How to choose best imputed data for further analysis in r

Question

I have a multivariate time series dataset (almost 30 years) with random missing values.

T	S	po4	si	din
9.00000	NA	0.290	5.310	18.51
8.45000	NA	0.130	6.180	14.74
13.60000	36.46000	0.010	0.500	1.86
23.20000	32.12000	0.010	6.580	0.81
26.00000	32.13000	0.070	0.500	0.23
NA	35.41400	0.010	1.670	0.72
24.80000	36.42000	0.000	3.540
24.20000	33.16000	0.110	2.020
22.50000	37.60000	0.040	0.400
16.32000	36.01000	0.020	2.900
17.60000	38.04000	0.010	0.970
9.70000	36.36000	0.120	7.950
13.80000	38.33000	0.010	5.760
7.90000	35.51000	0.060	2.350
11.90000	38.33000	0.030	3.410
24.10000	36.30000	0.020	0.730
25.20000	35.77000	0.020	1.370
24.70000	37.54000	0.330	0.700
5.75000	33.26000	0.120	0.860
13.30000	33.14000	0.000	0.000
13.60000	38.21265	0.000	0.190
15.70000	28.33000	0.040	11.500	41.64

I would like to fill the gaps in order to have a constant frequency (I have a monthly frequency with missing values) to try different techniques in the content of a time series analysis. I have tried to use the mice package in r and to decide which imputed dataset to use with with() and pool(),but I don't want to use all of them in a model, but obtain the most correct one and use that one for further analysis. How can I do that? How can I find the best one?

Otherwise, can you suggest me another method as an alternative to mice?

Thank you very much in advance

Answer 1

If you have a strong time correlation you can use the imputets package for time series imputation.

library(imputeTS)
na_kalman(your_dataframe)

There are also several other methods included in the package. As for mice, the whole point of multiple imputation is to have several imputed datasets. You would perform your analysis separately on each of them. Then you can compare the results. Since with imputation there always comes some uncertainty along (since your missing data replacements are only estimations). This technique enables you to model / get a feeling for the uncertainty.

If you don't want to do multiple analysis and do single imputation you can use any of these datasets, they are equally valid/there is no best one.

Or you could also use a single imputation package like misssForest.