I am doing an analysis of my big data on diabetic patients (n= 168615). The dependent variable is HbA1c blood test. I want to run binary regression on two groups of the dependent variable (HbA1c ≤ 7% coded Zero OR HbA1c > 7% coded One). The predictors are: weekly average maximum weather temperature (continuous) and the categories of the weekly average maximum weather temperature (High temperature coded 1, moderate temperature coded 2, and low temperature coded 3).

NOW: when I run the analysis for each predictor separately, I found statistically significant. But when I run the regression that predictor section contains both (continuous and categorical variables) some of them were not statistically significant. The attached photos are for what I have done: 1- Regression of weekly average maximum weather temperature (continuous) predictor alone. 2- Regression of the categories of weekly average maximum weather temperature predictor alone. 3- Regression of combined both (continuous and categorical variables).

So, please if anyone can help me in this analysis to examine if the model fit or not, which the right one I have done, and how to interpret the results. I would appreciate for any little help.

Photo (1):

Photo (2):

Photo (3):

  • $\begingroup$ What is the purpose of this analysis? Are you just trying to predict the outcome from temperature? Are you trying to estimate a specific parameter? Are you trying to build a causal model? Are you trying to describe a relationship? We need this information to be able to give you a good answer. $\endgroup$
    – Noah
    Jun 13 '19 at 0:20
  • $\begingroup$ Thank you some much for replying. I need to see if there is an effect of weather temperature on HbA1c level (relationship) and then to be able to predict. $\endgroup$ Jun 13 '19 at 0:23

First, a quick remark: plese stop using the buzz-word "big data" with $n= 168615$ I don't know if you expect to have trillions of patients in the future, but for now, this does not look like a big data problem.

Now to the actual question: statistical significance is nothing more than an arbitrary threshold put at $\alpha=0.05$, but there is nothing special about that $5\%$, so I wouldn't worry too much. What I would make sure of is that the performance of the multi-feature model is better.

You can try that by splitting your data into training and testing sets, fitting your model with the training data and then checking for how well they predict outputs in the testing data. There are multiple comparison metrics (plain accuracy for balanced classes, a precision-recall compromise for unbalanced ones...), so go for the one that suits you the most.

  • $\begingroup$ Thank you so much David for your comment, but I am not an expert in the statistics so I did not understand well what you mean. Sorry. $\endgroup$ Jun 13 '19 at 9:22
  • $\begingroup$ it's simple. Don't judge the models only on statistical significance os variables. Instead, try to see how well they perform (split your data into two groups, fit the models with the first group and see if their predictions for the rest of the data are good or not) $\endgroup$
    – David
    Jun 13 '19 at 11:35
  • $\begingroup$ Thank you so much, David, for the useful answer. I appreciate that for you. $\endgroup$ Jun 14 '19 at 8:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.