Absolute Effects and Relative Outcome Changes

Two Ways to Describe an Outcome Change

A regression estimate can describe a change in an outcome in its original units, such as a difference of 2.4 beats per minute or $2,400 per year. That is an absolute change. The same difference can sometimes be expressed as a percentage of a baseline outcome. That is a relative change. These descriptions answer different questions: the first asks how many outcome units separate two predictions; the second asks how large that difference is compared with a stated reference amount.

A percentage change is not determined by the difference alone. It also depends on the baseline used as the denominator. A difference of 5 units is 10% of a baseline of 50, but 5% of a baseline of 100. Therefore, a relative change should always be reported with its baseline, and the absolute difference should usually be reported alongside it.

Definition: For a baseline outcome $Y_0$ and a comparison outcome $Y_1$, the absolute change is $Y_1-Y_0$. When $Y_0$ is positive and a percentage comparison is meaningful, the relative percentage change is $\frac{Y_1-Y_0}{Y_0}\times 100\%$. The denominator is the baseline outcome, not the comparison outcome.

Regression Contrasts on the Outcome Scale

Consider a fitted regression model with outcome $Y$, predictor $x$, and another predictor $z$:

E(Y\mid x,z)=\beta_0+\beta_1x+\beta_2z.

Here, $E(Y\mid x,z)$ is the modelled mean outcome for observations with predictor values $x$ and $z$. In this additive model, changing $x$ by $\Delta x$ while holding $z$ fixed changes the fitted mean by $\beta_1\Delta x$. In practice, we use the estimated coefficient $\hat{\beta}_1$, so the estimated absolute contrast is $\hat{\Delta}=\hat{\beta}_1\Delta x$. Its units are outcome units. For example, if $Y$ is measured in dollars per year, the contrast is also in dollars per year.

The absolute contrast can be converted to a relative percentage only after a baseline mean has been specified. That baseline might be an observed reference mean or a fitted mean for a stated predictor profile. The comparison outcome is the baseline plus the contrast. The percentage calculation uses the baseline as its denominator.

Model conditions for this interpretation: The conditional mean is assumed to follow the stated regression form over the predictor values being compared, with errors having mean zero given the predictors. The additive equation also assumes the effect of changing $x$ is constant across values of $z$; if an interaction is included, the contrast must instead be calculated from the full model at stated values of the other predictors. Standard regression inference additionally relies on appropriate sampling and error conditions for the method used. A fitted contrast describes a modelled difference; by itself, it does not establish a causal effect.

The percentage formula follows directly from expressing the absolute difference as a fraction of the reference amount. If the baseline is $Y_0$ and the comparison is $Y_1$, then the difference is $Y_1-Y_0$. Dividing by $Y_0$ asks what fraction of the baseline that difference represents; multiplying by 100 converts the fraction to a percentage:

\text{Relative percentage change} = \frac{Y_1-Y_0}{Y_0}\times 100\%.

Equivalently, if the relative change is $r\%$, then $Y_1=Y_0(1+r/100)$. For a decrease, $r$ is negative. For instance, a change of $-5\%$ means the comparison outcome is 95% of the baseline. This arithmetic does not make the change causal; interpretation still depends on the study design and model.

Worked Example: An Absolute Difference and a Relative Decrease

Suppose a hypothetical regression model estimates that, for each additional hour of weekly exercise, mean resting pulse is 0.8 beats per minute lower, with other predictors held fixed. For a comparison involving 3 additional hours, the fitted baseline mean for the stated reference profile is 76 beats per minute. Calculate and interpret both changes.

Absolute contrast. The estimated slope is $-0.8$ beats per minute per additional exercise hour. For $\Delta x=3$ hours:

\hat{\Delta}=(-0.8\ \text{beats per minute per hour})(3\ \text{hours}) =-2.4\ \text{beats per minute}.

The comparison prediction is $76-2.4=73.6$ beats per minute. The modelled mean is therefore 2.4 beats per minute lower for the predictor contrast, at the same values of the other predictors.

Relative contrast. The baseline is 76 beats per minute, so:

\frac{73.6-76}{76}\times 100\% = \frac{-2.4}{76}\times 100\% \approx -3.16\%.

Thus, the fitted mean is 2.4 beats per minute lower, a decrease of about 3.16% relative to the stated baseline of 76 beats per minute. This is a modelled association, not evidence that adding exercise causes a pulse reduction.

Choosing and Reporting the Baseline

The baseline must correspond to the comparison being described. In regression, a clear approach is to state the reference predictor profile, identify its fitted mean outcome, and then calculate the percentage relative to that mean. If a different profile has a different baseline mean, the same absolute contrast can produce a different relative percentage. That is not an arithmetic contradiction: the denominator has changed.

