Fitting a regression line is only the first half of the job. Regression validation is the set of practices engineers use to ask whether a fitted model is actually trustworthy: Do the assumptions behind the least-squares machinery hold for this data? Does the model describe the sample well, and will it still perform acceptably on data it has never seen? These two questions — reliability on the data you have and generalizability to data you don't — are distinct, and a model can succeed at one while failing badly at the other.
In this course, validation matters because engineering decisions get made from model predictions: sizing a component, forecasting a load, setting a process tolerance. A model with an impressive R-squared can still produce systematically wrong predictions if the residuals show curvature, if variance grows with the fitted values, or if the model has simply memorized noise in a small sample. Residual analysis is the main diagnostic tool for the first family of problems; resampling methods such as cross-validation address the second. Together they feed a cycle: fit, diagnose, refine, re-check.
As you watch the videos below, keep three questions in mind. First, what specific quantity is being used to judge the model, and what does it actually measure? Second, which plot or procedure would reveal each core assumption violation — linearity, constant variance, independence, normality? Third, how does a validation finding translate into a concrete next step, such as transforming a variable or changing the functional form?
This video walks through common ways of evaluating how well a regression model performs, including R-squared and its interpretation as the proportion of variance in the outcome explained by the model. It is likely to discuss what such summary measures do and do not tell you about model quality, and how prediction error is assessed. Expect attention to the gap between fitting the data at hand and predicting new observations.
It directly supports the "What Is Regression Validation?" and "Reliability vs. Generalizability" subtopics by making clear that a goodness-of-fit number is not by itself evidence that a model is valid.
The opening lecture of Stanford's cross-validation unit, taught in the Hastie and Tibshirani statistical learning sequence, introduces resampling as a way to estimate how a model will perform on data it was not trained on. It is likely to cover the train/test split idea, why training error understates true error, and the motivation for systematic cross-validation procedures. This sets up the distinction between holdout, k-fold, and leave-one-out approaches.
It comes from the authors of the standard statistical learning text and gives an authoritative, carefully motivated treatment of generalizability — the half of validation that residual plots alone cannot address.
A tutor from the University of Liverpool's statistics support service demonstrates, step by step, how to check the assumptions behind a linear regression model using SPSS. The walkthrough is likely to cover the standard diagnostic outputs — residual plots, normality checks of residuals, and indications of non-constant variance or problematic observations — and how to read them. The emphasis is practical: which output to request and what a violation looks like on screen.
It grounds the "Core Assumptions" and "Role of Residual Analysis" subtopics in an actual diagnostic workflow, so you see what assumption checking looks like in practice rather than only in the abstract.
The Liverpool SPSS demonstration maps most directly onto MO1 and MO2, showing how residual diagnostic plots are produced and read to detect violations of linearity, homoscedasticity, independence, and normality, and it also touches the outlier and influence concerns in MO4. The Stanford cross-validation lecture establishes the conceptual basis for MO3 by explaining why out-of-sample estimation is necessary before the specific holdout, k-fold, and leave-one-out procedures are compared. The evaluation-metrics video ties the two strands together, clarifying what R-squared and error measures actually report, which is the judgment students need in order to decide — per MO5 — whether a model should be refined through variable removal, transformation, or a change in functional form. Taken as a set, they frame validation as the repeating diagnose-and-refine cycle this module is built around.
These videos were chosen from YouTube on October 2, 2026. Each has captions; the player turns them on. Videos on YouTube can be changed or removed — if one no longer plays, tell your instructor. If a player does not appear, use its Watch on YouTube link.
Fitting a regression line is only half the job. The least-squares estimates, the standard errors, the t-tests, the F-test, and every prediction interval you build all rest on a set of assumptions about how the data were generated: that the mean response is a linear function of the predictors, that the error variance is constant across the range of fitted values (homoscedasticity), that the errors are independent of one another, and that they are approximately normally distributed. If those assumptions hold, the inference you report is trustworthy. If they do not, the fitted line may still minimize squared error while the p-values, confidence intervals, and predictions it supports are misleading.
