Getting started  –  How to use coefplot

Basic usage

The basic procedure is to compute one or more sets of estimates (e.g. regression models) and then apply coefplot to these estimation sets to draw a plot displaying the point estimates and their confidence intervals. Estimation commands store their results in the so-called e() returns (type ereturn list after running an estimation command to see a list of what has been stored). By default, coefplot retrieves the point estimates from (the first equation in) vector e(b) and computes confidence intervals from the variance estimates found in matrix e(V). See the Estimates and Confidence intervals examples for information on how to change these defaults. Furthermore, coefplot can also read results from matrices that are not stored as part of an estimation set; see Plotting results from matrices below.

The syntax to produce a plot of the coefficients of a single model is

coefplot [name] [, options]

where name is the name of a stored model (see help estimates store), or . or empty string for the active model. For example, to plot the point estimates and 95% confidence intervals for the most recent model, type:

. sysuse auto, clear
(1978 automobile data)

. regress price mpg trunk length turn

      Source |       SS           df       MS      Number of obs   =        74
-------------+----------------------------------   F(4, 69)        =      5.79
       Model |   159570047         4  39892511.8   Prob > F        =    0.0004
    Residual |   475495349        69  6891236.94   R-squared       =    0.2513
-------------+----------------------------------   Adj R-squared   =    0.2079
       Total |   635065396        73  8699525.97   Root MSE        =    2625.1

------------------------------------------------------------------------------
       price | Coefficient  Std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
         mpg |  -186.8417   88.17601    -2.12   0.038     -362.748   -10.93533
       trunk |  -12.72642   104.8785    -0.12   0.904    -221.9534    196.5005
      length |   54.55294   35.56248     1.53   0.130    -16.39227    125.4981
        turn |  -200.3248   140.0166    -1.43   0.157    -479.6502    79.00066
       _cons |   8009.893   6205.538     1.29   0.201    -4369.817     20389.6
------------------------------------------------------------------------------

. coefplot, drop(_cons) xline(0)

Option drop(_cons) has been added to exclude the constant of the model; option xline(0) has been added to draw a reference line at zero so one can better see which coefficients are significantly different from zero.

By default, coefplot uses a horizontal layout in which the names of the coefficients are placed on the Y-axis and the estimates and their confidence intervals are plotted along the X-axis. Specify option vertical to use a vertical layout:

. coefplot, vertical drop(_cons) yline(0)

Note that, because the axes were flipped, we now have to use yline(0) instead of xline(0).

Keep and drop

By default, coefplot displays all coefficients from the first equation of a model. Alternatively, options keep() and drop() can be used to specify the elements to be displayed. For example, above, option drop(_cons) was used to exclude the constant. Furthermore, coefplot automatically excluded coefficients that are flagged as "omitted" or as "base levels". To include such coefficients in the plot, specify options omitted and baselevels.

For example, if you want to display all equations from a multinomial logit model (including the equation for the base outcome for which all coefficients are zero by definition), type:

. sysuse auto, clear
(1978 automobile data)

. gen mpp = mpg/8

. mlogit rep78 mpp i.foreign if rep>=3
(output omitted)

. coefplot, nolabel drop(_cons) keep(*:) omitted baselevels

Option keep(*:) selects all equations for display, not just first. For detailed information on the syntax, see the description of the keep() option in the help file.

Here is a further example that illustrates how keep() can be used to select different coefficients depending on equation:

. coefplot, nolabel keep(3:*.foreign 4:mpp 5:mpp _cons) omitted baselevels

Plotting multiple models

Models as separate series

The syntax to include multiple models as separate series in the same graph is

coefplot (name [, plotopts]) (name [, plotopts]) ... [, globalopts]

where plotopts are options that apply to a single series. These options specify the information to be collected, affect the rendition of the series, and provide a label for the series in the legend. globalopts are options that apply to the overall graph, such as titles or axis labels, but may also contain any options allowed as plot options to provide defaults for the single series. A basic example is as follows:

. sysuse auto, clear
(1978 automobile data)

. regress price mpg trunk length turn if foreign==0
(output omitted)

. estimates store D

. regress price mpg trunk length turn if foreign==1
(output omitted)

. estimates store F

. coefplot D F, drop(_cons) xline(0)

To specify separate options for an individual model, enclose the model and its options in parentheses. For example, to add a label for each plot in the legend, to use alternative plot styles, and to change the marker symbol, you could type:

. coefplot (D, label(Domestic Cars) pstyle(p3))  ///
>          (F, label(Foreign Cars)  pstyle(p4))  ///
>          , drop(_cons) xline(0) msymbol(S)

Option msymbol() is specified as a global option so that the same symbol is used in both series. To use different symbols, include an individual msymbol() option for each model.

