laboratories and protocols. Pooling treats all covariance as one structure, while separate analyses make recurrence difficult to see.

A concrete situation

Three ovarian-cancer expression studies may share one pathway pattern and contain platform-specific covariance.

What the mathematics does not license

A larger pooled sample does not erase systematic study differences. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Create a study-by-variable manifest with units, platforms, filters and independent sample counts. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

2. The shared-variable contract

The central mathematical idea is concrete. MSFA requires compatible variables across studies, although samples are different. Variable identity, orientation and transformation must be locked before modelling.

A concrete situation

A gene identifier mapped to two probes needs a documented aggregation rule.

What the mathematics does not license

Matching column names without matching biological meaning creates false common structure. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Trace five variables from raw files through mapping and transformation. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

3. The factor-analysis foundation

Here the geometry becomes operational. Ordinary factor analysis represents an observed vector as loadings times latent scores plus residual noise. The low-rank term explains structured covariance.

A concrete situation

Two correlated genes can load on one latent factor even when the factor itself is not directly measured.

What the mathematics does not license

A factor is a modelled direction, not a discovered physical object. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Multiply a short loading vector by three scores and reconstruct the rank-one matrix. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

4. Common loading structure

This step deserves a careful pause. MSFA includes factors whose loading patterns are shared across studies. Each study has its own sample scores along those common loadings.

A concrete situation

A coordinated immune-expression pattern can recur with different score distributions in each cohort.

What the mathematics does not license

Shared loadings do not require equal prevalence or equal mean score. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Sketch one common loading and different study score histograms. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

5. Study-specific loading structure

The useful question is not only how to calculate, but what the calculation assumes. Each study can also have low-rank covariance not represented by the common factors. These terms protect local structure from being forced into the shared model.

A concrete situation

One laboratory may show an RNA-quality pattern absent elsewhere.

What the mathematics does not license

Study-specific does not mean technical or disposable. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Relate specific-factor scores to platform, site and biological metadata. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

6. A two-study covariance example

A small example makes the mechanism visible. Suppose both studies share loading a=(1,1,0), while study two also has b=(0,1,-1). Outer products aa-transpose and bb-transpose show which variable pairs covary.

A concrete situation

The common term raises covariance of variables one and two in both studies; the specific term changes variables two and three only in study two.

What the mathematics does not license

Real data include residual variances and uncertain ranks. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Calculate both outer products and add them for study two. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

7. Covariance, not identical observations

This is where modelling choices begin to affect evidence. Because the studies contain different people, the model joins them through recurring covariance rather than row-by-row correspondence.

A concrete situation

A loading can recur even when no subject appears in more than one cohort.

What the mathematics does not license

This differs from JIVE or AJIVE applied to multiple blocks on the same subjects. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Explain the difference between aligned blocks and independent studies in one paragraph. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

8. Residual variances and scale

The central mathematical idea is concrete. Feature-specific residual variance captures variability outside the low-rank factors. Scaling choices change the covariance target and loading interpretation.

A concrete situation

A high-variance gene can dominate an unscaled covariance model.

What the mathematics does not license

Unit variance scaling may amplify noisy variables. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Compare covariance and correlation modelling on a three-variable example. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

9. Expectation Conditional-Maximisation

Here the geometry becomes operational. The ECM algorithm uses expected latent quantities and conditionally maximises parameter groups. Iteration continues until the objective stabilises under a stated rule.

A concrete situation

Loadings and residual variances can be updated in separate conditional steps.

What the mathematics does not license

Convergence to a stationary point does not guarantee the best global solution. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Run multiple initialisations and compare objectives and loading subspaces. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

10. Rotations, signs and identifiability

This step deserves a careful pause. Factor models can represent the same covariance with rotated or sign-flipped loadings. Interpretation should focus on identified subspaces or use explicit constraints.

A concrete situation

A loading vector and its negative reconstruct the same rank-one covariance.

What the mathematics does not license

Comparing raw columns across fits can exaggerate instability. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Align factors by correlations or principal angles before judging recurrence. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

11. Choosing common and specific ranks

The useful question is not only how to calculate, but what the calculation assumes. Too few factors leave structured residuals; too many factors fit noise and blur the common-specific boundary. Rank selection should combine fit, stability and external evidence.

A concrete situation

One common factor may recur while a second appears only under one initialisation.

What the mathematics does not license

Training likelihood rewards additional complexity. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Plot held-out likelihood and subspace stability across nearby rank combinations. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

12. Held-out studies as a stronger test

A small example makes the mechanism visible. Leaving out one study tests whether common loadings transport beyond the datasets that selected them. This is stronger than random sample splitting.

A concrete situation

Fit on three cohorts and project a fourth using locked variables and scaling.

