Hi, I am working with panel data (cross country, time series). If I have a non-stationary issue (having done panel unit root tests), someone said that adding time trend would not solve this issue.
Non-stationary data leads to problems of spurious correlations.
So, the way I understand it is that we add time trend (t=0,1,2..T) as an independent variable if we want to solve a spurious problem, controlling for exogenous increases in the independent and dependent variables.
Then, having added time trend, our data is still non-stationary but we no longer have the spurious problem? Is it correct?
So what is the relevance of adding time trend to stationarity? I dont think I have a co-integration analysis, so do I need to worry about the non-stationary issue? Should we eliminate the problem of non-stationarity? Having made the time series stationary (by whatever means), what`s the point of the time trend? (It is not going to spurious because we have made it stationary, is that right?)
Can someone help me clarify?
Hi, you should consider the problem of whether the underling trend in the data is stochastic or deterministic, and based on that make a conclusion on whether to include deterministic (usually linear) trend or whether to differentiate the data - what you should do in case of stochastic trend. This simple presentation should make things clearer: http://hedibert.org/wp-content/uploads/2015/04/DT-or-ST.pdf. Also for panel data you need to distinguish between individual and common unit root processes. All these problems can be dealt with standard unit root stationarity tests. And yes even if you do not use the cointegration approach you need to make sure that the time series you are using is stationary.
Typically, just removing the trend is not a good idea. Typical non-stationary data (stock price index, GDP per capita) have an exponential trend. However, just removing the trend would not be a good idea, because variance of the data and the beginning o the sample would be small, but variance of the data at the end of the sample would be large. In this particular case, calculating percentage changes is the way to go. However, what exactly you should do depends on your specific data.
Adding time trend would not solve the problem.Making data stationary is important before proceeding further.
Thanks very much for the answers! They are really helpful!
So, I tried to check whether my data is stationary in Eviews, but I am a bit confused on how to specify the tests.
Q1 Should I test for "Individual Intercept" OR "Intercept and trend"? What do you base your decision on?
Q2 Also, when it comes to spectral estimation, which Kernel should be used? "Bartlett, Parzen or Quadratic"? I have no idea what this is but changing them changes the results very significantly.
Q3 Also, is it possible that the Levin, Lin, Chu and Breitung (both assumes common unit root) have different results (0.0000 and 0.968 respectively)?
Q4 Finally, I checked for deterministic trend ( Yt=B+X(t-1)+e ) and Y(t-1) is significant so it means that there is a deterministic trend. So now, having discovered this, what would be the next step to make the data stationary? (Should I regress the original regression including Y(t-1)?)
Can someone help me choose and give me a bit of explanation which one to give more weight to? (I have done econometrics before but this was not included so I really don't know where to go).
Thanks very much in advance.
Hi, it depends on what package you are using. but in general you can do it on your own. Create a vector of consecutive integers from 1 to T, where T is your total number of observations:
trend
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series