Alternatively, you can also use p1(), p2(), etc. to specify options for the different series:

. coefplot D F, drop(_cons) xline(0) msymbol(S) ///
>     p1(label(Domestic Cars) pstyle(p3))       ///
>     p2(label(Foreign Cars)  pstyle(p4))

Changing the offsets

coefplot offsets the plot positions of the coefficients so that the confidence spikes do not overlap. To deactivate the automatic offsets, you can specify global option nooffsets. Alternatively, custom offsets may be specified by the offset() option (if offset() is specified for at least one model, automatic offsets are disabled). The spacing between coefficients is one unit, so usually offsets between –0.5 and 0.5 make sense. For example, if you want to use smaller offsets than the default, you could type:

. sysuse auto, clear
(1978 automobile data)

. regress price mpg trunk length turn if foreign==0
(output omitted)

. estimates store D

. regress price mpg trunk length turn if foreign==1
(output omitted)

. estimates store F

. coefplot (D, offset(0.05)) (F, offset(-0.05)), drop(_cons) xline(0)

Using multiple axes

If the dependent variables of the models you want to include in the graph have different scales, it can be useful to employ the axis() plot option to assign specific axes to the models. For example, to include a regression on price and a regression on weight in the same graph, type:

. sysuse auto, clear
(1978 automobile data)

. regress price mpg trunk length turn
(output omitted)

. estimates store Price

. regress weight mpg trunk length turn
(output omitted)

. estimates store Weight

. coefplot Price (Weight, axis(2)), drop(_cons) xtitle(Price) xtitle(Weight, axis(2))

Appending models

The syntax to merge multiple models into the same series is

coefplot (namelist [, plotopts]) ...

or, more precisely,

coefplot (namelist [, modelopts] \ namelist [, modelopts] \ ... [, plotopts]) ...

where modelopts are options that apply to a single model. For example, if you want to draw a graph comparing bivariate and multivariate effects, you could type:

. sysuse auto, clear
(1978 automobile data)

. regress price mpg trunk length turn
(output omitted)

. estimates store multivariate

. foreach var in mpg trunk length turn {
  2.     quietly regress price `var'
  3.     estimates store `var'
  4. }

. coefplot (mpg trunk length turn, label(bivariate)) ///
>          (multivariate)                            ///
>          , drop(_cons) xline(0)

When merging multiple models you may need to apply some renaming of coefficients, because coefficients that have the same name will be printed on top of each other. This can be achieved by applying the rename() option to the individual models. An alternative approach is presented in Model names as coefficient names.

Models as subgraphs

The syntax to create subgraphs is

coefplot plotlist [, subgropts] || plotlist [, subgropts] || ... [, globalopts]

where plotlist is a list of models as above and subgropts are options that apply to a single subgraph. An example with one model per subgraph is as follows:

. sysuse auto, clear
(1978 automobile data)

. regress price mpg trunk length turn if foreign==0
(output omitted)

. estimates store D

. regress price mpg trunk length turn if foreign==1
(output omitted)

. estimates store F

. coefplot D, bylabel(Domestic Cars) ///
>       || F, bylabel(Foreign Cars)  ///
>       ||, drop(_cons) xline(0)

Multiple models per subgraph

An example with multiple models per subgraph is:

. sysuse auto, clear
(1978 automobile data)

. regress price mpg trunk length turn if foreign==0
(output omitted)

. estimates store D

. regress price mpg trunk length turn if foreign==1
(output omitted)

. estimates store F

. regress weight mpg trunk length turn if foreign==0
(output omitted)

. estimates store D_weight

. regress weight mpg trunk length turn if foreign==1
(output omitted)

. estimates store F_weight

. coefplot (D, label(Domestic)) (F, label(Foreign)), bylabel(Price)   ///
>       || (D_weight)           (F_weight)         , bylabel(Weight)  ///
>       ||, drop(_cons) xline(0) byopts(xrescale)

Option byopts(xrescale) has been added so that the two subgraphs can have different scales. Furthermore, the plot labels for the legend were set within the first subgraph. They could also have been specified within the second subgraph, as plot styles are recycled with each new subgraph and plot options are collected across subgraphs (unless norecycle is specified; see below).

