Modeling Lingerie Appeal with Multiple Regression

Introduction

Welcome to our first lesson on supervised learning. In the previous module, we explored unsupervised methods, using PCA and hierarchical clustering to discover inherent structures and group your lingerie pieces into meaningful style families like 'minimalist' or 'boudoir'. We answered the question, "What groups exist in my data?" Now, we pivot to a different question: "How do specific design features influence a lingerie piece's success?"

Today, we will build a predictive model using one of the most fundamental tools in statistics: Multiple Linear Regression (MLR). You will learn how to develop and interpret an Ordinary Least Squares (OLS) model that quantifies the effect of tangible design features—such as the percentage of lace used or the saturation of its color—on a continuous outcome, like a photoshoot appeal score. This moves us from simply categorizing styles to understanding the drivers of their performance. Given your background in statistics, we will focus on the practical application and interpretation in the context of lingerie design, rather than re-deriving the underlying theory.

From Simple to Multiple Regression

You are familiar with simple linear regression, which models the relationship between a single predictor and an outcome. Multiple linear regression is the natural extension of this, allowing us to model an outcome based on several predictors simultaneously. This is crucial for our work, as the appeal of a lingerie piece is rarely driven by a single feature. Instead, it's a combination of fabric, cut, color, and construction.

The model takes the following form:

Y = β_{0} + β_{1} X_{1} + β_{2} X_{2} + \dots + β_{p} X_{p} + ϵ

In our lingerie design context, this translates to:

$Y$ : The outcome variable we want to predict, such as an Appeal_Score (e.g., on a scale of 1-100).
$X_{1}, X_{2}, \dots, X_{p}$ : The predictor variables, which are our quantifiable design features (e.g., Lace_Percentage, Strap_Width_mm, etc.).
$β_{0}$ : The intercept, representing the baseline appeal score when all predictor variables are zero.
$β_{1}, β_{2}, \dots, β_{p}$ : The coefficients, which quantify the relationship between each design feature and the appeal score.
$ϵ$ : The error term, capturing the variation in the appeal score not explained by our model.

The "Complete Guide to Multiple Linear Regression" provides a concise overview of this extension from simple to multiple regression.

Multiple Linear Regression: Complete Guide with Formulas ...

This guide will serve as a quick refresher on the core concepts of MLR, its advantages, and its mathematical foundation.

First, read the introductory section, Multiple linear regression. This part frames the model as a tool for understanding complex, multi-factor relationships. Next, review the Model Specification and Matrix Notation. Given your background, the matrix form \mathbf{y} = \mathbf{X}\boldsymbol{\beta} + \boldsymbol{\epsilon} will be familiar. Pay close attention to the structure of the design matrix \mathbf{X}, particularly the leading column of ones for the intercept. Finally, look at the OLS solution, \hat{\boldsymbol{\beta}} = (\mathbf{X}^\top \mathbf{X})^{-1} \mathbf{X}^\top \mathbf{y}, to refresh your memory on how these coefficients are estimated.

The Art of Interpreting Coefficients

The true power of regression in a design context comes from interpreting the coefficients. Each coefficient, $β_{j}$ , tells us the average change in the outcome variable for a one-unit increase in the predictor variable $X_{j}$ , holding all other predictor variables in the model constant. This last part—"holding all other variables constant"—is the most critical and powerful aspect of MLR. It allows us to isolate the impact of a single design choice.

The following short video provides an excellent and memorable rule for coefficient interpretation.

How to Interpret a Coefficient in a Regression

This video from Dr. HK offers a clear, universal template for interpreting any regression coefficient.

Watch the first two minutes, focusing on the main sentence to memorize for interpretation from the beginning and the explanation of what "controlling for" other variables means in practice, from this section.

A Lingerie Design Example

Let's imagine we've collected data on 50 lingerie pieces for a boudoir photoshoot and we fit a model to predict their Appeal_Score (rated 1-100 by a stylist). Our predictors are:

Sheerness_pct: The percentage of the garment that is sheer (0-100).
Strap_Width_mm: The width of the bra straps in millimeters (e.g., 5mm to 25mm).
Is_Balconette: A binary (dummy) variable, where 1 = Balconette style and 0 = Other style.

Suppose our OLS regression output gives us the following equation:

Appeal_Score = 45.2 + 0.3 \times (Sheerness_pct) - 1.5 \times (Strap_Width_mm) + 12.1 \times (Is_Balconette)

Here is how we would interpret each component:

Intercept ( $β_{0} = 45.2$ ): The model predicts a baseline appeal score of 45.2 for a lingerie piece that is not a balconette, has 0% sheerness, and straps of 0mm width. As is often the case, the intercept here is a statistical baseline rather than a practically meaningful scenario.
Sheerness ( $β_{1} = 0.3$ ): For every one-percentage-point increase in sheerness, the appeal score is expected to increase by 0.3 points, holding strap width and bra style constant. This suggests that, all else being equal, more sheer designs are slightly more appealing.
Strap Width ( $β_{2} = - 1.5$ ): For every 1mm increase in strap width, the appeal score is expected to decrease by 1.5 points, holding sheerness and bra style constant. This implies a strong preference for more delicate, thinner straps in this context.
Balconette Style ( $β_{3} = 12.1$ ): A balconette bra is predicted to have an appeal score 12.1 points higher than a non-balconette bra, assuming they have the same sheerness and strap width. This quantifies the powerful premium this particular style has on appeal.

