Least Trimmed Squares¶
A robust regression estimator that minimizes the sum of the h smallest squared residuals, reselecting the retained cases for each trial fit.
Core Idea¶
Least trimmed squares (LTS) fits a regression by minimizing the sum of the \(h\) smallest squared residuals. For each candidate coefficient vector, it recalculates all residuals and reselects which \(h\) enter the sum. This is a specific robust-regression objective, not a one-time deletion of cases. When \(h=n\), it becomes ordinary least squares (OLS). With a suitable smaller \(h\), it can resist contamination that would strongly pull OLS, but its breakdown guarantee depends on \(h\), model dimension and data conditions.[^ref-97b476460bf7]
Scope of Application¶
The original method was applied to actuarial line fits involving insurance, pension, health-insurance and inflation data. A later study used LTS diagnostics in multivariable cross-country growth regression. The same roles recur: observations and a regression model, trial-dependent squared residuals, retained count \(h\), trimmed objective, selected fit and substantive interpretation of unusual cases.[ref-97b476460bf7][ref-38adbfd0a276]
Live Robust Regression is the proposed strict DAG parent because its family definition explicitly includes LTS. Live Robustness supplies the broader stability-under-disturbance idea, but not the LTS formula.
Clarity¶
“Trimmed” does not mean first fit OLS, then permanently throw away its largest residuals. The retained subset can change with every trial fit. Nor does a large residual prove an observation is faulty. It may be a recording error, a member of another population or an informative genuine event. Least median of squares, least absolute deviations and FAST-LTS are related but different: the first minimizes a middle residual, the second changes loss, and the last is a computational search algorithm for LTS.[ref-97b476460bf7][ref-38adbfd0a276][^ref-6d2ec88eab57]
Manages Complexity¶
The method concentrates the fitting criterion on an \(h\)-case majority instead of letting all \(n\) squared discrepancies influence every candidate fit. That can make a dominant relation easier to see when a few cases are grossly discordant. The benefit has costs: smaller \(h\) uses fewer clean observations, rank changes make exact optimization harder, and a valid minority process may be overlooked. FAST-LTS can speed the search but may approximate rather than certify the global optimum on large data.[ref-97b476460bf7][ref-6d2ec88eab57]
Abstract Reasoning¶
Declare the model, number of observations \(n\), parameter count and coverage \(h\). For every trial coefficient vector, compute and order the squared residuals, sum the smallest \(h\), and seek the minimum. Report the algorithm and whether its result is exact or approximate. Then investigate large-residual or leverage cases against substantive evidence. A difference between LTS and OLS signals sensitivity, not by itself which cases or model should govern the final inference.[ref-97b476460bf7][ref-38adbfd0a276]
Knowledge Transfer¶
In an actuarial line fit, LTS uses the \(h\) smallest squared discrepancies from each trial line to find a majority relation without allowing a few atypical insurance records to dictate it. In a 61-country growth model, Zaman, Rousseeuw and Orhan used LTS residuals to inspect the multivariable fit and then separately compared OLS results with one high-residual country excluded. The ordered-residual construction transfers literally; the meaning of an exceptional policy versus an exceptional country does not. A later OLS exclusion is an analytic follow-up, not the LTS definition.[ref-97b476460bf7][ref-38adbfd0a276]
[^ref-97b476460bf7]: Peter J. Rousseeuw, B. Daniels and A. Leroy, “Applying robust regression to insurance,” Insurance: Mathematics and Economics 3 (1984), 67–72; equations (1)–(4), actuarial applications and Appendix. https://wis.kuleuven.be/stat/robust/papers/publications-1984/rousseeuwdanielsleroy-robustregressioninsurance-im.pdf [^ref-38adbfd0a276]: Asad Zaman, Peter J. Rousseeuw and Mehmet Orhan, “Econometric applications of high-breakdown robust regression techniques,” Economics Letters 71 (2001), 1–8, §2–3 and Table 1. https://wis.kuleuven.be/stat/robust/papers/2001/zamanrousseeuworhan-econometricapplications-econom.pdf [^ref-6d2ec88eab57]: Peter J. Rousseeuw and Katrien Van Driessen, “Computing LTS Regression for Large Data Sets,” Data Mining and Knowledge Discovery 12 (2006), 29–45, original-author abstract only. https://doi.org/10.1007/s10618-005-0024-4
Relationships to Other Abstractions¶
Current abstraction Least Trimmed Squares Domain-specific
Parents (1) — more general patterns this builds on
-
Least Trimmed Squares is a kind of Robust Regression Domain-specific
LTS is a specified high-breakdown robust regression estimator.
Hierarchy paths (13) — routes to 7 parentless roots
- Least Trimmed Squares → Robust Regression → Regression → Signal Extraction
- Least Trimmed Squares → Robust Regression → Regression → Function (Mapping)
- Least Trimmed Squares → Robust Regression → Regression → Statistical Inference → Inductive Reasoning
- Least Trimmed Squares → Robust Regression → Regression → Statistical Inference → Uncertainty
- Least Trimmed Squares → Robust Regression → Regression → Distributional Assumption → Assumption → Epistemic Mode Of A Proposition
- Least Trimmed Squares → Robust Regression → Regression → Distributional Assumption → Statistical Inference → Inductive Reasoning
- Least Trimmed Squares → Robust Regression → Regression → Distributional Assumption → Statistical Inference → Uncertainty
- Least Trimmed Squares → Robust Regression → Regression → Distributional Assumption → Probability → Measure → Set and Membership
- Least Trimmed Squares → Robust Regression → Regression → Statistical Inference → Probability → Measure → Set and Membership
- Least Trimmed Squares → Robust Regression → Regression → Distributional Assumption → Probability → Measure → Aggregation → Micro Macro Linkage
- Least Trimmed Squares → Robust Regression → Regression → Statistical Inference → Probability → Measure → Aggregation → Micro Macro Linkage
- Least Trimmed Squares → Robust Regression → Regression → Distributional Assumption → Statistical Inference → Probability → Measure → Set and Membership
- Least Trimmed Squares → Robust Regression → Regression → Distributional Assumption → Statistical Inference → Probability → Measure → Aggregation → Micro Macro Linkage
Neighborhood in Abstraction Space¶
Least Trimmed Squares sits in a sparse region of the domain-specific corpus (74th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Causal Inference & Regression Modeling (15 abstractions)
Nearest neighbors
- Residual Sum of Squares — 0.86
- Regression — 0.84
- Nonlinear Least Squares — 0.83
- FWL theorem — 0.82
- Hat matrix — 0.82
Computed from structural-signature embeddings · 2026-10-08