If the subgraphs do not contain the same number of models, it may be necessary to insert "empty" models to achieve the correct alignment. This can be achieved by typing _skip:

. coefplot (D, label(Domestic)) (F, label(Foreign)), bylabel(Price)   ///
>       || _skip                (F_weight)         , bylabel(Weight)  ///
>       ||, drop(_cons) xline(0) byopts(xrescale)

Different plot styles per subgraph

As evident in the last example, coefplot recycles plot styles within each subgraph. If you want each subgraph to use its own set of styles, apply the norecycle option:

. sysuse auto, clear
(1978 automobile data)

. forvalues i = 2/5 {
  2.     quietly regress price mpg trunk length turn if rep78==`i'
  3.     estimates store rep`i'
  4. }

. coefplot (rep2, label(rep78=2)) (rep3, label(rep78=3)), bylabel(Low record)  ///
>       || (rep4, label(rep78=4)) (rep5, label(rep78=5)), bylabel(High record) ///
>       ||, drop(_cons) xline(0) norecycle legend(colfirst)

How subgraphs are combined

Use option byopts(byopts) to determine how subgraphs are combined. See help by_option for available byopts. For example, to use a compact style and stack the subgraphs in one column, you could type:

. sysuse auto, clear
(1978 automobile data)

. regress price mpg trunk length turn if foreign==0
(output omitted)

. estimates store D

. regress price mpg trunk length turn if foreign==1
(output omitted)

. estimates store F

. coefplot D, bylabel(Domestic Cars) || F, bylabel(Foreign Cars) ///
>       ||, drop(_cons) xline(0) byopts(compact cols(1))

Note that Stata renders the titles of the subgraphs as "subtitles". Hence, you can use the subtitle() option to change their styling:

. coefplot D, bylabel(Domestic Cars) || F, bylabel(Foreign Cars) ///
>     ||, drop(_cons) xline(0) byopts(compact cols(1)) ///
>     subtitle(, size(vlarge) margin(medium) justification(left) ///
>                color(white) bcolor(black) bmargin(top_bottom))

Subgraphs by coefficients

Sometimes it makes sense to arrange coefficients in separate subgraphs with individual scales, as the size of coefficients may vary considerably. For example, when comparing results by subgroups or estimation techniques, the focus usually lies on differences across models and less on differences within models, so that it appears natural to use individuals subgraphs for the different coefficients.

Creating subgraphs by coefficients requires lengthy commands as for each coefficient a separate piece of subgraph syntax has to be put together. To avoid this extra typing you can use the bycoefs option. Technically, bycoefs flips coefficients and subgraphs, that is, the coefficients are treated as "subgraphs" and what was specified as subgraphs is treated as "coefficients". This seems difficult to understand, but should become clear in the following example:

. sysuse auto, clear
(1978 automobile data)

. forv i = 3/5 {
  2.     quietly regress price mpg headroom weight turn if rep78==`i'
  3.     estimate store rep78_`i'
  4. }

. coefplot rep78_3 || rep78_4 || rep78_5, drop(_cons) xline(0) ///
>     bycoefs byopts(xrescale)

As some people prefer vertical mode for such a graph, you might want to specify the vertical option:

. coefplot rep78_3 || rep78_4 || rep78_5, drop(_cons) yline(0) ///
>     bycoefs byopts(yrescale) vertical

Here is an example that adds another dimension. Displayed are the means of some variables by repair record and car type:

. sysuse auto, clear
(1978 automobile data)

. forv s = 0/1 {
  2.     forv i = 3/5 {
  3.         quietly mean price mpg headroom weight if foreign==`s' & rep78==`i'
  4.         estimate store m`s'_`i'
  5.     }
  6. }

. coefplot m0_3 m1_3, bylabel(rep78=3)  ///
>       || m0_4 m1_4, bylabel(rep78=4)  ///
>       || m0_5 m1_5, bylabel(rep78=5)  ///
>       || , bycoefs byopts(xrescale)   ///
>            plotlabels("Domestic cars" "Foreign cars")

Using wildcards in model names

Instead of providing distinct model names to coefplot, you can also specify a name pattern containing * (any string) and ? (any nonzero character) wildcards. coefplot will then plot the results from all matching models.

If a name pattern is specified within parentheses, the results from the matching models will be combined into the same plot. An example is as follows:

. sysuse auto, clear
(1978 automobile data)

. foreach var of varlist mpg trunk length turn {
  2.     quietly regress price `var' if foreign==0
  3.     estimates store d_`var'
  4.     quietly regress price `var' if foreign==1
  5.     estimates store f_`var'
  6. }

. estimates dir

----------------------------------------------------------------
             |           Dependent  Number of        
        Name | Command    variable     param.  Title 
-------------+--------------------------------------------------
       d_mpg | regress       price          2  Linear regression
       f_mpg | regress       price          2  Linear regression
     d_trunk | regress       price          2  Linear regression
     f_trunk | regress       price          2  Linear regression
    d_length | regress       price          2  Linear regression
    f_length | regress       price          2  Linear regression
      d_turn | regress       price          2  Linear regression
      f_turn | regress       price          2  Linear regression
----------------------------------------------------------------

. coefplot (d*, label(domestic)) (f*, label(foreign)) ///
>     , drop(_cons) xline(0) title("Bivariate effects on price by car type")

