Just wondering if anyone can save me another 8 hours of headache trying to figure out what stats to run on my data..
I am trying to understand whether a number of different factors, such as age (measured continuously) and BMI (also continuously), are predictors for a certain type of pain (binary, Y/N). I have this binary data also represented ordinally (Likert scale 1-5); I have been hesitant to analyse it in this way, but if it makes the stats easier, I may consider it.
Any help would be enormously appreciated!
In case you have a binary response, you can fit a logistic regression model. In your case it would look like this:
logit(P(Y)=1)= beta_0 + beta_1*Age + beta_2*BMI
where logit(X) = ln(X)/(1-ln(X))
the code in R looks like this, but take a look at the help-file (?glm) or one of the many tutorials on logistic regression in R:
output
In case you have a binary response, you can fit a logistic regression model. In your case it would look like this:
logit(P(Y)=1)= beta_0 + beta_1*Age + beta_2*BMI
where logit(X) = ln(X)/(1-ln(X))
the code in R looks like this, but take a look at the help-file (?glm) or one of the many tutorials on logistic regression in R:
output
Simple way is to assume that there exists a linear relation between the target variable and input variables. In this case, you can use linear regression analysis, then check out the p-value. But it is hardly true to be linear, you can use decision tree to make the prediction. The decision tree has the feature importance measure which helps you understand what is the most important features to some extent. Single decision tree usually has large variance, you can use decision tree ensembles such as Random Forest or Gradient boosting machine.
Hope it helps. Best luck.
If your Dependent variable is binary, Logistic regression would be fine. U can also validate results using "test" data
If you have rating data then reducing it to binary will probably lose information unless the rating data are very sparse. I would consider ordinal logistic regression using either the ordinal package in R or possibly the Bayesian package brms (more flexible and better options for graphics). Both ordinal and brms use syntax very similar to the glm function Juergen suggested.
If you understand logistic regression then the ordinal model is not so tricky. In essence it is like fitting a sequence of binary logistic regressions with a changing threshold (so not 0/1 but 1/2, 2/3, 3/4/, 4/5). With a 5 point scale this gives you 4 models that fit not the probability of an outcome but the cumulative probability of each rating. In practice these are fitted simultaneously.
If you are not familiar with logistic regression I'd suggest getting a stats advisor on board who is familiar.
Well guys, as far as I understand the question: Brooke is looking for some "statistical test", right? Not for a predictive model. Well, suggesting that she uses a predictive model and a) interprets the summary output of R in order to find the most influencial independent vars or b) to evaluate/validate the model against some "test case" is a little bit laborious, isn't it? And additionally requires assumptions about the linearity of dependencies etc.
How, about using simple frequentist statistical tests? Such as testing for the "difference of means" of both groups. But we actually do not know what Brooke likes to test, whether she likes to test if there is a difference in the mean, the variance or the distribution of the two groups (well, and if she changes her mind to test for all 5 values of the scale, ANOVA could be used for testing many MEANS). We also do not know whether the samples of the groups are imbalanced (a regression based approch may run into problems then) and how many samples she has available?
Actually, I am not a fan of frequentist testing. Because of the many unknowns to us, I would suggest to use Kruschke's BEST-Test (http://www.indiana.edu/~kruschke/BEST/BEST.pdf) there is an implementation for R and Python (whether there is a implementation for SPSS I don't know.). The output is easy to interpret (if you have read the paper).
Thank you for all the helpful responses! It looks like binomial logistic regression is exactly what I'm looking for, although I am now confused on how best to check the assumption of linearity for this test. Any suggestions would be wonderful!
Linearity is normally indicated by the value of R square-the coefficient of determination. If there are more independent variables Adj R square is desirable. Logistic regression is on log-linear format.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS