18.2 Interpreting and Communicating Results

Interpreting Linear Model Results

Understanding Coefficients

The slope coefficient tells you how much the response variable changes for a one-unit increase in the predictor, holding all other predictors constant. In a model predicting house prices from square footage, a slope of 50 means each additional square foot adds $50 to the predicted price, with other factors held fixed.

The intercept is the predicted value of the response when every predictor equals zero. In the house price example, an intercept of $100,000 would be the predicted price at zero square feet. That's not a realistic house, which highlights an important point: the intercept often serves as a mathematical anchor for the regression line rather than a meaningful standalone prediction. Always check whether zero is a sensible value for your predictors before interpreting the intercept literally.

Assessing Model Fit and Performance

• $R^2$ (coefficient of determination) measures the proportion of variance in the response variable explained by the model. An $R^2$ of 0.75 means the predictors account for 75% of the variability in the response.
• Adjusted $R^2$ corrects for the number of predictors. Plain $R^2$ can only increase (or stay the same) when you add a predictor, even a useless one. Adjusted $R^2$ penalizes unnecessary predictors, so it gives a more honest picture of explanatory power when comparing models of different sizes.
• Standard error of the estimate measures the average distance between observed and predicted values. Smaller is better. If your model predicting house prices has a standard error of $15,000, that's the typical size of its prediction errors.

Significance Testing and Confidence Intervals

The F-test evaluates the model as a whole. It asks: does this set of predictors, taken together, explain a significant portion of the variance in the response? A small p-value (typically < 0.05) means the model fits meaningfully better than a model with no predictors at all.

Individual t-tests do the same thing for each predictor separately. A small p-value for a particular coefficient means that predictor contributes significantly to the model, after controlling for the other predictors. A large p-value doesn't necessarily mean the variable is unimportant in general; it means it doesn't add much given the other variables already in the model.

Confidence intervals give a range of plausible values for the true population coefficient. A 95% confidence interval of [40, 60] for a slope means that, based on the sample data, you can be 95% confident the true slope falls in that range. If the interval doesn't contain zero, the coefficient is statistically significant at the 0.05 level.

Communicating Linear Model Findings

Tailoring Presentations to the Audience

Different audiences need different levels of detail:

• Non-technical audiences care about practical implications. Lead with what the results mean for decisions, and use real-world language. Instead of "the coefficient is 0.34 with $p < 0.01$," say "improving service quality by one point on our scale is associated with a meaningful increase in customer satisfaction."
• Technical audiences expect detail on the modeling process, assumption checks, and statistical measures. You can reference $R^2$, p-values, and diagnostics directly.

Regardless of audience, start with a clear summary of the research question, the data (source, sample size, key variables), and the methods used. State the objective up front so the audience knows what problem you're solving before you present numbers.

Interpreting and Discussing Results

Translate coefficients into the language of the problem. In a model predicting customer satisfaction from service quality factors, don't just report that "responsiveness has a coefficient of 0.45." Explain that a one-unit improvement in responsiveness is associated with a 0.45-point increase in satisfaction, holding other factors constant, and discuss whether that's a practically meaningful change.

Every model has limitations, and acknowledging them builds credibility:

• Note any assumption violations (non-linearity, heteroscedasticity, non-normal residuals) and how they might affect your conclusions.
• Address data limitations such as sample size, how the sample was collected, and whether the results generalize beyond the study population.
• Be honest about what the model cannot tell you. Correlation from a regression does not establish causation unless the study design supports causal claims.

Close with actionable implications. How should the findings inform strategy, policy, or future research? What follow-up questions does the analysis raise?

Engaging the Audience

Visuals make regression results far more accessible:

• Scatter plots with fitted lines show the relationship between predictor and response at a glance.
• Coefficient plots (point estimates with confidence interval bars) let viewers quickly compare the size and significance of different predictors.

Build in time for questions and discussion. Invite the audience to consider how the findings apply to their own context. This turns a one-way presentation into a conversation and often surfaces practical insights the analyst might miss.

