Interpreting Logistic Regression Coefficients as Odds Ratios

Introduction

In our previous session, we established the foundation for logistic regression, understanding why it's the appropriate tool for binary outcomes like classifying a lingerie piece as a 'Top Pick' or not. We formulated the model, which links the log-odds of success to a linear combination of design features, and practiced calculating a predicted probability. Now, we'll move to the crucial next step: interpreting what the model's coefficients tell us.

Your expertise in econometrics and statistical modeling provides a strong basis for this. Just as you would interpret coefficients in a linear model to understand marginal effects, we will learn to interpret logistic regression coefficients to understand their impact on the odds of an outcome. This lesson focuses on converting the model's $β$ coefficients into odds ratios, a powerful and intuitive way to explain how each design feature influences a lingerie piece's chances of being selected for the photoshoot.

From Log-Odds to Odds Ratios

Recall the core equation of our logistic regression model:

ln (\frac{P}{1 - P}) = ln (Odds) = β_{0} + β_{1} X_{1} + \dots + β_{k} X_{k}

The coefficients ( $β_{1}, \dots, β_{k}$ ) represent the change in the log-odds for a one-unit change in the corresponding predictor variable, $X_{k}$ , holding all other predictors constant.

While mathematically correct, "change in log-odds" is not very intuitive for communicating insights. We want to talk about the change in odds. To do this, we use the inverse of the natural logarithm: exponentiation. If we consider the effect of a single predictor $X_{1}$ , a one-unit increase in $X_{1}$ changes the log-odds by $β_{1}$ . To see how this affects the odds themselves, we exponentiate the coefficient.

The resulting value, $e^{β_{1}}$ , is the Odds Ratio (OR). It represents the factor by which the odds of the outcome change for a one-unit increase in the predictor.

If $β_{1} > 0$ , then $e^{β_{1}} > 1$ , meaning the odds of the outcome increase.
If $β_{1} < 0$ , then $e^{β_{1}} < 1$ , meaning the odds of the outcome decrease.
If $β_{1} = 0$ , then $e^{β_{1}} = 1$ , meaning the predictor has no effect on the odds.

The following resource provides a clear mathematical derivation showing how the odds ratio is simply the exponential of the predictor's coefficient.

A Deeper Dive into Odds Ratios Using Logistic Regression

This article section walks through the algebra of how the odds ratio is derived from the logistic regression equation.

Please read the section titled From Logit to Odds Ratio. The key insight is seeing how dividing the odds for an "exposed" group (e.g., X=1) by the odds for an "unexposed" group (X=0) isolates e^{\beta_1} as the odds ratio.

Let's apply this to our lingerie design model from the previous lesson, using the same fitted coefficients:

ln (Odds of Top Pick) = - 4.0 + 0.05 \cdot Lace_Coverage_Pct + 1.5 \cdot Is_Underwired - 0.5 \cdot Style_Is_Balconette

We will now interpret the coefficient for each type of predictor.

Interpreting Odds Ratios for Different Predictors

The interpretation of an odds ratio depends on whether the predictor variable is binary (like Is_Underwired) or continuous (like Lace_Coverage_Pct). The UCLA Statistics FAQ page is an excellent guide for this.

FAQ: How do I interpret odds ratios in logistic regression?

This guide provides exceptionally clear, step-by-step examples for interpreting odds ratios for different types of variables.

First, briefly review the section From probability to odds to refresh the core concepts. Then, carefully read the next two sections: The section on dichotomous predictors explains how the odds ratio compares the two groups of a binary variable. The section on continuous predictors shows how to interpret the odds ratio as the effect of a one-unit increase.

Let's use the logic from these readings to interpret our lingerie model's coefficients.

1. For a Binary Predictor: `Is_Underwired`

The coefficient for Is_Underwired is $β_{2} = + 1.5$ .

The odds ratio is $e^{1.5} \approx 4.48$ .

Interpretation: Holding lace coverage and style constant, a lingerie piece that is underwired has 4.48 times the odds of being a 'Top Pick' compared to a piece that is not underwired. You could also express this as a percentage change: the odds of being a 'Top Pick' are $(4.48 - 1) \times 100% = 348%$ higher for an underwired style.

2. For a Continuous Predictor: `Lace_Coverage_Pct`

The coefficient for Lace_Coverage_Pct is $β_{1} = + 0.05$ .

The odds ratio is $e^{0.05} \approx 1.051$ .

Interpretation: For each one-percentage-point increase in lace coverage, the odds of the piece being a 'Top Pick' are multiplied by a factor of 1.051 (an increase of about 5.1%), assuming Is_Underwired and Style_Is_Balconette are held constant. It is critical to specify the "one-unit" change—in this case, one percentage point.