What the mathematics does not license

A held-out study may differ in platform coverage and require a prespecified mapping. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Define the transport contract before looking at its outcomes. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

13. Interpreting loadings responsibly

This is where modelling choices begin to affect evidence. High-loading variables define a covariance direction and can suggest pathways or systems. Correlated features may exchange ranks without changing the factor.

A concrete situation

A gene set can be stable even when exact top genes move across bootstraps.

What the mathematics does not license

Loadings do not establish regulation or causation. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Report group-level and exact-variable stability separately. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

14. Comparing MSFA with nearby methods

The central mathematical idea is concrete. Pooled factor analysis ignores study-specific covariance; separate models do not estimate a common component; MOFA handles multiple views; SVA targets hidden nuisance variation.

A concrete situation

Method choice follows sample alignment and the scientific question.

What the mathematics does not license

Similar score plots can arise from different objectives. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Write a table of sample structure, shared object, outputs and main failure mode. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

15. A student simulation

Here the geometry becomes operational. Generate two studies from one common loading, add a specific factor to the second and vary residual noise.

A concrete situation

Known truth allows recovery, rank and rotation errors to be seen.

What the mathematics does not license

A clean simulation is easier than heterogeneous field data. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Increase confounding and unequal sample size to find the recovery boundary. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

16. Governance, careers and communication

This step deserves a careful pause. Archive mappings, ranks, initialisations, convergence, loadings, scores and held-out evidence. Multi-study work joins statistics, computing and domain knowledge.

A concrete situation

A team may include statisticians, data engineers, biologists and study investigators.

What the mathematics does not license

Learning the mathematics guarantees no admission, salary or scientific authority. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Create a model card that separates covariance evidence, interpretation and causal claims. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

A seven-day practice lab

1. Day 1: define the decision, observation unit, X variables and Y or reference condition.
2. Day 2: reproduce the smallest worked example and check every dimension, unit and sign.
3. Day 3: build a transparent baseline and write what it assumes.
4. Day 4: fit the main method inside a training-only pipeline; record every parameter.
5. Day 5: challenge one assumption with a deliberately difficult dataset or operating regime.
6. Day 6: compare validation metrics with the practical workload created by errors.
7. Day 7: present one supported conclusion, one unresolved uncertainty and one safe next experiment.

Questions for a student, parent or mentor

• Which part of the model is a mathematical definition, and which part is a choice?
• What small calculation would catch a sign, scale or dimension error?
• What evidence comes from training, validation and the untouched test?
• What benign condition could look like a fault or prediction failure?
• How would the answer change for another response, asset, school project or time period?
• What should a human review before acting on the score?

A method-specific audit before deployment

A Multi-Study Factor Analysis audit begins with a study manifest, compatible variables and independent samples. The model separates loadings shared across studies from loadings specific to each study, while each study retains its own factor scores. This is a statistical description of covariance. It does not automatically tell the analyst whether a common factor is biological, technical or created by common preprocessing.

Record the common rank, every study-specific rank, covariance assumptions, starting values, convergence checks and the exact variables used. The Expectation Conditional-Maximisation procedure improves an objective iteratively, so local solutions and weak identification deserve attention. Compare nearby rank choices and align factor signs or subspaces before declaring that a loading pattern is stable.

Validation should leave out people and, when transport matters, entire studies. A common factor earns a reproducibility claim when its loading subspace and scientific associations recur beyond the studies that selected it. A study-specific factor may still be useful; its label describes where the covariance pattern appears, not whether it is unimportant.

---

Frequently asked questions

What does Multi-Study Factor Analysis model?

It models covariance across compatible variables measured in several studies, separating loading patterns common to all studies from patterns specific to individual studies.

Are factor scores shared across studies?

The common loading structure is shared, while each study has scores for its own samples. Samples do not need to be the same people across studies.

How is this different from pooling every sample?

Pooling assumes one covariance structure. MSFA explicitly allows study-specific low-rank covariance in addition to common structure.

What does the ECM algorithm do?

It alternates conditional maximisation steps for parameter groups, using expected latent quantities, until the fitted objective meets a convergence rule.

Does a common factor prove the same cause?

No. It shows a recurring covariance direction under the model. Causal interpretation needs design, metadata and experiments.

How should MSFA be validated?

Use rank sensitivity, grouped resampling, held-out studies, loading-subspace stability, reconstruction checks and independent scientific evidence.

---

Primary source and fact-check note

Roberta De Vito, Ruggero Bellio, Lorenzo Trippa and Giovanni Parmigiani, “Multi-Study Factor Analysis,” Biometrics 75(1), 337–346 (2019), DOI 10.1111/biom.12974. Definitions and core claims in this article were checked against this primary publication. Calculations shown here were independently recomputed. Examples are educational illustrations; they are not reported industrial performance results.

