[R-sig-Geo] Spatial data when the missing mechanism is MNAR (non-ignorable)

[Amitha Puranik]

Dear Prof Roger,



This is in continuation to my previous query on spatial data imputation
with MNAR mechanism. I have gone through the references recommended by you
and have the following concerns which I request you to address.

1. The papers by Thomas Suesse, Takafumi Kato suggest likelihood based
approaches for predicting missing data in simultaneous autoregressive
models with an assumption of *missing at random* mechanism.

2. The additional reference provided by you, i.e. 'Missing Data in Wind
Farm Time Series: Properties and Effect on Forecasts' by Tawn et al.,
assume *missing not at random* mechanism in an autoregressive framework and
have applied mean imputation and multiple imputation methods.



I am presently looking for a technique to deal with MNAR in spatially
autocorrelated data. Would it be reasonable to apply the methods
recommended by Suesse or Kato in this scenario by ignoring the missing
mechanism?

>From what I understand, using conventional methods that are effective for
MAR case would produce biased estimates when data is MNAR. Can the approach
applied by Tawn et al. (i.e. mean imputation or multiple imputation) be
used on spatial data with MNAR mechanism?



Any comment/ suggestion will be appreciated.


Thanks in advance,

Amitha Puranik.












On Mon, Oct 5, 2020 at 2:13 PM Amitha Puranik <puranik.amitha using gmail.com>
wrote:

> Dear Roger,
>
> Thank you for the quick response. I shall refer to the articles that you recommended.
> Thanks again!
>
> Regards,
> Amitha Puranik.
>
> On Mon, Oct 5, 2020 at 1:48 PM Roger Bivand <Roger.Bivand using nhh.no> wrote:
>
>> On Sun, 4 Oct 2020, Amitha Puranik wrote:
>>
>> > Is it possible to impute missing values in spatial data when the
>> > missingness is *MNAR (non-ignorable)*? Can pattern mixture model or
>> > selection model be modified to incorporate autocorrelation property and
>> > used in this context?
>> > Any suggestion/opinion is appreciated.
>>
>> MNAR possibly means "missing not at random". I see
>> https://doi.org/10.1186/1476-072X-14-1 for point support data using
>> INLA.
>> For lattice data, see perhaps https://doi.org/10.1007/s10109-019-00316-z
>> and work by Thomas Suesse https://doi.org/10.1016/j.csda.2017.11.004
>> https://doi.org/10.1080/00949655.2017.1286495. This might be relevant:
>> https://doi.org/10.1016/j.epsr.2020.106640, but be extremely careful of
>> imputation in training/test settings with spatial data, as the spatial
>> (or
>> temporal or both) lead to information leaking between the training and
>> test data because they are no longer independent.
>>
>> Hope this helps,
>>
>> Roger
>>
>> >
>> > Thanks in advance,
>> > Amitha Puranik
>> >
>> >       [[alternative HTML version deleted]]
>> >
>> > _______________________________________________
>> > R-sig-Geo mailing list
>> > R-sig-Geo using r-project.org
>> > https://stat.ethz.ch/mailman/listinfo/r-sig-geo
>> >
>>
>> --
>> Roger Bivand
>> Department of Economics, Norwegian School of Economics,
>> Helleveien 30, N-5045 Bergen, Norway.
>> voice: +47 55 95 93 55; e-mail: Roger.Bivand using nhh.no
>> https://orcid.org/0000-0003-2392-6140
>> https://scholar.google.no/citations?user=AWeghB0AAAAJ&hl=en
>>
>

	[[alternative HTML version deleted]]