What is not built, and how far to trust what is
Read this before you rely on a number.
What we have
- checked against a real R (4.6.1): 34 of 46 translations, plus
svy: regress,svy: tabulate, the survival results,ologit, the decomposition and the moving average - new estimators are compared with older checked ones and with independent implementations written separately in the tests
Coming later
- nothing has ever been compared with real Stata — there is no Stata licence on the machine this was built on
- panel data (
xtset,xtreg), the largest single gap - Stata date formats;
arima,dfullerand forecasting predict,test,marginsand the rest of postestimationsvy:for models other thanregressandtabulate
How the results have been checked
Nothing here has been compared with real Stata. There is no Stata licence on the machine this was built on, so no output has ever been placed side by side with Stata's. Anywhere you see a claim about accuracy in this manual, it means one of these:
- Checked against R. The command was run in a real R (4.6.1) and the numbers matched. This covers 34 of the 46 translations, plus
svy: regress,svy: tabulate(the F test exactly), the survival results,ologit(including the sign convention and cut points), the decomposition and the moving average. - Checked against itself. A new estimator is compared with an older one that was checked, and with an independent brute-force implementation written separately in the test suite.
- Not checked. Said plainly wherever it applies.
The difference matters. R and Stata agree on almost everything, but "our numbers match R" and "our numbers match Stata" are different sentences, and only the first one is true here.
Not built yet
Panel data. xtset and xtreg do not exist. This is the largest gap.
Dates. A time variable is a plain whole number — 1, 2, 3, or 2019, 2020. Stata's date formats and functions are not implemented.
Time series. No arima, no Holt-Winters or exponential smoothing, no dfuller, no forecasting.
Survival. No delayed entry, no several records per subject, no time-varying covariates, no stratified or parametric models, no weights, no Nelson-Aalen, and no test other than the log-rank. Each of these is refused with a message rather than silently ignored.
Survey. svy: covers mean, total, proportion, ratio, regress and tabulate. Other models are not available under it. No replicate weights, no pooling of plausible values.
Models. No instrumental variables.
Charts. 21 of the 105 chart types in the catalogue are built. The chart gallery marks which.
Things to keep in mind
set mincell is small-cell suppression within a single table. It is not full disclosure control.
A do-file can run arbitrary Python, by design — set language python inside a do-file is a feature. Treat a do-file from someone else the way you would treat any script from someone else.
tssmooth ma's behaviour at the ends of a series has not been compared with Stata.
The name
"Stata" is a trademark of StataCorp LLC. Virata is not affiliated with, endorsed by, or connected to StataCorp in any way. It reads a language that looks like Stata's; it is not Stata and does not claim to produce Stata's numbers.