{smcl}
{vieweralsosee "effecttab" "help effecttab"}{...}
{viewerjumpto "Package overview" "regtab##package"}{...}
{viewerjumpto "Syntax" "regtab##syntax"}{...}
{viewerjumpto "Description" "regtab##description"}{...}
{viewerjumpto "Options" "regtab##options"}{...}
{viewerjumpto "Remarks" "regtab##remarks"}{...}
{viewerjumpto "Examples" "regtab##examples"}{...}
{viewerjumpto "Stored results" "regtab##stored"}{...}
{viewerjumpto "Also see" "regtab##seealso"}{...}
{viewerjumpto "Author" "regtab##author"}{...}
{title:Title}

{p2colset 5 15 17 2}{...}
{p2col:{cmd:regtab} {hline 2}}Format collected regression results for publication-ready Excel tables{p_end}
{p2colreset}{...}


{marker package}{...}
{title:Package}

{pstd}
{cmd:regtab} is part of the {helpb tabtools} suite. See also {helpb effecttab}
for treatment effects and margins tables.

{hline}


{title:regtab}

{pstd}Format {helpb collect}ed regression results into a polished Excel table.{p_end}

{marker syntax}{title:Syntax}

{p 4 8 2}{cmd:regtab}, [{opt xlsx(filename)} {opt excel(filename)}
{opt sheet(string)} {opt sep(string)} {opt models(string)}
{opt coef(string)} {opt title(string)} {opt noint:ercept} {opt keepi:ntercept}
{opt nore:effects} {opt stats(string asis)} {opt relab:el} {opt digits(#)}
{opt l:evel(#)}
{opt foot:note(string)} {opt open} {opt zebra} {opt headers:hade}
{opt high:light(#)} {opt bold:p(#)} {opt border:style(string)}
{opt font(string)} {opt fontsize(#)} {opt headerc:olor(string)} {opt zebrac:olor(string)}
{opt csv(string)} {opt mark:down(filename)} {opt mdapp:end} {opt fra:me(name)}
{opt eplotf:rame(name[, replace])} {opt keep(varlist)}
{opt drop(varlist)} {opt labelm:atch} {opt dimnon:sig} {opt factorl:abel} {opt ref:cat(string)}
{opt omitl:abel(string)} {opt emptyl:abel(string)} {opt notestl:abel(string)}
{opt absentl:abel(string)} {opt cnsl:abel(string)}
{opt cutl:abels(string)} {opt comp:act} {opt nop:value} {opt stars}
{opt starsl:evels(numlist)} {opt addr:ow(string asis)} {opt pdp(#)} {opt highpdp(#)} {opt cdisc} {opt labelw:idth(#)}
{opt cf:ormat(%fmt)} {opt reft:op} {opt celln:ote(spec)} {opt minc:ount(#)}
{opt trans:pose} {opt exposurel:abel(string)} {opt cil:abel(string)}
{opt pl:abel(string)} {opt addc:ol(string asis)} {opt statl:abels(string asis)}
{opt coll:abels(string asis)}]{p_end}

{p 4 8 2}{cmd:tabtools fitcount}, {opt ev:ents(varname)} [{opt peo:ple(varname)}
{opt exp:osure(varname)} {opt terms} {opt nam:e(collection)}]{p_end}

{pstd}Required: an active {helpb collect} with items {cmd:_r_b}, {cmd:_r_ci},
and {cmd:_r_p} and dimensions including {cmd:colname} and {cmd:cmdset}.{p_end}

{marker description}{title:Description}

{pstd}{cmd:regtab} reads the current {helpb collect} table and writes a clean Excel sheet with,
for each model (each {cmd:cmdset}), columns for the point estimate ({cmd:_r_b}),
confidence interval ({cmd:_r_ci}), and p-value ({cmd:_r_p}). Use {opt nopvalue} to suppress the
p-value column in the rendered output. It preserves raw coefficient names
while applying labels, number formats, and row selection, optionally merges
model headers, writes to your target workbook/sheet, and styles borders,
alignment, fonts, and column widths. Title text can be written to cell {cmd:A1}; the
main table begins at {cmd:B2}.{p_end}

{marker options}{title:Options}

{synoptset 27 tabbed}{...}
{synoptline}
{synopt:{opt xlsx(string)}}output Excel filename (must end with .xlsx){p_end}
{synopt:{opt sheet(string)}}target sheet to create/replace in xlsx(){p_end}
{synopt:{opt sep(string)}}text between the CI limits; default {cmd:", "}{p_end}
{synopt:{opt models(string)}}set merged model labels{p_end}
{synopt:{opt coef(string)}}header for the estimate column{p_end}
{synopt:{opt title(string)}}title written to A1, merged across the table{p_end}
{synopt:{opt noint:ercept}}drop intercept, cutpoint, and ancillary rows{p_end}
{synopt:{opt keepi:ntercept}}retain the intercept row{p_end}
{synopt:{opt nore:effects}}omit random-effects rows{p_end}
{synopt:{opt stats(string asis)}}model-fit statistics, built-in or {cmd:e(}{it:name}{cmd:)}{p_end}
{synopt:{opt statl:abels(string asis)}}relabel built-in {opt stats()} rows{p_end}
{synopt:{opt digits(#)}}set decimals for coefficients and CIs{p_end}
{synopt:{opt cf:ormat(%fmt)}}display format for estimate and CI bounds{p_end}
{synopt:{opt level(#)}}verify the collection's confidence level{p_end}
{synopt:{opt labelw:idth(#)}}cap the label-column width{p_end}
{synopt:{opt foot:note(string)}}add italic footnote text{p_end}
{synopt:{opt open}}open the Excel file after export{p_end}
{synopt:{opt zebra}}apply alternating light gray row shading{p_end}
{synopt:{opt high:light(#)}}yellow fill for rows where p-value < #{p_end}
{synopt:{opt bold:p(#)}}bold p-value cells below #{p_end}
{synopt:{opt headers:hade}}apply background fill to the header row{p_end}
{synopt:{opt font(string)}}set the Excel font family{p_end}
{synopt:{opt fontsize(#)}}set the Excel font size in points{p_end}
{synopt:{opt border:style(string)}}set the table border style{p_end}
{synopt:{opt cdisc}}apply CDISC labels and defaults{p_end}
{synopt:{opt relab:el}}relabel random effects{p_end}
{synopt:{opt stars}}add coefficient significance stars{p_end}
{synopt:{opt starsl:evels(numlist)}}custom p-value thresholds for stars{p_end}
{synopt:{opt headerc:olor(string)}}set the header fill color{p_end}
{synopt:{opt zebrac:olor(string)}}set alternating-row fill color{p_end}
{synopt:{opt csv(filename)}}also export the table as a CSV file{p_end}
{synopt:{opt markdown(filename)}}export the table as GitHub-Flavored Markdown{p_end}
{synopt:{opt mdappend}}append to an existing Markdown file{p_end}
{synopt:{opt fra:me(name[, replace flat keys])}}store output in a named frame{p_end}
{synopt:{opt eplotf:rame(name[, replace])}}save a graph-ready companion frame{p_end}
{synopt:{opt keep(varlist)}}keep exact variable or factor names{p_end}
{synopt:{opt drop(varlist)}}drop exact variable or factor names{p_end}
{synopt:{opt labelm:atch}}match label substrings with keep()/drop(){p_end}
{synopt:{opt dimnon:sig}}gray out non-significant rows (see Remarks){p_end}
{synopt:{opt factorl:abel}}render labels for factor-variable levels{p_end}
{synopt:{opt ref:cat(string)}}label for reference-category rows{p_end}
{synopt:{opt omitl:abel(string)}}label for coefficients the model dropped{p_end}
{synopt:{opt emptyl:abel(string)}}label for cells identifying no observations{p_end}
{synopt:{opt notestl:abel(string)}}label for levels not estimable{p_end}
{synopt:{opt absentl:abel(string)}}label for levels a model leaves out{p_end}
{synopt:{opt cnsl:abel(string)}}interval text of a constrained coefficient{p_end}
{synopt:{opt cutl:abels(string)}}relabel ordered-model cutpoints{p_end}
{synopt:{opt comp:act}}combine estimate and CI per model{p_end}
{synopt:{opt nop:value}}suppress p-value columns{p_end}
{synopt:{opt addr:ow(string asis)}}append or place custom rows{p_end}
{synopt:{opt addc:ol(string asis)}}append custom columns to a transposed table{p_end}
{synopt:{opt coll:abels(string asis)}}transposed column headers by term{p_end}
{synopt:{opt pdp(#)}}decimal places for p < 0.10{p_end}
{synopt:{opt highpdp(#)}}decimal places for p >= 0.10{p_end}
{synopt:{opt reft:op}}print reference level first in its block{p_end}
{synopt:{opt celln:ote(spec)}}replace one model's cell on a named row{p_end}
{synopt:{opt minc:ount(#)}}mask factor levels with fewer than # events{p_end}
{synopt:{opt trans:pose}}models as rows; terms, stats as columns{p_end}
{synopt:{opt exposurel:abel(string)}}label of the {cmd:stats(exposure)} row{p_end}
{synopt:{opt cil:abel(string)}}header text of the interval column{p_end}
{synopt:{opt pl:abel(string)}}header text of the p-value column{p_end}
{synoptline}

{pstd}{bf:Automatic Median Odds Ratio / Median Hazard Ratio}{p_end}

{pstd}When the model type is {cmd:melogit}, {cmd:regtab} automatically converts the random
intercept variance to a {bf:Median Odds Ratio (MOR)} using the formula MOR =
exp(sqrt(2 * {it:sigma}^2) * invnormal(0.75)), eq. (5) of Austin and Merlo
({it:Statistics in Medicine} 36: 3257-3277, sec. 4.5.2; see {help regtab##references:References}). For {cmd:mecloglog}, and for
{cmd:mestreg} in the log-hazard metric, the conversion produces a
{bf:Median Hazard Ratio (MHR)} by the same formula, eq. (4) of Austin, Wagner, and
Merlo ({it:Statistics in Medicine} 36: 928-938, sec. 2.2), given there for a
proportional hazards model with a normally distributed random intercept on the
log hazard. An {cmd:mestreg} fit in the time metric
({cmd:time}; TR header) keeps its random-intercept row as the variance it is,
labelled {cmd:var(_cons[}{it:group}{cmd:])} (or "Variance: {it:Group} (Intercept)"
under {opt relabel}): its random effect acts on log time, so a median hazard
ratio would misname it. The collected CI bounds are
transformed on the same scale. In multi-level models, each transformed
random-intercept row keeps its own grouping label, so the output reads, for
example, "Median Odds Ratio (District)" and "Median Odds Ratio
(School)". MOR/MHR values and other random effects (slopes, covariances,
residual) follow the requested {opt digits()} precision. When a transformed
bound cannot be represented (the MOR of a near-zero variance whose upper
bound is ~1e+29 overflows double precision), the CI cell is left blank; the
same applies to an exponentiated fixed-effect interval, so a cell never shows
the untransformed collect text. Use {opt nore} to suppress all
random-effects rows if desired.{p_end}


{pstd}
{it:Detailed option contracts}{p_end}

{phang}
{opt addr:ow(string asis)} append custom label/value rows below the table (see
Remarks for syntax). Values go to the models' columns in order; fewer values
leave the remaining models blank, and more values than models is an error
(r(198)), as for {opt addcol()}. A row ending in {cmd:, after(}{it:term}{cmd:)} is placed inside the table
body instead: right after the coefficient row whose raw name is {it:term}, matched
exactly and case-sensitively as {opt keep()} matches ({cmd:after(3.rep78)}, {cmd:after(mpg)}),
or, when {it:term} is a factor variable, after the last level of that factor's
block ({cmd:after(rep78)}). Rows given the same {it:term} keep their order; an inserted
row takes the indentation of the row it follows, and in Excel its value spans
the model's cells without the rule an appended row gets. {it:term} takes {opt keep()}'s
form: {cmd:c.} and level markers do not matter, so {cmd:after(1.foreign#c.mpg)} and
{cmd:after(1b.rep78)} name the rows {opt keep()} would. A {it:term} that matches no row,
several rows, or several blocks of one factor is an error (r(198)). A
multi-equation table ({cmd:mlogit}) lists each term once per equation, so {cmd:after()}
cannot place a row in it. {cmd:r(table)} does not include addrow() rows, and
{opt eplotframe()}'s {cmd:source_row} follows the rows as placed. A specification wrapped
whole in one more layer of quotes, as a program passes {cmd:addrow(`"`spec'"')}, is
unwrapped: {cmd:addrow(`""P trend" 0.03"')} is {cmd:addrow("P trend" 0.03)}.{p_end}

