Accounting for the Multiple Natures of Missing Values in Label-Free Quantitative Proteomics Data Sets to Compare Imputation Strategies

Revision as of 09:50, 25 February 2020 by Bwday (talk | contribs) (Study design and evidence level)

Accounting for the Multiple Natures of Missing Values in Label-Free Quantitative Proteomics Data Sets to Compare Imputation Strategies

Lazar, C., Gatto, L., Ferro, M., Bruley, C., and Burger, T. (2016): Accounting for the Multiple Natures of Missing Values in Label-Free Quantitative Proteomics Data Sets to Compare Imputation Strategies. Journal of Proteome Research, 15:1116–1125.

https://doi.org/10.1021/acs.jproteome.5b00981


Summary

In this paper 5 imputation algorithms are evaluated depending on the number of missing values and randomness of the data to set practical guideless in choosing an appropriate imputation method which accounts for the specific type of missingness mechanism.

Study outcomes

Following outcomes can be drawn from the paper:

Outcome O1

Imputation performs better with fewer missing values.

Outcome O2

There exist MNAR-devoted methods and MCAR-devoted methods (see Figures 2 and 3). Depending on the MNAR ratio of a specific data set, one should privilege a MNAR/MCAR-devoted method (see Figure 4).

Outcome O3

MNAR-devoted methods perform worse the more missing values and the more random the missing values are (see Figures 2 and 3).

MCAR-devoted methods perform worse the more missing values and the more NOT at random the missing values are (see Figures 2 and 3).

Outcome O4

On average MCAR-devoted methods outperform MNAR-devoted methods, so that MCAR-devoted methods are recommended if the randomness of missing values is not known.

Outcome O5

Peptide-level imputation is more accuarte (Figure 6).

Study design and evidence level

The consideration of simulated data as well as real data, plus the application on protein level as well as on peptide level, makes the result sound and reliable.

The great variations of missing value incorporation, 11 rates of MV and 11 rates of MNAR values, result in 121 simulated datasets which give a broad representation of different missingness mechanisms.

Imputation was performed with 3 MCAR-devoted methods (kNN, SVDimpute, MLE) and 2 MNAR-devoted methods (MinDet, MinProb) which is not many but still shows the performance difference between MCAR/MNAR-devoted methods.

Further comments and aspects

References

The list of cited or related literature is placed here.