BIO 500 Week 6 DQ 2: What Regression Minimises

The answer to the literal question is short: the fitted line minimises the sum of the squared residuals, where each residual is the observed value of y minus the value the line predicts for that same x. You should state that precisely in your opening sentence, using the word residual and the word squared, because both are load-bearing. But a one-sentence answer is not a discussion post, and since everyone on the board can produce that sentence, the grading falls entirely on what you do next. The productive move is to treat the question as asking why that particular quantity is the one being minimised, which turns a definition into an argument and gives you four or five paragraphs of real material. The prompt supports that reading, since it puts the word fit in quotation marks and so signals that fitting is the thing under examination rather than a step to be reported.

Start with what the residual measures, since the word predicted in the prompt is a clue. The line produces a fitted value for every x, and the residual is the vertical gap between the observation and that fitted value. Vertical is not an incidental detail. Regression treats x as known and asks how well y is predicted from it, so error is defined only in the y direction; the perpendicular distance from a point to the line, which many people picture, belongs to a different technique that treats both variables as measured with error. Saying this explicitly is one of the cheapest ways to demonstrate genuine understanding, because it addresses a mental image most students carry without examining. It also explains something that would otherwise look arbitrary, which is why swapping the roles of x and y produces a different line rather than the same one rearranged.

Then explain why the deviations are squared rather than simply added, and be careful to give both reasons rather than one. The first is that residuals are signed and, for a line fitted through the data, sum to zero, so minimising their plain total would be satisfied by any number of terrible lines. Some transformation that removes the sign is therefore required. The second is that squaring is a choice among the available transformations: it penalises a large miss disproportionately more than several small ones, and it produces a smooth criterion with a unique closed-form solution, which is why least squares was tractable long before computers. Absolute deviations would also remove the sign and give a different, perfectly respectable line. That method exists and is used when resistance to outliers matters more than closed-form convenience, so squaring is a trade rather than a truth.

The most valuable thing you can add is the caution that minimising this quantity does not mean the line is any good. Least squares always returns an answer. It will fit a straight line to a curve, to a cloud with no relationship at all, and to a dataset dominated by one extreme point, and in each case it returns the best straight line available while the straight line itself is the wrong model. This is why a fitted regression has to be followed by inspection: a residual plot that should show no pattern, an assessment of whether an outlier is dragging the slope, and a look back at the scatter plot the prompt itself mentions. The prompt puts the word fit in quotation marks, which invites exactly this observation. Anscombe's four datasets are the standard demonstration, since they share a regression line and summary statistics while looking nothing like one another on a scatter plot.

Be precise about the coefficient of determination if you mention it, since it is where good posts often slip. R squared is the proportion of variance in y accounted for by the model, and it is related directly to the quantity being minimised — it compares the residual sum of squares against the total sum of squares around the mean of y. What it is not is a test of whether a linear model is appropriate, and a high value can coexist with a badly curved residual pattern. Stating the relationship between R squared and the minimised sum is a genuinely on-topic contribution, since it shows the same quantity reappearing as a summary rather than introducing an unrelated statistic. It also keeps you inside the question, which asked about a minimised relation rather than about goodness of fit in general.

Finally, connect the criterion to what regression is for in a public health course, because the abstract answer benefits from one concrete anchor. When a fitted line is used to predict an outcome from an exposure or a biomarker, the residual sum of squares is what the estimate of prediction error is built from, and the standard error of the slope — which decides whether the relationship is distinguishable from noise — descends from the same quantity. So the thing being minimised is not merely a fitting rule; it is the raw material for every subsequent statement about how much the line can be trusted. One sentence of that kind stops the post reading as a definition recited and starts it reading as a method understood. It also gives your classmates something to reply to, which is the practical difference between a post that closes a topic and one that opens it.