Plotting adjusted means as a bar graph

[Lucy Hiscox]

Hi,

I am trying to plot the adjusted means (in the form of a bar graph as both predictors are categorical) from a linear regression model.

Say I run the following:

regress y i.a##i.b c.covariable c.covariable i.covariable

The interaction is significant, so therefore I want to plot the adjusted means for each level of a (2-levels) and b (2 levels).

To see the adjusted means, I do:
margins a#b

And to get the individual predicted values I do:

predict predicted_y

I figured the best way to plot bar graphs was using the following code:

cibar predicted_y, over1(a) over2(b)

However, the means on the bar graphs do not match up to the values obtained from margins a#b

Any ideas on why this might be? Or perhaps another way I can directly plot the outcome from the interaction in a bar graph?

Many thanks,
Lucy

[Andrew Musau]

After the margins command:

marginsplot, recast(bar) by(b)

[Lucy Hiscox]

Thanks! but i can't seem to edit the colours based on (a). The levels of b are in two separate panels, which is great - but is it possible to change the colours for each level of a?

[Andrew Musau]

It will be difficult to do this with marginsplot. But there are several posts here on differentiating bars or plots in other contexts. See, e.g., https://www.statalist.org/forums/for...e-along-x-axis and https://www.statalist.org/forums/for...insplot-to-bar. If it was my problem, I would install coefplot from SSC. Here is an example (you can change the axis labels, naturally):

webuse lbw, clear
regress bwt age lwt i.smoke#i.ui i.race
margins i.smoke#i.ui, post
estimates store marg
coefplot (marg, recast(bar) keep(0.smoke#0.ui 0.smoke#1.ui) bcolor(red%50)) ///
         (marg, recast(bar) keep(1.smoke#0.ui 1.smoke#1.ui) bcolor(blue%50)), ///
         nokey scheme(s1mono) xsc(r(0,.)) xlab(0(1000)3500)

One extra detail is that bars should start at zero, and coefplot truncates part of it. Some referees may be very particular on this point, so specify the axis labels such that the bars start at zero (highlighted).


Res.:

[Nick Cox]

A bar chart has no obvious advantage here.

In Andrew Musau's helpful example, the bars are truncated at 2000, and indeed there is no obvious point in starting the bars at zero as the comparison is of birthweights, with each other, not with zero. Yet bar length (height) encodes (birthweight MINUS 2000), which I presume is arbitrary and here a side-effect of the nice axis labels and the fact that the underlying command won't force bars to start at zero.

We have no way of commenting on @Lucy Hiscox's values as y is thoroughly anonymous and there is no data example.

I suggest as principles:

Bars starting at zero make sense if and only if comparisons with zero are germane.

Otherwise bars starting anywhere else need a rationale for "anywhere else".

Otherwise the most direct and informative plot represents coefficient estimates by point markers and uncertainty by capped or uncapped spikes.

[Lucy Hiscox]

Hi both, many thanks for your replies.

To add context - here is my data.

Perhaps it would be best to present coefficient estimates by point markers, as you suggest, with: marginsplot, x(ipv_any_recent) ?

 margins child_sex#ipv_any_recent, post

Predictive margins                                         Number of obs = 143
Model VCE: OLS

Expression: Linear prediction, predict()

------------------------------------------------------------------------------------------
                         |            Delta-method
                         |     Margin   std. err.      t    P>|t|     [95% conf. interval]
-------------------------+----------------------------------------------------------------
child_sex#ipv_any_recent |
   Male#Below threshold  |   1078.921    6.52852   165.26   0.000     1066.006    1091.836
   Male#Above threshold  |   1092.304   7.588248   143.95   0.000     1077.293    1107.316
 Female#Below threshold  |   1083.772   6.838856   158.47   0.000     1070.243    1097.301
 Female#Above threshold  |   1066.813   7.471749   142.78   0.000     1052.032    1081.594
------------------------------------------------------------------------------------------ 

marginsplot, recast(bar) by(ipv_any_recent)

[Andrew Musau]

Try:

margins child_sex#ipv_any_recent, post
marginsplot, recast(scatter) by(ipv_any_recent) xsc(r(-0.5 1.5))

[Nick Cox]

Fortunately, or fortuitously, the results make my point. If I understand correctly the t tests indicate massively that the estimated quantities are not zero, which seems unlikely to be scientific news. (I still don't know what the outcome variable is.)

