[Rupali Vashisht]

Dear Joao Santos Silva,

Thank you, this is very helpful!

Kind regards,
Rupali

[Cleo Rasmus]

Dear all,

I have a panel data with 52 countries stretching over 20 years. I ran 14 regressions with the same independent variables and different outcome variables. I used xtqreg with jackknife correction (bootstrapping with 20 replications) and without jackknife correction, and my results vary, but it is not clear that 1 approach gives more significant results across all 14 regressions. I was wondering, which is the better approach? Joao Santos Silva, in your paper, "quantiles via moments" you state that "In practice, the comparison between the MM-QR estimates and their bias-corrected counterparts should give a reasonable indication of the need to use the jackknife." To me it is not clear which model I should prefer, since the significance is not higher in one of the models, the coefficients and significance just vary. I also use a time fixed effect in my command (i.Year) in case this is relevant. Also, I read in your paper that for n/T up to 10 the jackknife correction does not need to be used. Am I correct to go for the model without jackknife correction then (since n/T is far below 10 for me)?

Further, for 1 of my regressions, I get a warning that more than 7% of the values of the scale function are negative. I read that always some values of the scale function are negative, but is 7% too much? Or, is there a way to reduce the percentage?

Thank you.
Kind regards,
Cleo

[Joao Santos Silva]

Dear Cleo Rasmus,

Indeed, you probably do not need the JK correction in this case. I would, however, use many more that 20 bootstrap replicas.

Best wishes,

Joao

[Cleo Rasmus]

Dear Joao Santos Silva,

thank you for your advice.

Best regards,
Cleo

[David Xiong]

Dear @Joao Santos Silva

I'm writing to kindly request some clarification on the divergent results I had using mmqreg in Stata VS rqpd in R. Specifically, the Stata command is as follows: mmqreg tr_diff ret_squared ret exp_ratio turn_ratio share_class ln_mtna ln_age, abs(year) nols. The dependent variable is the change in tracking error and I'm interested in the coefficient of ret_squared(quadratic term for benchmark-adjusted return). The R code is here for reference: rqpd::rqpd(tr_diff ~
ret_squared+ret+exp_ratio+turn_ratio+ln_mtna+
ln_age+share_class |
as.factor(year), panel(taus=.5, method="pfe", tauw=1, lambda=5),
data=regression.qr)

As you explained on Statalist, Koenker's estimator imposes that the fixed effects have the same impact in all quantiles, while the estimator based on moments does not impose this restriction, but assumes a location-scale model. Not surprisingly, results can be divergent. However, the question I have in mind is: which estimation is more "appropriate"? Stately differently, how can I be sure which method suits the data (my panel data is annual with roughly 28000 observations in regressions)? I got highly statistically significant results for ret_squared using rqpd whereas the results are not significant using mmqreg. I am really puzzled on how I should proceed.

Any clarification/assistance will be highly appreciated!

Best,

David

[Joao Santos Silva]

Dear David Xiong,

I understand from our email exchange that you only have 14 years of data the code you provide above suggests that you want to use fixed effects for years. If this us correct, just use qreg2 to estimate a model with dummies for years and cluster the standard errors.

Best wishes,

Joao