Hi,
I have a pretty basic question I think but am going round in circles with it.
I have relatively large dataset (2500+ cases). I'm working with categorical predictors and want to carry out numerous logistic regressions. Some IVs have many missing data responses (secondary data source - variables such as 'vulnerable' where 1 means that the variable was present, 0 absent, and missing data may mean that the variable is absent or that we just don't know if it's present).
I've looked into dealing with missing data in logistic regression but what I'm reading doesn't seem to meet my purpose. I don't want to carry out listwise deletion of cases. I'm wondering whether there are any rules around which variables to exclude from my analysis (due to a certain % of missing data) and if anyone could point me towards a good resource?
Thanks,
Freya
With categorical predictors, a simple fix is to include "missing" as an additional category. Then you don't have to delete any cases at all (except those missing on the response variable).
You'd have vulnerable coded 0 for no, 1 for yes, and 2 for missing. Of course, then it's a nominal-level variable, so you can't interpret it as ordinal or continuous. You'll have two slopes instead of one, for each predictor that would otherwise be dichotomous, but with a sample as large as you've got, you shouldn't have a problem of over-fitting due to estimating twice as many slopes. The second slope would be interpretable in terms of how the missing-data cases differ from the other two categories.
You can impute the missing values. If your independent variable is categorical, you can use K-NN imputation method.
The k nearest neighbors is an algorithm that is used for simple classification. The algorithm uses ‘feature similarity’ to predict the values of any new data points. This means that the new point is assigned a value based on how closely it resembles the points in the training set. This can be very useful in making predictions about the missing values by finding the k’s closest neighbors to the observation with missing data and then imputing them based on the non-missing values in the neighborhood. However K-NN is quite sensitive to outliers in the data.
For details please check this link:
https://towardsdatascience.com/6-different-ways-to-compensate-for-missing-values-data-imputation-with-examples-6022d9ca0779
Thank you both. Burke D. Grandjean I’ve tried this method out after your recommendation and it seems to be a feasible approach. Tanvir Ahammed I‘ve just read that article - one method I did think was an option was to impute the missing values to zeros so that they could be missing or absent. This might be a sensible option considering the ‘real world’, archival data I am looking at. Would this be a poor substitute to the model you’ve suggested?
With the 0,1,2, coding I gave as an example, if the slope for "missing" (category 2) is not significantly, nor substantively, different from zero (or with other coding, not different from the slope for "absent"), that would offer empirical support for your guess that in the real-world archival data the absence of a code usually means the absence of the trait being coded.
Hi,
In SPSS software there are two pallette items: Missing value analysis and Multiple imputation.
Hope it solves your problem.
Hi, I though you already have solution for your question bur just in case of you have to deal with this issues again but in different scenarios. You can use this material as a future reference.
https://www.bu.edu/sph/files/2014/05/Marina-tech-report.pdf
However, K-NN is also one solution you could try to solve this issue.
- Badges
- Science topic
- Similar topics
- Agricultural Science
- Agronomy