{phang}
{opt addc:ol(string asis)} the {opt addrow()} of a transposed table; requires
{opt transpose} (r(198) otherwise). Each specification
{cmd:"}{it:label}{cmd:"} {it:value1} {it:value2} ..., separated by {cmd:\},
adds one column after the last one, headed by {it:label}, with {it:value1}
in the first model's row, {it:value2} in the second's, and so on; fewer values
leave the remaining rows blank and more values than models is an error
(r(198)): {cmd:regtab, transpose addcol("P trend" 0.032 0.041)}. A
specification ending in {cmd:, after(}{it:term}{cmd:)} places its column
right after the columns of coefficient {it:term} (its raw name, as
{opt keep()} takes it) instead of last, after any column already placed
there: {cmd:addcol("P trend" 0.032 0.041, after(2.drug))}. A {it:term} with no
column is an error (r(198)). The column reaches every sink; {cmd:r(table)}
and {opt eplotframe()} do not include it.{p_end}

{phang}
{opt coll:abels(string asis)} sets the first header of a transposed table's columns
by term: {it:name} {cmd:"}{it:label}{cmd:"} pairs, {it:name} a raw coefficient name in {opt keep()}'s form,
matched case-sensitively,
e.g. {cmd:collabels(age "Age (years)" 2.drug "Drug B")}. In a multi-equation table
each header keeps its equation ("4: Miles"); {it:eq}{cmd::}{it:name}, {it:eq} as the headers show
it, relabels one equation's column only ({cmd:collabels(4:mpg "Miles")}). The label
replaces the term's header over both its estimate and p-value columns (the
second header, "HR (95% CI)" or {opt plabel()}, stays). Requires {opt transpose}; a name
that matches no coefficient row, or a missing label, is an error
(r(198)). {opt stats()} columns take their labels from {opt stats()} and {opt statlabels()}.{p_end}

{phang}
{opt bold:p(#)} bold p-value cells below #{p_end}

{phang}
{opt border:style(string)} border style: {cmd:default}, {cmd:thin}, {cmd:medium}, or {cmd:academic}
(default {cmd:thin}){p_end}

{phang}
{opt cil:abel(string)} replaces the confidence-interval header text, by default
{cmd:"95% CI"} (the collection's level), in every layout and sink: the interval
column's header, the {opt compact} header ({cmd:"HR 95% CI"} becomes {cmd:"HR} {it:string}{cmd:"}), the
{opt transpose} header (the text inside {cmd:"HR (95% CI)"}), the {opt frame()} rows and {cmd:flat}
variable labels, Excel, CSV, and Markdown. For example
{cmd:cilabel("95% confidence interval")}. The text is used as
given: {cmd:cilabel("(95% CI)")} gives a compact header {cmd:"HR (95% CI)"}.{p_end}

{phang}
{opt cdisc} CDISC mode: digits 4, coef label "Estimate", forces {cmd:stats(n)}{p_end}

{phang}
{opt celln:ote(spec)} replaces one model's cell on one coefficient row with text, for
a term that is shown but whose estimate must not be. {it:spec} is {cmd:"}{it:row label}{cmd:"} {it:#}
{cmd:"}{it:text}{cmd:"}, repeated with {cmd:\} between
specifications: {cmd:cellnote("Repleted, age 55 and over" 3 "12 / 3" \ "Smoking" 1 "Not fitted")}. {it:#}
is the model's position in the table. The text replaces the estimate, and that
model's interval and p-value on the row are blank; the cell is left out of
{cmd:r(table)} and of {opt eplotframe()}, and in Excel the model's columns on that row are
merged. The row label must equal exactly one coefficient row's displayed label
(trimmed, case-sensitive; factor levels are matched by their level label, such
as {cmd:"Foreign"}); a label matching no row or several rows, a model number outside
the table, or a specification without its text, is an error (r(198)), never a
guess; {cmd:""} as the text blanks the cell deliberately. {opt cellnote()} is applied
after {opt mincount()} and before {opt transpose}; it cannot target {opt stats()} or {opt addrow()}
rows. Labels and texts may be in plain or compound quotes and may hold spaces,
backslashes, and quotation marks; the {cmd:\} between specifications may touch its
neighbours ({cmd:"a""Smoking"}); and a specification wrapped whole in one more layer
of quotes, as a program passes {cmd:cellnote(`"`spec'"')}, is unwrapped, as in
{opt addrow()}.{p_end}

{phang}
{opt cf:ormat(%fmt)} is a full Stata numeric display format for the estimate and both
confidence bounds, for example {cmd:%12.0fc} (whole numbers with thousands
separators) or {cmd:%9.3f}. Each value is written as {cmd:strtrim(string(}{it:x}{cmd:, "}{it:%fmt}{cmd:"))},
after any exponentiation, so an odds ratio is formatted on the odds-ratio
scale. {opt digits()} is the shorthand for {cmd:%}{it:#}{cmd:.}{it:d}{cmd:f} and cannot be combined with
{opt cformat()} (r(198)). Under {opt digits()}, each confidence bound is rounded
as the estimate is, by {cmd:round()} at the displayed decimals, so one number
prints alike in both cells (0.125 and 0.85 included); a {opt cformat()} format
is applied as given. Fixed ({cmd:f}), general ({cmd:g}), and exponential ({cmd:e}) formats, with
or without {cmd:c}, are accepted; date, string, and hexadecimal formats are refused
(r(198)). P-values, {opt stats()} rows, {cmd:r(table)}, and {opt eplotframe()} are
unaffected. With a {cmd:c} format the bounds contain commas, so a {opt sep()} other than
the default, such as {cmd:sep(" to ")}, reads better. A decimal-comma format ({cmd:%9,2f})
with a {opt sep()} that contains a comma, the default {cmd:", "} included, is refused
(r(198)): the bounds could not be told apart; use, for example, {cmd:sep(" to ")} or
{cmd:sep("; ")}.{p_end}

{phang}
{opt coef(string)} header for the estimate column; auto-detected per model scale if
omitted (see Remarks){p_end}

{phang}
{opt comp:act} merge estimate and CI into one column per model{p_end}

{phang}
{opt csv(filename)} also export the table as a CSV file. {it:filename} must end in {cmd:.csv}
and differ from the Excel and Markdown targets. The CSV mirrors the workbook
with {opt title()} written as the first row and {opt footnote()} as the last row, both in
the first column and the table body between them.{p_end}

{phang}
{opt cutl:abels(string)} custom labels for ordered-model cutpoint rows,
backslash-separated{p_end}

{phang}
{opt dimnon:sig} gray out non-significant rows (see Remarks){p_end}

{phang}
{opt drop(varlist)} drops rows by exact, case-sensitive variable or factor-component
name; not with {opt keep()}. See {opt keep()} for matching rules.{p_end}

{phang}
{opt eplotf:rame(name[, replace])} store a graph-ready companion frame for {helpb eplot} (see
Remarks){p_end}

{phang}
{opt factorl:abel} replace factor-variable prefixes (e.g., {it:3.rep78}) with value labels{p_end}

{phang}
{opt foot:note(string)} add a footnote below the table in smaller italic font. The
token {cmd:" \ "} (a backslash with a space on each side) separates
paragraphs: {cmd:footnote("Data: national registers. \ Rates per 1,000 person-years.")}
writes one merged, wrapped row per paragraph in Excel, and the CSV and
Markdown writers receive the text unchanged and split it the same way.{p_end}

{phang}
{opt fra:me(name[, replace flat])} stores the output in a named frame; {cmd:replace}
replaces an existing frame. Without {cmd:flat} the frame is the display
dataset: variables {cmd:title}, {cmd:A}, {cmd:c1}, {cmd:ref1}, ... with the title, model-name, and
statistic-header rows on top. With {cmd:flat}, the frame holds one row per body line
(coefficient, {opt stats()}, and {opt addrow()} rows) and nothing else: {cmd:rowlabel}, the row
label as printed (factor levels keep their leading spaces), then one string
variable per printed column, {cmd:c1}, {cmd:c2}, ..., whose variable label is the
statistic header the column prints under its model's name ({cmd:"HR"}, {cmd:"95% CI"}, {cmd:"P"}
after {cmd:plabel("P")}), or the model name when that is blank. The model name is
kept in {cmd:char c}{it:#}{cmd:[tabtools_block]} (and per model in
{cmd:_dta[tabtools_model_label_}{it:#}{cmd:]}), and the full printed header, the model name and
the statistic joined by ", " ({cmd:"Model 1, HR"}), in
{cmd:char c}{it:#}{cmd:[tabtools_header]}. Under {opt transpose} a column's block is its term, so its
label keeps both ({cmd:"Mileage (mpg), P"}). There are no title or header rows and
no reference-marker columns, so
{cmd:frame f: puttab rowlabel c* using t.xlsx, varlabels} writes the body under one
row of statistic headers without any {cmd:drop}; with several models, name them from
{cmd:char c}{it:#}{cmd:[tabtools_block]} or use the full header in {cmd:char c}{it:#}{cmd:[tabtools_header]}, or
let {cmd:puttab, blockheader} write each model's name as a spanning row over its
columns; {cmd:char c}{it:#}{cmd:[tabtools_block_id]} identifies each column's block, so two
models with the same name stay apart. A variable label holds at most 80
characters; a longer header is truncated in the label (with a note) and kept
whole in {cmd:char c}{it:#}{cmd:[tabtools_header]}. The frame's {cmd:_dta[tabtools_layout]}
characteristic is {cmd:flat}.{p_end}

{pmore}
{cmd:keys}, with {cmd:flat} only (r(198) otherwise), adds key variables after the printed
columns, so a program can find rows and cells by what they are rather than by
their printed text: {cmd:_order} (the row order, 1, 2, ...: give a new row 3.5 and
{cmd:sort _order} to put it between rows 3 and 4), {cmd:_term}, {cmd:_rowtype}, and {cmd:_state}{it:#} for
each printed column {cmd:c}{it:#}. {cmd:_term} is the row's raw coefficient name in {opt keep()}'s
form, factor operators and level markers dropped ({cmd:9.edss}, {cmd:age}, {cmd:1.a#2.b},
{cmd:1.g#x}), preceded in a multi-equation table by the equation as the collection
names it ({cmd:4:mpg}, {cmd:4:1.foreign}) and on a random-effects row by its level
({cmd:district:var(_cons)}); a factor's heading row holds the factor ({cmd:edss}, {cmd:a#b}), a
{opt stats()} row {cmd:stat:}{it:item}, and an {opt addrow()} row {cmd:addrow:}{it:k}, {it:k} its position in
{opt addrow()}. For a built-in statistic {it:item} is its word as listed under {opt stats()}
({cmd:stat:n}, also for {cmd:n_sub} and {cmd:subjects}; {cmd:stat:aic}, also when that row shows
QICu; {cmd:stat:F}), for an {cmd:e()} item the item as typed without its options or label
({cmd:stat:e(N_fail)}, {cmd:stat:e(N_fail|N_sub)}), and for the {it:k}th {cmd:text()} item {cmd:text(}{it:k}{cmd:)}
({cmd:stat:text(1)}); so {cmd:stats(F e(F))} gives two distinct keys, and no two rows
share one. {cmd:_rowtype} is {cmd:factor} (a heading row), {cmd:level} (a factor level or
interaction cell), {cmd:var} (any other coefficient: a covariate, intercept,
cutpoint, or ancillary parameter), {cmd:re} (a random effect), {cmd:stat}, {cmd:addrow}, or
{cmd:header} (a body row without a coefficient name, which no standard layout
prints). {cmd:_state}{it:#} is the state of the model cell that column {cmd:c}{it:#} belongs to, the
same for a model's estimate, interval, and p-value columns: {cmd:est} (an estimate),
{cmd:ref} ({opt refcat()}), {cmd:omit} ({opt omitlabel()}), {cmd:notest} ({opt notestlabel()}), {cmd:masked} (the
{opt emptylabel()} of a level {opt mincount()} masks for too few events), {cmd:absent}
({opt absentlabel()}), {cmd:constrained} (the estimate with {opt cnslabel()}), {cmd:note}
({opt cellnote()}), {cmd:stat} or {cmd:text} (a {opt stats()} or {opt addrow()} value), or {cmd:empty} (a blank
cell: the model holds no such coefficient, or the row is a heading). Each key
variable carries {cmd:char} {it:var}{cmd:[tabtools_key]} {cmd:1}, which tells {helpb puttab} to leave it out
unless its varlist names it, and the frame's {cmd:_dta[tabtools_keys]} lists them. A
frame without {cmd:keys} is unchanged. {cmd:keys} cannot be combined with {opt transpose}
(r(198)): a transposed table's columns are terms and its rows models, which
these keys do not describe. See {it:Editing a flat frame} under Examples.{p_end}

