# Calculating error/accuracy in Logistic Regression

I am using Logistic Regression to create a binary classifier in R. Wrote a predictor for test data, and calculated the mean of my predicted result (I guess it can be translated to the efficiency of my method). I did find some references to training and test errors calculation, but can't seem to understand how to calculate them from the predicted fit. Any reference will help!

    glm.fit=glm(salary~age+workclass+fnlwgt+education+education.num+marital.status
+occupation+relationship+race+sex+capital.gain+capital.loss
+hours.per.week+native.country,
summary(glm.fit)
glm.probs[1:10]
glm.pred=ifelse(glm.probs>0.5,"2","1")
table(glm.pred,salary)
mean(glm.pred==salary)
plot(glm.probs)
boxplot(glm.probs)


Result of summary:

Deviance Residuals:
Min       1Q   Median       3Q      Max
-3.6016  -0.6337  -0.3539  -0.0786   3.3131

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept)    -1.433467   0.029795 -48.111  < 2e-16 ***
age             0.437952   0.026454  16.555  < 2e-16 ***
workclass      -0.104634   0.021628  -4.838 1.31e-06 ***
fnlwgt          0.052655   0.023348   2.255 0.024117 *
education       0.055044   0.028739   1.915 0.055455 .
education.num   0.851533   0.026941  31.607  < 2e-16 ***
marital.status -0.342673   0.026837 -12.769  < 2e-16 ***
occupation     -0.027290   0.023192  -1.177 0.239308
relationship   -0.223131   0.033368  -6.687 2.28e-11 ***
race            0.100934   0.026330   3.833 0.000126 ***
sex             0.404090   0.034456  11.728  < 2e-16 ***
capital.gain    2.392861   0.108381  22.078  < 2e-16 ***
capital.loss    0.259485   0.019611  13.232  < 2e-16 ***
hours.per.week  0.335252   0.025005  13.407  < 2e-16 ***
native.country -0.005913   0.024925  -0.237 0.812480
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 16832  on 15080  degrees of freedom
Residual deviance: 11829  on 15066  degrees of freedom
AIC: 11859

Number of Fisher Scoring iterations: 7

• Can you show your code? – Peter Flom - Reinstate Monica Mar 12 '14 at 17:21
• Added the code... – Divi Mar 12 '14 at 17:29

I think what you might be looking for is the precision and recall measures, and the corresponding F score.

P = precision = # true positives / # predicted positives

R = recall = # true positives / # actual positives

F = 2PR/(P+R)

Here is a Coursera video that explains how these are used.

• Added the result, still not sure how to interpret it! Please advise. – Divi Mar 12 '14 at 17:53
• webdocs.cs.ualberta.ca/~eisner/measures.html... I guess this will work. Cleared by doubt! – Divi Mar 12 '14 at 17:59
• Yes, that link spells out accuracy, as well. – John Yetter Mar 12 '14 at 18:26