Visualizing Linear Model Outcomes

Diagnostic Plots

Before trusting your model's results, check its assumptions visually:

1. Scatter plots with the fitted line help you assess whether the relationship is actually linear. Look for curves or clusters that a straight line misses.
2. Residuals vs. fitted values is the workhorse diagnostic plot. You want a random cloud of points centered around zero. Patterns (funnels, curves, clusters) signal problems like heteroscedasticity or non-linearity.
3. Q-Q plots compare residual quantiles against a theoretical normal distribution. Points falling along the diagonal line indicate approximately normal residuals. Systematic departures at the tails suggest skewness or heavy tails.
4. Scale-location plots ($\sqrt{|\text{standardized residuals}|}$ vs. fitted values) specifically help detect non-constant variance. A flat trend line is what you want; an upward slope suggests variance increases with fitted values.

Visualizing Variable Relationships and Effects

• Correlation heat maps display pairwise correlations among predictors using color gradients. They're a quick way to spot multicollinearity. If two predictors are highly correlated (say $|r| > 0.8$), their individual coefficients become unstable and hard to interpret.
• Coefficient plots / forest plots show each predictor's point estimate and confidence interval side by side. This makes it easy to compare effect sizes and see which predictors are significant (intervals that don't cross zero).
• Interaction plots visualize how the effect of one predictor on the response changes across levels of another predictor. Non-parallel lines indicate an interaction. These are especially useful for explaining interaction terms to audiences who find the raw coefficients abstract.

Effective Communication through Visualizations

Good visuals require attention to detail:

• Use clear, descriptive titles and axis labels. "Residuals vs. Fitted Values" is better than "Plot 2."
• Choose color schemes that are readable in grayscale and accessible to colorblind viewers.
• Add concise annotations to highlight key takeaways directly on the plot (e.g., labeling an outlier or marking where a confidence interval crosses zero).
• Avoid chart junk. Every element on the plot should help the viewer understand the data, not just decorate the slide.

Simplifying Linear Modeling Concepts

Explaining Key Terms and Concepts

A linear relationship means the response and predictor are connected by a straight-line pattern with a constant rate of change. Think of study time and exam scores: if each extra hour of studying adds roughly the same number of points to your score, that's a linear relationship.

The slope is that constant rate of change. In a model predicting sales from advertising spend, a slope of 10 means every additional dollar of advertising is associated with 10 more units sold.

The intercept is the predicted baseline. In the sales example, an intercept of 1,000 represents expected sales when advertising spend is zero. But be careful: if zero advertising is unrealistic in your data, the intercept is just a mathematical necessity of the line, not a meaningful prediction.

Goodness-of-Fit and Hypothesis Testing

$R^2$ answers the question: how much of the variation in the response does this model capture? Think of it as a percentage score for your model. An $R^2$ of 0.80 means the model explains 80% of the variability; the remaining 20% is unexplained noise or the influence of variables not in the model.

P-values address a different question: could these results have happened by chance alone? Formally, a p-value is the probability of seeing results at least as extreme as yours if the null hypothesis (no real effect) were true. A small p-value means chance is an unlikely explanation. The analogy of flipping a coin helps here: getting 10 heads in a row is possible with a fair coin, but the probability is so low ($p \approx 0.001$) that you'd start suspecting the coin is rigged. That's the logic behind significance testing.

Communicating Uncertainty and Limitations

Confidence intervals are like the margin of error in a poll. A point estimate gives you one number; the confidence interval tells you the range of values that are plausible given sampling variability. Reporting "the slope is 50, with a 95% CI of [40, 60]" is more informative than reporting the slope alone, because it conveys how precise (or imprecise) the estimate is.

When presenting to non-technical audiences, swap jargon for plain language:

• Say "strength of the relationship" instead of "coefficient of determination."
• Say "the range of likely values" instead of "confidence interval."
• Define any technical terms you do use, briefly and in context.

Visuals like annotated charts or simple infographics can make uncertainty tangible. A bar chart with error bars, for instance, shows both the estimate and its uncertainty in a single image, which is far more intuitive than a table of numbers.