WHY MATHEMATICS?

Why Mathematics? | M-Quantile Regression: Asymmetric Huber Scores, Robust Conditional Location, IRLS and Tuning

How tilted bounded scores bridge order-based quantiles and magnitude-sensitive expectiles, why the teaching root balances weighted influence, and which tuning choices must be disclosed.

Why mathematics is important becomes easier to see when a real decision depends on it. This guide follows the mechanism from definitions and dimensions to a worked calculation, validation, practical learning and responsible use. It keeps factual evidence separate from interpretation and treats every illustrative scenario as an example rather than a promise of results.

Choose a reading route

Standardise residuals with a declared scale
Apply a bounded score and asymmetric q weights
Solve the score equation or IRLS regression system
Check tuning, leverage, curve order, uncertainty and decision meaning

Chapter map

1. Level
2. Scale
3. Cutoff
4. Residuals
5. Huber score
6. Asymmetry
7. Weighted scores
8. Equation
9. Root
10. Quantile limit
11. Expectile limit
12. Regression
13. IRLS
14. Uncertainty
15. Validation
16. Claim

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

1. Level

A small example makes the mechanism visible. Choose q in (0,1).

A concrete situation

Use q=.75.

What the mathematics does not license

q is not automatically a coverage probability. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

State meaning. 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. Scale

This is where modelling choices begin to affect evidence. Choose positive scale.

A concrete situation

Use teaching s=1.

What the mathematics does not license

Scale changes standardized residuals and clipping. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Label illustration. 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. Cutoff

The central mathematical idea is concrete. Choose Huber c.

A concrete situation

Use c=1.345.

What the mathematics does not license

This is a convention, not a universal outlier threshold. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Run sensitivity. 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. Residuals

Here the geometry becomes operational. For candidate m compute u=(y-m)/s.

A concrete situation

At m=4.1725 get five residuals.

What the mathematics does not license

Sign at zero needs a convention. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

State rule. 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.

Did You Know? In this chapter, the most important number is useful only together with its construction rule. Changing scaling, sample membership or validation design can change the numerical answer without changing the arithmetic.

5. Huber score

This step deserves a careful pause. Clip u to plus or minus c.

A concrete situation

First two and last residual clip.

What the mathematics does not license

Clipping is not deletion. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Keep scores. 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. Asymmetry

The useful question is not only how to calculate, but what the calculation assumes. Weight positives by .75 and negatives by .25.

A concrete situation

The final positive score gets .75.

What the mathematics does not license

Reversing weights targets another tail. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Audit direction. 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. Weighted scores

A small example makes the mechanism visible. Multiply weight and Huber score.

A concrete situation

Values end at 1.00875.

What the mathematics does not license

Rounded components may not sum exactly. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Keep precision. 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. Equation

This is where modelling choices begin to affect evidence. Add weighted scores and set total to zero.

A concrete situation

Sum is approximately zero.

What the mathematics does not license

A zero score is an estimating equation, not zero error. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Explain target. 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. Root

The central mathematical idea is concrete. Solve for m.

A concrete situation

m=4.1725.

What the mathematics does not license

Root depends on scale and c. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Record solver. 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. Quantile limit

Here the geometry becomes operational. Compare with sign-dominated scores.

A concrete situation

The target moves toward order-based quantiles.

What the mathematics does not license

Limits require exact normalization. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Avoid slogans. 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. Expectile limit

This step deserves a careful pause. Let c grow so psi(u)=u.

A concrete situation

Magnitude regains full influence.

What the mathematics does not license

Tail sensitivity increases. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Show equation. 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. Regression

The useful question is not only how to calculate, but what the calculation assumes. Replace m by x beta_q.

A concrete situation

Solve vector estimating equations.

What the mathematics does not license

Design rank and leverage matter. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Check 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.

13. IRLS

A small example makes the mechanism visible. Update weights, coefficients and perhaps scale.

A concrete situation

Record every iteration.