{phang}
{opt headers:hade} apply background fill to the header row{p_end}

{phang}
{opt high:light(#)} yellow fill for rows where p-value < #{p_end}

{phang}
{opt highpdp(#)} max decimal places for large p-values (p >= 0.10); default 2{p_end}

{phang}
{opt keep(varlist)} shows only rows matching exact, case-sensitive raw
names; not with {opt drop()}. Labels do not affect name matching. For example,
{cmd:keep(age)} selects {cmd:age}, its factor levels, and interactions containing
it, but never {cmd:stage} or {cmd:Age}. {cmd:keep(2.arm)} selects that level and
interactions containing it, but not {cmd:20.arm}. Factor operators such as
{cmd:i.arm} and {cmd:c.age} are accepted; an interaction specification such as
{cmd:1.arm#c.age} selects that interaction exactly.{p_end}

{phang}
{opt labelm:atch} makes {opt keep()} or {opt drop()} match case-insensitive substrings of
displayed row labels instead of raw names. For example, {cmd:keep(Fuel) labelmatch}
selects a row labelled {it:Fuel efficiency}. Requires {opt keep()} or {opt drop()}.{p_end}

{phang}
{opt keepi:ntercept} force display of the intercept row even for exponentiated models{p_end}

{phang}
{opt labelw:idth(#)} maximum width of the label column in characters (default
45); longer labels wrap{p_end}

{phang}
{opt level(#)} verifies confidence-level provenance. By default, {cmd:regtab}
uses the level stored in the active collection, and an explicit value must
match that stored level or the command fails before writing output. Not every
Stata version records the level in the collection: Stata 17 does, Stata 19 does
not. When it is absent and {opt level()} is omitted, {cmd:regtab} warns and uses
the current {helpb set level} for interval labels. Because that is the session
setting at render time rather than necessarily the model-time level, specify
{opt level()} whenever the collected models were fit at a different level. The
resolved level labels intervals in headers and methods text, is returned in
{cmd:r(ci_level)}, and is stored on display and eplot frames; it affects no
computed quantity.{p_end}

{phang}
{opt markdown(filename)} export the table as GitHub-Flavored Markdown. Factor-level
rows keep their indentation: each leading space of a row label is written as
{cmd:&nbsp;}, which survives Markdown cell trimming{p_end}

{phang}
{opt mdappend} append the Markdown table to an existing file; requires {opt markdown()}{p_end}

{phang}
{opt models(string)} labels merged above each model's columns,
backslash-separated; auto-generated if omitted{p_end}

{phang}
{opt noint:ercept} drop intercept, cutpoint, and ancillary rows; auto-enabled for
all-ratio-scale models. Rows are identified by the coefficient's equation and
name in the collection, never by its display label: {cmd:_cons} in any equation,
{cmd:cut}{it:#} in the ancillary ({cmd:/}) equation, and the ancillary parameters {cmd:lnalpha},
{cmd:alpha}, {cmd:ln_p}, {cmd:p}, and {cmd:1/p} of the {cmd:/} and derived {cmd:_diparm} equations. Other
ancillary parameters, such as a lognormal's {cmd:lnsigma}/{cmd:sigma} or a loglogistic's
{cmd:lngamma}/{cmd:gamma}, stay visible. In the multi-equation layout the whole ancillary
block and any {cmd:lnsigma} scale equation are dropped. A covariate named or
labelled {cmd:p}, {cmd:alpha}, {cmd:constant}, {cmd:cut1}, {cmd:Intercept}, or {cmd:/x} is an ordinary
coefficient: it is kept and exponentiated with the others. A covariate that
shares its name with an ancillary parameter or cutpoint of its own model or of
another model in the collection (a covariate named {cmd:alpha} beside {cmd:nbreg}'s {cmd:alpha},
{cmd:ln_p} beside a Weibull {cmd:streg}, {cmd:lnsigma} beside a lognormal one, {cmd:cut1} beside
{cmd:ologit}) is shown as two rows, told apart by their equations: the covariate
among the covariates, on its model's scale, blank in the other models, and the
parameter, labelled with its own name and on its own scale, among the
ancillary rows. The collection itself is not changed.{p_end}

{phang}
{opt nop:value} suppress p-value columns; stars and highlighting still use p-values
internally{p_end}

{phang}
{opt nore:effects} drop all random-effects rows (variances, covariances, SDs){p_end}

{phang}
{opt open} open the Excel file after export; requires {opt xlsx()} or {opt excel()}{p_end}

{phang}
{opt pl:abel(string)} replaces the p-value header text, by default {cmd:"p-value"}, in
every layout and sink, as {opt cilabel()} does: {cmd:plabel("P")}.{p_end}

{phang}
{opt pdp(#)} max decimal places for small p-values (p < 0.10); default 3{p_end}

{phang}
{opt ref:cat(string)} label for reference-category rows. Default {cmd:"Reference"}{p_end}

{phang}
{opt reft:op} prints, within each factor's block of levels, the level a model
holds as its base first, directly under the factor's header row; the other
levels keep their order. Without it the levels appear in code order, so the
base of {cmd:ib3.x} sits third. If the models in the table hold different
levels of one factor as base ({cmd:ib3.x} in one model, {cmd:i.x} in another),
there is no single top row and {cmd:regtab} exits with r(198). Factors with
no base level ({cmd:ibn.}) are unchanged. {cmd:r(table)} and
{opt eplotframe()} follow the displayed order.{p_end}

{phang}
{opt minc:ount(#)} masks factor levels that are too thin to report. In each model, a
factor-level row (including interactions with a factor level), or the row of a
0/1 indicator entered as a plain variable, is shown as {opt emptylabel()} with a
blank interval and p-value when that level has fewer than {it:#} events in the
model's estimation sample (0 events included), and as {opt notestlabel()} when the
fit estimated no positive, finite variance for it (a level dropped as
collinear or empty included) or when the level is missing from the model
although the model includes the factor (a level no observation of a subgroup's
sample holds, so its {cmd:e(b)} has no column for it: shown as {opt notestlabel()}, not as
a blank cell). An indicator's level is 1. A model's own base level keeps its
{opt refcat()} label, and other continuous terms are never masked. The event counts
are those {cmd:tabtools fitcount, events() terms} stored right after each fit (see
{it:Fit-time counts} under Remarks): every model in the table needs them, and a
level a model shows that its counts do not cover is an error (r(459)); a model
with estimates but without counts is refused with r(198). A model with no
estimate at all has nothing to mask and needs no counts: its column is left
blank and a note names it. Such a column is how a fit that failed is kept in
its place: a failed estimation command can leave the previous model's results
in {cmd:e()}, so clear them before collecting ({cmd:ereturn clear}, then
{cmd:collect get e(), tags(cmdset[}{it:#}{cmd:])}) or collect a placeholder result instead
({cmd:collect get failed = "Did not converge", tags(cmdset[}{it:#}{cmd:])}); a {cmd:collect get e()}
straight after the error collects the previous model a second time, and {cmd:regtab}
notes a model that repeats an earlier one. Under {opt mincount()} the default
{opt emptylabel()} is a dash (U+2013), and {opt notestlabel()} defaults to it, so empty,
omitted, not estimable, and masked levels all read "–"; specify {opt emptylabel()}
and {opt notestlabel()} to choose other texts. A level of a factor block a model
includes that is missing from the model's {cmd:e(b)} altogether is told apart by its
estimation sample, which {cmd:tabtools fitcount, terms} records: a level no
observation of the sample holds (an {cmd:if} restriction or subgroup leaves it out
by design, such as {cmd:if edss != 9}) is shown as {opt absentlabel()}, blank by
default; a level the sample holds that the model does not estimate is shown as
{opt notestlabel()}. Counts stored by an older {cmd:fitcount}, which does not record the
sample's levels, cannot tell the two apart: such a level is shown as
{opt notestlabel()}, the reading that does not claim the model left it out, and a
note names it. A model's own base level keeps its {opt refcat()} label and a
coefficient a constraint fixes its value: neither is masked. Masked cells are
left out of {cmd:r(table)} and {opt eplotframe()}; {cmd:r(N_masked)} counts them, as in 2.3.1,
for the levels a model holds; {cmd:r(N_absent)} counts the levels missing from a
model, whichever label they are shown with. This is reporting hygiene for
unstable estimates, not disclosure control.{p_end}

{phang}
{opt omitl:abel(string)} label for a coefficient the model dropped, which Stata
reports as {cmd:(omitted)} -- most often a level or term dropped for
collinearity. Default {cmd:"Omitted"}{p_end}

{phang}
{opt emptyl:abel(string)} the default text of {opt notestlabel()}, and under {opt mincount()} the
label of a level masked for having too few events; default {cmd:"Empty"}, or a dash
(U+2013) under {opt mincount()}.{p_end}

{phang}
{opt notestl:abel(string)} label for a factor level the model holds but could not
estimate: a cell that identifies no observations in the estimation sample
(Stata's {cmd:(empty)}), and a level the model estimated whose variance is zero or
missing (a {cmd:stcox} level whose coefficient diverged under separation prints a
hazard ratio with blank standard error). Under {opt mincount()} it also labels a
level the model omitted and a level of a factor block the model includes that
is missing from the model although its estimation sample holds it (one the
sample does not hold is {opt absentlabel()}). The default is the {opt emptylabel()} text,
which these cells carried before {opt notestlabel()} existed, so a table that does
not specify it is unchanged; specify it to tell them apart from levels masked
for too few events, e.g. {cmd:mincount(5) notestlabel("NE")}. It must differ from
{opt refcat()} and {opt omitlabel()} (r(198)). Such a level is never the reference, and a
base level whose cell identifies no observations is such a level; see
{it:Constrained rows} under Remarks.{p_end}

{phang}
{opt absentl:abel(string)} under {opt mincount()}, the label of a level of a factor block a
model includes that no observation of that model's estimation sample holds, so
the model leaves it out by design: the "Not recorded" level of a model fitted
{cmd:if edss != 9}. The default is blank, the cell such a level had before
2.4.0; {cmd:absentlabel("–")} gives the dash 2.4.0 printed. A level the sample holds
but the model does not estimate is {opt notestlabel()} instead (see
{opt mincount()}). Requires {opt mincount()} and must differ from {opt refcat()} (r(198)).{p_end}

{phang}
{opt cnsl:abel(string)} the text printed where the interval would be for a
coefficient that a constraint fixes ({helpb constraint} define, then the estimator's
{cmd:constraints()}, {cmd:cnsreg}): its estimate, the fixed value on the table's scale, is
shown, then this text, and its p-value is blank, as Stata's own table prints
{cmd:(constrained)}. A coefficient a constraint fixes at 0 is one Stata prints
{cmd:(omitted)}, and so does {cmd:regtab} ({opt omitlabel()}); a factor level fixed so is never
the reference (see {it:Constrained rows} under Remarks). Default
{cmd:"(constrained)"}; with {opt compact} the cell reads {cmd:1.05 (constrained)}. A constrained
coefficient is known from the fit's own e(b) and {cmd:e(Cns)}: from the record
{cmd:tabtools fitcount} stores, or from the active estimation results when they are
the model; without either it cannot be told from a coefficient estimated with
a zero variance, and is shown as such (see {it:Constrained rows} under Remarks). It
is included in {cmd:r(table)} and posted to {opt eplotframe()} with rowtype
{cmd:constrained}; {opt mincount()} does not mask it.{p_end}

{phang}
{opt relab:el} relabel random effects using variable labels and parameter types (see
Remarks){p_end}

{phang}
{opt sep(string)} the text between the two confidence bounds, default
{cmd:", "}: {cmd:sep(" to ")} prints {cmd:(1.02 to 1.31)}. The text prints as
typed in every sink (console, Excel, CSV, Markdown, {opt frame()} with or
without {cmd:flat}); {cmd:r(table)} and {opt eplotframe()} hold the bounds as
numbers; see {help tabtools##sep:interval separators}.{p_end}

