balanceplot help file

The installed help file, rendered for the web

This page is generated from balanceplot.sthlp by build.do, so it matches the installed version. In Stata, type help balanceplot.

Title

balanceplot – Plot covariate imbalance for categorical or continuous treatments

Syntax

balanceplot varlist ifin, group(varname) [outcome(varname) base(#) level(#) noci matchweight(varname) cohensh fadens sort absolute threshold(#) graphop(options) leg1(string) leg2(string) plotcommand table tablefull store(stubname[, replace]) leftmargin(#) decimals(#) width(#) labwidth(#)]

balanceplot varlist ifin, contreat(varname) [outcome(varname) level(#) noci fadens sort absolute threshold(#) graphop(options) plotcommand table tablefull store(stubname[, replace]) leftmargin(#) decimals(#) width(#) labwidth(#)]

balanceplot [varlist], tebalance [sort absolute threshold(#) graphop(options) plotcommand leftmargin(#)]

Description

Balanceplot calculates and plots covariate imbalance. Specify group(), contreat(), or tebalance. With group(), standardized imbalance is calculated across all observed categories of a categorical grouping variable. With contreat(), correlations are calculated between a continuous treatment and each covariate row. After a supported teffects or stteffects model, tebalance compares the raw and matched or weighted standardized differences returned by tebalance summarize.

All variables in varlist are included in the balance analysis.

Factor syntax is required in varlist. Bare variables and variables specified with c. are treated as continuous. Variables specified with i. or ib#. are treated as binary/nominal.

A binary categorical covariate contributes only its nonreference category. A nominal categorical covariate contributes every category. With group(), the nominal reference category is a zero-valued placeholder row labeled (ref). With contreat(), every nominal category, including the reference category, receives an estimated correlation. Nominal covariates are displayed under a heading; continuous and binary covariates are not. Use factor-variable base syntax, such as ib2.race, to change the displayed covariate reference category.

With group(), every nonreference row is reported as a standardized mean difference: (mean_group - mean_base) / sqrt((sd_group^2 + sd_base^2)/2). With cohensh, binary and nominal variables specified with i. or ib#. are instead reported as Cohen’s h: 2*asin(sqrt(p_group)) - 2*asin(sqrt(p_base)). Continuous variables are always reported on the standardized-mean-difference scale.

With contreat(), polychoric is called separately for each covariate row so its standard error can be used for confidence intervals and p-values. Continuous covariates receive Pearson correlations; binary indicators receive polyserial correlations. Nominal covariates are converted to category indicators and receive one polyserial correlation per category. The user-written polychoric command must be installed.

Confidence intervals are plotted by default. Non-significant imbalance can be visually emphasized on the graph with fadens.

Balance plots are often used to examine balance before and after matching. matchweight() plots the unweighted and weighted results together. See details below.

Options

group(varname) specifies the categorical grouping variable. Imbalance statistics are calculated across all categories of this variable. group() and contreat() may not be combined.

contreat(varname) specifies a continuous treatment variable. The graph reports its correlation with each covariate row. contreat() may not be combined with group(), base(), cohensh, or matchweight().

tebalance is a postestimation mode for supported teffects and stteffects estimators. It quietly calls tebalance summarize and plots the first two columns of its r(table) matrix as unweighted and matched or weighted standardized differences. An optional varlist restricts the plot to selected treatment-model covariates. With a multivalued treatment, each nonbase treatment comparison is displayed as a separate covariate block under its treatment-category heading. Confidence intervals and fadens are unavailable because tebalance summarize does not return standard errors for these statistics. sort, absolute, threshold(), graphop(), and plotcommand are supported.

base(#) selects the base/reference category of group() and may not be used with contreat(). Each nonbase group is compared with the group base. For a binary grouping variable, the lowest category is the default. For three or more categories, the largest complete-case category is the default.

cohensh reports Cohen’s h statistics for binary and nominal variables when group() is specified. Imbalance statistics for continuous variables are unchanged. cohensH is a synonym.

outcome(varname) allows a hypothetical outcome variable which will be used in listwise deletion to determine the sample for balanceplot, but is not shown in the plot. Useful when you want balanceplot’s sample to match a subsequent model with all of varlist and an outcome().

matchweight(varname) compares unweighted and weighted balance in the same graph. Matching weights are applied to the base category of group() only; all nonbase categories receive weight 1. Each group comparison retains one color, while marker symbols distinguish unweighted and weighted estimates. Missing or zero base-group weights are retained in the unweighted results and excluded from the weighted results. It may not be combined with contreat(). This matches ATT-style output from psmatch2, where controls are ordinarily coded 0 and treated observations are coded 1. If the control category is not the command’s default base, specify it with base().

