I am currently working on a paper that evaluates impact of government expenditures on economic growth in Afghanistan. The research proposal was structured based on time series data. at the moment, I only access the data from 2003-2018. Is it possible to run regression? If not, what is your recommendation?
Thanks
It is possible, a similar question has been asked (and answered) here:
https://stats.stackexchange.com/questions/91103/suggest-models-for-prediction-based-on-small-sample-data
Increasing your sample size might be desirable and perhaps you could try, if the data is available, to do it by disaggregating your time series by city, region, or something similar.
Dear Ahmad Walid Barlas
I think that is possible, and that is depending on the linear relationship between tow variables and.
It is certainly possible but you probably need to consider other variables.
Ahmad -
If you are talking about regression, then you would have other variables used for predictor data besides the time series. But if, as I suspect, you really just have forecasting from a time series, things can change (change points) over 16 years anyway. You may want to use exponential smoothing for forecasting.
Best wishes - Jim
Dear Rick Manner I proposed annual growth rate as dependent variable and education, healthcare, infrastructure expenditures as independent variables.
Dear Ahmad,
A rule of thumb for the sample size is that regression analysis requires at least 20 cases per independent variable in the analysis.
So, the ideal would be to have a 20 years sample.
But I think that your 16 years sample is not very far from the iideal sample size and that you can run the regression model.
You could look into a technique called SiZer, which is more friendly with time series data.
Ahmad, I suggest that you plot your data to look for patterns. I expect there will be a lag between the time expenditures are made and the growth occurs. Results are rarely instantaneous.
SiZer papers:
Sonderegger DL, Wang H, Clements WH, Noon BR. 2009. Using
SiZer to detect thresholds in ecological data. Front Ecol Environ.
7(4):190–195.
Rondonotti V, Marron JS, Park C. 2007. SiZer for time series: a new
approach to the analysis of trends. Electron J Stat. 1:268–289.
Zhang JT, Marron JS. 2005. SiZer for smoothing splines.
Computation Stat. 20:481–502.
It is possible that you will have regression model results but that would be spurious. Consider d fact that your degrees of freedom will be great affected due to limited number of observations, (assuming you have say 3 independent variables or more). Therefore, you may use multiple imputation to increase your N as someone already suggested. Stata has that option, dont know about other softwares. If you can get monthly observations for the 16 yrs, that will increase your N and your regression output will be more reliable.
Of course, you are allowed to run a country-specific regression with 16 observations. I would recommend to run the regression without imputation.
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series