What the mathematics does not license

Different starts or scale rules can differ. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Report convergence. 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. Uncertainty

This is where modelling choices begin to affect evidence. Use compatible sandwich or bootstrap inference.

A concrete situation

Resample sampling units.

What the mathematics does not license

Quantile formulas do not transfer automatically. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Name method. 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. Validation

The central mathematical idea is concrete. Inspect score balance, held-out loss and curve order.

A concrete situation

Check several q values.

What the mathematics does not license

One fitted level cannot describe a distribution. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Keep diagnostics. 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. Claim

Here the geometry becomes operational. State q,c,scale,score and limits.

A concrete situation

Teaching root is 4.1725.

What the mathematics does not license

Do not call it a guaranteed percentile. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Bound conclusion. 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

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

M-quantiles blend asymmetric quantile thinking with an M-estimation score. Start with a symmetric convex loss rho and its score psi, then tilt the score so positive and negative residuals receive different weights controlled by q. With a bounded Huber psi, very large standardized residuals contribute a capped score. The resulting conditional-location family lies conceptually between ordinary quantiles, which use signs, and expectiles, which use residual magnitudes without clipping.

For y=1,2,3,4,20, q=.75, fixed teaching scale one and Huber cutoff c=1.345, solve sum w_q(u_i)psi(u_i)=0 with weight .75 for positive residuals and .25 for negative residuals. The root is m=4.1725. Residuals are -3.1725,-2.1725,-1.1725,-.1725,15.8275. Weighted scores are -.33625,-.33625,-.293125,-.043125 and 1.00875, which sum to zero apart from rounding.

The teaching result depends on q, c and scale. If c tends to infinity, the score becomes linear and the construction approaches an expectile-type equation. With a sign-like limiting score it approaches quantile behaviour. A finite Huber cutoff offers a tunable compromise, but it creates no universal best q or c. Regression adds covariates and usually solves the estimating equations by iterative reweighted least squares, so start, scale update, tolerance and convergence become part of reproducibility.

M-quantile regression does not inherit every property of quantiles or expectiles. Its fitted level is not automatically a probability statement, and the order of several fitted M-quantile curves can still need checking. Bounded residual scores do not automatically bound leverage from the design. Report the exact asymmetric score, scale, cutoff, level, algorithm, covariance or bootstrap method, leverage checks and intended decision meaning.

Frequently asked questions

What is an M-quantile?

An asymmetric M-estimation location defined by a tilted score equation.

Why use a Huber score?

It is linear near zero and caps large standardized residual scores.

What does q control?

The relative weight of positive and negative residual scores.

Is q a guaranteed coverage probability?

No. The fitted M-quantile needs its own decision interpretation and validation.

Why does scale matter?

It determines which residuals reach the Huber cutoff.

What must be reported?

q, score, cutoff, scale, IRLS rules, leverage, uncertainty and curve-order checks.

Primary source and fact-check note

Jens Breckling and Ray Chambers, “M-Quantiles,” Biometrika 75(4), 761–771 (1988), DOI 10.1093/biomet/75.4.761. 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

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: topic-specific validation and responsible use

Design audit

Freeze sampling units, response orientation, predictor support, dependence, missingness and repeated observations.

Objective audit

Write every residual, asymmetric weight, absolute value, check loss, score and constraint before fitting.

Optimisation audit

Record the linear program or IRLS start, scale, tuning, tolerance, solver status and tied optima.

Coherence audit

Check coefficient uniqueness, quantile order, curve crossing, interval width and extrapolation support.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Robustness audit

Separate vertical residual resistance from predictor-space leverage and sampling bias.

Uncertainty audit

Use compatible rank, sandwich or bootstrap methods and repeat every repair or tuning step.

Validation audit

Evaluate untouched observations by level, loss, calibration, width and subgroup relevance.