{phang}
{opt sheet(string)} target sheet to create/replace in {opt xlsx()}. Default
{cmd:"Regression"}. After {cmd:tabtools set workbook} (or {cmd:markdown}), a call that gives
{opt sheet()} without {opt xlsx()} (or {opt markdown()}) writes to the session target and says
so in the log; the first write of the session workbook starts it over (see
{helpb tabtools}). An explicit {opt xlsx()} or {opt markdown()} wins. When {opt sheet()} is given and no
workbook resolves, {cmd:regtab} prints
{cmd:(tabtools: sheet() ignored; no xlsx() and no session workbook)}.{p_end}

{phang}
{opt stars} add significance stars to coefficients (*, **, ***){p_end}

{phang}
{opt starsl:evels(numlist)} custom p-value thresholds for stars; exactly 3 values
(default 0.05 0.01 0.001){p_end}

{phang}
{opt stats(string asis)} model-fit statistics rows: {cmd:n}, {cmd:obs}, {cmd:events},
{cmd:people}, {cmd:exposure},
{cmd:groups}, {cmd:mi_m}, {cmd:aic}, {cmd:qic}, {cmd:bic}, {cmd:ll}, {cmd:icc},
{cmd:r2}, {cmd:r2_a}, {cmd:rmse}, {cmd:F}, {cmd:fmi}. The {cmd:qic} API name displays Pan's
fixed-penalty approximation as {cmd:QICu} for {cmd:xtgee} fits whose
dispersion is fixed at 1 (see Remarks). Rows always appear in that order,
whatever the order of the tokens: counts first (Observations or Subjects,
Observations ({cmd:obs}), Events, People, Person-time, Groups, Imputations), then likelihood criteria (AIC, QICu, BIC,
Log-likelihood), then ICC, R², and the linear-model rows (Adjusted
R², Root MSE, F statistic), with Largest FMI last. A row whose statistic no
collected model reports is omitted, with a console note naming it, and a model
that does not report it has a blank cell:{p_end}
{p2colset 9 24 26 2}{...}
{p2col:{cmd:n}}"Observations" ({cmd:e(N)}), or "Subjects" ({cmd:e(N_sub)})
for survival models (%12.0fc). After {cmd:mi estimate, esampvaryok}, where
{cmd:e(N)} is missing, it is {cmd:e(N_mi)}, the number of observations
{cmd:mi estimate} prints{p_end}
{p2col:{cmd:obs}}"Observations" (%12.0fc): {cmd:e(N)} (or {cmd:e(N_mi)} as for
{cmd:n}) for every model, including survival models, where it counts records
or intervals while {cmd:n} prints subjects; {cmd:r(obs_}{it:#}{cmd:)}{p_end}
{p2col:{cmd:events}}"Events" (%12.0fc): the events {cmd:tabtools fitcount}
counted for the model, for any estimator; otherwise {cmd:e(N_fail)}, so
survival models ({cmd:stcox}, {cmd:streg}, {cmd:stcrreg}) need no
{cmd:fitcount}{p_end}
{p2col:{cmd:people}}"People" (%12.0fc): distinct {opt people()} values in the
estimation sample, from {cmd:tabtools fitcount}; blank for a model without
them{p_end}
{p2col:{cmd:exposure}}{opt exposurelabel()}, default "Person-time"
(%12.0fc): the summed {opt exposure()} of the estimation sample, from
{cmd:tabtools fitcount}{p_end}
{p2col:{cmd:groups}}"Groups" (%12.0fc): {cmd:e(N_g)}, as {cmd:xtreg}, {cmd:xtlogit},
{cmd:mixed}, and {cmd:melogit} store it; blank for a model without it. {cmd:clogit}
stores its group variable but no count, and the collection holds
no estimation sample to count it from: when no collected model reports a count
and one is such a fit, the table is still built without a Groups row and a
console note names the model, the estimator, and its group variable. Run
{cmd:tabtools fitcount, events(}{it:depvar}{cmd:) people(}{it:groupvar}{cmd:)}
right after the fit and request
{cmd:stats(people) statlabels(people "Groups")}. When another model reports a
count, the {cmd:clogit} model has a blank cell and a console note says so{p_end}
{p2col:{cmd:r2_a}}"Adjusted R²", {cmd:e(r2_a)} (%5.3f); the {cmd:r2} token
still falls back to it when a model has neither {cmd:e(r2)} nor {cmd:e(r2_p)}{p_end}
{p2col:{cmd:rmse}}"Root MSE", {cmd:e(rmse)} (%9.3f), the N-k divisor Stata
reports; blank for models without it{p_end}
{p2col:{cmd:F}}"F statistic", {cmd:e(F)} (%9.2f) for linear-model F tests
({cmd:regress}, {cmd:anova}, {cmd:areg}, {cmd:xtreg}, {cmd:ivregress},
{cmd:cnsreg}); blank for every other model, including a {cmd:svy: logit} that
stores a design-based {cmd:e(F)}{p_end}
{p2col:{cmd:mi_m}}"Imputations", {cmd:e(M_mi)} (%12.0fc) after {cmd:mi estimate}{p_end}
{p2col:{cmd:fmi}}"Largest FMI", {cmd:e(fmi_max_mi)} (%6.4f), the largest
fraction of missing information across coefficients, as {cmd:mi estimate}
reports it{p_end}
{p2col:{cmd:e(}{it:name}{cmd:)}}any other scalar the collection holds, as
{cmd:e(}{it:name}{cmd:)} or {cmd:e(}{it:name}{cmd:)="}{it:label}{cmd:"}
(compound quotes also work), for example
{cmd:stats(n e(N_lc)="Left-censored" e(N_rc)="Right-censored")}. Each is a
row of its own after the built-in rows, in the order given, labelled
{it:label} (default {it:name}): integers with thousands separators
(%12.0fc), other values to three decimals (%12.3f). The value comes from the
collection, never from the active {cmd:e()}. A model that does not hold the
scalar has a blank cell; a scalar no collected model holds is an error
(r(111)), as is an item requested twice (r(198)). Returned, unmasked, as
{cmd:r(e_}{it:name}{cmd:_}{it:#}{cmd:)} when that name fits in 32
characters. Options follow a comma inside the parentheses: {cmd:e(}{it:name}{cmd:, %}{it:fmt}{cmd:)}
or
{cmd:e(}{it:name}{cmd:, fmt(%}{it:fmt}{cmd:))} formats the row with a numeric
display format ({cmd:e(N, %9.0fc)}); {cmd:mincell(}{it:#}{cmd:)}, # >= 2,
prints a value from 1 to #-1 as {cmd:<}{it:#}, the small-cell text of
{helpb tabcell} and {helpb ratetab} ({cmd:e(N_fail, mincell(5))}), and
{cmd:r(N_stats_masked)} counts such cells. A statistic named in several
items of one {opt stats()} is masked in every one of them with the strictest
{cmd:mincell()} given for it in any, with no declaration needed: {cmd:stats(e(ev, mincell(5)) e(tot|ev))}
prints {cmd:<5} in both rows,
{cmd:13 (<5)}, never {cmd:13 (3)}, so one table never prints a count it
masks elsewhere. {cmd:e(}{it:name1}{cmd:|}{it:name2}{cmd:)}
is one row "value1 (value2)",
each part formatted with the item's format and masked against its
{cmd:mincell()} threshold on its own (so either part can read {cmd:<}{it:#}),
"value1" alone where a model lacks {it:name2}, and a blank cell where it
lacks {it:name1}; default label "{it:name1} ({it:name2})": {cmd:e(N_fail|N_sub)="Events (subjects)"}
prints {cmd:31 (48)}. The counts
{cmd:tabtools fitcount} stores are collection results too, so
{cmd:e(tt_events|tt_people)="Events (people)"} combines them. {cmd:maskwith()},
a further option of {cmd:e()} items, is described below the
list{p_end}
{p2col:{cmd:text(}{it:label} {it:value1} {it:value2} ...{cmd:)}}a row of
text, one value per model in order, among the {cmd:e()} rows in the order
given: {cmd:stats(n text("Weighting" "Unweighted" "IPW"))}. Values with spaces
go in quotes; fewer values than models leave cells blank and more is an error
(r(198)). Unlike {opt addrow()}, it sits with the statistics and becomes a
column under {opt transpose}{p_end}
{p2colreset}{...}

{pmore}
{cmd:maskwith(}{it:names}{cmd:)} declares rows that would give a masked count back, so they are
masked with it ({it:complementary} masking within this table). {it:names} are other {cmd:e()}
names in {opt stats()}. The item's own names and {it:names} form one group, and groups
that share a name join. In a model where any member of a group prints {cmd:<}{it:#} under
its {cmd:mincell()}, every other member of that group with a value prints {cmd:–} (an en
dash) in that model, never a number, whatever its value (0 included). For
example, when people and people without an event are both printed, a masked
event count is their difference, so
{cmd:e(tt_events, mincell(5))="Events" e(tt_people)="People" e(n_noev, mincell(5) maskwith(tt_events))="Without an event"}
masks the third row wherever the first is masked, and the first wherever the
third is; people stay visible, since masking one of the dependent rows is
enough. Name in a group only the rows that must be withheld together. No
dependency is assumed: rows are linked only through {cmd:maskwith()}, and the two
parts of {cmd:e(}{it:a}{cmd:|}{it:b}{cmd:)} are linked only when the pair item itself has a {cmd:maskwith()}
(then both names join the group), because a pair such as people (events) is a
layout, not a derivation. Two counts whose sum is printed in another row, such
as cases (controls) under a total, need
one: {cmd:e(n_case|n_ctrl, mincell(5) maskwith(n_case))}. A percentage next to a
count whose denominator is known is linked from a row of its own, since
{cmd:mincell()} would also mask a small
percentage: {cmd:e(pct_ev, %4.1f maskwith(tt_events))="Events, %"}. Statistics that
are not {cmd:e()} scalars enter the collection with {cmd:collect get}
{it:name}{cmd: = (}{it:#}{cmd:), tags(cmdset[}{it:#}{cmd:])} and are then {cmd:e()} items like any other. A masked or
linked cell is masked in every sink and in {opt frame()}; {cmd:r()} keeps the
numbers. {cmd:r(N_stats_masked)} counts the {cmd:<}{it:#} cells and {cmd:r(N_stats_linked)} the {cmd:–}
cells. A group in which no name has a {cmd:mincell()} masks nothing, and a note says
so. The scope is this one table: differencing against counts printed in
another table or document is not covered.{p_end}

{phang}
{opt statl:abels(string asis)} relabels built-in {opt stats()} rows with {it:stat} {cmd:"}{it:label}{cmd:"}
pairs, {it:stat} being one of {cmd:n}, {cmd:obs}, {cmd:events}, {cmd:people}, {cmd:exposure}, {cmd:groups}, {cmd:mi_m},
{cmd:aic}, {cmd:qic}, {cmd:bic}, {cmd:ll}, {cmd:icc}, {cmd:r2}, {cmd:r2_a}, {cmd:rmse}, {cmd:F},
{cmd:fmi}: {cmd:stats(n events) statlabels(n "Patients" events "Deaths")}. A statistic
that is not requested in {opt stats()}, an unknown one, or a missing label is an
error (r(198)). {cmd:statlabels(exposure ...)} is {opt exposurelabel()}; the two cannot
both be given. {cmd:qic} also labels the QICu row that {cmd:aic} falls back to for
{cmd:xtgee}. Labels the {cmd:e(}{it:name}{cmd:)} items of {opt stats()} carry their own.{p_end}

{phang}
{opt title(string)} title written to {cmd:A1}, merged across the table; blank if omitted{p_end}

{phang}
{opt trans:pose} lays the table out with one row per model (labelled by {opt models()}, or
Model 1, Model 2, ...) for specification and sensitivity tables. The columns
are, in order, one per {opt stats()} statistic (its label over each model's value)
and, for every coefficient row with an estimate in some model, one "estimate
(CI)" column and, unless {opt nopvalue}, one p-value column. Use {opt keep()} to choose
the terms. A column's first header is the term's label, {it:Factor}{cmd:: }{it:level} for a
factor level ("Age group: 40-54"), preceded in a multi-equation model by the
equation as the untransposed row label shows it ("2: Car origin: Foreign"),
its second the estimate header with the interval ("HR (95% CI)") or
"p-value"; {cmd:frame(}{it:name}{cmd:, flat)} joins them. A model without the term has a blank
cell; reference, omitted, masked, and {opt cellnote()} cells show their
labels; {opt collabels()} renames the term headers. {opt compact} is implied. {opt addrow()},
{opt dimnonsig}, {opt highlight()}, and {opt boldp()} act on rows that a transposed table does
not have and are refused (r(198)); {opt addcol()} adds columns instead. {cmd:r(table)},
{cmd:r(N_models)}, and {opt eplotframe()} keep their usual, untransposed meaning.{p_end}