This is equivalent to the following command using explicit names:

. coefplot (d_mpg d_trunk d_length d_turn, label(domestic)) ///
>          (f_mpg f_trunk f_length f_turn, label(foreign))  ///
>     , drop(_cons) xline(0) title("Bivariate effects on price by car type")

If multiple patterns are specified, options attached to a specific pattern will be applied to all matching models. Example:

. coefplot (d*, asequation(Domestic) \ f*, asequation(Foreign) \ , pstyle(p4)) ///
>     , drop(_cons) xline(0) title("Bivariate effects on price by car type")

This is equivalent to the following command using explicit names:

. coefplot (d_mpg d_trunk d_length d_turn, asequation(Domestic) \ ///
>           f_mpg f_trunk f_length f_turn, asequation(Foreign)  \ ///
>           , pstyle(p4)) ///
>     , drop(_cons) xline(0) title("Bivariate effects on price by car type")

If a name pattern is specified without parentheses, the matching models will be treated as separate series:

. sysuse nlsw88, clear
(NLSW, 1988 extract)

. generate lnwage = ln(wage)

. forvalues i=0/1 {
  2.     forvalues j=1/2 {
  3.         quietly regress lnwage grade ttl_exp tenure if south==`i' & race==`j' 
  4.         estimates store est`i'_`j'
  5.     }
  6. }

. coefplot est0* || est1*, drop(_cons) xline(0) ///
>     plotlabels("White" "Black") bylabels("North" "South")

This is equivalent to the following command using explicit names:

. coefplot est0_1 est0_2 || est1_1 est1_2, drop(_cons) xline(0) ///
>     plotlabels("White" "Black") bylabels("North" "South")

When using a name pattern that is expanded into multiple series, you need to use p1(), p2(), etc. to provide separate options for the different series:

. sysuse auto, clear
(1978 automobile data)

. forvalues i=3/5 {
  2.     quietly regress price mpg trunk if rep78==`i'
  3.     estimates store rep_`i'
  4. }

. coefplot rep*, drop(_cons) xline(0) ///
>     p1(pstyle(p3) label("Rep=3")) ///
>     p2(pstyle(p4) label("Rep=4")) ///
>     p3(pstyle(p5) label("Rep=5"))

How coefficients are matched

The default for coefplot is to use the first (nonzero) equation from each model and match coefficients across models by their names (ignoring equation names). For example, regress returns one (unnamed) equation containing the regression coefficients whereas tobit returns two equations, an equation named after the dependent variable containing the regression coefficients and equation / containing the variance of the error term (this is true for Stata 16 or newer; in Stata 15 or lower, tobit uses different equation names). Hence, the default for coefplot is to match the regression coefficients from the two models and ignore equation / from the Tobit model:

. webuse laborsub, clear

. regress whrs kl6 k618 wa we

      Source |       SS           df       MS      Number of obs   =       250
-------------+----------------------------------   F(4, 245)       =      5.27
       Model |  16526046.1         4  4131511.52   Prob > F        =    0.0004
    Residual |   192218058       245    784563.5   R-squared       =    0.0792
-------------+----------------------------------   Adj R-squared   =    0.0641
       Total |   208744104       249  838329.733   Root MSE        =    885.76

------------------------------------------------------------------------------
        whrs | Coefficient  Std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
         kl6 |  -462.1233   124.6768    -3.71   0.000    -707.6985   -216.5481
        k618 |    -91.141   45.85001    -1.99   0.048    -181.4515   -.8305151
          wa |   -13.1577   8.334958    -1.58   0.116    -29.57502    3.259612
          we |   53.26156   26.09369     2.04   0.042     1.864986    104.6581
       _cons |   940.0593   530.7197     1.77   0.078     -105.296    1985.415
------------------------------------------------------------------------------

. estimate store regress

. tobit whrs kl6 k618 wa we, ll(0)

Refining starting values:

Grid node 0:   log likelihood = -1402.6764

Fitting full model:

Iteration 0:   log likelihood = -1402.6764  
Iteration 1:   log likelihood = -1371.1868  
Iteration 2:   log likelihood = -1367.1258  
Iteration 3:   log likelihood = -1367.0904  
Iteration 4:   log likelihood = -1367.0903  

Tobit regression                                    Number of obs     =    250
                                                           Uncensored =    150
Limits: Lower =    0                                    Left-censored =    100
        Upper = +inf                                   Right-censored =      0

                                                    LR chi2(4)        =  23.03
                                                    Prob > chi2       = 0.0001
Log likelihood = -1367.0903                         Pseudo R2         = 0.0084

