The vce option: changing standard error reports for quirky data
The vce() option causes Stata to change the way standard error is calculated. The vce option has three major types of variance estimators: likelihood-based, replication-based and sandwich estimators.
The two likelihood estimator subcommands are vce(oim) short for observed information matrix and vce(opg) short for outer product of the gradient vectors. Both refer to the matrices and math which underly the procedure.
The two replication-based estimators are vce(bootstrap) and vce(jackknife) To oversimplify, bootstrapping takes a series of random samples from the sample and uses this constant sampling with replacement to calculate standard error. This makes it useful for populations whose normality is uncertain. A brief explanation (with minimal math) of bootstrapping can be found here. Jackknifing is a somewhat similar procedure but where bootstrapping does relatively infinite sampling with replacement, Jackknifing does resampling equal to n and each iteration takes exactly one person out of the sample and recalculates the desired statistic.
The two sandwich estimator subcommands are , vce(robust) which uses a Huber/Whites/sandwich estimator and , vce (cluster [cluster variable]. You can learn more about what the robust estimate entails by looking at the Stata FAQ on the matter and at the additional resources it suggests. Stata also offers a brief discussion of why it might be preferable to the regular estimates. Using the ,vce (cluster [cluster variable] command negates the need for independent observations, requiring only that from cluster to cluster the observations are independent. Additionally, the Stata User's Guide [U] has a subsection specifically on robust variance estimates and the logic behind them. Both of these adjustments alter the precise interpretation of your data, so be aware of the implications (also discussed in [U]) if you use them.
The , vce option is available with "most estimation commands" according the Stata Reference Book [R} Q-Z, for example regression gpa weight panda, vce(robust)
Whether or not a given specific command has , vce options will be specified if you type help [command]