3. For a Predictor with a Negative Coefficient: `Style_Is_Balconette`

The coefficient for Style_Is_Balconette is $β_{3} = - 0.5$ .

The odds ratio is $e^{- 0.5} \approx 0.607$ .

Interpretation: Since the odds ratio is less than 1, this feature decreases the odds. Holding the other features constant, a balconette style has 0.607 times the odds of being a 'Top Pick' compared to a non-balconette style. To state this as a percentage decrease, we calculate $(1 - 0.607) \times 100% = 39.3%$ . So, being a balconette style is associated with a 39.3% reduction in the odds of being selected.

Adjusted Odds Ratios and Interactions

A key point that your background in multivariate analysis makes intuitive is that in a model with multiple predictors, each odds ratio is an adjusted odds ratio. This means the effect of one predictor (e.g., Lace_Coverage_Pct) is estimated while holding the other predictors (Is_Underwired, Style_Is_Balconette) constant. This is analogous to interpreting coefficients in multiple linear regression and is a crucial part of isolating the independent contribution of each design feature.

For a deeper dive into this, and to see how interpretation becomes more nuanced with multiple variables and potential interactions, the following resources are useful.

A Deeper Dive into Odds Ratios Using Logistic Regression

This reading expands on our discussion to cover multiple variables and interaction terms.

Please read the sections on multiple variables and on interaction effects. Notice how the odds ratio for a variable can change when another predictor is added to the model. The explanation of interaction effects shows how the influence of one feature might depend on the level of another—for instance, the effect of lace might be different for an underwired bra versus a bralette.

This concept of interaction is powerful in design. For example, a high Lace_Coverage_Pct might significantly increase the odds of a bralette being a 'Top Pick' but have a much smaller effect on a structured push-up bra. Modeling this interaction (Lace_Coverage_Pct * Style_Is_Bralette) would allow you to quantify these conditional effects.

Numerical Exercise

Now, let's put your interpretation skills to the test. Imagine you have fit a new logistic regression model to predict Top_Pick status based on a different set of features.

Model:

ln (Odds of Top Pick) = - 2.5 + 1.8 \cdot Color_Is_Red - 0.2 \cdot Strap_Width_mm

Where:

Color_Is_Red is a binary variable (1 if red, 0 otherwise).
Strap_Width_mm is a continuous variable measuring the width of the shoulder straps in millimeters.

Your Task:

Calculate the odds ratio for a lingerie piece being red (Color_Is_Red). Interpret this value in a single sentence related to the photoshoot project.
Calculate the odds ratio for Strap_Width_mm. Interpret this value in a single sentence, explaining the effect of a 1mm increase in strap width.
Briefly explain why one odds ratio is greater than 1 and the other is less than 1, based on their coefficients.

Click for the solution.

1. Interpretation for Color_Is_Red:

Calculation: The coefficient is +1.8. The odds ratio is $e^{1.8} \approx 6.05$ .
Interpretation: Holding strap width constant, a red lingerie piece has 6.05 times the odds of being a 'Top Pick' compared to a non-red piece.

2. Interpretation for Strap_Width_mm:

Calculation: The coefficient is -0.2. The odds ratio is $e^{- 0.2} \approx 0.819$ .
Interpretation: For every 1mm increase in strap width, the odds of a piece being a 'Top Pick' are multiplied by 0.819 (i.e., they decrease by 18.1%), holding color constant.

3. Why the difference?

The odds ratio for Color_Is_Red is greater than 1 because its coefficient (+1.8) is positive, indicating that being red increases the log-odds of being a 'Top Pick'.
The odds ratio for Strap_Width_mm is less than 1 because its coefficient (-0.2) is negative, indicating that increasing strap width decreases the log-odds of being a 'Top Pick'.

Conclusion

This lesson provided the tools to unlock the insights hidden within a logistic regression model. You have learned how to translate the abstract log-odds coefficients into tangible odds ratios, which provide a clear and communicable measure of a feature's influence.

Key Takeaways:

The odds ratio for a predictor is calculated by exponentiating its coefficient: $O R = e^{β}$ .
For a binary predictor, the OR compares the odds of the outcome for the two groups.
For a continuous predictor, the OR represents the multiplicative change in odds for a one-unit increase in the predictor.
In a multiple logistic regression model, these are adjusted odds ratios, meaning they account for the other variables in the model.

In our next lesson, we will complete our work on logistic regression by using a fitted model to calculate the predicted probability of selection for a new, unseen lingerie style, bringing together everything we've learned in this module.

Recommended resources