------------------------------------------------------------------------------
        whrs | Coefficient  Std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
         kl6 |  -827.7655   214.7521    -3.85   0.000    -1250.753   -404.7781
        k618 |  -140.0191   74.22719    -1.89   0.060     -286.221    6.182766
          wa |  -24.97918   13.25715    -1.88   0.061    -51.09118     1.13281
          we |   103.6896   41.82629     2.48   0.014     21.30625    186.0729
       _cons |   589.0002   841.5952     0.70   0.485    -1068.651    2246.652
-------------+----------------------------------------------------------------
  var(e.whrs)|    1715859   216775.7                       1337864     2200650
------------------------------------------------------------------------------

. mat list e(b)

e(b)[1,6]
           whrs:        whrs:        whrs:        whrs:        whrs:           /:
            kl6         k618           wa           we        _cons  var(e.whrs)
y1   -827.76553   -140.01914   -24.979183    103.68958    589.00023    1715858.7

. estimate store tobit

. coefplot regress tobit, xline(0) 

Even though the collected results from regress and tobit have different equation names (_ and whrs, respectively), coefplot matches their coefficients, that is, the equation names are ignored. This is the default if only one equation per model is collected. If you want to take equation names into account nonetheless, you can specify the eqstrict option:

. coefplot regress tobit, xline(0) eqstrict

Although eqstrict causes equation names to be relevant, the second equation from the tobit model is still ignored. To include all equations, specify keep(*:):

. coefplot regress tobit, xline(0) keep(*:) ///
>     transform(var(e.whrs) = sqrt(@)) rename(var(e.whrs) = Sigma)

(Option transform() has been applied to take the square root of the error variance, so that the coefficients are on a similar scale. Option rename() has been specified to replace var(e.whrs) by Sigma.)

Furthermore, to match the coefficients from regress with the first equation from tobit and also print the second equation from tobit, you can use asequation() to set the equation name of regress to whrs:

. coefplot (regress, asequation(whrs)) ///
>          (tobit, keep(*:) transform(var(e.whrs) = sqrt(@)) ///
>                  rename(var(e.whrs) = Sigma)) ///
>          , xline(0)

Alternatively, you could also use eqrename(_ = whrs) to rename equation _ to whrs or eqrename(whrs = _) to rename equation whrs to _.

The following example further illustrates how you can get rid of the equation labels:

. coefplot (regress, asequation(whrs))  ///
>          (tobit, keep(*:) transform(var(e.whrs) = sqrt(@)) ///
>                  rename(var(e.whrs) = Sigma)) ///
>          , xline(0) noeqlabels

As illustrated above, coefplot matches coefficients based on their names. Use the rename() option if you want to match coefficients that have different names in the input models. Here is an example that illustrates the effect of measurement error in regression models:

. drop _all

. matrix C = ( 1, .5, 0 \ .5, 1, .3 \ 0, .3, 1 )

. drawnorm x1 x2 x3, n(10000) corr(C)
(obs 10,000)

. generate y = 1 + x1 + x2 + x3 + 5 * invnorm(uniform())

. regress y x1 x2 x3
(output omitted)

. estimates store m1

. generate x1err = x1 + 2 * invnorm(uniform())

. regress y x1err x2 x3
(output omitted)

. estimates store m2

. coefplot (m1, label(Without error))   ///
>          (m2, label(With error))      ///
>          , xline(1) rename(x1err = x1)

We can see how measurement error in x1 distorts all slope coefficients in the model, even for variable x3 that is uncorrelated with x1 (due to the indirect correlation through x2).

Ordering and sorting

Automatic order of coefficients

In general, coefficients are plotted in the same order (from top to bottom) as they appear in the input models. However, coefficients appearing only in later models are placed after coefficients from earlier models (with the exception of _cons, which is always placed last). Have a look at the following example:

. sysuse auto, clear
(1978 automobile data)

. label variable mpg    "1. mpg"

. label variable trunk  "{bf:2. trunk}"

. label variable length "{bf:3. length}"

. label variable turn   "4. turn"

. regress price mpg length
(output omitted)

. estimate store m1

. regress price mpg trunk turn
(output omitted)

. estimate store m2

. regress price mpg trunk length turn
(output omitted)

. estimate store m3

. coefplot m1 || m2 || m3, xline(0) drop(_cons) byopts(row(1))

Even though in the full model (m3) trunk comes before length, the order of the two coefficients is reversed in the plot. This is because length but not trunk is part of the first model. That is, because trunk only appears in the later models, it is placed after length that appears already in the first model.

To establish an order as in the full model, you can use the orderby() option:

. coefplot m1 || m2 || m3, xline(0) drop(_cons) byopts(row(1)) orderby(3:)

Typing orderby(3:) instructs coefplot to use the model displayed in the third subgraph to determine the order of the coefficients. (Typing orderby(3) would refer to the third model in the first subgraph, which doesn't exist in the example above; orderby(3:3) would refer to the third model in the third subgraph.)