Useful next reading

• MOFA and multi-omics factor analysis
• AJIVE and shared score subspaces
• Surrogate Variable Analysis

Keep the options open. Mathematics education supports clearer questions, transferable problem-solving skills and informed choices across science, engineering, computing, AI, statistics, design and many other careers. It does not guarantee intelligence, income, admission or employment; those outcomes depend on far more than one subject or model.

Extended laboratory: validation, transfer and responsible use

Study manifest

Record study ownership, recruitment, platform, variable mapping, exclusions, sample size and outcome availability. The manifest explains whether common covariance is scientifically comparable. It also prevents accidental duplication of subjects or reference samples across studies from being counted as independent recurrence.

Rank grid

Evaluate common and study-specific ranks together. Save held-out likelihood, reconstruction, residual dependence and loading-subspace stability for every plausible combination. If several choices are effectively tied, prefer the smaller stable model and report the ambiguity rather than presenting one rank as certain.

Initialisation challenge

Run the ECM procedure from several deterministic and random starts. Compare final objectives and projection matrices rather than raw signed columns. Disagreement can reveal local solutions or weak identification. Archive seeds so another analyst can reproduce both the preferred and competing fits.

Study-size sensitivity

Repeat analysis with study-balanced weighting or subsampling where justified. A large study can determine common loadings while smaller studies merely receive scores. Report whether the central subspace survives. This test distinguishes broad recurrence from dominance by sample count.

Residual audit

Plot residual covariance by study and variable group. Remaining blocks can indicate an omitted factor, nonlinear dependence, heteroscedastic noise or mapping error. Residual structure is evidence about model limits, not an inconvenience to delete.

Held-out projection

Lock variables, scaling and common loadings, then project an untouched study. Report coverage, reconstruction and scientific associations without refitting. If transport requires remapping or refitting, label it as a new model rather than external validation.

Common-factor naming

Use neutral labels until a factor survives metadata challenges, negative controls and external evidence. A common technical protocol can generate a common factor across laboratories. The word common describes recurrence, not cause.

Specific-factor value

Study-specific factors can diagnose platform behaviour or reveal population-specific biology. Preserve them and test explanations. Removing every specific component before downstream analysis can discard exactly the contextual evidence that explains failed transport.

Uncertainty reporting

Bootstrap independent samples within studies and, when enough studies exist, resample studies. Report principal angles or projector distances with intervals. Exact loading ranks are often less stable than the subspace or variable group.

Reproducible computation

Version software, convergence tolerance, maximum iterations, seeds and floating-point environment. Save objective traces and parameter hashes. A second analyst should reconstruct one covariance matrix from archived common, specific and residual terms.

Student extension

Let learners rotate two factor columns and verify that the reconstructed covariance does not change. Then align the rotated spaces using principal angles. This reveals why factor interpretation requires constraints and why subspace comparisons are often more honest.

Reader checkpoint

A complete claim identifies studies, shared variables, common and specific ranks, estimation, convergence, held-out evidence and interpretation limits. Without those facts, a recurring factor plot cannot support a reproducibility claim.

Deep dive: how to stress-test the mathematical claim

Deep dive 1

Test whether common structure survives removing one study at a time. Compare projector distances and variable-group enrichment. If one study controls the result, label the factor pooled rather than broadly reproducible.

Deep dive 2

Use simulated residuals matched to each study to examine false common rank. Unequal noise can make a high-variance study produce apparent recurrence. Rank evidence should be calibrated to the fitted heteroscedastic setting.

Deep dive 3

Check residual covariance for blocks that mirror the proposed factor. An omitted nonlinear or correlated-noise structure can be misrepresented as extra low-rank signal.

Deep dive 4

Compare factor analysis on covariance and correlation matrices. A factor stable under both supports a broad relational claim; disagreement shows that measurement scale is part of the result.

Deep dive 5

When exact loadings are unstable, compare subspaces and variable sets. Rotational freedom means that column-by-column matching can create a false impression of failure.

Deep dive 6

Use negative-control variables or simulated null features. They should not dominate common loadings. If they do, inspect scaling, missingness and preprocessing references shared across studies.

Deep dive 7

Assess whether study-specific factors predict known platform variables. This can support a technical interpretation, but correlation alone still does not prove the platform caused the factor.

Deep dive 8

Hold out a study with a different population as a transport stress test. Failure may identify a real boundary rather than invalidate the model everywhere.

Deep dive 9

Report the denominator of evidence: studies and independent samples, not variables. Millions of gene values do not create millions of independent replications.

Deep dive 10

End with a claim ladder covering covariance reconstruction, recurrence, metadata association and causal explanation. Each step requires new evidence.

Worked evidence notebook

Matrix dimensions

For study s with n_s samples and p shared variables, X_s is n_s by p. Common and study-specific score matrices have n_s rows, while loading matrices have p rows. Write every dimension before coding to catch transposes.

