3

It always create a doubt to me, whether we can apply linear or non linear multiple regression on time series data. If yes, should I consider year also an independent variable.

Thanks,
Arushi

Arushi
  • 1,027
  • 4
  • 14
  • 22
  • 1
    You might enjoy seeing the examples of both regression and time series analysis applied to the same dataset at http://stats.stackexchange.com/questions/18538. Additional debate in the comments compares and contrasts the methods. Pay attention to some of the caveats: time series methods are indicated when the regression *residuals* are strongly correlated, for instance, which they were not in this example. – whuber Feb 07 '14 at 22:32
  • Sure. The three issues you'll have to consider are (1) heteroskedacity (2) autocorrelation and (3) seasonality. There are a number of methods of addressing all three concerns. For short time series seasonality is less of an issue, for longer time series the limitations of using static seasonal components may become more apparent. – charles Feb 08 '14 at 15:06

1 Answers1

1

Yes, you can. The forecast::tslm function was written to help you with that. You may also read on generalized least squares to fight correlations in residuals that are common and expected in time series regression problems. This should give you better estimates of the standard errors of the regression parameters.

Hibernating
  • 3,723
  • 2
  • 21
  • 34
  • Thanks for quick response. Can you provide me some references, I am finding any better article on the mentioned method. It would be really helpful. – Arushi Feb 09 '14 at 12:21
  • tslm is just a wrapper for lm, with "trend" and "season" predefined for convenience. As to the better references, see chapter 5 "Regression" of "Introductory time series with R" by Cowpertwait and Metcalfe. Also search for "time series regression" to find more. I hope it helps. – Hibernating Feb 09 '14 at 14:57