flowchart LR A[Choose] --> B[Fit] B --> C[Assess] C --> D[Use] C --> A
Multiple Logistic Regression
Lucy D’Agostino McGowan
Types of statistical models
| response | predictor(s) | model |
|---|---|---|
| quantitative | one quantitative | simple linear regression |
| quantitative | two or more (of either kind) | multiple linear regression |
| binary | one (of either kind) | simple logistic regression |
| binary | two or more (of either kind) | multiple logistic regression |
Types of statistical models
| response | predictor(s) | model |
|---|---|---|
| quantitative | one quantitative | simple linear regression |
| quantitative | two or more (of either kind) | multiple linear regression |
| binary | one (of either kind) | simple logistic regression |
| binary | two or more (of either kind) | multiple logistic regression |
Types of statistical models
| variables | predictor | ordinary regression | logistic regression |
|---|---|---|---|
| one: \(x\) | \(\beta_0 + \beta_1 x\) | Response \(y\) | \(\textrm{logit}(\pi)=\log\left(\frac{\pi}{1-\pi}\right)\) |
| several: \(x_1,x_2,\dots,x_k\) | \(\beta_0 + \beta_1x_1 + \dots+\beta_kx_k\) | Response \(y\) | \(\textrm{logit}(\pi)=\log\left(\frac{\pi}{1-\pi}\right)\) |
Multiple logistic regression
- ✌️ forms
| Form | Model |
|---|---|
| Logit form | \(\log\left(\frac{\pi}{1-\pi}\right) = \beta_0 + \beta_1x_1 + \beta_2x_2 + \dots \beta_kx_k\) |
| Probability form | \(\Large\pi = \frac{e^{\beta_0 + \beta_1x_1 + \beta_2x_2 + \dots \beta_kx_k}}{1+e^{\beta_0 + \beta_1x_1 + \beta_2x_2 + \dots \beta_kx_k}}\) |
Steps for modeling
Fit
Call: glm(formula = Acceptance ~ MCAT + GPA, family = "binomial", data = MedGPA)
Coefficients:
(Intercept) MCAT GPA
-22.3727 0.1645 4.6765
Degrees of Freedom: 54 Total (i.e. Null); 52 Residual
Null Deviance: 75.79
Residual Deviance: 54.01 AIC: 60.01Fit
Call:
glm(formula = Acceptance ~ MCAT + GPA, family = "binomial", data = MedGPA)
Deviance Residuals:
Min 1Q Median 3Q Max
-1.7132 -0.8132 0.3136 0.7663 1.9933
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -22.3727 6.4538 -3.467 0.000527 ***
MCAT 0.1645 0.1032 1.595 0.110786
GPA 4.6765 1.6416 2.849 0.004389 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 75.791 on 54 degrees of freedom
Residual deviance: 54.014 on 52 degrees of freedom
AIC: 60.014
Number of Fisher Scoring iterations: 5Fit
How do we get a confidence interval?
Call:
glm(formula = Acceptance ~ MCAT + GPA, family = "binomial", data = MedGPA)
Deviance Residuals:
Min 1Q Median 3Q Max
-1.7132 -0.8132 0.3136 0.7663 1.9933
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -22.3727 6.4538 -3.467 0.000527 ***
MCAT 0.1645 0.1032 1.595 0.110786
GPA 4.6765 1.6416 2.849 0.004389 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 75.791 on 54 degrees of freedom
Residual deviance: 54.014 on 52 degrees of freedom
AIC: 60.014
Number of Fisher Scoring iterations: 5Fit
How do we convert this to an odds ratio from the log odds scale?
Fit
How do we convert this to an odds ratio from the log odds scale?
Assess
What are the assumptions of multiple logistic regression?
- Linearity
- Independence
- Randomness
Assess
How do you determine whether the conditions are met?
- Linearity
- Independence
- Randomness
Assess
How do you determine whether the conditions are met?
- Linearity: empirical logit plots
- Independence: look at the data generation process
- Randomness: look at the data generation process (does the outcome come from a Bernoulli distribution?)
Assess
If I have two nested models, how do you think I can determine if the full model is significantly better than the reduced?
