Skip to content

debugging.md

Latest commit

 

History

History
155 lines (112 loc) · 5.92 KB

debugging.md

File metadata and controls

155 lines (112 loc) · 5.92 KB
Raw

Debugging Hayashi Scripts

Hayashi ships with a built-in Debug Adapter Protocol (DAP) server, so you can debug .hay scripts directly from VS Code.

Starting a debug session

  1. Install the Hayashi VS Code extension.
  2. Make sure the hay binary is in your PATH.
  3. Open a .hay file, go to the Run and Debug panel, and select "Debug Hayashi script".
  4. Set breakpoints and start debugging (F5).

Behind the scenes, VS Code launches:

hay dap path/to/script.hay

The DAP server supports breakpoints, single stepping (step over, step in, step out, continue), scopes (Locals / Globals), and variable inspection.

Inspecting model results

When you pause on a breakpoint after fitting a model, model objects expand in the Variables panel with a concise summary plus structured children. For example, an OLS result appears as:

result: OLS(k=2, n=10000), R2=1.0000
  coefficients    DataFrame(2 rows, 7 cols)
  fit             Dict(13 entries)
  residuals       Series(residuals: 10000 values)
  fitted_values   Series(fitted_values: 10000 values)
  params          Series(params: 2 values)
  std_errors      Series(std_errors: 2 values)
  test_values     Series(test_values: 2 values)
  p_values        Series(p_values: 2 values)
  conf_lower      Series(conf_lower: 2 values)
  conf_upper      Series(conf_upper: 2 values)

coefficients

A DataFrame containing the coefficient table:

variable coef std_err t / z p_value conf_low conf_high
x1 ... ... ... ... ... ...
x2 ... ... ... ... ... ...

The test statistic column is labelled t for t-based inference and z for z-based inference; both map to the same test_values Series in the debugger.

fit

A Dict with model-specific fit statistics:

  • OLS / IV / panel OLS-style: r2, adj_r2 (where available), f_stat, prob_f, aic, bic, log_lik, sigma, n_obs, df_model, df_resid, cov_type, inference.
  • Binary choice (logit, probit): model_name, pseudo_r2, log_lik, iterations, inference.
  • Count data (poisson, nbreg, zip, zinb): log_lik, deviance, null_deviance, aic, bic, pseudo_r2, pearson_chi2, n_obs, df_resid, df_model, iterations, converged, plus model-specific entries such as alpha (NegBin / ZINB) and dispersion (GLM).
  • Quantile (qreg): tau, r2, iterations.
  • Tobit: sigma, log_lik, n_obs, n_censored, df_resid, iterations.
  • Random effects (re): r2_overall, sigma_u, sigma_e, theta, inference.
  • GMM: j_stat, j_p_value, n_obs, df_model, df_overid.
  • Arellano-Bond / System GMM: sargan_stat, sargan_pvalue, sargan_df, n_obs, n_entities, n_instruments, max_lags, step, m1_stat, m1_pval, m2_stat, m2_pval.
  • PCSE / PanelGLS: r2, n_obs, n_entities, t_periods, df_resid, sigma (and panels for PanelGLS).
  • GLSAR: r2, n_obs, df_resid.
  • Mixed: log_lik, aic, bic, n_obs, n_groups, var_resid, iterations, converged.
  • Beta: precision_param, log_lik, aic, bic, pseudo_r2, n_obs, iterations, converged.
  • Zero-inflated (zip / zinb): separate count_coefficients and inflate_coefficients DataFrames plus a fit Dict.

params, std_errors, test_values, p_values

These Series expose the raw coefficient vectors. They are useful when you want to inspect one vector at a time or copy a value from the debugger.

conf_lower / conf_upper

Confidence interval bounds for each parameter. Only shown when the underlying model stores them.

Model-specific views

Panel models (fe, re, pcse, xtgls, feiv)

result: Panel(k=2, n=1000, N=50), R2=0.8123
  coefficients  DataFrame(2 rows, 7 cols)
  fit           Dict(...)
  params        Series(params: 2 values)
  ...

Panel and FE2SLS report n and N (entities). RandomEffects reports sigma_u, sigma_e, and theta.

Binary choice

result: Logit(k=2), pseudoR2=0.3124
  coefficients  DataFrame(2 rows, 5 cols)
  fit           Dict(5 entries)
  ...

Count data

result: Poisson(k=2, n=100), pseudoR2=0.8890
  coefficients  DataFrame(2 rows, 7 cols)
  fit           Dict(13 entries)
  ...

Zero-inflated models (zip, zinb)

result: ZeroInflated(count=2, inflate=2, n=100), logLik=-134.2
  count_coefficients   DataFrame(2 rows, 5 cols)
  inflate_coefficients DataFrame(2 rows, 5 cols)
  fit                  Dict(...)

Mixed models (mixed)

result: Mixed(fixed=2, n=100, groups=10), logLik=-123.4
  fixed_effects    DataFrame(2 rows, 5 cols)
  random_effects   DataFrame(...)
  fit              Dict(...)

The fixed_effects DataFrame contains the population-level coefficients; random_effects contains one column per group.

SUR and 3SLS

For system estimators, each equation is exposed as a child variable:

result: SUR(eqs=2), sysR2=0.8543
  system_r2  Float
  sigma_cross  DataFrame(...)
  equation_0 / gnp   Dict { coefficients, fit }
  equation_1 / inv   Dict { coefficients, fit }

3SLS follows the same pattern but omits the system-level system_r2 and sigma_cross nodes.

Limitations

  • Not all estimators are debuggable yet. The expansion covers the most common cross-sectional, panel, count, binary, and system models.
  • Time-series models (ARIMA, GARCH, VAR, etc.) currently display as plain text without structured children.
  • Very large Series/DataFrame children are truncated by VS Code (namedVariables: 100), but you can request more values through the Variables panel.

Command-line debugging

You can also start the DAP server manually for testing or integration with other editors:

hay dap script.hay

It reads DAP messages from stdin and writes responses to stdout. The VS Code extension handles the protocol automatically.