Command reference — Stata-class syntax, GPU-accelerated by Apple Silicon.
Every command follows one grammar:
command [varlist] [if] [in] [, options]
Load a dataset into memory. Format is inferred from the extension: .parquet, .csv, .tsv, .json, .arrow, and Stata .dta (formats 117–119).
use filename
. use sales.parquet . use ~/data/panel.dta
Loading goes through DuckDB for columnar formats and a native reader for .dta. The sandboxed app reads user-selected files and its own container.
Load a sample dataset shipped with Metrika. Bare `sysuse` lists what is available.
sysuse [name]
. sysuse sales . sysuse mtcars
sales: 200-firm × 5-year panel (revenue, price, region, employees). mtcars: the R classic, with a string model column.
Write the working dataset to disk (.parquet, .csv, or Stata .dta format 118).
save filename [, replace]
| replace | overwrite an existing file |
. save results.dta, replace
Drop the working dataset from memory.
clear
List variables with their storage types and missing counts.
describe
Print observations (first 50 shown).
list [varlist] [if] [in]
. list revenue price in 1/10
Count observations satisfying a condition.
count [if]
. count if revenue > 100 & !missing(price)
Create a new variable from an expression. Observations excluded by `if` become missing.
generate newvar = expression [if] [in]
. gen log_rev = ln(revenue) . gen high = revenue > 120 if !missing(revenue)
Functions: ln, log, log10, exp, sqrt, abs, floor, ceil, round, int, min, max, missing. Missing values propagate through arithmetic; division by zero and ln of a nonpositive number yield missing.
Overwrite values of an existing variable; reports the number of real changes.
replace varname = expression [if] [in]
. replace flag = 0 if missing(flag)
Remove variables, or observations with `if`/`in`.
drop varlist | drop if expression | drop in range
. drop temp1 temp2 . drop if price < 0
Keep only the listed variables, or only observations satisfying a condition.
keep varlist | keep if expression | keep in range
. keep in 1/1000
Means, standard deviations, and ranges; `detail` adds percentiles, skewness, and kurtosis (Stata definitions).
summarize [varlist] [if] [in] [, detail]
| detail | percentiles p1–p99, skewness, kurtosis |
. summarize revenue price, detail
One-way frequency table or two-way cross-tabulation with totals.
tabulate varname [varname2] [if] [, missing]
| missing | include missing values as a category |
. tab region . tabulate region purchase
Pearson correlation matrix with listwise deletion.
correlate [varlist] [if]
. correlate revenue price orders
Ordinary least squares via LAPACK QR. Factor variables (i.var) expand to indicators; c.a#c.b forms continuous interactions.
regress depvar [indepvars] [if] [in] [, robust hc2 hc3 cluster(varname) noconstant level(#)]
| robust | HC1 heteroskedasticity-consistent SE (Stata's robust) |
| hc2 / hc3 | leverage-adjusted sandwich estimators |
| cluster(var) | cluster-robust SE, t on G−1 df |
| noconstant | suppress the intercept |
| level(#) | confidence level, default 95 |
. reg log_rev price, robust . reg log_rev price i.region, cluster(firm_id) . reg wage c.age#c.age education
Results match R to 1e-10. Listwise deletion is reported explicitly.
Logistic regression by Fisher scoring; reports LR χ², McFadden pseudo-R², z statistics.
logit depvar indepvars [if] [, robust cluster(varname) level(#)]
. logit purchase price, robust
Probit regression (normal link), same options as logit.
probit depvar indepvars [if] [, robust cluster(varname)]
. probit purchase price
Poisson regression for counts (log link).
poisson depvar indepvars [if] [, robust cluster(varname)]
. poisson orders price, robust
Two-stage least squares. Endogenous regressors and their instruments go in the parenthesized group.
ivregress 2sls depvar [exogvars] (endogvars = instruments) [, robust cluster(varname)]
. ivregress 2sls log_rev (price = z1 z2), robust
Residuals come from the original regressors; inference follows Stata's `small` convention (t statistics on N−K df).
Panel fixed-effects (within) estimator. Declare the panel with xtset first.
xtreg depvar indepvars, fe [cluster(panelvar)]
| fe | fixed effects (required — the only estimator so far) |
| cluster(panelvar) | panel-clustered SE, t on G−1 df |
. xtset firm_id . xtreg log_rev price, fe cluster(firm_id)
L1-penalized linear regression (coordinate descent, glmnet conventions). Selects variables by zeroing coefficients.
lasso depvar indepvars, lambda(#)
| lambda(#) | penalty strength; omit it to see lambda_max for your data |
. lasso log_rev price z1 z2 orders, lambda(0.05)
Elastic-net linear regression mixing L1 and L2 penalties.
elasticnet depvar indepvars, lambda(#) [alpha(#)]
| alpha(#) | 1 = lasso, 0 = ridge; default 1 |
| lambda(#) | penalty strength |
. elasticnet log_rev price z1 z2, alpha(0.4) lambda(0.02)
Matches R glmnet, including its gaussian y-standardization convention (ridge penalties scale with sd of the response).
