中文

Postestimation and diagnostics

Start with the expanded example for each command. Open the variants for options and alternative R approaches. Read the qualifications before adapting the code.

brier

Calculate the overall Brier mean probability score

Stata

brier b p

R Package:stats

mean((data$b-data$p)^2)

Differs from Stata Returns the overall score only; Stata brier also reports decomposition components.

Stata reference · R reference

dfbeta

Obtain standardized deletion influence on one slope

Stata

dfbeta x

R Package:stats

stats::dfbetas(fit)[,"x"]

Differs from Stata Returns standardized coefficient-deletion diagnostics from dfbetas for x; coefficient ordering, missing rows and diagnostic naming must be aligned with the Stata model.

Stata reference · R reference

estat

Test heteroskedasticity against one predictor using a studentized statistic

Stata

estat hettest x, iid

R Package:lmtest

lmtest::bptest(fit,varformula=~x,data=data,studentize=TRUE)

Differs from Stata Uses the studentized Koenker version of the Breusch-Pagan test, corresponding to the iid variant shown. Match the auxiliary regressors and estimation sample; the default bptest differs from default Stata hettest.

Stata reference · R reference

9 more examples and alternatives

Run the classical Breusch-Pagan heteroskedasticity test

Stata

estat hettest, rhs

R Package:lmtest

lmtest::bptest(fit,studentize=FALSE)

Differs from Stata studentize=FALSE selects the normal-theory version; default bptest uses the Koenker version.

Stata reference · R reference

Compare penalized-likelihood model criteria

Stata

estat ic

R Package:stats

c(AIC=stats::AIC(fit),BIC=stats::BIC(fit))

R keeps fitted model objects; model names are fixed workflow labels.

Stata reference · R reference

Run a White-style auxiliary-regression heteroskedasticity test

Stata

estat imtest, white

R Package:lmtest

lmtest::bptest(fit,varformula=~x+z+I(x^2)+I(z^2)+x:z,data=data,studentize=TRUE)

This covers the White heteroskedasticity component, not the complete information-matrix decomposition.

Stata reference · R reference

Run Ramsey RESET using powers of fitted values

Stata

estat ovtest

R Package:lmtest

lmtest::resettest(fit,power=2:4,type="fitted")

Differs from Stata Uses RESET powers two through four of the fitted values or regressors as shown. Regressor-factor expansion and rank-deficient auxiliary terms may differ from Stata.

Stata reference · R reference

Run RESET using powers of regressors

Stata

estat ovtest, rhs

R Package:lmtest

lmtest::resettest(fit,power=2:4,type="regressor")

Differs from Stata Uses RESET powers two through four of the fitted values or regressors as shown. Regressor-factor expansion and rank-deficient auxiliary terms may differ from Stata.

Stata reference · R reference

Summarize variables in the actual estimation sample

Stata

estat summarize

R Package:stats

summary(stats::model.frame(fit))

R keeps fitted model objects; model names are fixed workflow labels.

Stata reference · R reference

Extract the coefficient covariance matrix

Stata

estat vce

R Package:stats

stats::vcov(fit)

R keeps fitted model objects; model names are fixed workflow labels.

Stata reference · R reference

Extract correlations among coefficient estimates

Stata

estat vce, correlation

R Package:stats

stats::cov2cor(stats::vcov(fit))

R keeps fitted model objects; model names are fixed workflow labels.

Stata reference · R reference

Compute variance inflation factors after OLS

Stata

estat vif

R Package:car

car::vif(fit)

Differs from Stata car returns ordinary VIFs for one-degree-of-freedom terms and generalized VIFs for multi-parameter terms. Stata estat vif reports column-level VIFs, so factor-term output is not directly interchangeable.

Stata reference · R reference

estimates

List stored model names

Stata

estimates dir

R Package:stats

names(models)

R keeps fitted model objects; model names are fixed workflow labels.