{phang}
{opt xlsx(string)} output Excel filename (must end with {cmd:.xlsx}); {opt excel()} is a synonym{p_end}

{phang}
{opt zebra} apply alternating light gray row shading{p_end}

{phang}
{opt exposurel:abel(string)} the label of the {cmd:stats(exposure)} row, for example
{cmd:exposurelabel("Person-years")}; default {cmd:"Person-time"}. Requires
{cmd:stats(exposure)}.{p_end}


{phang}
{opt headerc:olor(string)} custom header color (Stata color name or RGB
triplet; default {cmd:"219 229 241"}){p_end}

{phang}

{phang}
{opt zebrac:olor(string)} custom zebra color (Stata color name or RGB triplet; default
{cmd:"237 242 249"}){p_end}

{marker remarks}{title:Remarks}

{pstd}{bf:Fit-time counts: tabtools fitcount}{p_end}

{pstd}A collection keeps a model's coefficients and {cmd:e()} scalars but not its
estimation sample, so counts of events, people, and person-time in that sample
must be taken while the fit is still the active estimation result. Run
{cmd:tabtools fitcount} right after each {cmd:collect:} fit:{p_end}

{p 8 12 2}{cmd:tabtools fitcount}, {opt ev:ents(varname)} [{opt peo:ple(varname)} {opt exp:osure(varname)} {opt terms}
{opt nam:e(collection)}]{p_end}

{pstd}On {cmd:e(sample)} it counts the events (the sum of {opt events()}, a 0/1 failure
indicator such as {cmd:_d} or a nonnegative integer event count), the people
(distinct values of {opt people()}, numeric or string), and the person-time (the sum
of {opt exposure()}). With {opt terms} it also counts, for every factor-level coefficient
of {cmd:e(b)} (an interaction counts the observations that match all of its factor
levels) and every plain variable that is 0/1 in the estimation sample (its
level 1), the events in that level and whether the fit estimated a positive,
finite variance for it. In a multi-equation model, an equation in which
nothing is estimated (the base outcome of {cmd:mlogit}) does not count toward that
variance check. The counts are of observations: {cmd:pweight}s and {cmd:aweight}s are
ignored, and {cmd:fweight}s and {cmd:iweight}s, under which an observation is not one
observation, are refused (r(101)). {cmd:svy, subpop()} results are refused (r(459)),
because their {cmd:e(sample)} holds the whole design sample rather than the
subpopulation. {cmd:collect, name(}{it:c}{cmd:):} does not change the current collection and
leaves no trace in {cmd:e()}, so for such a fit give {opt name(c)}: the counts are written
into collection {it:c}, which is current only while {cmd:fitcount} runs (r(111) if it
does not exist), and {cmd:r(collection)} names it. Without {opt name()}, the current
collection must hold the fit, or {cmd:fitcount} exits with r(459). {cmd:regtab} reads the
current collection, so {cmd:collect set} {it:c} before it. The results are stored with
{cmd:collect get} in the active collection under the model's own {cmd:cmdset}, as results
{cmd:tt_events}, {cmd:tt_people}, {cmd:tt_exposure}, and {cmd:tt_terms}; {cmd:regtab} reads them for
{cmd:stats(events people exposure)} and {opt mincount()}. {cmd:fitcount} also stores {cmd:tt_cns}, the
fit's own notes for its base, empty, omitted, and constrained coefficients
(from the {cmd:e(b)} stripe and {cmd:e(Cns)}), which {cmd:regtab} reads to label them (see
{it:Constrained rows} below), and, with {opt terms}, {cmd:tt_levels}, the level combinations
each factor term of {cmd:e(b)} holds in {cmd:e(sample)}, which tells a level a model
leaves out from one it holds but does not estimate (see {opt mincount()} and
{opt absentlabel()}). They do not appear in {cmd:regtab}'s coefficient rows, but a layout
of your own naming {cmd:result} levels can show them.{p_end}

{pstd}The fit is identified exactly: the collection's latest {cmd:cmdset} must carry the
active {cmd:e(cmdline)} and {cmd:e(N)}; every coefficient of the active {cmd:e(b)} that the
collected model also holds under the same equation and name must equal its
collected {cmd:_r_b} (as is, or exponentiated where the collection holds an eform
estimate) to a relative difference below 1e-12, and at least one must
match; and {cmd:e(sample)} must still hold {cmd:e(N)} observations of the data in
memory. A refit of the same command line on changed data is therefore
refused. Otherwise {cmd:fitcount} exits with r(459) (r(119) when nothing is
collected, r(301) with no estimation results) rather than attach counts to the
wrong model. {cmd:mi estimate} results, which have no single estimation sample, are
refused (r(198)). Missing, negative, or noninteger {opt events()} values, or missing
or negative {opt exposure()} values, in the estimation sample are errors
(r(459)). {cmd:fitcount} leaves the data and its sort order unchanged and returns
{cmd:r(events)}, {cmd:r(people)}, {cmd:r(exposure)}, {cmd:r(N)}, {cmd:r(cmdset)}, {cmd:r(collection)}, and with
{opt terms} {cmd:r(n_terms)} and {cmd:r(terms)} ({it:key}{cmd:=}{it:events}{cmd:|}{it:variance ok}, separated by
semicolons).{p_end}

{pstd}Prerequisites and expectations{p_end}
{p 4 8 2}- Run your models inside {cmd:collect:} or otherwise ensure the
relevant results are in the active {helpb collect}. {cmd:regtab} does not run models.{p_end}
{p 4 8 2}- {cmd:regtab} expects dimensions including {cmd:colname} and {cmd:cmdset}, and result items
{cmd:_r_b}, {cmd:_r_ci}, {cmd:_r_p}. It applies cell styles: {cmd:_r_b} as %4.2fc, {cmd:_r_ci} as
{cmd:sformat("(%s")} with {cmd:cidelimiter(", ")}, and {cmd:_r_p} as %5.4f.{p_end}
{p 4 8 2}- Because {cmd:regtab} works through the active {cmd:collect}, it
intentionally updates the labels of {cmd:_r_b}, {cmd:_r_ci}, and {cmd:_r_p},
the cell styles, and the layout before export. The labels of every other
result it reads (the command, command line, and {cmd:vce} metadata, and the
{opt stats()} results) are restored exactly as found. If you need the
original collection layout unchanged for later commands, save or rebuild that
collection before running {cmd:regtab}.{p_end}
{p 4 8 2}- The text between the CI limits in {cmd:regtab}'s tables is
{opt sep()}; default {cmd:", "}. Example alternative: {cmd:sep("; ")}. It is
never handed to {cmd:collect}, so the collection's own {cmd:cidelimiter()}
stays {cmd:", "}.{p_end}
{p 4 8 2}- If {opt coef()} is not provided, {cmd:regtab} detects the display
scale per collected model from the collected command metadata and fills the
estimate header automatically. When models use different scales, estimate
headers are set per model and {cmd:r(coef_label)} returns {cmd:mixed}.{p_end}
{p 4 8 2}- Multi-equation models such as {cmd:mlogit}, {cmd:zip}, {cmd:zinb},
and {cmd:churdle} use the equation/outcome dimension in the row labels, so rows
read like {it:Partial response: Age z-score} or
{it:Inflation equation: Prior events} instead of collapsing repeated covariate
names across equations. For labeled multinomial outcomes, value labels are used
when available. {cmd:mlogit} is displayed as relative risk ratios (RRR) by
default; zero-inflated and hurdle models remain on their native coefficient
scale unless {opt coef()} and collection styling are supplied by the user. A
factor covariate keeps its header row inside each equation block, as in the
single-equation layout: {it:4: Sex} above the indented level rows
{it:4:   Male} (Reference) and {it:4:   Female}.{p_end}
{p 4 8 2}- {cmd:r(methods)} describes each model from its collected metadata,
not from the estimate header: the estimates the model reports (Odds ratios,
Hazard ratios, Coefficients, ...; a {opt coef()} or {opt cdisc} relabel does
not change it), the model named from the command, the {cmd:glm}/{cmd:xtgee}
family and link, and the survival metric (probit regression, negative binomial
regression, complementary log-log regression, gamma regression with a log link,
zero-inflated Poisson regression, ordered logistic regression, linear
mixed-effects regression, mixed-effects logistic regression, generalized
estimating equation (GEE) logistic regression, Fine-Gray competing-risks
regression, heteroskedastic probit regression, quantile regression,
instrumental-variables regression, ...), and "univariable" (one predictor
variable) or "multivariable", counting a factor variable once and counting
the covariates of a variance equation (such as {cmd:hetprobit}'s
{cmd:het()}). Prefixes are named: survey-weighted
..., ... with multiple imputation, ... with bootstrap (or
jackknife) standard errors. Several models list each distinct model once
("univariable and multivariable logistic regression across 2 models"). Collections
mixing estimate scales keep the generic sentence "Collected
regression estimates with #% confidence intervals across # models.".{p_end}
{p 4 8 2}- Model header labels are auto-generated unless {opt models()} supplies
explicit names. {opt models()} values are split on the backslash character.{p_end}
{p 4 8 2}- {opt coef()}: if omitted, the estimate-column header and scale are
auto-detected per collected model: {cmd:logit}/{cmd:logistic}/{cmd:clogit} {it:->} OR,
{cmd:mlogit} {it:->} RRR, {cmd:stcox} {it:->} HR, {cmd:poisson}/{cmd:nbreg}
{it:->} IRR, {cmd:stcrreg} {it:->} SHR, {cmd:streg}/{cmd:stintreg}/{cmd:mestreg} {it:->} HR in
the log-hazard metric and TR in the log-time metric, {cmd:regress}/{cmd:mixed}
{it:->} Coef. A ratio family is always shown on its ratio scale: a fit displayed
as coefficients ({cmd:logit} or {cmd:clogit} without {cmd:or}, {cmd:logistic} with
{cmd:coef}, {cmd:stcox}, {cmd:streg}, or {cmd:stintreg} with {cmd:nohr}, {cmd:stcrreg} with
{cmd:noshr}, or a log-time {cmd:streg} or {cmd:stintreg} without {cmd:tratio}) is exponentiated, and a fit Stata already
exponentiated is left as it is, so the header always names the numbers under
it. Display options are read only from the option list after the command's
comma, with the estimator's own abbreviations ({cmd:ir}, {cmd:rr}, {cmd:ti},
{cmd:tr}), so a covariate named {cmd:or} is never mistaken for the option. Estimation
prefixes are set aside, alone or nested: {cmd:svy:},
{cmd:bootstrap:} ({cmd:bs:}, {cmd:bstrap:}), {cmd:jackknife:} ({cmd:jknife:}),
and {cmd:mi estimate:}. {cmd:svy: logit}, {cmd:bootstrap: logit}, and
{cmd:jackknife: poisson} are classified like {cmd:logit} and {cmd:poisson}
(OR or IRR, intercept suppressed), and a prefix's own options, such as
{cmd:subpop()} or {cmd:reps()}, are never read as the command's display
options. An eform option given to {cmd:bootstrap} or {cmd:jackknife}
({cmd:bootstrap, eform: logit}) counts like the command's own {cmd:or}. {cmd:mi estimate}
reports the coefficient metric whatever the fitted command's
display options, unless {cmd:mi estimate} itself is given an eform option
({cmd:or}, {cmd:hr}, {cmd:irr}, ...; see {helpb mi estimate}), so
{cmd:regtab} reads that rule from {cmd:e(cmdline_mi)}: {cmd:mi estimate: logit}
and {cmd:mi estimate, or: logit} both show odds ratios, never exponentiated
twice. Intervals are the ones {cmd:mi estimate} computed, on a t reference
distribution with each coefficient's own degrees of freedom
({cmd:e(df_mi)}); {cmd:mi estimate} reports no log likelihood, so the
{cmd:ll}, {cmd:aic}, and {cmd:bic} rows stay blank. A {cmd:level()} given to
{cmd:mi estimate} or {cmd:bootstrap} counts as the model's level.{p_end}
{p 4 8 2}- {cmd:streg}, {cmd:stintreg}, and {cmd:mestreg} follow Stata's metric rules
(exponential and Weibull fit in the log-hazard metric, HR, unless {cmd:time} or
{cmd:tratio} is given; Gompertz is log-hazard only; lognormal, loglogistic, and
generalized gamma are log-time only), and the log-time metric is shown as TR,
the exponentiated log-time coefficients: {cmd:stintreg, distribution(ggamma)}
and {cmd:stintreg, distribution(ggamma) tratio} both print time ratios under
"TR", and {cmd:r(methods)} reads "Time ratios ... from ... interval-censored
accelerated failure-time survival regression" ("Hazard ratios ... interval-censored
parametric proportional hazards survival regression" in the
log-hazard metric). {cmd:stintreg} names its distribution in {cmd:e(cmd)}, as
{cmd:streg} does.{p_end}
{p 4 8 2}- {cmd:glm} and {cmd:xtgee} are resolved from their
{opt family()} and {opt link()} rather than the command name, with glm's own
abbreviations ({cmd:f(b)}, {cmd:fam(bin)}, {cmd:l(logit)}, {cmd:ef}), so
{cmd:family(binomial)} or {cmd:family(bernoulli)} with the default or
{cmd:link(logit)} {it:->} OR, and {cmd:family(poisson)} with the default or
{cmd:link(log)} {it:->} IRR. Both are shown exponentiated and drop the
intercept row. Another family/link fitted with {cmd:eform} keeps glm's
exponentiated values under RR (binomial, log link), IRR (negative binomial,
log link), or exp(b); without {cmd:eform} it stays on the coefficient scale
with the {cmd:Coef.} header. {cmd:xtgee} takes the same options and defaults, so a
GEE fit gets the header, scale, and intercept rule that {cmd:glm} gets for the
same family and link. Supply {opt coef()} to label it yourself.{p_end}
{p 4 8 2}- All collected models must share one confidence level, because one
"#% CI" header labels every model. Models fitted with different {opt level()}
values (a model without {opt level()} counts as the current {cmd:set level})
are refused with an error.{p_end}
{p 4 8 2}- {opt relab:el}: relabels random effects using variable labels and explicit
parameter types. For single-level models {cmd:var(_cons)} becomes
{it:Variance: GroupLabel (Intercept)} and {cmd:cov(x,_cons)} becomes
{it:Covariance: GroupLabel (X label, Intercept)}; multi-level models label each
level separately. The {cmd:me}{it:*} estimators are labelled the same way: {cmd:var(x[clinic])}
becomes {it:Variance: Clinic label (X label)} and
{cmd:cov(x[clinic],_cons[clinic])} becomes
{it:Covariance: Clinic label (X label, Intercept)}. With differing grouping structures and no {opt relab:el},
the generic collection labels such as {cmd:var(_cons)} and {cmd:var(e)} are
retained rather than assigning a group label from another model. If {opt relab:el}
is requested with ambiguous random-effects metadata, {cmd:regtab} exits with
error 459 unless {opt nore:effects} suppresses random-effects rows.{p_end}
{p 4 8 2}- {opt eplotframe()}: stores a graph-ready companion frame for {helpb eplot} containing
{cmd:label}, {cmd:estimate}, {cmd:ll}, {cmd:ul}, {cmd:pvalue}, {cmd:model}, {cmd:model_label}, {cmd:rowtype}, and source-row
metadata. When {opt frame()} is also set, the display frame records the companion in
{cmd:_dta[tabtools_eplotframe]}.{p_end}
{p 4 8 2}- Frame provenance: requested display and eplot frames store the CI
level, ordered statistic IDs, model count, and per-model command identity,
outcome identity, effect scale, and display label as {cmd:_dta[tabtools_*]}
characteristics. {helpb comptab} uses these identities to align compatible
sources and rejects ambiguous or conflicting metadata.{p_end}
{p 4 8 2}- {opt addrow()}: appends custom label/value rows below the table body; separate
multiple rows with a backslash, e.g.,
{cmd:addrow("P trend" 0.032 0.041 \ "P interaction" 0.15 0.22)}. End a row
with {cmd:, after(}{it:term}{cmd:)} to place it inside the body, e.g. {cmd:addrow("P trend" 0.032 0.041, after(agecat))}
after the last level of
{cmd:agecat}.{p_end}
{p 4 8 2}- {opt labelwidth()}: caps the label-column width (in characters, default
45); labels longer than the cap wrap onto extra lines rather than being
clipped by the adjacent estimate cell.{p_end}
{p 4 8 2}- {opt dimnonsig}: dims rows whose every displayed fixed-effect CI
includes the null (1 for ratio scales, 0 for coefficients); reference, omitted,
and empty rows are always dimmed and category headers dim unless a level is
significant.{p_end}