In this module the main diagnostic tool is the residual — the difference between an observed response and the value the model predicts. Residuals are what we actually have access to, and the way they are patterned carries information about the unobservable errors. A plot of residuals against fitted values reveals curvature (a linearity problem) and fanning or funneling (a variance problem); a plot of residuals against time or observation order reveals dependence; a histogram or normal probability plot of residuals addresses the normality condition. Learning to read these plots is a skill you build by looking at many of them.
As you watch the videos below, pay attention to three things: what each specific plot is designed to detect, what the "good" (random, structureless) version of that plot looks like, and what the failure modes look like. Also notice how the same diagnostic logic is carried out in different software environments — the computations differ, but the questions being asked of the data are identical. Finally, keep in mind the consequence side: a violation of homoscedasticity damages your standard errors and intervals, while a violation of linearity damages the fitted relationship itself, and the two call for different remedies.
A concise conceptual walkthrough of the conditions required for simple linear regression and how each one is checked in practice. Based on its title and description, it should lay out the linearity, constant-variance, independence, and normality conditions and show which graphical check corresponds to each, using residual plots as the central tool. The focus is on understanding what the conditions mean rather than on software mechanics.
Coming from a university-produced introductory statistics course, it gives a clean, accurate framing of all four assumptions in one short video, which makes it the natural starting point before moving to software-based diagnostics.
A tutor from the Stats@Liverpool support service demonstrates, step by step, how to test linear regression assumptions using Stata. Over fifteen minutes it should move through fitting a model and then producing and interpreting the diagnostic output — residual-versus-fitted plots, normality checks on residuals, and related tests — commenting on what the results indicate about the model.
It is the most detailed of the three and shows a full diagnostic workflow rather than a single plot, which is useful for seeing how linearity, homoscedasticity, and normality checks fit together in one sitting.
A short, practical demonstration of building and reading residual plots in Excel 2016, continuing from earlier videos on scatterplots and correlation. It should show where the residual output options live in Excel's regression tool and how to use the resulting plots to judge whether the assumptions are reasonable.
Excel is often the first tool engineering students reach for, and this video connects the abstract assumption checks to concrete button-clicks and to residuals you can compute yourself, directly supporting hands-on residual analysis.
Together these three videos cover the diagnostic core of the module. The conceptual video establishes what linearity, homoscedasticity, independence, and normality mean and which plot detects each, supporting MO1. The Excel and Stata demonstrations show residuals being generated and read from actual output, which builds the computation-and-interpretation skill in MO2 and the systematic plot-based workflow habit required for MO4. Because the Stata walkthrough interprets diagnostic results and comments on what they imply about the fitted model, it also starts you toward the remediation decisions in MO5. Note that these videos address assumption checking rather than resampling; the holdout, k-fold, and leave-one-out comparison in MO3 is covered by other materials in this module.
These videos were chosen from YouTube on October 2, 2026. Each has captions; the player turns them on. Videos on YouTube can be changed or removed — if one no longer plays, tell your instructor. If a player does not appear, use its Watch on YouTube link.
A regression line gives you a prediction for every observation, but the quality of a model lives in what the line misses. A residual is simply the observed value minus the predicted value — the vertical distance from a data point to the fitted line. Individually, residuals tell you how far off a single prediction was; collectively, their pattern tells you whether the model you chose was the right one at all.
In this course, residual analysis is the main practical tool for checking the assumptions behind least-squares regression. Fitting a line is easy and software will always produce one, even when a straight line is a poor description of the relationship, even when the spread of errors grows with the predictor, and even when one unusual point is dragging the slope around. Residual plots — most commonly residuals against fitted values or against the predictor — make these problems visible, because a well-behaved model should leave behind residuals that look like structureless, evenly scattered noise centered on zero.
As you watch the videos below, keep a short checklist in mind. Is there curvature or any systematic shape in the residual plot (a sign of non-linearity)? Does the vertical spread stay roughly constant across the range of fitted values (homoscedasticity) or does it fan out? Are there a few points with unusually large residuals, and would removing them noticeably change the fit (outliers versus influential points)? And do the residuals look approximately normally distributed when you histogram them or plot them on a normal probability plot? Also pay attention to the mechanics: how residuals are actually computed and how software produces these plots for you.
