reference class failed

[Sophialai]

Hi you all, I have a question toward relevel the reference class in Logistic Regression. I used exactly the same codes as my professor giving but the reference class still not functioned as I set, anyone knows what is the reason for that?

Below are my code script. For some reason the forum does not accept my csv file, I did not mean to be rude, but if you guys could recognisize something, that would be great. Thanks a lot!

# install and acticate the following package
#      "regclass"
# R-Studio logistic regression with OrganicsData.csv data
OrganicsData <- read.csv
ogdata <- OrganicsData

summary(ogdata)
attach(ogdata)

## relevel DemGender & PromClass
DemGender <- relevel(DemGender, ref = "M")
PromClass <- relevel(PromClass, ref = "Silver")

## since the outcome variable TargetBuy is stored as integer
## convert the column into factors
TargetBuy <- as.factor(TargetBuy)

# fit a logistic regression model with the given set of variables
finalmodel <- glm(TargetBuy ~.-ï..ID,  
                  data = ogdata, family = binomial)

# View estimation results
summary(finalmodel)


# compute and display odds ratios
oddsratios <- exp(coef(finalmodel))
oddsratios
# predicting the probabilities for individual cases from model
predicted_probabilities <- predict(finalmodel, ogdata, type= "response")

View(predicted_probabilities)

## setting a threshold to then predict 1 or 0
threshold <- 0.5
predictions <- factor( ifelse(predicted_probabilities >= threshold, 1, 0))

odata <- ogdata
odata <- cbind(odata, predicted_probabilities, predictions)

View(odata)

# Confusion Matrix for the Logistic Regression model
# install and acticate the following package
#      regclass
## use the confusion_matrix() function from this package
confusion_matrix(finalmodel)

[technocrat]

Hurried suggestion: take a look at this S/O post; may be a zero indexing issue

[Sophialai]

Hi technocrat,

Thanks a lot for the contributing.
I did follow the instructions to set the reference level. However, my question still is: I set "Male" as reference level and it is still existed in the final predicted output.
A bit different for PromClass, I set "Silver" as reference class. But at the end the predicted output referred to PromClassGold which I am very curious why that happens?

[technocrat]

Is osdata available online anywhere?

[Sophialai]

I only saw on this：
https://www.coursehero.com/file/p738ol9/The-ORGANICS-data-set-available-in-SAS-Metadata-Repository-in-library-AAEM-the/

Thanks,
Sophia

[technocrat]

Thanks, Sophia. Unfortunately, I can't unlock that without a SAS installation, which I don't have. I'll try to replicate a similar database and see what I can figure out.

[Sophialai]

Hi,

I have sent data file through your LinkedIn message. I really appreciate your help!

Good night,
Sophia

[technocrat]

Thanks. I would post this as a gist, so others can follow along, but I'm not sure of the IP restrictions.

Rather than attaching ogdata, I changed the values in place to avoid namespace clashes between the attached fields changes and ogdata, which does not reflect those changes, e.g.,

> class(TargetBuy)
[1] "factor"
> class(ogdata$TargetBuy)
[1] "integer"

To get to your question, however, I have to get past the specification of finalmodel

finalmodel <- glm(TargetBuy ~.-ï..ID,  
                  data = ogdata, family = binomial)

doesn't work, obviously. Do you mean

finalmodel <- glm(TargetBuy ~ DemGender + PromClass + ID,  
                  data = ogdata, family = binomial)

or

finalmodel <- glm(TargetBuy ~ .,  
                  data = ogdata, family = binomial)

which give quite different coefficients.

[Sophialai]

What I meant was predict TargetBuy with all the independent variables except ï..ID. that's why I put :
finalmodel <- glm(TargetBuy ~.-ï..ID,
data = ogdata, family = binomial)
"-ï..ID" means excluding that variable.

Thanks a lot,
Sophia

[technocrat]

Thanks, I was confused by the umlaut over i

[Sophialai]

Sure, I think it is R makes the change of the name of "ID" to "ï..ID" when I import the dataset.

[technocrat]

I'm going to be using

finalmodel <- glm(TargetBuy ~ DemAffl + DemAge + DemGender + PromClass+ PromSpend + PromTime,  data = ogdata, family = binomial)

for two reasons:

1. From Zen of Python: explicit is better than implicit
2. Including all except ID brings in TargetBuy on both sides

and I'm also using a function to add confidence bands to odds ratios

odr <- function(x) {
  exp(cbind(OR = coef(x), confint(x)))
}

[Sophialai]

Hi,

Looking cool!
My question is : is your output the same as you expected? The outputs of DemGender and PromClass won't show the base class which are "Male" & "Silver" respectively?

Thanks,
Sophia

[technocrat]

OK, got it. The reference levels should be quoted as strings

ogdata <- ogdata %>% mutate(DemGender = relevel(DemGender, ref = "M"))
ogdata <- ogdata %>% mutate(PromClass = relevel(PromClass, ref = "Silver"))

So,

1. Don't attach
2. Don't include TargetBuy on both sides of the tilde ~ operator
3. Quote the reference class names as strings

Before you go further into predict, review the concept of newdata.

Good luck!

[Sophialai]

Hi technocrat,

Gotcha! Thanks for the helping! Have an awesome Sunday!

Sophia