Stata reference · R reference

6 more examples and alternatives

Redisplay a fitted linear model summary

Stata

estimates replay

R Package:stats

summary(fit)

R keeps fitted model objects; model names are fixed workflow labels.

Stata reference · R reference

Restore one model from a model collection

Stata

estimates restore model1

R Package:stats

fit <- models[["model1"]]

R keeps fitted model objects; model names are fixed workflow labels.

Stata reference · R reference

Save and reload a fitted model object

Stata

estimates save model1

R Package:base

{ path <- tempfile(fileext=".rds"); saveRDS(fit,path); readRDS(path) }

Writes to an R temporary file; choose an explicit project path for persistent storage.

Stata reference · R reference

Compare AIC across stored model objects

Stata

estimates stats model1 model2

R Package:stats

do.call(stats::AIC,unname(models))

R keeps fitted model objects; model names are fixed workflow labels.

Stata reference · R reference

Store a fitted model under a name

Stata

estimates store model1

R Package:stats

model1 <- fit

R keeps fitted model objects; model names are fixed workflow labels.

Stata reference · R reference

Compare coefficients and standard errors across models

Stata

estimates table model1 model2, b se

R Package:stats

lapply(models,function(m) summary(m)$coefficients[,1:2,drop=FALSE])

R keeps fitted model objects; model names are fixed workflow labels.

Stata reference · R reference

lincom

Estimate the sum of two slopes with its standard error

Stata

lincom x + z

R Package:stats

{ L <- c(0,1,1); c(estimate=sum(L*stats::coef(fit)),se=sqrt(drop(t(L)%*%stats::vcov(fit)%*%L))) }

Differs from Stata The contrast vector assumes coefficient order intercept, x, z. Returns the estimate and its covariance-based standard error only; Stata also prints tests and confidence limits.

Stata reference · R reference

1 more examples and alternatives

Estimate a difference of slopes with its standard error

Stata

lincom x - z

R Package:stats

{ L <- c(0,1,-1); c(estimate=sum(L*stats::coef(fit)),se=sqrt(drop(t(L)%*%stats::vcov(fit)%*%L))) }

Differs from Stata The contrast vector assumes coefficient order intercept, x, z. Returns the estimate and its covariance-based standard error only; Stata also prints tests and confidence limits.

Stata reference · R reference

linktest

Regress an outcome on fitted values and their square

Stata

linktest

R Package:stats

{ data$hat <- stats::fitted(fit); stats::lm(y ~ hat+I(hat^2),data=data) }

Shown for OLS; link tests after other model families must refit the appropriate family.

Stata reference · R reference

lrtest

Compare nested maximum-likelihood models

Stata

lrtest reduced full

R Package:lmtest

lmtest::lrtest(reduced,full)

Differs from Stata Fixed workflow.

Stata reference · R reference

margins

Predict at specified covariate values

Stata

margins, at(x=(2 3 4))

R Package:stats

vapply(c(2,3,4),function(value) { newdata <- data; newdata$x <- value; mean(stats::predict(fit,newdata=newdata,type="response")) },numeric(1))

Differs from Stata At each x value, predictions are averaged over the observed distribution of z.

Stata reference · R reference

3 more examples and alternatives

Average a continuous-predictor marginal effect

Stata

margins, dydx(x)

R Package:stats

{ X <- stats::model.matrix(fit); beta <- stats::coef(fit); p <- stats::fitted(fit); gradient <- beta["x"]*colMeans(X*as.vector(p*(1-p)*(1-2*p))); gradient["x"] <- gradient["x"]+mean(p*(1-p)); c(estimate=unname(beta["x"]*mean(p*(1-p))),se=sqrt(drop(t(gradient)%*%stats::vcov(fit)%*%gradient))) }

Differs from Stata Uses the fitted GLM and its estimation data; reference grids and averaging rules should be chosen explicitly.

