In R confused about inferences I'm getting from logistic regression Currently, I'm using Wisconsin Breast Cancer Dataset and running the logistic regression on the dataset. When I'm using the whole dataset for the logisitc regression I've almost 30 features which are significant (p-value less like between 0.1 & 0.001 ). But when I use a subsample say 80% of the data for fitting all my variables are not significant.
fit.glm<-glm(diagnosis~.,data = data1, family = binomial)
summary(fit.glm)
smp_size <- floor(0.8* nrow(data1))


set.seed(123)
train_ind <- sample(seq_len(nrow(data1)), size = smp_size)

train <- data1[train_ind, ]
test <- data1[-train_ind, ]

fit.glm1<-glm(formula=diagnosis~.,binomial(link="logit"),data=train)
summary(fit.glm1)

This is the code I wrote for it.
This is the result using all the data:

This is the result using 80% of data:

You can see that all variables are becoming insignificant after using 80% of data and I also tried with 95% of the dataset and the same thing happened. I'm curious to understand why it' happening
 A: This is probably coming from perfect or near-perfect separation in your logistic regression analysis of the full data set.
With a large number of predictors it is common to find that some combination of them happens perfectly to distinguish the 2 classes in your particular data set. That can lead to difficulties in performing the analysis at all, very large apparent regression coefficients, or instability of results from subsample to subsample.
So in your example I would suspect separation in your full data set, leading to the extremely large coefficients. As MichaelM points out in a comment, you really do have to pay attention to the magnitudes of coefficients; do you really think that it's reasonable to expect a single predictor to have a million-fold influence on the log-odds of class membership? That's what a coefficient of 1e+06 means. It's not good practice just to focus on p-values.
Then when you look at a subset the (highly data-dependent) separation disappears, as might be expected.
Penalized regression approaches like ridge or LASSO can help work around this type of problem.