Explicit order of coefficients

Alternatively, you can specify an explicit order of coefficients using the order() option:

. coefplot m1 || m2 || m3, xline(0) drop(_cons) byopts(row(1))    ///
>     order(mpg trunk length)

Within order(), you can use the * (any string) and ? (any nonzero character) wildcards. Furthermore, you can type . to insert gaps. Example:

. label variable mpg

. label variable trunk

. label variable length

. label variable turn

. coefplot m1 || m2 || m3, xline(0) drop(_cons) byopts(row(1))    ///
>     order(. mpg . t* . length .)

In case of multiple equations, the specified order of coefficients applies to each equation:

. sysuse auto, clear
(1978 automobile data)

. gen mpp = mpg/8

. mlogit rep78 mpp if rep>=3
(output omitted)

. estimates store m1

. mlogit rep78 mpp i.foreign if rep>=3
(output omitted)

. estimates store m2

. coefplot m1 || m2, xline(0) nolabel keep(*:) order(_cons 1.foreign mpp)

Alternatively, to change the order of equations without changing the order of coefficients, type:

. coefplot m1 || m2, xline(0) nolabel keep(*:) order(5: 4:)

You can also specify a separate order for each equation or even take equations apart, as in the following example:

. coefplot m1 || m2, xline(0) nolabel keep(*:) ///
>     order(4:1.foreign 5:1.foreign 4:_cons mpp)

Sort coefficients by size

Coefficients can be ordered by size using the sort() option. Here is an example that displays average wages by industry, from lowest to highest:

. sysuse nlsw88, clear
(NLSW, 1988 extract)

. drop if inlist(industry,2)
(4 observations deleted)

. regress wage ibn.industry, nocons noheader
-----------------------------------------------------------------------------------------
                   wage | Coefficient  Std. err.      t    P>|t|     [95% conf. interval]
------------------------+----------------------------------------------------------------
               industry |
 Ag/Forestry/Fisheries  |   5.621121   1.348538     4.17   0.000     2.976592     8.26565
          Construction  |   7.564934   1.032496     7.33   0.000     5.540173    9.589695
         Manufacturing  |   7.501578   .2902381    25.85   0.000     6.932411    8.070745
Transport/Comm/Utility  |   11.44335   .5860926    19.52   0.000       10.294     12.5927
Wholesale/Retail trade  |   6.125897   .3046951    20.11   0.000     5.528379    6.723414
Finance/Ins/Real est..  |   9.843174   .4012702    24.53   0.000     9.056269    10.63008
   Business/Repair svc  |    7.51579   .5995678    12.54   0.000     6.340017    8.691564
     Personal services  |   4.401093    .564549     7.80   0.000     3.293993    5.508193
 Entertainment/Rec svc  |   6.724409   1.348538     4.99   0.000      4.07988    9.368938
 Professional services  |   7.871186   .1936975    40.64   0.000     7.491338    8.251033
 Public administration  |   9.148407   .4191131    21.83   0.000     8.326512    9.970302
-----------------------------------------------------------------------------------------

. coefplot, sort

To sort from highest to lowest, specify the descending suboption:

. coefplot, sort(, descending)

sort() has a by() suboption to select the statistic by which the coefficients are ordered. For example, to sort by estimation precision (standard errors) you could type:

. coefplot, sort(, by(se))

Note how confidence intervals increase from top to bottom.

If a graph contains multiple series, it usually makes sense to select a specific series for sorting the coefficients (the default is to take all available estimates into account; this is equivalent to sort coefficients based on their minimums across series). An example is as follows:

. sysuse nlsw88, clear
(NLSW, 1988 extract)

. drop if inlist(industry,2)
(4 observations deleted)

. regress wage ibn.industry, nocons noheader
(output omitted)

. estimates store overall

. regress wage ibn.industry if union==0, nocons
(output omitted)

. estimates store nonunion

. regress wage ibn.industry if union==1, nocons
(output omitted)

. estimates store union

. coefplot overall, nokey ///
>       || nonunion union, bylabel(by union status) ///
>       || , norecycle byopts(legend(position(5))) sort(1, descending)

In the example, the first series (overall means) is used for sorting. To sort by the third series (wages of the unionized; green markers in the right panel), you could type:

. coefplot overall, nokey ///
>       || nonunion union, bylabel(by union status) ///
>       || , norecycle byopts(legend(position(5))) sort(3, descending)

Because in this example the norecycle option has been specified, the series are uniquely identified across the subgraphs. If norecycle is omitted, series are repeated by subgraph. Hence, to sort by a series in a specific subgraph in this case you need to provide both the subgraph number and the series number, as in the following example:

. sysuse nlsw88, clear
(NLSW, 1988 extract)