This lesson introduces residuals as the gap between observed and predicted values and then builds up to residual plots as a tool for judging whether a linear model is appropriate. Based on its title and accompanying handout, it works through computing residuals from a fitted line and reading residual plots for leftover patterns, using a real context involving income and SAT/ACT scores. Expect worked examples with an emphasis on interpretation rather than formulas.
Skew The Script is known for clear, context-driven statistics instruction, and this video directly covers the foundational subtopics — what residuals are, how to compute them, and how to read residuals versus fitted values to detect non-linearity.
Produced by chemical engineering faculty at CU Boulder, this video frames residual analysis as a check on model accuracy for fitted engineering data. Given the title and source, it likely demonstrates plotting residuals, identifying systematic trends that indicate the wrong functional form, and recognizing non-constant spread or unusual points as signs the model needs revision. It is aimed at an engineering audience working with experimental data.
LearnChemE is a well-established engineering education channel, so the framing and examples match this course's audience and connect residual diagnostics to deciding whether a model is adequate or needs adjustment.
A short, hands-on walkthrough of generating residual plots in Excel 2016 to check regression assumptions. It continues from earlier videos on scatterplots and correlation, so it focuses on the software steps for producing residual output and then on reading those plots for violations of linearity and constant variance.
It supplies the practical, tool-based side of the topic — actually producing residuals and residual plots in software you are likely to use — which complements the conceptual videos and supports assessing homoscedasticity and spotting outliers visually.
Together these three videos move from definition to diagnosis to execution. The first establishes what residuals are and how to compute and plot them, supporting MO2 and the linearity portion of MO1. The second applies residual analysis in an engineering modeling context, reinforcing MO1 and MO2 and pointing toward the judgment calls in MO4 and MO5 about whether a model's functional form should change. The third gives you the software workflow to generate residual plots yourself, which is the practical foundation for the systematic diagnostic routines in MO1 and MO4. Note that cross-validation (MO3) is addressed elsewhere in this module; these videos focus on the residual-based half of model validation.
These videos were chosen from YouTube on October 2, 2026. Each has captions; the player turns them on. Videos on YouTube can be changed or removed — if one no longer plays, tell your instructor. If a player does not appear, use its Watch on YouTube link.
When you fit a regression model, the fit statistics you compute on the data used for fitting (R-squared, residual standard error, training error) are optimistic. The model has already seen those points, so it can partly memorize their noise. Cross-validation is the family of techniques that answers a different and more useful engineering question: how accurately would this model predict a new observation it has never seen?
This topic covers the main approaches in increasing sophistication. The holdout method splits the data once into a training set and a test set. K-fold cross-validation rotates the role of the test set across k partitions so every observation is predicted exactly once out-of-sample, then averages the error. Leave-one-out cross-validation (LOOCV) is the extreme case where k equals the sample size. Each choice trades off bias, variance, and computational cost, and each is appropriate in different situations — a distinction you are expected to be able to make.
While watching, focus on three things: the mechanics of how data are split and reused, why the estimated test error behaves differently than the training error, and how the resulting cross-validation curve is read when comparing candidate models of different complexity. That last point connects directly to model refinement — choosing which predictors or functional forms to keep is a decision you can justify with cross-validated error rather than in-sample fit alone.
Trevor Hastie introduces cross-validation as a way to estimate test error when a separate test set is not available, contrasting it with training error and explaining why training error understates prediction error. Based on the title and the Statistical Learning curriculum, this opening lecture should cover the validation-set (holdout) approach and its drawbacks, including the variability that comes from a single arbitrary split.
It is the authoritative first lecture on the subject from Stanford Online, taught by one of the authors of the standard text, and it maps directly onto the "Why Cross-Validation Is Needed" and "Holdout Method" subtopics.
