Bootstrap error with NON PANEL data

[Federica Vassalli]

Hi guys,

I'm experiencing some trobles when running a bootstrap over a program. I got this double error message

repeated time values in sample an error occurred when bootstrap executed multi_mfi, posting missing values insufficient observations to compute bootstrap standard errors no results will be saved

. I've googled the error and find that it is typical of panel data, but this is not my case. My dataset is not a panel dataset. The data is monthly from 2003m1 to 2021m12, I will drop the program here:

cap program drop multi_mfi
program multi_mfi, rclass
    
    args X Y Z
    qui {
    sort Date
    tset Date, monthly
    forv h = 12/12 {
    * X cumulative
        newey c`X'_`h' lc`X' target_factor ts_factor tf_ECB_totasset_IV_dem tsf_ECB_totasset_IV_dem  l.(lgdp hicp corpborr_cost LoantoDepSpread Capital LoanR Liquidity), lag(`h')
 
     cap drop `X'_ab `X'_ab2
    gen `X'_ab = e(b)[1,2]*-1+e(b)[1,4]*-1
    gen `X'_ab2 = e(b)[1,3]*-1+e(b)[1,5]*-1
    
    * Y cumulative
        newey c`Y'_`h' lc`Y' target_factor ts_factor tf_ECB_totasset_IV_dem tsf_ECB_totasset_IV_dem  l.(lgdp hicp corpborr_cost LoantoDepSpread Capital LoanR Liquidity), lag(`h')
 
    cap drop `Y'_ab `Y'_ab2
    gen `Y'_ab = e(b)[1,2]*-1+e(b)[1,4]*-1
    gen `Y'_ab2 = e(b)[1,3]*-1+e(b)[1,5]*-1
    
    * Z cumulative
        newey c`Z'_`h' lc`Z' target_factor ts_factor tf_ECB_totasset_IV_dem tsf_ECB_totasset_IV_dem  l.(lgdp hicp corpborr_cost LoantoDepSpread Capital LoanR Liquidity), lag(`h')
 
    cap drop `Z'_ab `Z'_ab2
    gen `Z'_ab = e(b)[1,2]*-1+e(b)[1,4]*-1
    gen `Z'_ab2 = e(b)[1,3]*-1+e(b)[1,5]*-1
    
    *** ECB tot asset
    cap drop ECB_totasset_`h'
    gen ECB_totasset_`h' = f`h'.ECB_totasset
    
    newey ECB_totasset_`h' target_factor ts_factor l.(lgdp hicp crisis), lag(`h')
        
        * conventional
    cap drop ECB_totasset_b ECB_totasset_b2
    gen ECB_totasset_b = e(b)[1,2]*-1
    gen ECB_totasset_b2 = e(b)[1,3]*-1
    }
    
    * coefficients conventional shock
    tempname MFI_loan_re_ab_conv
    gen `MFI_loan_re_ab_conv'=`X'_ab+`Y'_ab 
    * conventional multiplier
    return scalar  MFI_multi_re_conv=`MFI_loan_re_ab_conv'/ECB_totasset_b 
    return scalar MFI_multi_fin_conv=`Z'_ab/ECB_totasset_b 
    * coefficient unconventional shock
    tempname MFI_loan_re_ab_unconv
    gen `MFI_loan_re_ab_unconv'=`X'_ab2+`Y'_ab2     
    * unconventional multiplier
    return scalar MFI_multi_re_unconv=`MFI_loan_re_ab_unconv'/ECB_totasset_b2     
    return scalar  MFI_multi_fin_unconv=`Z'_ab2/ECB_totasset_b2
    }    
end

and this is the string for bootstrap

bootstrap r(MFI_multi_fin_unconv), reps(5) noisily:  multi_mfi MFI_loan_nfc MFI_loan_hh MFI_loan_nbfi

the variables preceded by return scalar are those for which I want to bootstrap the error term.
If you have any questions about the dataset I will be happy to reply.

Many thanks in advance,

Federica

[Clyde Schechter]

I think they say it is common with panel data, and don't say anything about time-series data is just, at least to judge by frequency of mention hereon Statalist, panel data is so much more common. But I think it is not possible to apply bootstrapping to time-series data at all, or at least not if you need to use time-series operators like lags in the calculations.

The reason is that bootstrapping involves sampling with replacement. This entails that in almost all bootstrap iterations, there will be some times that get represented more than once, leading to the impossibility of -tsset-ing the data. In fact, the problem is intractable for time series data as far as I can see. For panel data, you could sometimes work around this by bootstrapping full panels instead of individual observations (using the -cluster()- and -clusterid()- options). But with a single time series, I don't see any way out of the problem.

[Federica Vassalli]

Dear Clyde,
thank you for your reply. At least now I know what is the issue and I can look for it.
Many thanks.
Federica

[Rich Goldstein]

in Efron, B and Tibshirani, RJ (1993), An Introduction to the Bootstrap, Chapman & Hall, the authors discuss two methods of bootstrapping time series: (1) fit a model and bootstrap the residuals; (2) "moving blocks" bootstrap (pp. 92-104) - I never use time series analytic methods and cannot provide further details but possibly others, or tech support, can help