Statistical Methods for Bias Adjustment, "Analysis of Missing Data" Professor Takahiro Hoshino

Title:Statistical Methods for Bias Adjustment, "Analysis of Missing Data"
Professor Takahiro Hoshino, Department of Economics, Keio University

My focus research topics are statistical causal inference and its applications.
You may not be familiar with the term "causal inference," so let me give you an example.

Let's say we want to find out which is the better way to treat a certain illness: medication or surgery.
As a result of investigation, of the two groups, one medicated and one having had surgery, is it reasonable to conclude that surgery is the better approach to treatment in cases where it offers a far higher survival rate?

If only patients in good overall condition with no complications can undergo surgery, while many patients in poor condition with complications cannot, it may seem that the difference between the survival rates for medication and surgery may be due to the difference in the baseline condition of the patient.

If a patient who has undergone surgery could have also been cured by medication, perhaps medication would be a better approach to treatment than　placing a heavy burden on the body with surgery.
【True effects cannot be understood by simple comparison】----------------------------------

The same can also be said of verification of the effects of costly TV advertisements (TV Ads).
In fact, a comparison of two groups, one which has seen a TV Ad for a game application and one which has not,

reveals what first seems to be the opposite effect to the one intended, where the group that has seen the TV Ad used the application for less time and opened the application less times than the group that has not seen the TV Ad.

However, the group that has not seen the TV Ad spends more time using smartphones than watching TV, so actually, the result is natural. Really, the proper evaluation index is "how much application usage time would be decreased if the group that saw the TV Ad had not seen it."

"Usage time had not seen it" is a missing value, known by the term "potential outcome,".
Therefore analysis needs to be performed, factoring in this so-called potential outcome.
Looking at almost all problems in society, true effects cannot be obtained by simple comparison in areas such as evaluation of policies in economics, evaluation of marketing measures and the effects of teaching methods.

My research on related to the development and application of methodology for the performance of correct policy evaluation and statistical causal effect received the Japan Society for the 13th Promotion of Science Prize and Japan Statistical Society Research Achievement Award.
【The analysis of missing data that handles data that cannot be observed】-------------------

Statistical causal inference is one of important fields in missing data analysis that deals with unobservable data that we considered earlier, such as potential "usage time".
Recently, decreasing accuracy of government statistics has become problematic and this has led to calls for development of new indices that combine data from government surveys with big data acquired by companies.

However, because big data is missing data which is biased in that it contains only "in-house purchasing and behavior logs" of a company's own customers, I am working with the Statistics Bureau of the Ministry of Internal Affairs and Communications on the development of new indices that incorporate big data with bias corrected.

No matter how much big data is acquired, because bias that exists in the data may yield incorrect results, The development and application of missing data analysis and statistical data fusion methods are becoming ever-more important in fields such as academic research, government decision-making and corporate marketing practices.

http://www001.upp.so-net.ne.jp/bayesian/Eindex.html

Видео Statistical Methods for Bias Adjustment, "Analysis of Missing Data" Professor Takahiro Hoshino канала Keio University Mita Campus