coefplot: group results by DV and label by coefficient

[Tom Ford]

Hello,
Take the following example data

* create base data
sysuse auto,clear
gen brand = word(make, 1)
keep if inlist(brand, "Audi", "Fiat", "Olds", "Pont.")
encode brand, generate(brand_en)

* estimate regression of interest
reg price ibn.brand_en,nocons
eststo rpr
reg weight ibn.brand_en,nocons
eststo rwe
reg length ibn.brand_en,nocons
eststo lwe

I am estimating the effect of brands on price, weight and length.
I am using coefplot to present my results

coefplot (rpr,label(Price) pstyle(p3) msymbol(Oh) offset(-0.25)) (rwe,label(weight)  msymbol(S) pstyle(p2)) (lwe,label(Length)  msymbol(D) pstyle(p4) offset(0.25)), vert  ciopts(lwidth(medium) recast(rcap)) ylabel(,grid) plotregion(margin(vlarge))  //

This is the resulting plot:


To ease interpretation, I would like to change this plot. The x axis should consist of the three dependent variables, hence the coefficients for the same DV should be close to one another. Different colors should be used in each coefficient bar, to indicate the brands (Audi, Fiat, Ods, etc..). In other words, I would like to swap the regressor in the x-axis, with the DVs in the legend box.

I am not sure how to do this: does anyone knows how I could do this, with coefplot or in other ways?

thanks a lot in advance for your help

[Andrew Musau]

coefplot is from SSC, as you are asked to explain (FAQ Advice #12). suest is one way to combine estimates from regress. See

help suest

* create base data
sysuse auto,clear
gen brand = word(make, 1)
keep if inlist(brand, "Audi", "Fiat", "Olds", "Pont.")
encode brand, generate(brand_en)

* estimate regression of interest
reg price ibn.brand_en,nocons
eststo rpr
reg weight ibn.brand_en,nocons
eststo rwe
reg length ibn.brand_en,nocons
eststo lwe

suest rpr rwe lwe
est sto all
set scheme s1color
coefplot (all, keep(*:1.brand_en)) (all, keep(*:2.brand_en)) ///
(all, keep(*:3.brand_en)) (all, keep(*:4.brand_en)), ///
eqlabels(Price Weight Length) vert nokey

Res.:

[Tom Ford]

Thank you Andrew,
This is really helpful!

I was wondering, however, if there is a way to use suest using bootstrap standard errors.

In the actual data, I am running these regressions,

* estimate regression of interest
reg price ibn.brand_en,nocons vce(bootstrap, rep(99))
eststo rpr
reg weight ibn.brand_en,nocons vce(bootstrap, rep(99))
eststo rwe
reg length ibn.brand_en,nocons vce(bootstrap, rep(99))
eststo lwe
suest rpr rwe lwe

and if i run suest I get an error:

suest rpr rwe lwe
rpr was estimated with a nonstandard vce (bootstrap)

Is there a way to use suest for boostrap standard errors? otherwise are there other ways of obtaining this plot? perhaps using combomarginsplot? thanks a lot

[Andrew Musau]

bootstrap, reps(99): mvreg price weight length= ibn.brand_en, nocons

See

help mvreg

[Tom Ford]

thanks!

[Tom Ford]

Dear Andrew, one quick question, it appears that I get slightly different results if I do

bootstrap, reps(99): mvreg price weight length= ibn.brand_en, nocons or
bootstrap, reps(99): mvreg price = ibn.brand_en, nocons

Note that the results in this latter case are equivalent to running

reg price ibn.brand_en,nocons vce(bootstrap, rep(99) seed(123))

How is this possible? what can I do if I want to keep baseline estimation?

[Andrew Musau]

There is a random element inherent in the bootstrapping procedure. In each iteration of the bootstrap process, a sample is drawn from the original dataset by randomly selecting data points with replacement. This means that some data points may be included multiple times in a bootstrap sample, while others may not be included at all. You repeat the resampling process many times (indicated by -reps()-) to generate a distribution of the statistic of interest. That is why you set the seed if you want to reproduce the same results. See

help bootstrap

bootstrap, reps(99) seed(123): mvreg price weight length= ibn.brand_en, nocons or
bootstrap, reps(99) seed(123): mvreg price = ibn.brand_en, nocons

[Tom Ford]

So my issue does not appear to be with SEs, but rather with point estimates. Unfortunately I can't do this with the example data, but I will show you my results with real data.

Say I want to estimate this model

reg y1 ib4.x, cluster(factory_code) vce(bootstrap, rep(99) seed(123))

I get the following results

                                  (Replications based on 50 clusters in factory_code)
--------------------------------------------------------------------------------------
                     |   Observed   Bootstrap                         Normal-based
              y1 | coefficient  std. err.      z    P>|z|     [95% conf. interval]
---------------------+----------------------------------------------------------------
      biq_combined_c |
    x1  |   .0562682    .079509     0.71   0.479    -.0995665    .2121029
    x2 |   .2881812   .0613841     4.69   0.000     .1678705    .4084918
      x3  |   .3174499   .0634304     5.00   0.000     .1931286    .4417711
x5  |    .133072   .0721366     1.84   0.065    -.0083131    .2744572
                     |
               _cons |  -.1583722   .0584802    -2.71   0.007    -.2729913    -.043753
--------------------------------------------------------------------------------------

Note that the coefficient on x2 is .288

However if I estimate the same model as part of mvreg, I obtain a different coefficient

bootstrap, reps(99) seed(123) cluster(factory_code): mvreg y1 y2 y3 y4= ib4.x

I get the following results

                                   (Replications based on 50 clusters in factory_code)
--------------------------------------------------------------------------------------
                     |   Observed   Bootstrap                         Normal-based
                     | coefficient  std. err.      z    P>|z|     [95% conf. interval]
---------------------+----------------------------------------------------------------
y1               |
      biq_combined_c |
    x1  |   .0528582    .085247     0.62   0.535    -.1142229    .2199393
    x2  |   .2864781   .0589678     4.86   0.000     .1709033     .402053
      x3  |   .3137325   .0643007     4.88   0.000     .1877056    .4397595
x5  |   .1154333   .0741515     1.56   0.120     -.029901    .2607677
                     |
               _cons |  -.1518942   .0622998    -2.44   0.015    -.2739995   -.0297888

As you can see results from mvreg are different. The coefficient on x2 is 2.86, rather than 2.88. Given that the reason to run mvreg, is to plot my results for the baseline equations, I would want them to be identical. Do you know why this is happening?