Mixed-frequency VAR (EViews 11)
Posted: Mon Mar 02, 2020 8:18 am
Hello everyone,
I am trying to estimate the impact of monetary policy on income inequality in a mixed-frequency VAR (U-MIDAS) setting in EViews 11 (student lite version), using quarterly data on inflation, a short-term interest rate, and a stock price index. For the inequality variable I have annual Gini coefficients. All variables are stationary at I(1). The sample period is 1991Q1 to 2016Q4, meaning that I have 104 observations of the quarterly variables but only 26 observations of the Gini coefficient. Only 1 lag is included (with more than 1 lag I get a singular matrix) and also an exogenous crisis dummy that takes the value of 1 for the years 2007 and 2008. Inflation and the interest rate enter in differences, while the stock index and the Gini enter in log-differences.
I few questions arose when estimating the model. I would be really grateful if you can help me out.
1. Most importantly, when trying to do an impulse response analysis, I ALWAYS get the error "near singular matrix". This is not the case when including only 3 variables in the VAR, but due to the underlying theory I need these 4 variables. I could replace one of the variables but for this analysis I should at least have 4 variables in total. I then give the interest rate an impulse with Cholesky ordering (dof adj.) inflation, interest rate, stock index and Gini. Am I getting multicollinearity due to only having 26 observations of the Gini and therefore a near singular matrix? Any suggestions on how I could solve this problem, while keeping the sample period as it is?
2. The estimation output shows 4 variables for each of the high-frequency variables, one for every quarter in a year, and only one Gini variable, since its frequency is annual. How can I interpret these results? Is it possible to interpret the "whole" effect of a high-frequency variable (e.g. by adding the quarterly coefficients or so?) instead of having to interpret the quarterly coefficients individually? What are the implications in terms of significance (t-stats) of having the high-frequency variables divided into 4?
3. Related to both questions: Is it possible to give a "whole" high-frequency variable an impulse instead of having to give an impulse to each quarter of the variable? And again, what are the implications in terms of interpretation of the impulse response functions, since it does seem I can only give impulses to each quarter of high-frequency variables?
I apologize for the long text but I'm really struggling with this. I would highly appreciate any guidance! If you have any questions regarding the model let me know.
Thank you in advance and best regards,
Thomas
I am trying to estimate the impact of monetary policy on income inequality in a mixed-frequency VAR (U-MIDAS) setting in EViews 11 (student lite version), using quarterly data on inflation, a short-term interest rate, and a stock price index. For the inequality variable I have annual Gini coefficients. All variables are stationary at I(1). The sample period is 1991Q1 to 2016Q4, meaning that I have 104 observations of the quarterly variables but only 26 observations of the Gini coefficient. Only 1 lag is included (with more than 1 lag I get a singular matrix) and also an exogenous crisis dummy that takes the value of 1 for the years 2007 and 2008. Inflation and the interest rate enter in differences, while the stock index and the Gini enter in log-differences.
I few questions arose when estimating the model. I would be really grateful if you can help me out.
1. Most importantly, when trying to do an impulse response analysis, I ALWAYS get the error "near singular matrix". This is not the case when including only 3 variables in the VAR, but due to the underlying theory I need these 4 variables. I could replace one of the variables but for this analysis I should at least have 4 variables in total. I then give the interest rate an impulse with Cholesky ordering (dof adj.) inflation, interest rate, stock index and Gini. Am I getting multicollinearity due to only having 26 observations of the Gini and therefore a near singular matrix? Any suggestions on how I could solve this problem, while keeping the sample period as it is?
2. The estimation output shows 4 variables for each of the high-frequency variables, one for every quarter in a year, and only one Gini variable, since its frequency is annual. How can I interpret these results? Is it possible to interpret the "whole" effect of a high-frequency variable (e.g. by adding the quarterly coefficients or so?) instead of having to interpret the quarterly coefficients individually? What are the implications in terms of significance (t-stats) of having the high-frequency variables divided into 4?
3. Related to both questions: Is it possible to give a "whole" high-frequency variable an impulse instead of having to give an impulse to each quarter of the variable? And again, what are the implications in terms of interpretation of the impulse response functions, since it does seem I can only give impulses to each quarter of high-frequency variables?
I apologize for the long text but I'm really struggling with this. I would highly appreciate any guidance! If you have any questions regarding the model let me know.
Thank you in advance and best regards,
Thomas