level(#) sets the confidence interval level. The default is 95.

noci suppresses confidence intervals in the graph. Confidence intervals remain in the returned matrices and table (if requested).

fadens fades nonsignificant point estimates and confidence intervals on the graph. With matchweight(), unweighted and weighted estimates are faded according to their own p-values.

sort By default, balanceplot retains covariates in the order specified in varlist and keeps all categories of a nominal variable together under a heading. sort sorts variables from negative to positive imbalance. With absolute, it sorts from the smallest to largest absolute imbalance. With matchweight(), sorting is based on the weighted estimates. Variable headings are omitted when sort is specified because categories from the same nominal variable may no longer remain together.

absolute plots the absolute magnitude of each imbalance statistic. Returned matrices and printed tables retain the signed statistics. Confidence intervals are reflected onto the absolute scale for the graph; an interval that crosses zero begins at zero. abs is the shortest allowed abbreviation.

threshold(#) adds dashed light-gray reference lines at -# and #. With absolute, only the positive threshold line is shown. The threshold must be greater than zero. With contreat(), it must also be no greater than one. The zero reference line remains unchanged.

store(stubname[, replace]) posts results as separate estimation results for use with esttab, estout, or official estimates commands. For an unweighted two-category group() analysis, the generated names are stubname_mean_g# for each group and stubname_imbalance. With three or more categories, each nonbase comparison is stored as stubname_imb_g#. Mean results contain e(b) only. Imbalance results contain e(b) and a diagonal e(V), so significance information is calculated from the stored estimate and standard error.

With matchweight(), store() posts complete unweighted and weighted sets. The prefixes are stubname_unw_ and stubname_w_. For example, a binary analysis with store(bp) creates bp_unw_mean_g0, bp_unw_mean_g1, bp_unw_imbalance, bp_w_mean_g0, bp_w_mean_g1, and bp_w_imbalance. With contreat(), the correlation result is stored as stubname_correlation, with correlations in e(b) and squared polychoric standard errors on the diagonal of e(V).

The replace suboption replaces estimation results whose generated names already exist. Without replace, balanceplot reports an error before posting any results. For example, specify store(bp, replace) to replace results previously created from store(bp).

graphop(options) passes graph options to coefplot.

table displays one compact table for each group comparison. Its columns are the two group-specific means, Standardized Imbalance, and p-value. With matchweight(), an unweighted table is followed by a weighted table for each comparison. With contreat(), it displays the correlation and p-value. width() and labwidth() customize the table size.

tablefull displays the same table but with more details: the two group-specific means, Standardized Imbalance, standard error, lower confidence limit, upper confidence limit, and p-value. With matchweight(), an unweighted table is followed by a weighted table for each comparison. With contreat(), it displays the correlation, standard error, confidence limits, and p-value. table and tablefull may not be combined.

decimals(#) sets the number of digits after the decimal in printed tables. The default is 3.

width(#) sets the display width of every statistic column in printed tables. The default is 10.

labwidth(#) sets the display width of the leftmost column containing variable and category labels. The default is 24. Labels longer than this width are abbreviated.

Stored results

balanceplot stores the first nonbase comparison in r(bias1), the second in r(bias2), and subsequent comparisons in r(bias3), and so on. Each matrix contains mean_base, mean_ref, ttest_pval, std_diff, std_err, ci_low, and ci_high. Under cohensh, the last four columns are on the Cohen’s h scale for factor-indicator rows. With matchweight(), these established returns contain the weighted results; unweighted results are additionally stored in r(bias_unweighted1), r(bias_unweighted2), and so on.

Export-ready compact matrices are stored in r(table1), r(table2), and so on. Their columns are the two group-specific means, std_imbalance, and p_value. Detailed matrices are stored in r(tablefull1), r(tablefull2), and so on; they additionally contain std_err, ci_low, and ci_high. Mean-column names encode the actual group values, such as mean_g0 and mean_g1. All confidence limits occupy separate matrix columns. These matrices are returned whether or not a printed table is requested. With matchweight(), weighted results retain these names and unweighted counterparts are stored in r(table_unweighted#) and r(tablefull_unweighted#).

With store(stubname), group means and imbalance statistics are also posted as estimation results. For example, store(bp) with an unweighted binary grouping variable creates bp_mean_g0, bp_mean_g1, and bp_imbalance. The same coefficient rows appear in every result, so esttab bp_mean_g0 bp_mean_g1 bp_imbalance, se label mtitles produces three aligned columns. Group-mean results contain e(b) without a substantive e(V). Imbalance results contain the imbalance estimates in e(b) and squared standard errors on the diagonal of e(V). With cohensh, esttab derives significance from Cohen’s h and its stored standard error.