At the same time the intervals all overlap....

[Lucy Hiscox]

Fortunately, or fortuitously, the results make my point. If I understand correctly the t tests indicate massively that the estimated quantities are not zero, which seems unlikely to be scientific news. (I still don't know what the outcome variable is.)

At the same time the intervals all overlap....

Hi Nick,

The outcome is brain volume.

Here is the main omnibus test of the interaction:

  regress amygdala_corrected ib0.ipv_any_recent##ib0.child_sex i.clinic i.maternal_hiv i.bdi_total_threshold c.birthweight i.prenatal_cotin
> ine i.prenatal_alcohol_composite c.child_age 

      Source |       SS           df       MS      Number of obs   =       140
-------------+----------------------------------   F(11, 128)      =      3.44
       Model |  47257.5452        11  4296.14047   Prob > F        =    0.0003
    Residual |  159882.635       128  1249.08308   R-squared       =    0.2281
-------------+----------------------------------   Adj R-squared   =    0.1618
       Total |   207140.18       139  1490.21712   Root MSE        =    35.342

--------------------------------------------------------------------------------------------
        amygdala_corrected | Coefficient  Std. err.      t    P>|t|     [95% conf. interval]
---------------------------+----------------------------------------------------------------
            ipv_any_recent |
          Above threshold  |   11.55974   8.659454     1.33   0.184    -5.574472    28.69394
                           |
                 child_sex |
                   Female  |   6.939471   8.343113     0.83   0.407    -9.568803    23.44775
                           |
  ipv_any_recent#child_sex |
   Above threshold#Female  |  -27.22183    12.2964    -2.21   0.029    -51.55235   -2.891306
                           |
                    clinic |
                TC Newman  |  -9.942632   7.795401    -1.28   0.204    -25.36716    5.481902
                           |
              maternal_hiv |
                 Positive  |   .4519586   8.840126     0.05   0.959    -17.03974    17.94366
                           |
       bdi_total_threshold |
          Above threshold  |  -3.721508   6.851118    -0.54   0.588    -17.27762      9.8346
               birthweight |    -.02854   .0067888    -4.20   0.000    -.0419729   -.0151072
                           |
         prenatal_cotinine |
           Passive smoker  |   2.076213   7.555715     0.27   0.784    -12.87406    17.02649
            Active smoker  |   4.749271   8.759693     0.54   0.589    -12.58328    22.08182
                           |
prenatal_alcohol_composite |
                 Exposure  |    11.6201   8.300294     1.40   0.164    -4.803455    28.04365
                 child_age |  -9.187756   3.754138    -2.45   0.016    -16.61596   -1.759554
                     _cons |   1201.906   26.28029    45.73   0.000     1149.906    1253.906
-------------------------------------------------------------------------------------------- 

So we aren't looking at the t tests or how they differ from zero.

In sex-stratified results, I get:

 bysort child_sex: regress amygdala_corrected ib0.ipv_any_recent i.clinic i.maternal_hiv i.bdi_total_threshold c.birthweight i.prenatal_co
> tinine i.prenatal_alcohol_composite c.child_age

-------------------------------------------------------------------------------------------------------------------------------------------
-> child_sex = Male

      Source |       SS           df       MS      Number of obs   =        71
-------------+----------------------------------   F(9, 61)        =      3.01
       Model |  32976.7139         9  3664.07932   Prob > F        =    0.0049
    Residual |  74229.1089        61  1216.87064   R-squared       =    0.3076
-------------+----------------------------------   Adj R-squared   =    0.2054
       Total |  107205.823        70  1531.51175   Root MSE        =    34.884

