Problem with Multisynth

[ahmedchaudhryy]

Hi,
So I am running multisynth for one of my research projects and it is giving the following error:

Error in Xc[[j]][, (dj - ndim + 1):dj, drop = F] :
subscript out of bounds

There are no missing values in my outcome variable.

Thanks in advance.

[ebenmichael]

Could you give some more info? The stack trace, and the function call in particular?

[ahmedchaudhryy]

Thank you for your reply.

Here is the traceback call; I hope this helps.
Edit: Removing/changing the number of lags/leads also does not help. The same error persists.

Error in Xc[[j]][, (dj - ndim + 1):dj, drop = F] :
subscript out of bounds

traceback()

7: FUN(X[[i]], ...)
6: lapply(1:J, function(j) {
dj <- ncol(Xc[[j]])
ndim <- min(dj, n_lags)
max_dim <- min(d, n_lags)
mat <- Xc[[j]][, (dj - ndim + 1):dj, drop = F]
n0 <- nrow(mat)
zero_mat <- Matrix::Matrix(0, n0, max_dim - ndim)
Matrix::t(cbind(zero_mat, mat))
})
5: do.call(Matrix::bdiag, lapply(1:J, function(j) {
dj <- ncol(Xc[[j]])
ndim <- min(dj, n_lags)
max_dim <- min(d, n_lags)
mat <- Xc[[j]][, (dj - ndim + 1):dj, drop = F]
n0 <- nrow(mat)
zero_mat <- Matrix::Matrix(0, n0, max_dim - ndim)
Matrix::t(cbind(zero_mat, mat))
}))
4: make_constraint_mats(trt, grps, n_leads, n_lags, Xc, Zc, d, n1)
3: multisynth_qp(X = bal_mat, trt = wide$trt, mask = wide$mask,
Z = wide$Z[, !colnames(wide$Z) %in% wide$match_covariates,
drop = F], n_leads = n_leads, n_lags = n_lags, relative = relative,
nu = 0, lambda = lambda, V = V, time_cohort = time_cohort,
donors = donors, eps_rel = eps_rel, eps_abs = eps_abs, verbose = verbose)
2: multisynth_formatted(wide = wide, relative = T, n_leads = n_leads,
n_lags = n_lags, nu = nu, lambda = lambda, V = V, force = force,
n_factors = n_factors, scm = scm, time_cohort = time_cohort,
time_w = F, lambda_t = 0, fit_resids = TRUE, eps_abs = eps_abs,
eps_rel = eps_rel, verbose = verbose, long_df = long_df,
how_match = how_match, ...)
1: multisynth(minwage_binary ~ crb, state, year, cfri, fixedeff = TRUE,
n_lags = 6, n_leads = 6)

[frederickluser]

Dear Eli

Thank you for your amazing work and congrats for the publication!
I am getting exactly the same error for my project. Do you have any idea where the problem might be?

Best, Frederic

[sethneller]

I'm having this exact issue as well. I'm wondering if it has something to do with always-treated units as I can replicate the error message using the vignette data if I recode at least of the states to have cbr=1 for the entire period. Unfortunately, when I drop my always-treated units, R-studio encounters a fatal error so this doesn't really fix things. @frederickluser, do you also have always-treated units in your data?

I'm also attaching the data file that I'm using in case that's helpful in anyway.

Thanks to Eli and team for putting this package together!

analytic_file.txt

[frederickluser]

@sethneller Thanks for your comment, I played around with it and deleted all the treatments that took place in the first two and last two periods. Like this it would work.

[ebenmichael]

@sethneller thanks for pointing this out!

There are two issues going on. First, if there are always treated units then the method won't work. This seems to be the issue for @frederickluser and @ahmedch92. Second if you're including a fixed effect (fixedeff = TRUE which is the default) and a unit has only 1 pre-treatment outcome, then the optimization problem is ill-specified and R breaks. This was the other problem that @sethneller was having.

I just added some checking code that'll give a more interpretable error message for both these cases, pointing out which units are causing the problems. Let me know if that works.

[sethneller]

@ebenmichael, thank you so much for your help! That definitely fixed everything on my end and the error messages were super informative. Looking forward to using this method!

[ebenmichael]

Great, glad this worked!