Stata reference · R reference

Evaluate a marginal effect at mean covariates

Stata

margins, dydx(x) atmeans

R Package:stats

{ X <- colMeans(stats::model.matrix(fit)); beta <- stats::coef(fit); p <- stats::plogis(sum(X*beta)); c(estimate=unname(beta["x"]*p*(1-p))) }

Differs from Stata Returns the derivative at means; use a delta-method gradient if standard errors are needed.

Stata reference · R reference

Average predictions on the linear-index scale

Stata

margins, expression(predict(xb))

R Package:stats

{ L <- colMeans(stats::model.matrix(fit)); c(estimate=sum(L*stats::coef(fit)),se=sqrt(drop(t(L)%*%stats::vcov(fit)%*%L))) }

Differs from Stata Uses the fitted GLM and its estimation data; reference grids and averaging rules should be chosen explicitly.

Stata reference · R reference

nlcom

Use the delta method for a ratio of slopes

Stata

nlcom _b[x]/_b[z]

R Package:car

car::deltaMethod(fit,"x / z")

Differs from Stata Uses the first-order delta method on named fitted coefficients. Returns the transformed estimate and standard error; Stata also prints reference tests and confidence limits.

Stata reference · R reference

1 more examples and alternatives

Use the delta method for an exponentiated slope

Stata

nlcom exp(_b[x])

R Package:car

car::deltaMethod(fit,"exp(x)")

Differs from Stata Uses the first-order delta method on named fitted coefficients. Returns the transformed estimate and standard error; Stata also prints reference tests and confidence limits.

Stata reference · R reference

predict

Obtain Cook influence distances

Stata

predict result, cooksd

R Package:stats

data$result <- stats::cooks.distance(fit)

Differs from Stata Run after fit <- lm(y ~ x + z, data=data).

Stata reference · R reference

12 more examples and alternatives

Compute binary-model deviance residuals

Stata

predict result, deviance

R Package:stats

data$result <- stats::residuals(fit,type="deviance")

Differs from Stata After fitting a binary-response GLM. Prediction options are model-specific.

Stata reference · R reference

Obtain DFFITS influence statistics

Stata

predict result, dfits

R Package:stats

data$result <- stats::dffits(fit)

Differs from Stata Run after fit <- lm(y ~ x + z, data=data).

Stata reference · R reference

Obtain leverage values

Stata

predict result, hat

R Package:stats

data$result <- stats::hatvalues(fit)

Differs from Stata Run after fit <- lm(y ~ x + z, data=data).

Stata reference · R reference

Compute binary-model Pearson residuals

Stata

predict result, pearson

R Package:stats

data$result <- stats::residuals(fit,type="pearson")

Differs from Stata After fitting a binary-response GLM. Prediction options are model-specific.

Stata reference · R reference

Predict logistic probabilities

Stata

predict result, pr

R Package:stats

data$result <- stats::predict(fit,type="response")

Differs from Stata After fitting a binary-response GLM. Prediction options are model-specific.

Stata reference · R reference

Obtain raw OLS residuals

Stata

predict result, residuals

R Package:stats

data$result <- stats::residuals(fit)

Differs from Stata Run after fit <- lm(y ~ x + z, data=data).

Stata reference · R reference

Obtain internally standardized residuals

Stata

predict result, rstandard

R Package:stats

data$result <- stats::rstandard(fit)

Differs from Stata Run after fit <- lm(y ~ x + z, data=data).

Stata reference · R reference

Obtain externally studentized residuals

Stata

predict result, rstudent

R Package:stats

data$result <- stats::rstudent(fit)

Differs from Stata Run after fit <- lm(y ~ x + z, data=data).

Stata reference · R reference

Obtain forecast standard errors including residual variance

Stata

predict result, stdf

R Package:stats

data$result <- sqrt(stats::predict(fit,se.fit=TRUE)$se.fit^2+summary(fit)$sigma^2)

Differs from Stata Adds residual variance to the squared standard error of the fitted OLS mean. This is a standard error for a new individual prediction, not the fitted mean; assumes homoskedastic errors and matching model sample.