--------------------------------------------------------------------------------------------
        amygdala_corrected | Coefficient  Std. err.      t    P>|t|     [95% conf. interval]
---------------------------+----------------------------------------------------------------
            ipv_any_recent |
          Above threshold  |   12.17911   8.742984     1.39   0.169    -5.303563    29.66179
                           |
                    clinic |
                TC Newman  |  -15.20955   11.19939    -1.36   0.179    -37.60413     7.18502
                           |
              maternal_hiv |
                 Positive  |    5.48999   12.46346     0.44   0.661    -19.43224    30.41222
                           |
       bdi_total_threshold |
          Above threshold  |  -.8874126   10.31498    -0.09   0.932    -21.51349    19.73866
               birthweight |  -.0306118   .0092658    -3.30   0.002    -.0491399   -.0120837
                           |
         prenatal_cotinine |
           Passive smoker  |  -2.936045   10.62852    -0.28   0.783    -24.18909      18.317
            Active smoker  |  -3.984423   12.25012    -0.33   0.746    -28.48005     20.5112
                           |
prenatal_alcohol_composite |
                 Exposure  |   14.01245   11.68538     1.20   0.235    -9.353903    37.37881
                 child_age |   -12.5523   6.162457    -2.04   0.046     -24.8749   -.2297055
                     _cons |   1223.427   37.29894    32.80   0.000     1148.843    1298.011
--------------------------------------------------------------------------------------------

-------------------------------------------------------------------------------------------------------------------------------------------
-> child_sex = Female

      Source |       SS           df       MS      Number of obs   =        69
-------------+----------------------------------   F(9, 59)        =      1.59
       Model |  19053.9529         9  2117.10587   Prob > F        =    0.1388
    Residual |  78474.2361        59   1330.0718   R-squared       =    0.1954
-------------+----------------------------------   Adj R-squared   =    0.0726
       Total |   97528.189        68  1434.23807   Root MSE        =     36.47

--------------------------------------------------------------------------------------------
        amygdala_corrected | Coefficient  Std. err.      t    P>|t|     [95% conf. interval]
---------------------------+----------------------------------------------------------------
            ipv_any_recent |
          Above threshold  |  -20.29075   9.759249    -2.08   0.042    -39.81896   -.7625396
                           |
                    clinic |
                TC Newman  |   -.550149   11.95036    -0.05   0.963    -24.46277    23.36247
                           |
              maternal_hiv |
                 Positive  |  -7.854308   13.70646    -0.57   0.569    -35.28087    19.57226
                           |
       bdi_total_threshold |
          Above threshold  |  -6.252422   10.14643    -0.62   0.540    -26.55539    14.05055
               birthweight |  -.0247939    .010742    -2.31   0.025    -.0462885   -.0032992
                           |
         prenatal_cotinine |
           Passive smoker  |    10.4688   11.39933     0.92   0.362    -12.34119     33.2788
            Active smoker  |    9.10838   13.55255     0.67   0.504    -18.01021    36.22697
                           |
prenatal_alcohol_composite |
                 Exposure  |   12.35853    12.5019     0.99   0.327    -12.65772    37.37478
                 child_age |  -5.739353   5.094657    -1.13   0.264    -15.93374    4.455033
                     _cons |   1180.553   38.49096    30.67   0.000     1103.532    1257.573
-------------------------------------------------------------------------------------------- 

which suggests IPV effects females only. Hence the need to want to show adjusted means in the form of a bar graph.

Does that sound ok?

[Nick Cox]

I think discussing these results requires more medical expertise than I have, which is almost none. I am left with a feeling that you have rather a small dataset for this complicated a model, but I doubt that is news to you, and it's a cheap comment when often getting more data that remains comparable is often far easier said than done.