The reason you are getting these results is that the algorithm only has the reported data to work with and since all data in the range that you are looking to impute have been removed, it has no basis upon which to develop a prediction model. The only solution I can think of would be to use 2s where the data are missing and report that as a limitation of the study. If you had information about the data that were omitted, but without identifiers, you could get a mean and use the mean as a replacement. This would be better than my first suggestion, but would still result in an underestimate of variance.

Good luck!

@Erdem Erkoyun, you could also use median number as a replacement of missing numbers and rerun the model.

This link maybe useful

"Hands-on with Feature Engineering Techniques: Imputing Missing Values | by Younes Charfaoui | Heartbeat" https://heartbeat.comet.ml/hands-on-with-feature-engineering-techniques-imputing-missing-values-6c22b49d4060

Erdem, the intention of censoring and truncating data (also known as blinding) is to render it impossible to "reverse engineer" the actual true value that has been intentionally censored. This intentional blinding of certain granularity in the dataset (as you have observed to render it impossible to use the data to identify the individual person that the data is an attribute of) is common in healthcare, financial, and criminal records.

The very good news is that there are many well published mathematical method that have become available over the years, that effectively deal with this critically serious data integrity issue.

Replacing the missing data with the mean has been mentioned. Replacing the data with other descriptive statistics has also been used to advantage. If you have the luxury of a large N sample (i.e. anything more than 3,500 observations is large) you can easily find descriptive values for mean, median, and mode. An approach that I have taken is to compare what happens when the null values are replaced with mean, then with median, and finally with mode (replace the missing value with the value that has the largest mode in the distribution). I also run the analytical models that I intend to develop for the data set using only the unblinded data and compare those unblinded models to the mean, median, and mode "replacement value" models. What I have found in healthcare data mining is that when your model has a dichotomous dependent variable (No = 0, Yes = 1) such as inpatient mortality, the median replacement value tends to be more concordant with the "as is" uncensored data model, and especially when there are no outliers or extreme outliers in the distribution of the variable that is having its missing values replaced. In my opinion one of the best graphics in SPSS for showing a diagnostic view of the shape of a distribution is Tukey's Box and Whisker plots. If there are any outliers or extreme outliers in your data set, they are quickly identified in the default SPSS Box Plot with the observation number to make it very easy to navigate straight to the identified outliers.

An additional strategy that I have used, is to first remove all of the outlier values and then re-calculate the descriptive mean, median and mode, and substitute those and compare as described above. We will never be able to have data integrity that is as pristine as the actual raw data set that has no missing values. But, strategically replacing the missing values, and providing a detailed and articulated narrative documentation of the replacement value method that was used, can make the difference between getting a manuscript published (or not).

There are many methods for missing value replacement. What I have described seems to work quite well in health care data for such things as Age Years, Length of Stay, that are linear, but does not work as well (more noisy, chaotic) for variables that can be "non-linear" in some segments of their functions in a time-series (heart rate, FI02, temp, etc.).

I hope this helps. My suggestion is to try the median of the distribution as a replacement value for missing values.

Richard E. Gilder, MS RN-C

Bioinformatics Scientist

The Gilder Company

TTUHSC Adjunct Faculty

[email protected]