. drop if inlist(industry,1,2,3,10)
(67 observations deleted)

. regress wage ibn.industry if union==0 & south==0, nocons
(output omitted)

. estimates store nonunionnorth

. regress wage ibn.industry if union==1 & south==0, nocons
(output omitted)

. estimates store unionnorth

. regress wage ibn.industry if union==0 & south==1, nocons
(output omitted)

. estimates store nonunionsouth

. regress wage ibn.industry if union==1 & south==1, nocons
(output omitted)

. estimates store unionsouth

. coefplot nonunionnorth unionnorth, bylabel(North) ///
>       || nonunionsouth unionsouth, bylabel(South) ///
>       || , plotlabels("nonunion" "union") sort(2:1, descending)

sort(2:1) instructs coefplot to sort the coefficients according to the first series in the second subgraph (wages of nonunionized in the south; blue series in right panel).

Plotting results from matrices

By default, coefplot reads results from the e() returns of stored estimation sets. To read results from a matrix, type matrix(name) instead of name when referring to the results. coefplot will then collect the point estimates from the first row of matrix name instead of from e(b) of estimation set name. If plotting results from matrices, you also have to specify how to obtain the confidence intervals, using the v(), se(), or ci() option. For example, to plot medians and their confidence intervals as computed by centile you could type:

. sysuse auto, clear
(1978 automobile data)

. matrix median = J(1,3,.)

. matrix colnames median = mpg trunk turn 

. matrix CI = J(2,3,.)

. matrix colnames CI = mpg trunk turn 

. matrix rownames CI = ll95 ul95

. local i 0

. foreach v of var mpg trunk turn {
  2.     local ++ i
  3.     quietly centile `v'
  4.     matrix median[1, `i'] = r(c_1)
  5.     matrix CI[1, `i'] = r(lb_1) \ r(ub_1)
  6. }

. matrix list median

median[1,3]
      mpg  trunk   turn
r1     20     14     40

. matrix list CI

CI[2,3]
            mpg      trunk       turn
ll95         19         12  37.078729
ul95         22         16         42

. coefplot matrix(median), ci(CI)

By default coefplot reads the point estimates from the first row of the specified matrix, and the CIs from the first two rows of the matrix specified in ci(). Depending on the layout of your results matrices, you will need to specify the rows and columns to read from (see the next example, or the remark on Plotting results from matrices in the help file).

Results from estimation commands and from matrices can be combined in the same graph. Here is an example that displays means and medians of price by repair record:

. sysuse auto, clear
(1978 automobile data)

. matrix R = J(5, 3, .)

. matrix coln R = median ll95 ul95

. matrix rown R = 1 2 3 4 5

. forv i = 1/5 {
  2.     quietly centile price if rep78==`i'
  3.     matrix R[`i',1] = r(c_1), r(lb_1), r(ub_1)
  4. }

. matrix list R

R[5,3]
      median       ll95       ul95
1     4564.5       4195       4934
2       4638   3898.525    8993.35
3       4741  4484.8407  5714.9172
4     5751.5  4753.4403  7055.1933
5       5397  3930.5673  6988.0509

. mean price, over(rep78)

Mean estimation                              Number of obs = 69

---------------------------------------------------------------
              |       Mean   Std. err.     [95% conf. interval]
--------------+------------------------------------------------
c.price@rep78 |
           1  |     4564.5      369.5      3827.174    5301.826
           2  |   5967.625   1265.494      3442.372    8492.878
           3  |   6429.233   643.5995       5144.95    7713.516
           4  |     6071.5   402.9585      5267.409    6875.591
           5  |       5913   788.6821      4339.209    7486.791
---------------------------------------------------------------

. coefplot (., label(mean) rename(^.*([0-9])\..+$ = \1, regex)) ///
>          (matrix(R[,1]), ci((2 3)) label(median)) ///
>          , ytitle(Repair Record 1978) xtitle(Price)

(In this example, option rename() can be omitted in Stata 15 or lower, or if version is set to 15 or lower.)

Changing the plot type (recast)

By default, coefplot uses marker symbols for point estimates and spikes for confidence intervals. To change the plot types, use the recast() and ciopts(recast()) options. For example, to use bars for point estimates and capped spikes for confidence intervals, you could type:

. sysuse auto, clear
(1978 automobile data)

. regress price mpg trunk length if foreign==0
(output omitted)

. estimates store D

. regress price mpg trunk length if foreign==1
(output omitted)

. estimates store F

. coefplot (D, label(Domestic Cars)) (F, label(Foreign Cars))   ///
>     , drop(_cons) xline(0) recast(bar) ciopts(recast(rcap)) citop barwidt(0.3)

Option citop has been added so that the confidence spikes are plotted in front of the bars.