This follow-on lecture moves from the single holdout split to k-fold cross-validation, describing how the data are partitioned into k folds and how the fold-level errors are combined into one estimate of test error. Given the title and sequence, expect discussion of LOOCV as the special case k equals n, the common choices of k such as 5 or 10, and the bias-variance and computational trade-offs among them.
It covers the K-Fold and LOOCV subtopics in one place and gives you exactly the procedural comparison you need to distinguish the methods and justify a choice of k.
A compact, under-nine-minute overview that, per its description, walks through the basics of training and test data and then explains why cross-validation is needed before presenting leave-one-out and k-fold schemes. It is pitched as part of an introductory statistical learning course, so expect a straightforward, example-led treatment rather than heavy theory.
It is a fast, accessible summary that reinforces the same vocabulary as the Stanford lectures, making it useful as a preview before them or as a review afterward if the longer lectures move quickly.
These three videos concentrate on MO3: together they define the holdout split, k-fold cross-validation, and LOOCV, lay out the procedure for each, and compare them on variance, bias, and computational cost so you can state when each is appropriate. They also support MO5 indirectly, since the cross-validation error curve is the standard evidence base for deciding whether to drop a variable, change a functional form, or accept a simpler model — refinement decisions justified by out-of-sample performance rather than in-sample fit. The residual-diagnostic outcomes (MO1, MO2, MO4) are addressed by other materials in this module; use cross-validation alongside those plots, not in place of them, since a model can generalize reasonably well and still violate assumptions in ways residual analysis will reveal.
These videos were chosen from YouTube on October 2, 2026. Each has captions; the player turns them on. Videos on YouTube can be changed or removed — if one no longer plays, tell your instructor. If a player does not appear, use its Watch on YouTube link.
Fitting a regression model is only the first step; deciding whether that model can be trusted is the real engineering work. A model can have a high R-squared and still be built on unstable coefficients, be driven by a handful of unusual data points, or hide predictors that duplicate each other's information. This topic gives you a systematic way to look for those weaknesses before you report a result or use a model to make a design or process decision.
Three problems come up again and again. Multicollinearity occurs when predictor variables are strongly related to one another, which inflates standard errors and makes individual coefficient estimates unstable and hard to interpret; the variance inflation factor (VIF) is the standard numerical check. Outliers are observations with unusually large residuals — they sit far from the fitted line. Influential observations are points whose removal would noticeably change the fitted model, and these are usually points with high leverage (an unusual combination of predictor values) combined with a sizable residual. Outlier and influential are not the same thing, and keeping that distinction straight is one of the main goals of this week.
As you watch, build yourself a checklist: what quantity is being computed, what plot displays it, and what cutoff or rule of thumb signals a problem. Pay attention to how leverage, residuals, and influence relate to each other, and think about what you would actually do next — drop a variable, transform it, re-examine the data point, or collect more data. Also notice the consequences of ignoring these issues: misleading p-values, coefficients with the wrong sign, and predictions that fail on new data.
This lecture segment introduces collinearity among predictors and the variance inflation factor (VIF) as the standard diagnostic for it. Based on the title and description, it begins with exact (perfect) collinearity and a formal definition, then develops VIF and its connection to the R-squared obtained by regressing one predictor on the others. Expect the discussion to explain why correlated predictors inflate the variance of estimated coefficients.
It is a focused, university-lecture treatment of the multicollinearity subtopic, and tying VIF back to an auxiliary R-squared gives you the intuition behind the number rather than just a threshold to memorize.
A short lecture clip that separates two ideas students often merge: an outlier (a point with a large residual) and an influential observation (a point that actually moves the fitted line). The description indicates it works through when an observation has high leverage and when that leverage translates into influence on the regression fit. Expect conceptual explanation with reference to how the fitted line is constructed.
It directly addresses the "Detecting Outliers" and "Identifying Influential Observations" subtopics and clarifies a distinction that is essential before you can interpret diagnostic plots correctly.
This longer diagnostics segment centers on leverage and how it combines with residual size to change the estimated regression slope. The description points to the key case: a high-leverage point paired with a large residual produces a substantial shift in the fit. Expect coverage of leverage as a measurable quantity and its role in a broader set of regression diagnostics.
