I am doing panel data analysis. I have N= 103 firms in T= 5 years (that is around 507 observations with an “unbalance data” fixed effect). The study period is from 2008-2012. The period of the financial crisis is 2008.

I am not sure whether I can use a differences in differences model to get the effect of the crisis. I can’t see a control group because all companies were affected by the crisis.

Michael R. Roberts talks in his book "Endogeneity in Empirical Corporate Finance" about single time-series difference before and after treatment. Does he mean to interact my independent and moderating variables with a crisis dummy which is the year 2008? And is it right to interact the crisis dummy with all the observations to get the crisis effect?


If you have no control group, then you can't do Diff-in-Diff.

Since you are interested in estimating the effect of some other variable on an outcome and seeing whether the effect is different during and after the crisis, I think that you would want to run some regression like:

$y_{it} = \beta_0 + \beta_1 x_{it} + \beta_2 Crisis_t * x_{it} + \tau_t + c_i + \epsilon_{it}$

Here you are interested in the effect of $x_{it}$ on $y_{it}$ and if it is different during the crisis than it is after the crisis (this is equivalent to testing whether $\beta_2 = 0$. $Crisis_t$ is a dummy variable $=1$ in crisis years (2008) and $=0$ otherwise.

$\tau_t$ are dummy variables for each time period. These should soak up differences in outcomes related to the time period, and are probably important to include. $c_i$ are individual fixed effects.

| cite | improve this answer | |
  • $\begingroup$ thanks for your reply @bmciv ....... I am interested in number two; to get the effect of some variables(dependent and moderating variable) on the dependent during and after the crisis. > you may check "p.35" papers.ssrn.com/sol3/papers.cfm?abstract_id=1748604 $\endgroup$ – Ben May 21 '14 at 14:34
  • $\begingroup$ Edited my answer based on your reply. $\endgroup$ – bmciv May 21 '14 at 15:01
  • $\begingroup$ thanks a lot for the valuable reply dear @bmciv. my case is a little different which confused me I have four independent variables and one interaction variable. On top of that, I want to see the effect of the crisis for all the variables not only one which will look a long model if I use your way and I am not sure I could do that. I have first to multiple my interaction variable with the 4 independent variables "IVs" then to multiply the 4 IVs with the Crisis only and finally to do three ways interaction and multiply every IV*interaction var.*crisis $\endgroup$ – Ben May 25 '14 at 16:06
  • $\begingroup$ Hi Ben, the comment I marked (with 18 $\beta$s) is the most general regression that you could run in your case, I think. It would be a good starting point. With so many covariates, you may get some strange results though. If you can argue that some of these coefficients are 0, e.g. Mod may not matter in some cases or something like that, then that could enable you to drop quite a few terms which might be helpful. Without knowing the details of your particular project, it is hard to say what the best choice is. $\endgroup$ – bmciv May 27 '14 at 17:34
  • 1
    $\begingroup$ No, that is a modeling choice. It is certainly more general to include the additional interaction term (if there is no additional effect, the estimated coefficient would be close to 0), but I would say that it is more common for researchers to leave out interaction terms unless they specifically have a requirement for it. $\endgroup$ – bmciv May 28 '14 at 14:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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