The recast() option can also be used to select other plot types such as connected-line plots or area plots. Furthermore, different plot types can be combined in a single graph by specifying a separate recast() option for each series. Here is an example, in which proportions are displayed as bars (without confidence intervals using the noci option) and means are displayed as a connected-line plot with capped spikes for confidence intervals:

. sysuse auto, clear
(1978 automobile data)

. proportion rep78
(output omitted)

. estimates store prop

. mean price, over(rep78)
(output omitted)

. estimates store mean

. coefplot (prop, recast(bar) noci barwidth(0.5) color(*.6)) ///
>          (mean, recast(connected) ciopts(recast(rcap)) axis(2)) ///
>     , vertical nooffsets plotlabels("Proportion" "Price") ///
>       xtitle("Repair record") ytitle("Proportion") ytitle("Price", axis(2)) ///
>       rename(^.*([0-9])\..+$ = \1, regex)

The example also illustrates some other options. Option vertical has been specified to flip the axes, option nooffsets omits offsetting the plot positions so that bars and markers are both centered above the categories, option axis() has been applied to the second series so that proportions and means have different axes. Furthermore, option plotlabels() provides an alternative way to specify legend labels for the series (instead of specifying separate label() options).

Option rename() is applied because mean and proportion label the coefficients differently. In Stata 15 or lower, or if version is set to 15 or lower, the option can be omitted.

Using a continuous axis (at)

The coefficients provided to coefplot may represent estimates along a continuous dimension. Examples are predictive margins or marginal effects computed over values of a continuous variable. In such a case, use the at() option to provide the plot positions to coefplot. Here is an example where predictive margins of foreign are computed by level of mpg, once from a bivariate model and once from a multivariate model:

. sysuse auto, clear
(1978 automobile data)

. logit foreign mpg
(output omitted)

. margins, at(mpg=(10(2)40)) post
(output omitted)

. estimates store bivariate

. logit foreign mpg turn price
(output omitted)

. margins, at(mpg=(10(2)40)) post
(output omitted)

. estimates store multivariate

. coefplot bivariate multivariate, ytitle(Pr(foreign=1)) xtitle(Miles per Gallon) ///
>     at recast(line) lwidth(*2) ciopts(recast(rline) lpattern(dash))

The example also illustrates how to change the plot types such that the estimates and their confidence intervals are displayed as lines. Furthermore, note that automatic offsetting of plot positions is deactivated by default if the at() option is specified.

Types and placement of options

coefplot has four levels of options:

1. modelopts are options that apply to a single model (or matrix). They specify the information to be collected from the model.
2. plotopts are options that apply to a single plot (i.e., a single series of points displayed in the same style), possibly containing results from multiple models. They affect the rendition of markers and confidence intervals and provide a label for the plot.
3. subgropts are options that apply to a single subgraph, possibly containing multiple plots.
4. globalopts are options that apply to the overall graph. This also includes option byopts() to determine how subgraphs are combined.

The levels are nested in the sense that upper level options include all lower level options. That is, globalopts includes subgropts, plotopts, and modelopts; subgropts includes plotopts, and modelopts; plotopts includes modelopts. However, upper level options may not be specified at a lower level.

If lower level options are specified at an upper level, they serve as defaults for all included lower levels elements. For example, if you want to draw 99% and 95% confidence intervals for all included models, specify levels(99 95) as global option:

coefplot model1 model2 model3, levels(99 95)

Options specified with an individual element override the defaults set by upper level options. For example, if you want to draw 99% and 95% confidence intervals for model 1 and model 2 and 90% confidence intervals for model 3, you could type:

coefplot model1 model2 (model3, level(90)), levels(99 95)

There are some fine distinctions about the placement of options and how they are interpreted. For example, if you type

coefplot m1, opts1 || m2, opts2 opts3

then opts2 and opts3 are interpreted as global options. If you want to apply opts2 only to m2 then type

coefplot m1, opts1 || m2, opts2 ||, opts3

Similarly, if you type

coefplot (m1, opts1 \ m2, opts2)

then opts2 will be applied to both models. To apply opts2 only to m2 type

coefplot (m1, opts1 \ m2, opts2 \)

or, if you also want to include opts3 to be applied to both models, type

coefplot (m1, opts1 \ m2, opts2 \, opts3)

or

coefplot (m1, opts1 \ m2, opts2 \), opts3

In case of multiple subgraphs there is some ambiguity about where to specify the plot options (unless global option norecycle is specified). You can provide plot options within any of the subgraphs, as plot options are collected across subgraphs. However, in case of conflict, the plot options from the later subgraphs usually take precedence over earlier plot options. In addition, you can also use global options p1(), p2(), etc. to provide options for specific plots. In case of conflict, options specified within a plot take precedence over options provided via p1(), p2(), etc.