Change script to apply on a different dataset.

Hello, I want to apply a particular econometric methodology for my own research purpose. The script for the methodology has been made public by the author. How can I change it to apply it to my own data.

[Adam Danz]

If you plan on citing the script, you shouldn't change it at all. Instead, you should set up your data so it works with the inputs of the script. If you need to change the script you'll need to provide a detailed description of what changes were necessary for you to use it. If you make adaptations, your results may not be comparable with others' results who used the script.

[Saptorshee Chakraborty]

Yes, I intend to cite the script along with the paper. But I don't understand what else inputs I require except the data.

[Adam Danz]

That's impossible to answer without (much) more information.

[Saptorshee Chakraborty]

I have sent you the script on Matlab mail.

[Rik]

It is always best to include all relevant information in your question. That way everybody can help you instead of only the people you send your data by email. (just as a note: don't send me the files by email, it's more efficient to upload them here yourself)

[Saptorshee Chakraborty]

Sorry,

The code:

clear global p T N d pi; IC_needed = 1; cvx_solver mosek load('balancedPanelX1995.mat') X = [lagsaving, cpi, interest, gdp]; y = saving;

p = size(X, 2); T = 15; N = 56; d = 2; % number of IVs

K_max = 5; lamb.grid = 10; lamb.min = .2; lamb.max = 2.0; lamb_const = lamb.min * (lamb.max / lamb.min ).^( ( (1:lamb.grid) - 1) /( lamb.grid -1 ) ); % the constant for lambda. very important!!

lambda = lamb_const* T^(-1/3); numlam = length(lambda);

tol = 0.0001; R = 80; W = 1;

%% scale the dependent variable and its lag term [Dy, DX, Z ] = IV_generate(y, X); std_yi = std(Dy); Dy = bsxfun(@rdivide, Dy, std_yi) ; DX = bsxfun(@rdivide, DX, std(DX, 1)); Z(:,:,3:5) = DX(:,:,2:4);

beta_hat0 = zeros(N, p); for i = 1:N DXi = permute( DX(:,i,:), [1 3 2]); Zi = permute( Z(:,i,:), [1 3 2]); Dyi = Dy(:, i); beta_hat0(i,:) = pinv(DXi' * (Zi * Zi') * DXi) * (DXi' * (Zi * Zi') * Dyi); %initial value end beta_hat00 = beta_hat0;

T = 13; ds = dataset( kron( (1:N)', ones(T,1) ), repmat( (1:T)', [N 1]),... reshape(Dy, [N*T 1]), reshape(DX, [N*T size(DX,3)] ), reshape(Z, [N*T size(Z,3)]) ); ds.Properties.VarNames = {'N' 'T' 'y' 'X', 'Z'};

%% if IC_needed == 1 IC_value = zeros(K_max, numlam); IC = IC_PGMM(ds, ones(N,1)); % K = 1 case IC_value(1, :) = log( IC/(N*T) );

    for ll = 1:numlam
        disp(ll)
        for K =  2:K_max
            lam = lambda(ll)*var(ds.y);
            [b_K, hat.a] = PGMM_est(Dy, DX, Z, beta_hat0, W, K, lam, R, tol);
            [~, hat.b, ~ , group] = report_b( T, b_K, hat.a, K );
            sum(group)
            IC_kk = zeros(1,K);
            for  kk = 1:K
                IC_kk(kk) = IC_PGMM(ds, group(:,kk) );
            end
            IC_value(K,ll) = log( sum(IC_kk)/(N*T) ) 
        end
    end

    pen =  (2/3)*(N*T)^(-.5)* p*(1:K_max);
    IC_final = IC_value + repmat( pen, [numlam, 1])';
    [~, l_hat] = min(IC_final(2,:));
end

%% the information criterion selects % K = 2 and lambda(8)

K = 2; l_hat lam = lambda(l_hat)*var(ds.y);

[b_K, hat.a] = PGMM_est(Dy, DX, Z, beta_hat0, W, K, lam, R, tol); [~, hat.b, ~ , group] = report_b( T, b_K, hat.a, K ); H = hat_IC_PGMM( ds, hat.b, hat.a, K, group); sum(group)

est_post_lasso = zeros(p,6); pi = 0.05; for kk = 1:K this_group = group(:,kk); dat = ds; a_hat = hat.a(kk,:)';

    index = 1:N;
    g_index = index(this_group);
    g_data = dat( ismember(dat.N, g_index), : ); % group-specific data

    ky = g_data.y;
    kX = g_data.X;
    kZ = g_data.Z;
    Nk = sum(this_group);
    XZZX = (kX' * kZ) * pinv(kZ' * kZ) * (kZ' * kX)/(Nk*T)+ridge(pi, N, T);
    XZZy = (kX' * kZ) * pinv(kZ' * kZ) * (kZ' * ky)/(Nk*T);

    post_a = (pinv(XZZX) * XZZy)';

    [vari] = var_post_GMM(T, post_a', ky, kX, kZ, pi); % no ridge used
    se = sqrt(diag(vari));
    est_post_lasso( :, (3 * ( kk-1 ) + 1) :3 * kk ) = [post_a; se'; post_a./se' ]';

end

est_post_lasso = mat2dataset( est_post_lasso, 'VarNames', ... {'g1_coef', 'g1_sd', 'g1_t', 'g2_coef', 'g2_sd', 'g2_t'}); disp(est_post_lasso)

load('country56.mat') country(group(:,1)) country(group(:,2)) sum(group)

%% pooled estimation % when the two groups are put together, we found the IV's is weak % this is the heterogeneity induced weak IV. % we add a positive value to the matrix to stablize the inverse, as in % ridge regression

XZZX = ds.X' * ds.Z * pinv(ds.Z' * ds.Z) * ds.Z' * ds.X/(N*T) + ridge(pi, N, T); XZZy = ds.X' * ds.Z * pinv(ds.Z' * ds.Z) * ds.Z' * ds.y/(N*T) ;

alpha1 = XZZX \ XZZy;

[vari] = var_post_GMM(T, alpha1, ds.y, ds.X, ds.Z,pi); se = sqrt(diag(vari)); [alpha1, se]

group_PGMM = group; save('group_PGMM.mat', 'group_PGMM');