This level of interpretation provides directly actionable insights. If you were deciding between two otherwise identical designs, one with 10mm straps and one with 15mm straps, this model predicts the one with thinner straps would score $5 \times 1.5 = 7.5$ points higher.

Association is Not Causation

A crucial point, especially as we prepare for our later module on causal inference, is that these coefficients represent association, not necessarily causation. The model shows that balconette bras are associated with higher scores, but it doesn't prove the balconette design causes the higher score. There could be a confounding factor; perhaps balconettes are more often made with premium fabrics, and it's the fabric, not the style, driving the appeal.

This video makes this important distinction clear.

How to Interpret a Coefficient in a Regression

The same video also explains why we must be careful with our language when interpreting regression results.

Please watch the segment on causal effects. The speaker's emphasis on using "is associated with" instead of "causes" is a critical habit for rigorous analysis.

A Real-World Application in Aesthetics

This method isn't just theoretical. Researchers use it to understand what drives visual appeal. A study by Thömmes & Hübner (2020) did something very similar to our goal: they tried to quantify the drivers of aesthetic appeal in Instagram photos.

[PDF] A New Measure for Aesthetic Appeal Derived From Instagram Data

This research paper provides a direct parallel to our task. The authors create an "Image Aesthetic Appeal" (IAA) score and then use multiple regression to see which image features predict it.

Please read the final paragraph of the paper, in the section "Discussing the Empirical Validation." Start at the summary paragraph. Notice how they calculate a regression to predict IAA scores based on gender, balance, and curvature. They conclude that "gender has the largest effect on aesthetic appeal, while balance and curvature play a much smaller role." This is exactly the kind of quantitative conclusion we aim to draw for our lingerie designs.

This study validates our approach: by modeling an appeal score as a function of design attributes, we can gain a quantitative understanding of which features are most important.

Numerical Exercise

You have run a regression analysis on a small dataset of 15 lingerie sets to predict their Photoshoot_Rating (a score from 1 to 10 based on test shots). The model includes three features:

Color_Saturation: A measure from 0 (grayscale) to 1 (fully saturated).
Lace_Coverage_pct: The percentage of the surface area covered in lace (0-100).
Is_Thong: A dummy variable (1 if the bottom is a thong, 0 otherwise).

The fitted OLS model is:

Photoshoot_Rating = 2.1 + 4.5 \times (Color_Saturation) + 0.02 \times (Lace_Coverage_pct) + 1.8 \times (Is_Thong)

Questions:

What is the full interpretation of the coefficient for Is_Thong?
What is the full interpretation of the coefficient for Color_Saturation?
Based on this model, which has a larger impact on the rating: increasing lace coverage from 0% to 50%, or choosing a thong bottom over a different style? Justify your answer with calculations.
Calculate the predicted Photoshoot_Rating for a lingerie set with a color saturation of 0.8, 30% lace coverage, and which is not a thong.

Click for a detailed solution.

Interpretation of Is_Thong coefficient (1.8): Holding color saturation and lace coverage constant, a lingerie set with a thong bottom is predicted to have a photoshoot rating that is 1.8 points higher than a set with a non-thong bottom.
Interpretation of Color_Saturation coefficient (4.5): For every 0.1 unit increase in color saturation (e.g., from 0.5 to 0.6), the photoshoot rating is expected to increase by 0.45 points, holding lace coverage and bottom style constant. Alternatively, for a full 1-unit increase from grayscale (0) to fully saturated (1), the rating is expected to increase by 4.5 points.
Impact Comparison:
- Effect of Lace: Increasing lace coverage from 0% to 50% changes the rating by $50 \times 0.02 = 1.0$ point.
- Effect of Thong: Choosing a thong over a non-thong changes the rating by 1.8 points.
- Conclusion: Choosing a thong bottom has a larger predicted impact on the rating than adding 50% lace coverage.
Prediction Calculation:
- We use the equation with Color_Saturation = 0.8, Lace_Coverage_pct = 30, and Is_Thong = 0.
- $Rating = 2.1 + 4.5 \times (0.8) + 0.02 \times (30) + 1.8 \times (0)$
- $Rating = 2.1 + 3.6 + 0.6 + 0$
- $Rating = 6.3$
- The predicted rating for this set is 6.3.

Conclusion

In this lesson, you have learned how to construct and, more importantly, interpret a multiple linear regression model to quantify the influence of various design features on a lingerie piece's appeal. This technique transforms subjective design choices into measurable factors, allowing for data-informed curation and creation.

Key Takeaways:

Multiple linear regression models an outcome using several predictors simultaneously.
The OLS method finds the coefficients that best fit the data.
Each coefficient represents the expected change in the outcome for a one-unit change in a predictor, holding all other factors in the model constant.
This "ceteris paribus" interpretation allows you to isolate the impact of individual design decisions.
Remember that regression coefficients describe association, not causation.

Building a model is the first major step. However, a model is only useful if it is reliable. In our next lesson, we will address the critical task of model diagnostics. You will learn how to test for common issues like multicollinearity, heteroscedasticity, and non-normality of residuals, ensuring that the coefficients you interpret are stable and trustworthy.

Recommended resources