{pstd}Notes on output shaping{p_end}
{p 4 8 2}- Constrained rows: a base category, a term dropped for collinearity,
and a factor cell identifying no observations all reach the table as a 0 (a 1
after {cmd:eform}) with an empty CI and an empty p-value. {cmd:regtab} tells
them apart from the constraint class the collection records for each
coefficient and substitutes {it:Reference}, {it:Omitted}, or {it:Empty} in the
estimate column, spanning that model's CI and p-value cells. The class is read
per model, so a level one model dropped keeps its estimate in the models that
retained it. Change the words with {opt refcat()}, {opt omitlabel()}, and
{opt notestlabel()} (default: the {opt emptylabel()} text); {opt refcat()},
{opt omitlabel()}, and {opt emptylabel()} must differ. A base level whose
cell identifies no observations ({cmd:logit} drops the observations of a
perfectly predicted base level, and another level is then dropped to identify
the rest) is not estimable: {opt notestlabel()}, not {it:Reference}. A base
level {cmd:collect} records as empty while no other level of its factor is
constrained has observations (Stata 17 stamps it so after {cmd:nbreg} beside
an omitted continuous term) and is {it:Reference}.{p_end}
{p 8 8 2}Stata 17's {cmd:collect} records every constrained cell of some fits as empty --
{cmd:nbreg}, {cmd:zinb}, {cmd:intreg}, and {cmd:streg} and {cmd:stintreg} with an ancillary parameter --
although the model reports the level as {cmd:(base)}; a model whose only recorded
class is empty is treated as carrying no class. It records no class at all for
a coefficient a constraint fixes. For these, the fit's own notes decide
whenever they are known: the notes Stata prints in the estimator's table,
{cmd:(base)}, {cmd:(empty)}, {cmd:(omitted)}, and {cmd:(constrained)}, read from the {cmd:e(b)} column
stripe and {cmd:e(Cns)}, for main effects and interaction cells alike. They come
from the record {cmd:tabtools fitcount} stores with each model at fit time (result
{cmd:tt_cns}), or, when the model is the active fit ({cmd:regtab} checks the command line
and that every coefficient equals the collected one, as {cmd:tabtools fitcount}
does), from the active {cmd:e(b)}. With the record, the labels do not depend on the
data in memory: they hold after a variable is dropped, after {cmd:estimates restore}
of another model, and on changed data.{p_end}
{p 8 8 2}Without the fit's notes, the base level of a factor main effect is
decided by the model's own specification: {cmd:regtab} expands its varlist,
as {cmd:e(cmdline)} records it, with {helpb fvexpand} over the observations
holding the levels the model holds, so the default (lowest level),
{cmd:ib}{it:#}{cmd:.}, {cmd:ib(first)}, {cmd:ib(last)}, {cmd:ibn.}, and
{cmd:fvset base} resolve as they did at fit time; the level marked {cmd:b.} is
{it:Reference} and any other constrained level of that factor is {it:Omitted}
({opt omitlabel()}), as {cmd:regress} reports a collinear level: a main-effect
level with no observations is not in {cmd:e(b)} at all. A base that only
{cmd:fvset} names (not an {cmd:ib}{it:#}{cmd:.} in the command line) may have
been set after the fit, so when the factor has two or more constrained levels
none of them is called {it:Reference}: all read {opt notestlabel()} and a note
says so. A factor never shows two {it:Reference} cells: when the base cannot
be verified (the data in memory no longer hold the fit's levels, or the
variable was dropped) and two or more levels are constrained, they all read
{opt notestlabel()}, with the same note; a single constrained level is the
base. This reads the data in memory, which must hold the model's variables,
under your {cmd:varabbrev} setting; where it cannot decide -- the varlist does
not expand, the base depends on frequencies ({cmd:ib(freq).},
{cmd:fvset base frequent}), the model holds a single level of the factor, or
the level it names was estimated in the model -- the value rule below
applies. An interaction cell is a base cell, {it:Reference}, when one of its
factor levels is the base the model holds for that factor's main effect; any
other constrained cell is empty or omitted and reads {opt notestlabel()}. A
cell of a factor the model holds no decided base for (an interaction without
its main effects) cannot be told: it reads {opt notestlabel()} and a note says
so.{p_end}
{p 8 8 2}Only a cell holding the constrained value itself (0, or 1 when the
collection holds a ratio) can be a base, omitted, or empty coefficient. Any
other value with a blank CI and p-value is an estimate: one a constraint fixes
(shown with {opt cnslabel()} when the fit's notes say so) or one estimated with
a zero or missing variance (a {cmd:stcox} level whose coefficient diverged
under separation), shown as {opt notestlabel()}, never as the reference.{p_end}
{p 8 8 2}In a model whose classes {cmd:collect} did record, a constrained level it left
without a class (a level a constraint fixes at 0, which Stata prints
{cmd:(omitted)}) is {it:Reference} only when it is the factor's base: the base its
specification resolves, as above, or an {cmd:ib}{it:#}{cmd:.} of the command line (which holds
without the data in memory). Any other such level is {it:Omitted} when the factor's
base is known, and {opt notestlabel()}, with a note, when it is not. Where the
collection carries no class and nothing above decides, a constrained factor
level holding the constrained value is {it:Reference} only when it is the factor's
only constrained level. A collection whose raw row identities cannot be read
is rejected before output.{p_end}
{p 4 8 2}- Equations with nothing estimated: in a multi-equation model such as
{cmd:mlogit}, the base-outcome equation constrains every one of its
coefficients. It carries no information and is dropped whole, so its levels
cannot be misread as reference categories of the equations that were
estimated.{p_end}
{p 4 8 2}- Interaction rows: an interaction term heads a single parent row for
all of its level combinations, labelled from the interacted variables'
labels ({it:Car origin # Weight (lbs.)}), with each combination indented
beneath and named in full ({it:Foreign # Weight (lbs.)}, or
{it:Female # Black} for two factors), so it never repeats a main-effect row's
label. A level with no value label reads {it:variable label = level}. A
product of continuous variables has no parent row and reads
{it:Mileage (mpg) # Weight (lbs.)}. A label set on the collection with
{cmd:collect label levels colname} is kept.{p_end}
{p 4 8 2}- Random-effects variance components ({cmd:var()}, {cmd:cov()},
{cmd:sd()}) from {cmd:mixed}, {cmd:melogit}, {cmd:mepoisson}, and similar
commands use the same {opt digits()} precision as the main coefficient
rows. Random-effects rows can be removed entirely with {opt nore}.{p_end}
{p 4 8 2}- Intercept, ordered cutpoint, and ancillary-only rows can be removed with
{opt noint}. Use {opt keepintercept} plus {opt cutlabels()} if you intentionally want
ordered-model cutpoints displayed with publication-friendly labels.{p_end}
{p 4 8 2}- P-value columns can be removed from the rendered table with
{opt nopvalue}. If {opt stars} is also specified, significance stars are still
computed from the collected p-values before the p-value columns are dropped.{p_end}
{p 4 8 2}- By default, fonts are set to Arial 10, but this can be overridden by
{opt font()}, {opt fontsize()}, or session defaults set with
{helpb tabtools:set font} / {helpb tabtools:set fontsize}. Borders
are drawn around the table and model blocks. Column widths and row heights are
adjusted heuristically to fit labels and contents: each model's estimate,
confidence-interval, and p-value columns are sized from that model's own cells,
so a model with wide estimates does not widen the others, and a model header
longer than its block wraps onto extra header lines.{p_end}
{p 4 8 2}- The command writes Excel and Markdown output through the shared tabtools
Mata {cmd:xl()} backend and then applies formatting in the same workbook session.{p_end}
{p 4 8 2}- Model statistics ({opt stats()}): For multi-model tables, N, AIC, BIC, QICu,
log-likelihood, and groups are extracted per model from the {helpb collect} framework
and placed in each model's column. If collection-based statistics extraction
fails, {cmd:regtab} exits with error 459 rather than substituting active
{cmd:e()} values. Statistics that are not defined for a model family remain
blank. For GEE models
({cmd:xtgee}), AIC is undefined because GEE uses quasi-likelihood rather than full
maximum likelihood. When dispersion is fixed at 1, a requested {cmd:aic} therefore
falls back to QICu ({cmd:deviance + 2p}; Cui 2007, eq. 2); QICu can also be requested directly via
{cmd:stats(qic)}. Standard binomial and Poisson GEE fits use this fixed scale, and
continuous fits may request it explicitly with {cmd:scale(1)}. If {cmd:e(phi)}
is estimated or otherwise differs from 1, {cmd:regtab} leaves QICu unavailable
rather than divide each candidate model by a different scale. The {cmd:qic}
option and {cmd:r(qic_}{it:#}{cmd:)} names are retained for backward
compatibility. ICC is computed per model from variance components in the
collected results when that variance decomposition is defined. Latent-response
ICC uses the link-specific level-1 variance
(logit: {cmd:pi^2/3}; probit: 1; complementary log-log: {cmd:pi^2/6}), as in
{bf:[ME] estat icc}, Methods and formulas; Austin and Merlo (sec. 4.5.1; see
{help regtab##references:References}) give the logit and probit forms. For
model families without a defined level-1 variance, ICC is left blank rather
than guessed. If the collection yields no usable ICC components for supported
models, {cmd:regtab} exits with error 459 rather than substituting active
{cmd:e(b)} values.{p_end}

