Hello Experts,
I am working on a pooled country dataset (of 19 countries) to look at association of linear growth (both as continuous and binary variable) with IFA intake during pregnancy using linear and logistic regression.
I followed the process for model building as suggested by Homer & Lemeshow but the final predictor model had a p-value of
Could it simply be a result of the analysis that the overall explanatory power of these variables is low (or lower than expected)? Perhaps it just is what it is...
I tend to agree with Dr. Geiser. However Hosmer and Lemeshow use a stepwise regression based procedure which doesn't work. Since you apparently have a binary DV and hence a logistic regression if you wish to try something that does work I am attaching a couple of things. The first is a paper that describes using adaptive lasso logistic regression to build a good prediction model. Second is an R program that implemented the method. You could try it if you want to try one last time. Sorry I can't be more encouraging. Best wishes and good luck David Booth
It could of course be that your model explains a low proportion of variance in the observed data. But it could also be that your model isn't ideal. I don't know what the Homer & Lemeshow procedure is, but I have a few suggestions. First, I would plot each of the independent variables against each other and against the dependent variable, and calculate the correlations for each of these. This will give you some insight into the relationships among your variables. One consideration is to avoid including two independent variables that are highly correlated. Another consideration is to pay attention to the shape of the bivariate relationship between an IV and the DV. It may be that you might get a better fit if you transform some variables or add a quadratic term for some variables.
Many thanks to all of you for your useful insights.
David Eugene Booth, you have introduced me to a new concept which seems to be the new in. I cannot thank you enough for it. I don't want to start in R so I have started researching it in Stata.
Sal Mangiafico, I have done all the required basics in my preliminary steps. ALso, I have no continuous variable to transform.
Though I will know better once I have done the adaptive lasso but as Dr. Geiser suggested, maybe it is what it is.
The project is to look at fetal growth as a function of iron-folic acid supplementation? I am not clear on how to convert growth into a binary variable. If it is something like growth versus no-growth then I might look at the proportion of 0 and 1. If the data are highly skewed to one state (zero or one) then there may be problems.
Is this a cohesive study, or is this a meta-dataset? Either way, there are usually a suite of confounding factors (like ethnicity, age, income, education) that could generate sub-populations within the study and that would decrease R2 if ignored.
With a large number of independent variables one approach is to use some selection method (forwards, backwards, stepwise). An alternative is to do factor analysis on the independent variables and then run regression on the dependent variable versus the principal components. If this goes well, each principal component encapsulates the important aspect of a small number of the independent variables in some interpretable way.
The data are the data, but there are some serious problems in the way we (as the global scientific community) collect and process data. Here is one place to start: https://en.wikipedia.org/wiki/Reproducibility_Project#:~:text=The%20Reproducibility%20Project%3A%20Psychology%20was,the%20project%20in%20November%202011. Clearly understanding the issues in such work will help you better understand why your outcome is not as clear as often reported especially if you are using metadata or combining new data with old.
@Depali I'm attaching an R program for you that does all of these things. WRT.to Dr Ebert's question this program handles most common types of regression. R is not too hard to use just download a copy of Jared Lander R for everyone from the z-library. It has many useful examples too. Best wishes David Booth
@Dr Ebert you may find the attachment useful reading for your Statistical interest. The best program I know of to do it is in the previous post by me. Best wishes and step wisely everywhere. Please at least read the abstract and introduction.Best of luck as you step more wisely in the future. David Booth
Thank you Dr. Booth. I will have to start learning R as it is good for multilevel too. For now, my Stata is running 'lassologit'. Let us see if it works out. If not, I will blindly follow the above.
- Badges
- Science topic
- Similar topics
- Filing