Final claim

Translate the conditional location into original units without causal, coverage or guarantee overclaims.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Mechanism workshop: calculate, challenge and transfer

Workshop 1: Level

Begin with this mechanism: Choose q in (0,1). Recreate the concrete case: Use q=.75. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: q is not automatically a coverage probability. Complete the transfer step: State meaning. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 2: Scale

Begin with this mechanism: Choose positive scale. Recreate the concrete case: Use teaching s=1. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Scale changes standardized residuals and clipping. Complete the transfer step: Label illustration. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 3: Cutoff

Begin with this mechanism: Choose Huber c. Recreate the concrete case: Use c=1.345. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: This is a convention, not a universal outlier threshold. Complete the transfer step: Run sensitivity. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 4: Residuals

Begin with this mechanism: For candidate m compute u=(y-m)/s. Recreate the concrete case: At m=4.1725 get five residuals. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Sign at zero needs a convention. Complete the transfer step: State rule. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 5: Huber score

Begin with this mechanism: Clip u to plus or minus c. Recreate the concrete case: First two and last residual clip. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Clipping is not deletion. Complete the transfer step: Keep scores. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 6: Asymmetry

Begin with this mechanism: Weight positives by .75 and negatives by .25. Recreate the concrete case: The final positive score gets .75. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Reversing weights targets another tail. Complete the transfer step: Audit direction. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 7: Weighted scores

Begin with this mechanism: Multiply weight and Huber score. Recreate the concrete case: Values end at 1.00875. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Rounded components may not sum exactly. Complete the transfer step: Keep precision. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 8: Equation

Begin with this mechanism: Add weighted scores and set total to zero. Recreate the concrete case: Sum is approximately zero. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: A zero score is an estimating equation, not zero error. Complete the transfer step: Explain target. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 9: Root

Begin with this mechanism: Solve for m. Recreate the concrete case: m=4.1725. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Root depends on scale and c. Complete the transfer step: Record solver. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 10: Quantile limit

Begin with this mechanism: Compare with sign-dominated scores. Recreate the concrete case: The target moves toward order-based quantiles. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Limits require exact normalization. Complete the transfer step: Avoid slogans. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 11: Expectile limit

Begin with this mechanism: Let c grow so psi(u)=u. Recreate the concrete case: Magnitude regains full influence. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Tail sensitivity increases. Complete the transfer step: Show equation. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 12: Regression

Begin with this mechanism: Replace m by x beta_q. Recreate the concrete case: Solve vector estimating equations. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Design rank and leverage matter. Complete the transfer step: Check matrix. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 13: IRLS

Begin with this mechanism: Update weights, coefficients and perhaps scale. Recreate the concrete case: Record every iteration. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Different starts or scale rules can differ. Complete the transfer step: Report convergence. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 14: Uncertainty

Begin with this mechanism: Use compatible sandwich or bootstrap inference. Recreate the concrete case: Resample sampling units. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Quantile formulas do not transfer automatically. Complete the transfer step: Name method. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 15: Validation

Begin with this mechanism: Inspect score balance, held-out loss and curve order. Recreate the concrete case: Check several q values. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: One fitted level cannot describe a distribution. Complete the transfer step: Keep diagnostics. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 16: Claim

Begin with this mechanism: State q,c,scale,score and limits. Recreate the concrete case: Teaching root is 4.1725. Write levels, predictors, responses, fitted values, scales, standardized residuals, Huber scores, asymmetric weights, score sums, coefficients and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not call it a guaranteed percentile. Complete the transfer step: Bound conclusion. Change exactly one observation, predictor value, response value, residual sign, quantile level, loss weight, shared-slope restriction, solver tolerance, optimal vertex, scale, cutoff, starting coefficient, crossing point, rearrangement grid, bootstrap rule, leverage value or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, objective, optimisation, coherence, robustness, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the full calculation table and once to a parent with the decision in original units. Add a counterexample involving repeated x-values, censoring, dependence, a real minority group, high leverage, mixture structure or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.