Feature request: different ways to display coefficient digits in etable

[lyifa]

The LaTex output of etable shows an inconsistent number of decimal places for estimated coefficients, t values, and R squared values. Specifically, some of my coefficients have 1 decimal place while the others have about 4. The t values also show this inconsistent pattern. Meanwhile, the R squared values have 7 decimal places.

In my area (accounting and finance), all journal papers require numbers to have the same decimal places.

[lrberge]

That's not a bug, it's a choice. The argument digits refers to the number of significant digits (I've just clarified the help) and I chose not to affect the statistics which have their own formatting.

Since you're not the only one requesting this feature, I'll see what I can do. But it won't be before a while, so maybe you could use modelsummary instead?

[lyifa]

Thanks for the clarification. I understand this is a choice.
I think I will stick to the plm package for a while before this feature is implemented in fixest.

[lrberge]

Sure. I'm letting the issue open until the feature is added.

[lyifa]

One more thing, I noticed that some R- squared values use scientific notation while others don't. I am not sure if this is also a choice.

[waynelapierre]

I also encountered the scientific notation issue. Hopefully, the etable function will be refined. This seems to be the only factor keeping people from using it in place of plm and/or lfe.

[lrberge]

So here's an update: I've added a different way to display the decimals.
To have a fixed number of decimals, use the code digits = "rd" with d the number of digits: e.g. digits = "r2" will round at the second decimal and will always show two decimals--no more, no less.

I have separated the display for the fit statistics, now there's the new argument digits.stats. I also took the advantage to speed up the code.

Here's an example of the new behavior:

n = 1e3
base = data.frame(x1 = rnorm(n), x2 = rnorm(n))
base$y = base$x1 / 1000 + 10*base$x2 + rnorm(n, sd = 150)

res1 = feols(y ~ x1, base)
res2 = feols(y ~ I(1000*x1) + x2, base)

etable(res1, res2, digits = 2)
#>                       res1          res2
#> Dependent Var.:          y             y
#>                                         
#> (Intercept)     -4.4 (4.8)    -4.2 (4.8)
#> x1               5.5 (5.1)              
#> I(1000 x x1)               0.005 (0.005)
#> x2                             4.1 (5.1)
#> _______________ __________ _____________
#> S.E. type         Standard      Standard
#> Observations         1,000         1,000
#> R2                   0.001         0.002
#> Adj. R2             0.0002       -0.0001
etable(res1, res2, digits = "r2")
#>                         res1         res2
#> Dependent Var.:            y            y
#>                                          
#> (Intercept)     -4.39 (4.76) -4.16 (4.77)
#> x1               5.55 (5.08)             
#> I(1000 x x1)                  0.01 (0.01)
#> x2                            4.10 (5.07)
#> _______________ ____________ ____________
#> S.E. type           Standard     Standard
#> Observations           1,000        1,000
#> R2                     0.001        0.002
#> Adj. R2               0.0002      -0.0001
etable(res1, res2, digits = "r2", digits.stats = "r2")
#>                         res1         res2
#> Dependent Var.:            y            y
#>                                          
#> (Intercept)     -4.39 (4.76) -4.16 (4.77)
#> x1               5.55 (5.08)             
#> I(1000 x x1)                  0.01 (0.01)
#> x2                            4.10 (5.07)
#> _______________ ____________ ____________
#> S.E. type           Standard     Standard
#> Observations           1,000        1,000
#> R2                      0.00         0.00
#> Adj. R2                 0.00        -0.00

setFixest_etable(digits = "r2", digits.stats = 2)
etable(res1, res2)
#>                         res1         res2
#> Dependent Var.:            y            y
#>                                          
#> (Intercept)     -4.39 (4.76) -4.16 (4.77)
#> x1               5.55 (5.08)             
#> I(1000 x x1)                  0.01 (0.01)
#> x2                            4.10 (5.07)
#> _______________ ____________ ____________
#> S.E. type           Standard     Standard
#> Observations           1,000        1,000
#> R2                     0.001        0.002
#> Adj. R2               0.0002      -0.0001

The default behavior for digits is actually not exactly to display the first digits significant digits. It's the outcome of an algorithm that I have now detailed in the help pages.

It's in the dev branch but I'll submit the new version to the CRAN next week.

[lyifa]

Thanks for the update! Have you looked into the R squared value scientific notation issue?

[lrberge]

Yes, it was due to the initial format displaying many digits for the statistics--it has been naturally solved by the introduction of the new argument digits.stats.

Now the scientific notation will still be popping but only when digits is not round and |x| < 1e-5 (so it should be rare). But then the threshold at which it appears can be set with the powerBelow argument.
Note though that for very large numbers, i.e. |x| > 1e10 (like the likelihoods for instance), scientific notation will pop and there's no way around it (I really don't think it's readable without it anyway).

[lrberge]

Just to say that I'll submit to the CRAN this week, so if you have any other comment/suggestion on that issue I'll try to implement them before submission.

[lyifa]

I have tested it extensively. It works well.

[lrberge]

Great then, I'll let it that way.

[maswiebe]

It seems that powerBelow overrides digits='sd', but not digits='rd'.

When I use digits='r4', powerBelow=-9, very small numbers show up as 0.0000.

But when I use digits='s4', powerBelow=-9, the small number shows up as 0.00000005.

Ok, I see you have the rule:

For numbers lower than 1 (in absolute terms), the number of decimals displayed is equal to digits except if there are only 0s in which case the first significant digit is shown.

I'd prefer to see 0.0000 instead of scientific notation or 0.00000005. I don't want to use digits='rd' in this case because I have some large coefficients where four decimal places is awkward, eg. I want 5001.6 instead of 5001.6245.