{marker examples}{title:Examples}

{pstd}Logistic regression with odds ratios:{p_end}
{phang2}{cmd:. webuse nhanes2, clear}{p_end}
{phang2}{cmd:. collect clear}{p_end}
{phang2}{cmd:. collect: logit diabetes age female i.race bmi highbp}{p_end}
{phang2}{cmd:. regtab, xlsx(regression.xlsx) sheet("Diabetes") ///}{p_end}
{phang3}{cmd:title("Odds Ratios for Diabetes") coef(OR)}{p_end}

{pstd}Multinomial logistic regression with outcome-specific RRR rows:{p_end}
{phang2}{cmd:. sysuse auto, clear}{p_end}
{phang2}{cmd:. xtile price_group = price, nq(3)}{p_end}
{phang2}{cmd:. label define price_group 1 "Low price" 2 "Middle price" 3 "High price"}{p_end}
{phang2}{cmd:. label values price_group price_group}{p_end}
{phang2}{cmd:. collect clear}{p_end}
{phang2}{cmd:. collect: mlogit price_group mpg weight, baseoutcome(1)}{p_end}
{phang2}{cmd:. regtab, xlsx(regression.xlsx) sheet("Multinomial") title("Price group model")}{p_end}

{pstd}
Rows are labeled by outcome and term (for example, {it:Middle price: Mileage (mpg)}),
and the estimate header is auto-detected as {cmd:RRR}.{p_end}

{pstd}Ordered logit with custom cutpoint labels:{p_end}
{phang2}{cmd:. sysuse auto, clear}{p_end}
{phang2}{cmd:. keep if !missing(rep78)}{p_end}
{phang2}{cmd:. collect clear}{p_end}
{phang2}{cmd:. collect: ologit rep78 mpg weight}{p_end}
{phang2}{cmd:. regtab, xlsx(regression.xlsx) sheet("Ordered") keepintercept ///}{p_end}
{phang3}{cmd:cutlabels("1 to 2 \ 2 to 3 \ 3 to 4 \ 4 to 5")}{p_end}

{pstd}
The number of labels may match however many cutpoints the ordered model
returns. Without {opt keepintercept}, {cmd:regtab} treats cutpoints like ancillary rows
and omits them from ratio-scale presentation tables.{p_end}

{pstd}Two models with merged headers, dropping the intercept row:{p_end}
{phang2}{cmd:. webuse nhanes2, clear}{p_end}
{phang2}{cmd:. collect clear}{p_end}
{phang2}{cmd:. collect: logit diabetes age female}{p_end}
{phang2}{cmd:. collect: logit diabetes age female i.race bmi highbp}{p_end}
{phang2}{cmd:. regtab, xlsx(regression.xlsx) sheet("Table 2") ///}{p_end}
{phang3}{cmd:models("Unadj \ Adj") coef("OR") title("Table 2. Odds ratios") noint}{p_end}

{pstd}The multilevel and survival examples below are workflow sketches: they assume
a fitted dataset in memory with the named outcome, exposure, and grouping
variables (for example {cmd:provider}, {cmd:district}, {cmd:school}). Substitute your own
model; for runnable public-data fits see the {helpb melogit}, {helpb mixed}, and {helpb stcox} manual
examples.{p_end}

{pstd}Mixed-effects logistic model with Median Odds Ratio and ICC:{p_end}
{phang2}{cmd:. collect clear}{p_end}
{phang2}{cmd:. collect: melogit outcome treated age female || provider:}{p_end}
{phang2}{cmd:. regtab, xlsx(results.xlsx) sheet("Mixed") title("Multilevel Model") ///}{p_end}
{phang3}{cmd:relabel stats(n icc groups)}{p_end}

{pstd}
This produces a table where the random intercept variance is automatically
converted to a Median Odds Ratio (MOR), {opt relabel} translates
{cmd:var(_cons[provider])} into a readable label using the variable label of
{cmd:provider}, and {opt stats()} appends N, ICC, and number of groups at the
bottom.{p_end}

{pstd}Multi-level linear mixed model with nested random effects:{p_end}
{phang2}{cmd:. collect clear}{p_end}
{phang2}{cmd:. collect: mixed outcome treatment age || district: || school:}{p_end}
{phang2}{cmd:. regtab, xlsx(results.xlsx) sheet("Nested") title("Table 3") relabel stats(n icc)}{p_end}

{pstd}
With {opt relabel}, each grouping level gets a separate label derived from its
variable label and the parameter type: {it:Variance: District (Intercept)},
{it:Variance: School (Intercept)}, {it:Residual Variance}. Without {opt relabel}, the raw
bracket notation is shown: {cmd:var(_cons[district])}, {cmd:var(_cons[school])},
{cmd:var(e)}. Random effects are sorted by level (outermost first) after fixed
effects. Models with random slopes and covariance terms (e.g.,
{cmd:|| school: treatment, cov(unstructured)}) produce additional rows such as
{it:Variance: School (Treatment)} and
{it:Covariance: School (Treatment, Intercept)}. Residual terms that Stata
qualifies by equation, such as {cmd:var(e.y)} after {cmd:tobit} or
{cmd:meglm}, are labelled {it:Residual Variance (y)}, and
{cmd:sd(e.y)} is labelled {it:Residual SD (y)}.{p_end}

{pstd}Auto-detected coefficient labels and conditional formatting:{p_end}
{phang2}{cmd:. collect clear}{p_end}
{phang2}{cmd:. collect: stcox treated age female i.education}{p_end}
{phang2}{cmd:. regtab, xlsx(results.xlsx) sheet("Cox") title("Cox Model") ///}{p_end}
{phang3}{cmd:noint boldp(0.05) highlight(0.05)}{p_end}

{pstd}
When {opt coef()} is omitted, {cmd:regtab} auto-detects the label from the model
type: {cmd:logit}/{cmd:logistic} {it:->} OR, {cmd:stcox} {it:->} HR, {cmd:poisson}/{cmd:nbreg} {it:->} IRR,
{cmd:stcrreg} {it:->} SHR, {cmd:mlogit} {it:->} RRR, {cmd:zip}/{cmd:zinb}/{cmd:churdle} {it:->} Coef., {cmd:streg}
(log-hazard metric) {it:->} HR, {cmd:streg} (log-time metric) {it:->} TR, {cmd:regress}/{cmd:mixed} {it:->}
Coef. The {opt boldp()} option bolds p-value cells below the threshold, and
{opt highlight()} applies yellow fill to entire rows.{p_end}

{pstd}Thousands separators and "to" between the bounds (cost models):{p_end}
{phang2}{cmd:. sysuse auto, clear}{p_end}
{phang2}{cmd:. collect clear}{p_end}
{phang2}{cmd:. collect: regress price mpg weight i.foreign}{p_end}
{phang2}{cmd:. regtab, cformat(%12.0fc) sep(" to ") nopvalue}{p_end}

{pstd}A flat frame that {helpb puttab} writes as is, with the reference level of
{cmd:ib3.rep78} first in its block:{p_end}
{phang2}{cmd:. sysuse auto, clear}{p_end}
{phang2}{cmd:. collect clear}{p_end}
{phang2}{cmd:. collect: regress price mpg ib3.rep78}{p_end}
{phang2}{cmd:. regtab, frame(t2, replace flat) reftop compact}{p_end}
{phang2}{cmd:. frame t2: puttab rowlabel c* using regression.xlsx, sheet("Flat") varlabels}{p_end}

{pstd}{marker editflat}{...}
{bf:Editing a flat frame.} With {cmd:frame(}{it:name}{cmd:, flat keys)} the
frame carries each row's coefficient name, type, and order and each cell's
state, so the table can be edited by what its rows are. Here two factors print
the same level label ("Yes"), so a row label cannot pick one of them, but
{cmd:_term} can; a row goes in before the {cmd:heavy} block through
{cmd:_order}; and the rule above the first {opt stats()} row is computed from
{cmd:_term}:{p_end}
{phang2}{cmd:. sysuse auto, clear}{p_end}
{phang2}{cmd:. generate byte heavy = weight > 3000}{p_end}
{phang2}{cmd:. generate byte lengthy = length > 190}{p_end}
{phang2}{cmd:. label define yesno 0 "No" 1 "Yes"}{p_end}
{phang2}{cmd:. label values heavy lengthy yesno}{p_end}
{phang2}{cmd:. collect clear}{p_end}
{phang2}{cmd:. collect: regress price mpg i.heavy i.lengthy}{p_end}
{phang2}{cmd:. regtab, stats(n) compact frame(t, replace flat keys)}{p_end}
{phang2}{cmd:. frame t: replace c1 = "see note" if _term == "1.lengthy"}{p_end}
{phang2}{cmd:. frame t: summarize _order if _term == "heavy", meanonly}{p_end}
{phang2}{cmd:. local before = r(min) - 0.5}{p_end}
{phang2}{cmd:. frame t: insobs 1}{p_end}
{phang2}{cmd:. frame t: replace rowlabel = "Vehicle size" in l}{p_end}
{phang2}{cmd:. frame t: replace _order = `before' in l}{p_end}
{phang2}{cmd:. frame t: sort _order}{p_end}
{phang2}{cmd:. frame t: count if substr(_term, 1, 5) != "stat:"}{p_end}
{phang2}{cmd:. frame t: puttab rowlabel c* using regression.xlsx, sheet("Edited") varlabels hlines(`=r(N) + 1')}{p_end}

{pstd}Events, people, and person-time under each model, with thin levels
masked:{p_end}
{phang2}{cmd:. sysuse cancer, clear}{p_end}
{phang2}{cmd:. stset studytime, failure(died)}{p_end}
{phang2}{cmd:. generate id = _n}{p_end}
{phang2}{cmd:. generate pt = _t - _t0}{p_end}
{phang2}{cmd:. collect clear}{p_end}
{phang2}{cmd:. collect: stcox i.drug}{p_end}
{phang2}{cmd:. tabtools fitcount, events(_d) people(id) exposure(pt) terms}{p_end}
{phang2}{cmd:. collect: stcox i.drug age}{p_end}
{phang2}{cmd:. tabtools fitcount, events(_d) people(id) exposure(pt) terms}{p_end}
{phang2}{cmd:. regtab, stats(events people exposure) exposurelabel("Person-months") mincount(5)}{p_end}
{phang2}{cmd:. regtab, mincount(5) emptylabel("<5") notestlabel("NE") addrow("P trend" 0.04 0.03, after(drug))}{p_end}