Outer-product check

For a=(1,1,0), aa-transpose equals a three-by-three matrix with a two-by-two block of ones. For b=(0,1,-1), bb-transpose adds positive variances and negative covariance between variables two and three.

Rotation check

Choose a two-column loading matrix and multiply it by an orthogonal rotation. Verify that L L-transpose is unchanged. The covariance stays the same even though individual columns look different.

Rank evidence

Create a table of common rank, specific ranks, held-out fit and principal-angle stability. Highlight tied configurations instead of hiding them.

Transport note

Document which variables were missing or remapped in the held-out study and how projections were computed. Coverage is part of external validity.

Interpretation note

For each factor, list algebraic evidence, recurrence, metadata associations and unresolved alternatives. Do not merge these levels into one label.

A final worked scenario

A final multi-study scenario

Imagine four independent expression studies with the same mapped genes but different laboratories. The team fits one common factor and allows one study-specific factor in two cohorts. The common loading subspace recurs when each study is omitted, and a fifth untouched cohort projects into the same variable-group pattern. One specific factor tracks an older array platform, while another aligns with a population feature. The team reports both without claiming either cause is proven. A second common factor improves training likelihood but is unstable across initialisations and is left in the ambiguity register. This is a strong mathematical result because every layer is visible: covariance reconstruction, rank sensitivity, convergence, held-out transport and scientific alternatives. The model has not merged the studies into one imaginary population. It has shown which low-rank relationships recur, which remain contextual and where further measurement is needed.

Method checklist and conclusion boundary

A compact reproducibility checklist

Freeze variable mappings and study manifests, then save scaling, ranks, starts, convergence traces, common and specific loadings, scores and residuals. Compare reconstruction and projector stability across sample and study holdouts. Reproduce one covariance by summing common, specific and residual terms. Report tied rank solutions and failed starts. These steps let another analyst see whether recurrence belongs to the data, the estimator or one dominant study.

What a careful conclusion sounds like

A careful conclusion says that a loading subspace recurred across specified studies under documented mappings and rank assumptions. It identifies external projection evidence and alternative technical explanations. It does not claim that every factor column is uniquely identified, that a common pattern has one cause, or that a study-specific factor is unimportant.

Practice, misconception and transfer

From a school project to research practice

A student can begin with two small covariance matrices and ask which rank-one pattern appears in both. In research, the matrices may come from separate cohorts with thousands of variables, but the reasoning is the same: identify the shared structure, preserve local structure and check whether the result survives new evidence. The scale changes; the duty to state variables, units and assumptions does not.

Why rank is a scientific choice

Rank determines how many recurring and local patterns the model is allowed to describe. A rank selected only because it gives the best training fit usually tells a richer story than the evidence supports. Ask whether each added direction recurs under resampling, improves held-out reconstruction and connects to a prespecified scientific question. A modest model with one stable common factor can be more valuable than a detailed unstable catalogue.

What parents and teachers can emphasise

The accessible starting points are averages, covariance, vectors, outer products and matrix dimensions. Encourage the learner to reconstruct a tiny matrix by hand and explain why two studies can share a loading without sharing participants. This builds transfer from school algebra to modern data science. The goal is sound reasoning, not early memorisation of software commands.

A misconception to correct

MSFA does not average factor scores across studies to create one universal patient coordinate. Each study contains different samples and receives its own scores. What recurs is the loading pattern over compatible variables. Keeping this distinction clear prevents a common error and helps readers understand why held-out-study projection is meaningful.

When the method should be paused

Pause when variable mappings are doubtful, studies have incompatible measurements, sample overlap is hidden, or one study contains nearly all the evidence. Resolve data provenance before adding factors. A more elaborate optimiser cannot make mismatched columns scientifically comparable. The decision to stop and repair the study contract is mathematical maturity.

A useful next experiment

If a common factor aligns with a proposed pathway, design a targeted replication using independent samples and direct measurements of plausible drivers. Predefine the loading or variable set and test it without reopening rank selection. This turns an exploratory covariance pattern into a sharper hypothesis and shows how mathematics guides the next observation rather than replacing it.

Final evidence checks

Final evidence checks 1

Preserve raw study matrices and mappings; adjusted scores and factors are derived evidence.

Final evidence checks 2

Require a second analyst to reproduce one covariance reconstruction and one held-out projection.

Final evidence checks 3

Report common and specific ranks with nearby sensitivity rather than one unexplained setting.

Final evidence checks 4

Use subspace comparisons when rotational freedom makes individual columns unstable.

Final evidence checks 5

Keep study-specific factors because they may explain transport boundaries.

Final evidence checks 6

State that common covariance does not prove a common cause.