What happens when data go missing?
A missing measurement is a gap in our knowledge. Try filling a few gaps and see how the choice of method changes the answer.
- Hide measurements by changing the amount or pattern of missing data.
- Fill the gaps with a simple method and compare its estimates with the other methods.
- Reveal the original values to see what the estimates recovered.
Loading the sample dataset…
| Sample | Feature A | Feature B | Feature C | Feature D |
|---|
Dashed border = a gap or estimate. Colour indicates the value; every cell also has a number or “?”.
How close are the estimates?
Average absolute error on the hidden values, in a.u. Lower means closer to the originals.
Fill the gaps to compare the three methods on exactly the same hidden values.
How do these methods work?
Feature mean: use the average of that feature’s visible measurements.
Feature median: use their middle value.
Three nearest samples: compare samples using their shared, visible features, scale each feature by its visible range, and average the target values of up to three nearest donors. If there is no comparable donor, use the feature mean.
What does this tell us?
Here we know the original values because we deliberately hid them. That lets us measure the error. In a real experiment, the missing values are unknown, so the reason they are missing and the assumptions behind an algorithm matter.
This is a learning example with synthetic data and three simple methods. Its rankings describe this example and this hidden set. They are not the published imputomics benchmark or a recommendation for analysing a real dataset.
Read the imputomics story Explore the full imputomics application ↗