Including observable data during COVID19 period into DSGE model

Dear Johannes,
First thank you for your previous guidance, I am grateful.
I plan to include 2020 macroeconomic data covering COVID19 period into my DSGE model, however, I find values of this period’s data is very extreme like outliers, I am wondering that do we stick to the assumption of no structural break/stable structure for DSGE model estimation? Do two quarters (2020 Q1 and 2020 Q2) extreme values for DSGE model matter in estimating DSGE models? for instance, will these two quarters’ extreme values cause misspecification/bias/inefficiency problems?
Thank you and look forward to hearing from you.
Best regards,

That is a tricky issue. It may be best to not include these data. Please refer to the recent work of Giorgio Primiceri

Dear Johannes,
Thank you for your helpful guidance. But for 2008 Great Financial Crisis, we also include that period data without accounting for a structural break for 2008 crisis, why is data for COVID19 period an exception?
And 2 quarters data out of 254 quarters data is only a small proportion, will including the extreme 2 quarters data have serious effect upon model misspecification, estimation unbiasedness and efficiency?
Best regards,

For the GFC, there is for example the Stock/Watson paper ( arguing that there was no structural break, just a bigger shocks. It is hard to argue the same for the COVID19-period. Also, the current GDP drop is bigger and faster than previous ones. With normally distributed shocks, this should pretty much never happen. Thus, these outliers will massively affect estimation.

Hi professor
After two years of the pandemic and given the evolution of the data in most countries, what are your thoughts now on including or not including the years of crisis?

I estimated a model recently and the estimates are better when I include the Covid-19 years, and if you compare the value of the parameters, most of them don’t change much except for one or two. Even the shock distribution is almost the same.


I don’t think there is a consensus yet. I would tend to just do the exercise you did: compare the results with and without the last two years. If it does not make a big difference, I would leave it in an report that results are robust.

Hi everyone, I thought after an additional year of data I would bring this thread up again,

I’m currently estimating a medium-scale DSGE model for UK data and I wanted to ask: what is the emerging consensus in the literature / best practice in the field to deal with the 2020-Q2 and 2020-Q3 data points?

On a quarterly growth basis, we had -21% and 17% for gdp in 2020 Q2/3, any model will choke on that.

In the VAR and local projections literature, the default seems to be to use dummys/pandemic priors. Is there a similar accepted ‘quick fix’ for DSGE estimation?

Some solutions might be:

  • specify these two quarters as NaN (at least for GDP and expenditure components) and let the Kalman filter deal with it?
  • If one uses HP-filtered data instead of first-diff filtered data, maybe there is a way to somehow tweak the HP filter and shuffle some of the spikes into the ‘trend’ component? So one would end up with ‘normal’ cyclical data…

Any idea and/or any links to recent papers that dealt with this issue would be much appreciated :slight_smile:

Kind regards and thanks a lot

I don’t know if there is a consensus yet. You could try option 1. I would never use the HP filter due to being a non-causal filter.

1 Like