I have a large dataset of regional tourism rates for each year for the period 1990-2010. At the year 2000, an economic disaster had occurred, and I would like to determine whether there it had a statistically significant effect on the regional tourism rates.

I have considered running the following Fixed-Effects Regression Model: RegionalTourismRate = EconomicCrisis + Factor(Region)

I coded the Economic Crisis as a binary variable, where 0 is before the crisis and 1 is after the crisis. I coded the Regions as dummy variables.

Would this be an appropriate method to determine the effect of the crisis on regional tourism rates?

Code used for the Hausman Test:

phtest(dat$TourismRate ~ dat$Crisis,data = dat, model = c("within", "random"), index = c("Region","Year"), vcov = TRUE)
  • $\begingroup$ Welcome. Did the economic disaster affect all regions? $\endgroup$ Commented Jun 16, 2020 at 23:22
  • $\begingroup$ Thank you! Yes, it affected all regions. There are 10 regions in the sample. I'm happy to clarify further if necessary! $\endgroup$
    – dayleymart
    Commented Jun 16, 2020 at 23:24
  • $\begingroup$ Did the crisis impact tourism in some regions more than others? Did tourism rates itself influence the onset of this crisis in some way, or was it an exogenous event? $\endgroup$ Commented Jun 16, 2020 at 23:31
  • $\begingroup$ Yes, I am theoretically assuming the crisis impacted more in some regions compared with others. Having looked at the graph of trends in tourism, it also seems as though tourism in some areas was impacted more by the crisis. Tourism rates did not have an effect on the onset of the crisis. $\endgroup$
    – dayleymart
    Commented Jun 16, 2020 at 23:41

2 Answers 2


Your approach seems reasonable. Acknowledging Savage Henry's answer, your model restricts attention to within-region variation. I would like to add a few additional points:

  1. If you can acquire data on tourism rates for additional locales/regions unexposed to the economic crisis, then you could supplement your fixed effects analysis with other regression techniques such as difference-in-differences, which is a special case of fixed effects and a very powerful approach for estimating treatment effects. However, if the event (treatment) you are investigating affected all regions within a particular country, then this might not be a worthwhile endeavor.

  2. You should also consider modeling the passage of time. You could enter time (year) into your model as a continuous linear time index. However, incorporating year fixed effects (i.e., $T-1$ dummies for years) introduces problems as your binary treatment (i.e., economic crisis) exhibits the same pattern across all regions. If I interpreted your question correctly, the change in treatment status occurs at the same time for all regions in your sample. Thus, the binary treatment 'turns on' (i.e., changes from 0 to 1) in the year 2000 and 'stays on' until 2010. Depending upon the ordering of your input variables, software may omit one additional year effect (resulting in only $T-2$ year effects), or it may drop your treatment dummy entirely.

  3. You could also investigate possible differences in treatment intensity. Ultimately, this requires intimate knowledge of the treatment under evaluation. There might have been a greater dose of economic hardship experienced in a subset of regions. What about the lasting effects of the economic depression on tourism rates? You could investigate this with a good visual plot of the region-specific tourism trends.

I hope some of these suggestions help.

  • $\begingroup$ Thanks very much for the advice - it's certainly been useful. I further ran a Hausman test and found that the p value =1. Would I be right in assuming therefore that the RE model is in fact a better option? $\endgroup$
    – dayleymart
    Commented Jun 17, 2020 at 13:25
  • $\begingroup$ In practice, I have found that it is rare to fail to reject the null hypothesis (i.e., random effects is preferred) when performing a Hausman test. Often, we are concerned that the unique errors are correlated with the regressors. It is even more rare to observe a p-value equal to 1. I am hesitant to advise you without seeing your code. Your pseudocode above leads me to believe you’re using R. Did you save your estimates from your fixed and random effects models and run this: phtest(fixed, random)? I would recommend either posting your code, or submit a new question entirely. $\endgroup$ Commented Jun 17, 2020 at 16:25
  • $\begingroup$ Instead of blindly following the results of this test, you should focus on the goal(s) of your study. Do you want to generalize to other regions/countries outside of your sample, or do you only care about the units under observation? $\endgroup$ Commented Jun 17, 2020 at 16:30
  • $\begingroup$ I have edited my original question to include the code used to run the Hausman test in R. I am only interested in the effects within the units under observation - i am not interested in generalising the results. $\endgroup$
    – dayleymart
    Commented Jun 17, 2020 at 18:33
  • $\begingroup$ Run two models. One using model = "within" and another using model = "random". Is the coefficient on Crisis the same for both models? $\endgroup$ Commented Jun 17, 2020 at 19:19

There is nothing inherently wrong with this as an approach. General advice would be the standard modeling process: remember what you're model is doing -- fixed-effects for regions will eat up all the inter-region variation, leaving you just the intra-region variation, which is probably what you want since it helps reduce omitted variable bias. Also, think through whether each region might plausibly have a different slope from the others (maybe a Chow test?). But again, from a purely statistical view, fixed-effects here is plausible.


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.