I'm analyzing a monthly district-wise dataset of 10 years to determine factors influencing dengue incidence. The dataset is not stationary and the dengue incidence shows a seasonal pattern. My instructor suggested using panel data analysis, I tried pooled models, fixed effect models, first differenced models, etc but the model assumptions are still violated (heteroscedasticity and normality assumptions are violated in all cases). When setting the data as panel data I used the district as individuals and the Year-Month combination as the time component.

I'm not very familiar with panel data analysis so I'm not sure whether the data needs to be stationary to apply panel data analysis. I referred to several sources but most of them have not considered the stationarity of the data before applying panel data analysis. One source stated that stationarity must be considered only if the dataset is very large but my dataset only contains 2700 observations.

Can someone tell me whether the data needs to be stationary to apply panel data analysis? Are there any other methods I can use to analyze this dataset besides panel data analysis?

And also what test is used to test the multicollinearity in panel data. I used the VIF test but it shows an error every time. Is there any other specific test used for panel data?


1 Answer 1


Here are a few pointers on handling stationarity and multicollinearity in panel data analysis:

  • Stationarity is not an absolute requirement for panel data analysis, but it is preferred in many cases. With a short time series of 10 years, stationarity may not be critical.

  • For non-stationary data, using fixed effects or first differences models can help control for unobserved heterogeneity and trends over time. This seems like the right approach you are taking.

  • Other options include using time fixed effects to account for any time trends, or explicitly modeling the seasonal/cyclical patterns in the data.

  • Multicollinearity in panel data can be assessed using standard methods like VIF on the pooled data. However, VIF is more problematic in panel data.

  • Alternative approaches include:

  1. Looking at pair-wise correlations between independent variables
  2. Examining changes in coefficient estimates when dropping variables
  3. Using Lagrange multiplier tests for redundancy
  • If multicollinearity is a major issue, options include dropping problematic variables, combining related variables, or using penalized regression like ridge or LASSO.

  • Since you have a short time frame, you may just focus on a parsimonious model with a limited set of key explanatory variables.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.