At about eleven minutes it gives the most detailed treatment of leverage and influence here, supporting both the diagnostic-plot interpretation subtopic and the consequences of leaving such points unexamined.
Together these videos target MO4 most directly: the VIF lecture supplies the numerical tool for diagnosing multicollinearity, while the outlier-versus-influence clip and the leverage diagnostics lecture supply the concepts and measures needed to flag unusual and influential observations. Because leverage and influence are read off residual-based diagnostic plots, the latter two also reinforce MO1 and MO2 by connecting residual patterns to specific model problems. Finally, knowing which weakness you have — a redundant predictor versus a single distorting observation — is what lets you choose an appropriate remedy such as variable removal or transformation, supporting MO5.
These videos were chosen from YouTube on October 2, 2026. Each has captions; the player turns them on. Videos on YouTube can be changed or removed — if one no longer plays, tell your instructor. If a player does not appear, use its Watch on YouTube link.
Validation tells you whether a regression model is trustworthy; refinement is what you do about it. This topic closes the loop: once residual plots, diagnostic statistics, and cross-validation results have revealed a problem — curvature in the residuals, non-constant variance, a redundant or collinear predictor, an influential point, or poor out-of-sample performance — you have to choose a response. Typical responses include dropping or adding variables, transforming the response or a predictor, changing the functional form, or accepting the model as it stands because it is good enough for its intended use.
In this course, refinement matters because engineering models are built to support decisions, not to win a contest for the highest R-squared. A slightly less accurate model that uses measurable, inexpensive inputs and is easy to explain may be the better engineering choice. You will also see that refinement is iterative: each change produces a new model that must be re-validated, and you need a stopping rule so the process converges rather than drifting into overfitting the data you happen to have.
While watching the videos, pay attention to three things: what evidence triggers a particular refinement, how competing models are compared on a common footing (information criteria, adjusted fit measures, or out-of-sample error), and what the trade-off is between fit and complexity. Keep asking: would this change actually make the model more useful to the engineer who has to apply it?
Brandon Foltz works through a regression model before and after applying variable transformations, using the well-known Boston housing data as the running example. Expect to see how a transformation changes the fitted relationship and the diagnostic picture, and how you judge whether the transformed model is actually an improvement.
This is the most direct coverage of the "applying transformations and functional form adjustments" subtopic, and the before-and-after structure models exactly the diagnose-change-recheck loop that iterative refinement requires. Foltz's Statistics 101 series is widely used for its slow, worked-example pacing.
This video introduces AIC, AICc, and BIC as scores for comparing multiple regression models, explaining what the acronyms stand for and how the criteria penalize added complexity. It focuses on how to read and compare these numbers across candidate models rather than on heavy derivation.
Comparing competing models and balancing fit against simplicity is central to this topic, and information criteria give you a defensible, quantitative basis for that comparison instead of eyeballing R-squared.
From Stanford Online's Statistical Learning course with Trevor Hastie, this opening lecture of the model-selection unit motivates why we do not simply keep every available predictor and then presents best subset selection as a systematic search over candidate variable sets. It frames variable selection in terms of prediction accuracy and model interpretability.
It supplies the conceptual backbone for variable selection and model simplification from the authors of the standard reference text, which complements the more hands-on treatment in the other two videos.
These videos extend the diagnostic work of MO1, MO2, and MO4 into action: once residual patterns or collinearity have been identified, the Foltz transformation video shows how a functional-form change is applied and re-evaluated, and the Hastie lecture shows how redundant predictors are systematically removed — together supplying the specific refinement strategies called for in MO5. The AIC/AICc/BIC video gives you a principled way to compare the resulting candidate models and to decide when added complexity is not paying for itself, which reinforces the out-of-sample thinking behind MO3 and supports judgments about when a model is sufficient for its intended engineering application.
These videos were chosen from YouTube on October 2, 2026. Each has captions; the player turns them on. Videos on YouTube can be changed or removed — if one no longer plays, tell your instructor. If a player does not appear, use its Watch on YouTube link.