Panel VAR using STATA

[Klery Chikwede]

can panel VAR using xtvar2 work when the data has some variables with missing values. If not do I have to drop rows with missing values or impute using the mean? What is the best way to del with missing values if I am to use pane VAR. I want to check the responses of economic growth, government consumption, trade openness, capital formation, financial depth and inflation to four different types of disasters.

[Andrew Musau]

Missing data pose a problem if the data are missing non-randomly. Stata's xtvar command, introduced in StataNow 18.5, handles missing data, and how it does this is discussed on pages 27–29 of the manual (see https://www.stata.com/manuals/xtxtvar.pdf). I am not familiar with the workings of xtvar2 from SSC. If you have substantial missing data, this may present an issue with the xtvar command, as it relies on variable lags, which can amplify the amount of missing data in the regression equation. My recommendation would be to first explore alternative data sources to supplement the missing observations. It is not clear that other methods, such as interpolation, will necessarily improve the results.

[Nick Cox]

Historically the main use of interpolation was "reading between the lines", so going beyond printed tables of deterministic functions such as logarithms to get an extra decimal place. Younger readers may have no recollection of doing anything of the kind, but in my youth it was a standard small skill taught in early secondary school. Such interpolation depends on the function concerned being very smooth and the simplest methods are local and linear.

It's a big jump from that to interpolating data, such as time series, which are usually erratic on several time scales.

I've found programming interpolation methods to be an amusing challenge -- I am that way inclined -- but I've come to think of most interpolation of data as an exercise in wishful thinking. Put sceptically, and refutations are welcome, analysing data with any particular time step -- daily, weekly, monthly, yearly, whatever -- comes with a not that hidden assumption that the variables concerned may and usually will change from time point to time point with each step. Replacing a missing value with a value based on a smooth interpolation can't increase the information in the data.

Population size may often change smoothly -- but prices, production, consumption, employment???

I stick here to economic examples because that is what the question is about, and the topic of most similar threads. but similar stories could be told about e.g. temperature and rainfall.