Any advice? I prefer coefficient plots, but they can be hard to interpret without some guidance. What methods have you all found effective?
Essentially, how can I best illustrate
A) which variables are significant
B) the magnitude of their impact (size of coefficients)
Hi Nicholas!
When illustrating which variables are significant and the magnitude of their impact (size of coefficients) in your analysis, coefficient plots can indeed be very useful.
Here are some effective methods and considerations:
Coefficient Plots: Plotting coefficients with confidence intervals can visually depict which variables have statistically significant effects. Significant variables are typically those whose confidence intervals do not include zero.
P-Values: Include p-values alongside coefficient plots. Variables with p-values below a certain threshold (commonly 0.05) are considered statistically significant.
Forest Plots: Similar to coefficient plots but arranged vertically, forest plots display coefficients and their confidence intervals for each variable, making it easy to see which variables have significant effects.
Standardized Coefficients: Standardizing coefficients (e.g., using z-scores) can help directly compare the magnitude of impact across variables, as they represent the effect size in standard deviation units.
Partial Regression Plots: Also known as added variable plots or component plus residual plots, these plots show the relationship between a predictor and the response variable after accounting for the effects of other variables. This can help visualize the specific impact of each variable.
Bar Charts: For simpler visualizations, bar charts of coefficients (with error bars indicating confidence intervals) can show the relative size of each variable’s effect.
By using these methods, you can effectively communicate both the significance and the magnitude of variables in your analysis, providing clear insights to your audience about the factors influencing your outcomes.
I concur. Their comprehension and attention span will improve the more “infographics-like” they are. At times, I just obtain the data and outcomes and utilize Photoshop to create graphics with pertinent icons to better serve the audience (such as a logistic regression about diaper leaks; drawing some wet and other leaky diapers was a lot of fun and well-received).
To be more clear, don’t say “we fit a linear regression model to the data by solving for the coefficients that minimize the sum of the squared residuals.” Instead, say “this is a mathematical model that predicts the value of Y based on X.”
Based on my research, when presenting linear regression results to a non-technical audience, I would focus on visualizing the key insights in a clear, easy-to-understand way rather than diving into the technical details. I would start by showing a simple line chart that compares the predicted values from the regression model against the actual observed data. This allows the audience to quickly grasp how well the model is performing without needing to understand the underlying mathematics. I would also discuss the model’s RMSE (root mean squared error) in plain business terms, explaining whether the level of error is acceptable for their needs. For the key variables driving the regression, I would highlight the top 5-10 most impactful factors using a coefficient plot or bar chart. I would explain in plain language how a one-unit change in each variable affects the outcome, avoiding statistical jargon. Visuals like icons or infographics can also help convey the relationships intuitively. The goal is to focus on the practical implications and actionable insights from the regression analysis, rather than the technical details. By using clear, accessible language and impactful visualizations, I can effectively communicate the key findings to my non-technical stakeholders.
Ensure you have much to say about how the predicted associations you discovered came about and whether they may be leveraged to target higher NPS. There is nothing worse than an analyst who effectively states We have fitted a curve, here it is
To describe the fitted curve,
An ordered logistic regression model
I would calculate and visualize how the projected probability of the categories vary in response to relevant predictor contrasts and on average. There are no odds, odds ratios, etc.