A baseline should also be appropriate for the outcome's scale. Relative percentages are often useful for positive quantities where ratios have a natural interpretation, such as costs, counts with a meaningful positive reference, or durations. They are generally not suitable when the baseline is zero, negative, or so close to zero that a small absolute difference creates an enormous and unstable percentage.

Worked Example: A Salary Contrast Relative to a Stated Baseline

In a hypothetical regression analysis of annual salary, a model estimates a difference of $2,400 per year between two stated training levels, with the other predictors held fixed. The fitted baseline salary for the reference profile is $48,000 per year. Calculate the comparison prediction and its relative change.

Absolute contrast. The estimated salary difference is $+\$2,400$ per year. The comparison prediction is:

\$48{,}000+\$2{,}400=\$50{,}400\ \text{per year}.

Relative contrast. Using the reference salary of $48,000 per year as the denominator:

\frac{\$50{,}400-\$48{,}000}{\$48{,}000}\times 100\% = \frac{\$2{,}400}{\$48{,}000}\times 100\% =5.00\%.

The fitted difference is $2,400 per year, or 5.00% of the stated $48,000 baseline. The percentage is not an additional salary difference; it is the same difference expressed relative to that baseline. The calculation alone does not show that training caused the difference.

If the same $2,400 difference were compared with a $60,000 baseline, the relative change would be $2,400/60,000\times100\%=4.00\%$. Reporting only “a 5% increase” without identifying the baseline would leave an important part of the comparison unstated.

Outcomes Already Expressed as Percentages

When the outcome itself is a percentage or proportion, an absolute difference and a relative percentage change must still be distinguished. Subtracting two rates gives a difference in percentage points. Dividing that difference by the baseline rate gives a relative percentage change. These quantities have different denominators and different meanings.

Worked Example: Percentage Points Versus Relative Percent Change

Suppose a hypothetical regression model gives a fitted outcome of 12% for a reference profile and 15% for a comparison profile. Treat these as fitted rates for the stated profiles. Find the absolute difference and the relative change from the 12% baseline.

Absolute difference. Subtract the rates:

15\%-12\%=3\ \text{percentage points}.

Relative percentage change. The baseline rate is 12%, so:

\frac{15\%-12\%}{12\%}\times 100\% = \frac{3\%}{12\%}\times 100\% =25\%.

The fitted rate is 3 percentage points higher, which is a 25% increase relative to the 12% baseline. Saying “3% higher” would be ambiguous: it could be mistaken for a 3% relative increase rather than a difference of 3 percentage points.

When a Relative Percentage Is Not Useful

A zero baseline makes the relative percentage formula undefined because division by zero is not defined. A negative baseline can produce a numerical ratio, but the usual language of “percent increase” or “percent decrease” may be confusing because the sign and direction of the outcome do not behave like a positive reference quantity. With a baseline close to zero, even a small absolute difference can yield a very large relative percentage. In these cases, report the absolute contrast and explain why a relative percentage is not suitable.

For example, if a fitted count rises from 0 to 2 cases, the absolute difference is 2 cases, but $(2-0)/0$ has no defined value. It would be incorrect to describe this as a finite percentage increase from zero. The absolute scale communicates the change without inventing a denominator.

Common Mistakes

Leaving out the baseline. A relative change is incomplete unless readers know the amount against which the difference is compared. State the baseline value and how it was chosen.
Using the comparison outcome as the denominator. The standard relative change from a baseline uses $(Y_1-Y_0)/Y_0$, not $(Y_1-Y_0)/Y_1$.
Confusing percentage points with percent change. A change from 12% to 15% is 3 percentage points and a 25% relative increase from the 12% baseline. These are not interchangeable.
Reporting a percentage when the baseline is zero. Division by zero is undefined. Report the absolute difference instead.
Implying causation from a regression contrast. A modelled difference is not automatically a causal effect. Describe it as a fitted or estimated association unless the study design supports a causal interpretation.
Reporting only a relative percentage. A percentage can make a small absolute difference sound large, or conceal a substantial one. Give the absolute difference in outcome units as well.

Key takeaway: First calculate the regression contrast in the outcome's original units. Then, if a relative comparison is meaningful, divide by an explicitly stated baseline and multiply by 100%. Keep the absolute difference, the relative percentage, and the baseline distinct in both calculation and wording.

Check Your Understanding

For each question, state the baseline clearly and distinguish absolute from relative change.

A fitted outcome rises from 80 units to 84 units. What is the absolute change, and what is the relative percentage change from the baseline?
A rate changes from 7% to 9%. Express the difference in percentage points and as a relative percentage change from the 7% baseline.
A model estimates a difference of 6 minutes, with a reference mean of 120 minutes. What percentage of the baseline is the difference?
Why is a relative percentage change from a baseline of zero undefined, even when the comparison outcome is known?
Two groups have the same estimated absolute difference, but different positive baseline means. Explain why their relative percentage changes can differ.