{pstd}A specification table, one row per model, and a cell that must not show
an estimate:{p_end}
{phang2}{cmd:. sysuse auto, clear}{p_end}
{phang2}{cmd:. collect clear}{p_end}
{phang2}{cmd:. collect: regress price mpg i.foreign}{p_end}
{phang2}{cmd:. collect: regress price mpg i.foreign weight}{p_end}
{phang2}{cmd:. regtab, transpose keep(mpg 1.foreign) stats(n) nopvalue models("Crude \ Adjusted")}{p_end}
{phang2}{cmd:. regtab, transpose keep(mpg 1.foreign) stats(e(N, %9.0fc)="N" text("Adjustment" "None" "Weight")) collabels(mpg "Mileage") addcol("Note" a b, after(mpg)) nopvalue}{p_end}
{phang2}{cmd:. regtab, cellnote("Foreign" 2 "Not shown") nopvalue}{p_end}

{marker stored}{title:Stored results}

{pstd}{cmd:regtab} stores the following in {cmd:r()}:{p_end}

{synoptset 18 tabbed}{...}
{p2col 5 18 22 2: Scalars}{p_end}
{synopt:{cmd:r(N_rows)}}number of rows in output table{p_end}
{synopt:{cmd:r(N_cols)}}number of columns in output table{p_end}
{synopt:{cmd:r(N_models)}}number of models{p_end}
{synopt:{cmd:r(ci_level)}}confidence level carried by the collected intervals{p_end}
{synopt:{cmd:r(aic_}{it:#}{cmd:)}}AIC for model {it:#} when requested{p_end}
{synopt:{cmd:r(bic_}{it:#}{cmd:)}}BIC for model {it:#} when requested{p_end}
{synopt:{cmd:r(qic_}{it:#}{cmd:)}}QICu for fixed-scale {cmd:xtgee} model {it:#}, when available{p_end}
{synopt:{cmd:r(icc_}{it:#}{cmd:)}}ICC for model #, when available{p_end}
{synopt:{cmd:r(ll_}{it:#}{cmd:)}}log-likelihood for model {it:#} (when {cmd:stats(ll)}){p_end}
{synopt:{cmd:r(n_}{it:#}{cmd:)}}sample size for model {it:#} (when {cmd:stats(n)}){p_end}
{synopt:{cmd:r(groups_}{it:#}{cmd:)}}number of groups for model {it:#} (when {cmd:stats(groups)}){p_end}
{synopt:{cmd:r(events_}{it:#}{cmd:)}}events for model {it:#} (when {cmd:stats(events)}){p_end}
{synopt:{cmd:r(people_}{it:#}{cmd:)}}people for model {it:#} (when {cmd:stats(people)}){p_end}
{synopt:{cmd:r(exposure_}{it:#}{cmd:)}}person-time for model {it:#} (when {cmd:stats(exposure)}){p_end}
{synopt:{cmd:r(obs_}{it:#}{cmd:)}}{cmd:e(N)} for model {it:#} (when {cmd:stats(obs)}){p_end}
{synopt:{cmd:r(N_masked)}}cells masked by {opt mincount()} (when specified){p_end}
{synopt:{cmd:r(N_absent)}}levels missing from a model under {opt mincount()}{p_end}
{synopt:{cmd:r(N_stats_masked)}}{opt stats()} cells masked by {cmd:mincell()}{p_end}
{synopt:{cmd:r(N_stats_linked)}}{opt stats()} cells masked through {cmd:maskwith()} (when given){p_end}
{synopt:{cmd:r(mi_m_}{it:#}{cmd:)}}imputations for model {it:#} (when {cmd:stats(mi_m)}){p_end}
{synopt:{cmd:r(r2_a_}{it:#}{cmd:)}}adjusted R² for model {it:#} (when {cmd:stats(r2_a)}){p_end}
{synopt:{cmd:r(rmse_}{it:#}{cmd:)}}root MSE for model {it:#} (when {cmd:stats(rmse)}){p_end}
{synopt:{cmd:r(F_}{it:#}{cmd:)}}linear-model F for model {it:#} (when {cmd:stats(F)}){p_end}
{synopt:{cmd:r(fmi_}{it:#}{cmd:)}}largest FMI for model {it:#} (when {cmd:stats(fmi)}){p_end}
{synopt:{cmd:r(e_}{it:name}{cmd:_}{it:#}{cmd:)}}{cmd:e(}{it:name}{cmd:)} for model {it:#} (when {cmd:stats(e(}{it:name}{cmd:))}){p_end}
{synopt:{cmd:r(markdown_rows)}}body rows written to Markdown{p_end}
{synopt:{cmd:r(markdown_cols)}}columns written to Markdown{p_end}

{p2col 5 18 22 2: Macros}{p_end}
{synopt:{cmd:r(xlsx)}}Excel filename (if exported){p_end}
{synopt:{cmd:r(sheet)}}sheet name (if exported){p_end}
{synopt:{cmd:r(markdown)}}Markdown filename (if exported){p_end}
{synopt:{cmd:r(coef_label)}}shared or mixed coefficient label{p_end}
{synopt:{cmd:r(methods)}}auto-generated methods paragraph{p_end}
{synopt:{cmd:r(stars)}}stars option value{p_end}
{synopt:{cmd:r(frame)}}frame name (if {cmd:frame()} specified){p_end}
{synopt:{cmd:r(eplotframe)}}graph-ready companion frame name{p_end}

{p2col 5 18 22 2: Matrices}{p_end}
{synopt:{cmd:r(table)}}numeric displayed-body coefficients{p_end}
{p2colreset}{...}

{pstd}{cmd:r(table)} excludes the title and any appended stats/addrows. Row names are
derived from each variable's display label with periods, spaces, commas, and
colons replaced by underscores or stripped, then truncated to 32 characters. A
name that Stata's matrix row-name rules reject or rewrite (a bracketed
random-effects key such as {cmd:var(x[clinic])}, or {cmd:cov(x_cons)}, which Stata reads
back as {cmd:var(x_cons)}) instead has every character other than letters, digits,
and underscores replaced by an underscore. Names are then made unique in row
order: when a name repeats an earlier one (two variables labelled alike,
labels differing only in punctuation, or labels sharing their first 32
characters), it takes the first free suffix {cmd:_2}, {cmd:_3}, ..., with the base
shortened so the name stays within 32 characters. Row {it:j} of {cmd:r(table)} is always
row {it:j} of the displayed body.{p_end}

{pstd}The per-model statistic scalars ({cmd:r(aic_}{it:#}{cmd:)},
{cmd:r(bic_}{it:#}{cmd:)}, {cmd:r(qic_}{it:#}{cmd:)}, {cmd:r(icc_}{it:#}{cmd:)},
{cmd:r(ll_}{it:#}{cmd:)}, {cmd:r(n_}{it:#}{cmd:)}, {cmd:r(groups_}{it:#}{cmd:)})
carry the full-precision values; the corresponding rows in the displayed table
and {cmd:r(table)} are rounded for presentation. Only the scalars for the
statistics actually requested in {opt stats()} (and available for the model
family) are posted.{p_end}

{pstd}
{cmd:AIC} and {cmd:BIC} are recomputed from the log-likelihood, parameter count, and N as
{cmd:AIC = -2*ll + 2*k} and {cmd:BIC = -2*ll + k*ln(N)}, rather than read from
{cmd:e(aic)}/{cmd:e(bic)}, which {cmd:glm}/GEE backends store on an incomparable per-observation
or deviance scale. {cmd:k} is the number of estimated free parameters, whatever the
variance estimator. That is {cmd:e(rank)}, under a robust-type {cmd:vce()} too: it leaves
out base levels, omitted levels and the base outcome equation of {cmd:mlogit}, nets
out linear constraints, and gives the values {helpb estat ic} reports. The one
exception is a robust-type {cmd:vce()} ({cmd:cluster}, {cmd:robust}, {cmd:bootstrap}, {cmd:jackknife},
{cmd:linearized}) over too few clusters, GEE panels, or replications. The sandwich
variance then has rank at most G-1 (G clusters or panels) or R-1 (R
replications); when that cap is below the number of collected coefficients
with a nonzero standard error (outside the derived {cmd:_diparm} parameters; collect
stores the standard error of a base or omitted level, or of a coefficient
fixed by a constraint, as 0), {cmd:k} is that number. {bf:Divergence from estat ic:} in
that capped case {helpb estat ic} uses the capped {cmd:e(rank)} (for example df = 1 for a
{cmd:logit} with four coefficients and {cmd:vce(cluster)} on two clusters), while {cmd:regtab}
counts four, so its AIC and BIC equal those of the same model fitted with the
model-based variance. The likelihood itself does not depend on {cmd:vce()}. The
collection does not hold {cmd:e(Cns)}, so in the capped case a constraint fixing a
coefficient at a value is netted out (its standard error is 0) but an equality
constraint between coefficients is not: {cmd:constraints()} with {cmd:mpg = weight} and
{cmd:vce(cluster)} on two clusters counts four, where the model-based fit has
three. A robust variance that is rank deficient for another reason while the
cap is not below that number keeps {cmd:e(rank)}, as {helpb estat ic} does.{p_end}

{pstd}
{cmd:QICu} is {it:not} a likelihood criterion and does not come from
{helpb estat ic}. GEE fits a quasi-likelihood, so no log-likelihood is
available. For an {cmd:xtgee} fit with {cmd:e(phi)=1}, {cmd:regtab} reports
{cmd:deviance + 2*k}, where {cmd:k} is the number of regression parameters,
counted as for AIC above (Pan's penalty is 2p, p the number of parameters, so
a robust {cmd:vce()} over few panels does not shrink it). This
is a deviance-scale representation of QICu; it can differ from another
implementation's absolute QICu by a data-only additive constant, while
within-data model differences agree. Values are on a different scale from AIC
and BIC and must not be compared with them.{p_end}

{pstd}
{bf:Which criterion this is.} {cmd:deviance + 2*k} is Pan's {bf:QICu}, the fixed-penalty
approximation to QIC, {it:not} QIC itself. Pan's QIC is
{cmd:-2Q(b,I) + 2*trace(Omega*Sigma)}, whose penalty is a trace of the
independence-model information times the robust sandwich variance; {cmd:QICu}
replaces that trace with {cmd:k}, and that difference carries a real
restriction: {bf:QICu} compares only models sharing a working correlation
structure. Candidates must use the same outcome observations, weights, family,
link, clustering, and working correlation and differ only in the mean
specification. QICu must {it:not} be used to choose a working correlation
structure; that is precisely the job the trace penalty does and this
approximation drops.{p_end}

{pstd}
{bf:Dispersion boundary.} When dispersion is unknown, Pan's definition uses
one common scale estimate for every candidate model (typically from the
largest mean model). Dividing each model's quasi-likelihood by its own
{cmd:e(phi)} would make the comparison invalid. Because {cmd:regtab} formats an
existing collection and cannot prove that a common external scale was used, it
computes QICu only when each candidate's stored scale is 1. This includes
standard binomial and Poisson GEE fits and continuous fits estimated with
{cmd:scale(1)}. Otherwise it prints a note and leaves the QICu row and
{cmd:r(qic_}{it:#}{cmd:)} absent. Use a dedicated QIC implementation with one
explicit common scale for estimated-dispersion comparisons. The stored result
keeps the name {cmd:r(qic_}{it:#}{cmd:)} for backward compatibility; read it as
QICu.{p_end}

{pstd}
Cui (2007) gives QIC (eq. 1) and QICu (eq. 2) as Pan defined them.{p_end}


{marker references}{...}
{title:References}

{phang}
Austin, P. C., and J. Merlo. 2017. Intermediate and advanced topics in
multilevel logistic regression analysis. {it:Statistics in Medicine}
36: 3257-3277. {browse "https://doi.org/10.1002/sim.7336"}.{p_end}

{phang}
Austin, P. C., P. Wagner, and J. Merlo. 2017. The median hazard ratio: a useful
measure of variance and general contextual effects in multilevel survival
analysis. {it:Statistics in Medicine} 36: 928-938. {browse "https://doi.org/10.1002/sim.7188"}.{p_end}

{phang}
Cui, J. 2007. QIC program and model selection in GEE analyses. {it:Stata Journal}
7: 209-220. {browse "https://doi.org/10.1177/1536867X0700700205"}.{p_end}

{marker seealso}{...}
{title:Also see}

{pstd}{helpb effecttab} for treatment effects and margins tables{p_end}
{pstd}{helpb tabtools} for suite overview and persistent formatting defaults{p_end}
{pstd}{helpb collect} for the underlying collection framework{p_end}
{pstd}{helpb tabtools_tips} for quick reference{p_end}


{marker author}{...}
{title:Author}

{pstd}Timothy P Copeland, Karolinska Institutet{p_end}

{hline}