Stata reference · R reference

Obtain standard errors of fitted means

Stata

predict result, stdp

R Package:stats

data$result <- stats::predict(fit,se.fit=TRUE)$se.fit

Differs from Stata Run after fit <- lm(y ~ x + z, data=data).

Stata reference · R reference

Obtain linear fitted values

Stata

predict result, xb

R Package:stats

data$result <- stats::predict(fit)

Differs from Stata Run after fit <- lm(y ~ x + z, data=data).

Stata reference · R reference

Predict logistic linear indices

Stata

predict result, xb

R Package:stats

data$result <- stats::predict(fit,type="link")

Differs from Stata After fitting a binary-response GLM. Prediction options are model-specific.

Stata reference · R reference

roccomp

Compare paired marker ROC areas using DeLong inference

Stata

roccomp b x z

R Package:pROC

pROC::roc.test(pROC::roc(data$b,data$x,levels=c(0,1),direction="<",quiet=TRUE),pROC::roc(data$b,data$z,levels=c(0,1),direction="<",quiet=TRUE),paired=TRUE,method="delong")

Differs from Stata Both markers must be measured on the same complete-case subjects for paired inference.

Stata reference · R reference

roctab

Estimate a nonparametric ROC area for a marker

Stata

roctab b x

R Package:pROC

pROC::roc(data$b,data$x,levels=c(0,1),direction="<",quiet=TRUE)

Differs from Stata Explicit direction means larger markers indicate cases; fixing direction prevents data-driven reversal.

Stata reference · R reference

Tabulate sensitivity and specificity across ROC thresholds

Stata

roctab b x, detail

R Package:pROC

pROC::coords(pROC::roc(data$b,data$x,levels=c(0,1),direction="<",quiet=TRUE),x="all",ret=c("threshold","sensitivity","specificity"))

Differs from Stata Returns all pROC thresholds with sensitivity and specificity for explicit event direction. Threshold locations and endpoint conventions differ from Stata roctab detail cutpoints.

Stata reference · R reference

test

Test a single regression coefficient

Stata

test x = 0

R Package:car

car::linearHypothesis(fit,"x = 0")

Differs from Stata Tests the specified coefficient restriction on the existing fit. car selects its reference statistic from the model class; a GLM chi-square test differs from the OLS F test used by Stata regress postestimation.

Stata reference · R reference

2 more examples and alternatives

Test equality of two regression slopes

Stata

test x = z

R Package:car

car::linearHypothesis(fit,"x = z")

Differs from Stata Tests the specified coefficient restriction on the existing fit. car selects its reference statistic from the model class; a GLM chi-square test differs from the OLS F test used by Stata regress postestimation.

Stata reference · R reference

Jointly test two regression coefficients

Stata

test x z

R Package:car

car::linearHypothesis(fit,c("x = 0","z = 0"))

Differs from Stata Tests the specified coefficient restriction on the existing fit. car selects its reference statistic from the model class; a GLM chi-square test differs from the OLS F test used by Stata regress postestimation.

Stata reference · R reference

testnl

Test an exponentiated coefficient with a nonlinear Wald test

Stata

testnl exp(_b[x]) = 1

R Package:car

{ d <- car::deltaMethod(fit,"exp(x)"); w <- ((d$Estimate-1)/d$SE)^2; c(chisq=w,p.value=stats::pchisq(w,df=1,lower.tail=FALSE)) }

Differs from Stata Forms a one-degree-of-freedom Wald chi-square from a first-order delta-method standard error. The R vector omits Stata coefficient and confidence-interval output.

Stata reference · R reference

testparm

Jointly test a categorical predictor

Stata

testparm i.g

R Package:car

car::linearHypothesis(fit,c("g6 = 0","g8 = 0"))

Differs from Stata The restrictions name the fitted factor coefficients explicitly. Different reference levels or contrasts require different coefficient names; this is not an automatic expansion of any Stata factor-variable expression.

Stata reference · R reference