How to find the correct lag length in a fixed-effects panel regression?

[Silvia Haertl]

Dear all,

My work consists on the Paper of Jordà, Schularick and Taylor (2016) - Sovereigns vs. Banks: Credit Crisis and Consequences which is available under
https://academic.oup.com/jeea/articl...4/1/45/2319810

I am currently doing the following regression,

foreach v in `irvars5' {
    forvalues i =1/5 {
    
    reg `v'`i' pk_norm pk_fin excess_prv `rhs8a' d1-d17  ///
        if core==1 & (pk_fin==1|pk_norm==1) , noconstant vce(cluster iso)

where:

local c_rhs8a ///
c_dlrgdp   c_drprv  c_dlcpi  c_dlriy  c_drpub  c_stir  c_ltrate  c_cay ///
c_ldlrgdp c_ldrprv c_ldlcpi c_ldlriy c_ldrpub c_lstir c_lltrate c_lcay

local irvars5 lrgdp lriy lcpi rprv rpub

dlrgdp is the growth rate of real GDP per capita (annual, in percent)
drprv is the growth rate of real loans per capita (annual, in percent)
dlcpi is the consumer price index (CPI) inflation rate (annual)
dlriy is the growth rate of real investment per capita (annual, in percent)
drpub is the growth rate of real public debt per capita (annual, in percent)
stir are short-term interest rates on government bonds
ltrate are long-term interest rates
cay is the current-account-to-GDP-ratio

excess_prv is the growth of real private credit in the previous expansion before a financial crisis. This is my experimental variable. First I estimate the model with the mean-growth of private credit in prior expansions, than with the mean + 1 SD.

The authors Jordà, Schularick and Taylor used the lag-length of 1 in their paper, but didn´t clarify why. Now I try to find the appropriate lag length and found the stata command

estat ic

The problem with this command is, that in some way, for me it doesnt make sense. Because it produces the optimal lag length of 5 years, which - for me - seems a little bit too long in that case.

Does someone know maybe another command (not estat ic) to find the appropriate lag length for my model?
Maybe someone can help?

Thank you very much in advance

[Silvia Haertl]

Maybe someone has an idea?

[Jesse Wursten]

Finding the appropriate lag length in panels is very tricky. In time series there's a long tradition of looking at residuals and AIC/BIC indicators, in panels this seems much less common. From what I've seen, the arguments tend to come from theory - e.g. what lag structure makes sense. You can do some statistical testing, but I don't know if most readers/referees would even consider it added value (e.g. serial correlation testing, order of integration of the residuals).

[Silvia Haertl]

To Jesse:

Thank you very much for your response!
I was thinking a lot about that topic today and you are right. With my panel, I estimate up to 50 equations simultaneously, so finding the appropriate lag length is really not easy. In my work I considered that problem and tried to give a different explanation. As credit booms and public debt overhangs are typically considered phenomena that last for many years, I will include up to five annual lags of any regressor .. I haven´t done the estimation yet, but let´s see how the results differ with a lag-length of 5.

Thank you for your help!

[Jala Youssef]

Dear Silvia Haertl,

I am facing a similar issue. Were you able to find out how to choose the appropriate lag length with the Local Projection method?

Many thanks in advance for your help.