Gradient-boosted regression trees (xgboost-style exact greedy, squared loss). Deterministic — no subsampling.
boost depvar features, rounds(#) [eta(#) maxdepth(#) lambda(#)]
| rounds(#) | number of trees |
| eta(#) | learning rate, default 0.3 |
| maxdepth(#) | tree depth, default 6 |
| lambda(#) | L2 regularization on leaf weights, default 1 |
. boost log_rev z1 z2 orders, rounds(100) eta(0.1) maxdepth(3) . predict yhat
Matches R xgboost predictions on identical settings. Training R² is in-sample — expect it to be optimistic.
Pairs bootstrap of a regression. Large replication counts run batched on the Apple GPU; results are identical either way for a given seed.
bootstrap, reps(#) [seed(#)]: regress depvar indepvars
| reps(#) | number of replications |
| seed(#) | Philox seed — full reproducibility across CPU and GPU |
. bootstrap, reps(10000) seed(42): reg log_rev price
Replicates are counter-addressable: any subset recomputes identically regardless of chunking or backend. The planner dispatches ≥500 reps to the GPU when available.
Bayesian linear regression by Gibbs sampling: posterior means, standard deviations, and 95% credible intervals.
bayes [, mcmcsize(#) burnin(#) seed(#) normalprior(#)]: regress depvar indepvars
| mcmcsize(#) | posterior draws after burn-in (default 10000) |
| burnin(#) | discarded warm-up iterations (default 2500) |
| seed(#) | Philox seed — chains are exactly reproducible |
| normalprior(#) | prior variance of the N(0, #) coefficient priors (default 10000) |
. bayes, mcmcsize(20000) seed(42): reg log_rev price
Priors: coefficients N(0, normalprior), variance InvGamma(0.01, 0.01) — Stata's bayes defaults. With diffuse priors the posterior reproduces OLS.
Permutation test: the response is permuted, the model refit, and empirical two-sided p-values reported per coefficient.
permute, reps(#) [seed(#)]: regress depvar indepvars
. permute, reps(1000) seed(42): reg log_rev orders
Set session parameters. `set seed` fixes the Philox key used by bootstrap and permute.
set seed #
. set seed 42
Generate predictions from the last estimation over all current observations.
predict newvar [, xb residuals pr n]
| xb | linear prediction (default after regress/ivregress) |
| residuals | response residuals |
| pr | predicted probability (default after logit/probit) |
| n | predicted mean count (default after poisson) |
. reg log_rev price . predict yhat . predict e, residuals
Average marginal effects with delta-method standard errors. After OLS/IV the effect is the coefficient; after logit/probit/poisson it averages dμ/dx over the estimation sample.
margins, dydx(varlist) [level(#)]
. logit purchase price . margins, dydx(price)
Continuous regressors only for now; factor and interaction terms are rejected with a message.
Scatter plot (also available as `graph scatter`). `by()` splits into colored series.
scatter yvar xvar [if] [, by(varname)]
. scatter log_rev price, by(region)
Frequency histogram of one variable (Sturges bins by default).
histogram varname [if] [, bins(#)]
| bins(#) | override the bin count |
. histogram revenue, bins(12)
Kernel density estimate (Epanechnikov kernel, Silverman bandwidth).
kdensity varname [if]
. kdensity revenue
General plotting front end: graph scatter, graph line, graph histogram.
graph scatter|line yvar xvar [if] [, by(varname)]
. graph line gdp year
Declare the panel structure (unit and, optionally, time variable) for xtreg.
xtset panelvar [timevar]
. xtset firm_id year
Evaluate and print an expression.
display expression
. display 2 + 2 * 3 . display ln(100)
Record commands and output to a text file.
log using filename | log close
. log using session.log
Show this reference, or the entry for one command.
help [command]
. help regress
Sample native plugin: generates z_varname, the standardized version of a variable.
zscore varname
. zscore revenue
Native plugins are Swift types conforming to ZQCommandPlugin, compiled into the app. Script commands are .zyq files in ~/Library/Application Support/Metrika/Commands/ — an optional leading `args name…` line names positional arguments, referenced as `name' or `1' in the body.