With matchweight(), the unweighted and weighted stored sets use _unw_ and _w_ in their names. Each set can therefore be passed to esttab independently. With contreat(), stubname_correlation contains the signed correlations in e(b) and their squared standard errors in e(V). Stored results remain signed when absolute is specified.

The names created by store() are returned in r(stored_estimates); r(storestub) contains the requested stub and r(stored) indicates whether estimation results were posted.

With contreat(), results are returned in r(correlation). Its columns are correlation, std_err, ci_low, ci_high, and p_value. The command also returns r(N) and r(contreat). When store() is requested, the generated correlation estimate name is listed in r(stored_estimates).

With tebalance, r(table) and r(size) reproduce the matrices returned by tebalance summarize, while r(balance) contains its two standardized-difference columns. r(adjustment) identifies the second series as Matched or Weighted. Tebalance mode also returns r(nrows), r(absolute), r(threshold), r(width), r(labwidth), r(rownames), r(measure), r(xtitle), r(sortmode), r(tablemode), r(headings), r(coeflabels), r(plotcommand), and r(mode). Storage is not supported in tebalance mode.

Group and continuous-treatment modes return r(nrows), r(level), r(fadens), r(fadealpha), r(absolute), r(threshold), r(stored), r(storestub), r(stored_estimates), r(width), r(labwidth), r(rownames), r(outcome), r(measure), r(xtitle), r(graphnote), r(sortmode), r(tablemode), r(headings), r(plotcommand), r(mode), and r(contreat). Group mode additionally returns r(base), r(ngroups), r(cohensh), r(weighted), r(base_n), r(base_n_unweighted), r(base_sumw), r(groups), r(matrices), r(tablematrices), r(tablefullmatrices), r(unweighted_matrices), r(table_unweighted_matrices), r(tablefull_unweighted_matrices), and r(weightvar).

Examples

sysuse nlsw88, clear
balanceplot wage age i.married i.race tenure, group(union)
balanceplot wage age i.married i.race tenure, group(union) cohensh
balanceplot wage age i.married i.race tenure, group(union) fadens
balanceplot wage age i.married i.race tenure, group(union) cohensH noci
balanceplot wage age i.married i.race tenure, group(union) sort
balanceplot wage age i.married i.race tenure, group(union) absolute sort threshold(.1)
balanceplot wage age i.married i.race tenure, group(union) threshold(.1)
balanceplot wage age i.married i.race tenure, group(union) table
balanceplot wage age i.married i.race tenure, group(union) tablefull
balanceplot wage age i.married i.race tenure, group(union) table labwidth(30) width(11)
matrix list r(table1)
matrix list r(tablefull1)

Store results for esttab:

balanceplot wage age i.married i.race tenure, group(union) store(bp)
esttab bp_mean_g0 bp_mean_g1 bp_imbalance, se label mtitles
balanceplot wage age i.married tenure, group(race) store(br)
esttab br_mean_g1 br_mean_g2 br_mean_g3 br_imb_g2 br_imb_g3, se label mtitles

Continuous treatment:

balanceplot age i.union i.race tenure, contreat(wage)
balanceplot age i.union i.race tenure, contreat(wage) threshold(.1)
balanceplot age i.union i.race tenure, contreat(wage) abs sort threshold(.1)
balanceplot age i.union i.race tenure, contreat(wage) fadens sort tablefull
balanceplot age i.union i.race tenure, contreat(wage) store(ct)
esttab ct_correlation, se label mtitles
matrix list r(correlation)

More than two groups:

balanceplot wage age i.married tenure, group(race)

Matching-weight examples

After ATT matching with psmatch2, _weight equals 1 for treated observations and records how often or how strongly each matched control contributes. Unmatched controls have missing weights. matchweight() displays both the original complete-case balance and the weighted matched balance. With a conventional 0/1 treatment variable:

sysuse nlsw88, clear
psmatch2 union age tenure grade, neighbor(1) logit
balanceplot age tenure grade, group(union) matchweight(_weight)
balanceplot age tenure grade, group(union) matchweight(_weight) store(bm)
esttab bm_unw_mean_g0 bm_unw_mean_g1 bm_unw_imbalance, se label mtitles
esttab bm_w_mean_g0 bm_w_mean_g1 bm_w_imbalance, se label mtitles

Official teffects and stteffects estimators provide raw and matched or weighted standardized differences through tebalance summarize. balanceplot, tebalance calls that command internally and plots the results:

sysuse nlsw88, clear
teffects psmatch (wage) (union age tenure grade), atet
balanceplot, tebalance
balanceplot age tenure, tebalance absolute sort threshold(.1)