- We can compare values of \(-2 \log(\mathcal{L})\) (deviance) between the two models
- Calculate the “drop in deviance” the difference between \((-2 \log(\mathcal{L}_{reduced})) - ( -2 \log(\mathcal{L}_{full}))\)
- This is a “likelihood ratio test”
- This is \(\chi^2\) distributed with \(p\) degrees of freedom
- \(p\) is the difference in number of predictors between the full and reduced models
Assess
- We want to compare a model with GPA and MCAT to one with only GPA
Call:
glm(formula = Acceptance ~ GPA, family = binomial, data = MedGPA)
Deviance Residuals:
Min 1Q Median 3Q Max
-1.7805 -0.8522 0.4407 0.7819 2.0967
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -19.207 5.629 -3.412 0.000644 ***
GPA 5.454 1.579 3.454 0.000553 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 75.791 on 54 degrees of freedom
Residual deviance: 56.839 on 53 degrees of freedom
AIC: 60.839
Number of Fisher Scoring iterations: 4
Call:
glm(formula = Acceptance ~ GPA + MCAT, family = binomial, data = MedGPA)
Deviance Residuals:
Min 1Q Median 3Q Max
-1.7132 -0.8132 0.3136 0.7663 1.9933
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -22.3727 6.4538 -3.467 0.000527 ***
GPA 4.6765 1.6416 2.849 0.004389 **
MCAT 0.1645 0.1032 1.595 0.110786
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 75.791 on 54 degrees of freedom
Residual deviance: 54.014 on 52 degrees of freedom
AIC: 60.014
Number of Fisher Scoring iterations: 5Assess
- We want to compare a model with GPA and MCAT to one with only GPA
Call:
glm(formula = Acceptance ~ GPA, family = binomial, data = MedGPA)
Deviance Residuals:
Min 1Q Median 3Q Max
-1.7805 -0.8522 0.4407 0.7819 2.0967
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -19.207 5.629 -3.412 0.000644 ***
GPA 5.454 1.579 3.454 0.000553 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 75.791 on 54 degrees of freedom
Residual deviance: 56.839 on 53 degrees of freedom
AIC: 60.839
Number of Fisher Scoring iterations: 4
Call:
glm(formula = Acceptance ~ GPA + MCAT, family = binomial, data = MedGPA)
Deviance Residuals:
Min 1Q Median 3Q Max
-1.7132 -0.8132 0.3136 0.7663 1.9933
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -22.3727 6.4538 -3.467 0.000527 ***
GPA 4.6765 1.6416 2.849 0.004389 **
MCAT 0.1645 0.1032 1.595 0.110786
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 75.791 on 54 degrees of freedom
Residual deviance: 54.014 on 52 degrees of freedom
AIC: 60.014
Number of Fisher Scoring iterations: 5Assess
- We want to compare a model with GPA, MCAT, and number of applications to one with only GPA
Call:
glm(formula = Acceptance ~ GPA, family = binomial, data = MedGPA)
Deviance Residuals:
Min 1Q Median 3Q Max
-1.7805 -0.8522 0.4407 0.7819 2.0967
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -19.207 5.629 -3.412 0.000644 ***
GPA 5.454 1.579 3.454 0.000553 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 75.791 on 54 degrees of freedom
Residual deviance: 56.839 on 53 degrees of freedom
AIC: 60.839
Number of Fisher Scoring iterations: 4
Call:
glm(formula = Acceptance ~ GPA + MCAT + Apps, family = binomial,
data = MedGPA)
Deviance Residuals:
Min 1Q Median 3Q Max
-1.6949 -0.8309 0.2900 0.7926 1.8238
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -23.68942 7.02387 -3.373 0.000744 ***
GPA 4.86062 1.69441 2.869 0.004123 **
MCAT 0.17287 0.10537 1.641 0.100867
Apps 0.04379 0.07617 0.575 0.565412
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 75.791 on 54 degrees of freedom
Residual deviance: 53.682 on 51 degrees of freedom
AIC: 61.682
Number of Fisher Scoring iterations: 5Use
- We can calculate confidence intervals for the \(\beta\) coefficients: \(\hat\beta\pm z^*\times SE_{\hat\beta}\)
- To determine whether individual explanatory variables are statistically significant, we can calculate p-values based on the \(z\)-statistic of the \(\beta\) coefficients (using the normal distribution)
Use
How do you interpret these \(\beta\) coefficients?
Call:
glm(formula = Acceptance ~ MCAT + GPA, family = "binomial", data = MedGPA)
Deviance Residuals:
Min 1Q Median 3Q Max
-1.7132 -0.8132 0.3136 0.7663 1.9933
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -22.3727 6.4538 -3.467 0.000527 ***
MCAT 0.1645 0.1032 1.595 0.110786
GPA 4.6765 1.6416 2.849 0.004389 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 75.791 on 54 degrees of freedom
Residual deviance: 54.014 on 52 degrees of freedom
AIC: 60.014
Number of Fisher Scoring iterations: 5\(\hat\beta\) interpretation in multiple logistic regression
The coefficient for \(x\) is \(\hat\beta\) (95% CI: \(LB_\hat\beta, UB_\hat\beta\)). A one-unit increase in \(x\) yields a \(\hat\beta\) expected change in the log odds of y, holding all other variables constant.
\(e^{\hat\beta}\) interpretation in multiple logistic regression
The odds ratio for \(x\) is \(e^{\hat\beta}\) (95% CI: \(e^{LB_\hat\beta}, e^{UB_\hat\beta}\)). A one-unit increase in \(x\) yields a \(e^{\hat\beta}\)-fold expected change in the odds of y, holding all other variables constant.
Summary
| – | Ordinary regression | Logistic regression |
|---|---|---|
| test or interval for \(\beta\) | \(t = \frac{\hat\beta}{SE_{\hat\beta}}\) | \(z = \frac{\hat\beta}{SE_{\hat\beta}}\) |
| — | t-distribution | z-distribution |
| test for nested models | \(F = \frac{\Delta SSModel / p}{SSE_{full} / (n - k - 1)}\) | G = \(\Delta(-2\log\mathcal{L})\) |
| — | F-distribution | \(\chi^2\)-distribution |
