Regression Tables API

TexTables should be able to provide a basic regression table for any model that adheres to the RegressionModel API found in StatsBase and makes it easy to customize the tables with additional fit statistics or model information as you see fit. This section documents how to use and customize the regression tables functionality for models in your code, as well as how to override the default settings for a model in your Package.

Special Structure of Regression Tables

Regression tables in TexTables are constructed using a special API that is provided to ensure that the regression tables from different estimators (potentially from separate packages) can be merged together. You should not construct your tables directly if you want them to merge nicely with the standard regression tables. Instead, you should use the methods documented in this section.

Regression tables are divided into 3 separate row blocks:

  1. Coefficients: This block contains the parameter estimates and standard errors (possibly decorated with stars for p-values) and always appears first
  2. Metadata: This block is empty by default (and therefore will not be printed in the table), but can be populated by the user to include column/model specific metadata. For example, a user might want to denote whether or not they controlled for one of the variables in their data, or which estimator they used in each column (OLS/Fixed Effects/2SLS/etc...)
  3. Fit Statistics: This block contains fit statistics. It defaults to $R^2$ and the number of observations, but this can be changed by the user.

You can construct sub-blocks within each of these three layers, although this is turned off by default. In order to support these three layers and the possible addition of sublayers, TableCols that conform to this API must be subtypes of TableCol{3,M} where M. For convenience a typealias RegCol{M} = TableCol{3,M} is provided, along with a constructor for empty RegCols from just the desired header.

Adding Each Block

You can construct or add to each of the three blocks using the convenience methods setcoef!, setmeta!, and setstats!. All three have an identical syntax:

set[block]!(t::RegCol, key, val[, se]; level=1, name="")
set[block]!(t::RegCol, key=>val; level=1, name="")
set[block]!(t::RegCol, kv::Associative)

This will insert into t a key/value pair (possibly with a standard error) within the specified block. Like the TableCol constructor, the pairs can be passed as either individual key/value[/se] tuples or pairs, as several vectors of key/value[/se] pairs, or as an associative.

To add additional sub-blocks, use the level keyword argument. Integers less than 0 will appears in blocks above the standard block, and integers greater than 1 will appear below it.

To name the block or sub-block, pass a nonempty string as the name keyword argument.

For instance, if you wanted to construct a regression column with two coefficients 1.32 (0.89) and -0.21 (0.01), metadata that indicates that the underlying estimation routine used OLS, and an $R^2$ of 0.73, then you would run the following code:

col = RegCol("My Column")
setcoef!(col, "Coef 1"=>(1.32, 0.89), "Coef 2"=>(-0.21, 0.01))
setmeta!(col, :Estimator=>"OLS")
setstats!(col, "\$R^2\$"=>0.73)

# output
          | My Column
   Coef 1 |     1.320
          |   (0.890)
   Coef 2 |    -0.210
          |   (0.010)
Estimator |       OLS
    $R^2$ |     0.730

Robust Standard Errors

If you would like to override the standard stderror function for your table, use the stderror keyword argument. For instance, you might want to use the CovarianceMatrices package to compute robust standard errors. In this case, you would simply define a new function

using CovarianceMatrices
robust(m) = stderror(m, HC0)
TableCol("My Column", m; stderror=robust)

Note: This feature is relatively experimental and its usage may change in future releases.

Integrating TexTables into your own Estimation Package

Once you know how you would like your model's regression tables to look, it is extremely easy to built it with TexTables. For instance, the code to integrate TexTables with some of the basic StatsModels.jl RegressionModel types is extremely short, and quite instructive to examine:

function TableCol(header, m::RegressionModel;
                  stats=(:N=>Int∘nobs, "\$R^2\$"=>r2),
                  meta=(), stderror::Function=stderror, kwargs...)

    # Compute p-values
    pval(m) = ccdf.(FDist(1, dof_residual(m)),

    # Initialize the column
    col  = RegCol(header)

    # Add the coefficients
    for (name, val, se, p) in zip(coefnames(m), coef(m), stderror(m), pval(m))
        addcoef!(col, name, val, se)
        0.05 <  p <= .1  && star!(col[name], 1)
        0.01 <  p <= .05 && star!(col[name], 2)
                p <= .01 && star!(col[name], 3)

    # Add in the fit statistics
    addstats!(col, OrderedDict(p.first=>p.second(m) for p in stats))

    # Add in the metadata
    addmeta!(col, OrderedDict(p.first=>p.second(m) for p in meta))

    return col

Here, we

  1. Constructed an empty column with the header value passed by the user

  2. Looped through the coefficients, their names, their standard errors, and their p-values. On each iteration, we:

    a. Insert the coefficient value and its standard error into the table

    b. Check whether the p-values fall below the desired threshold (in descending order), and if so, call the function star!(x::FormattedNumber, num_stars) with the desired number of stars.

TexTables stores all of the table values internally with a FormattedNumber type, which contains the value, the standard error if appropriate, the number of stars the value should display, and a formatting string. As a result, it is probably easiest to set the table value first, and then add stars later with the star! function. However, we could also have constructed each value directly as:

if .05 < pval <= .1
    coef_block[name] = val, se, 1
elseif 0.01 < pval <= .05
    coef_block[name] = val, se, 2
elseif pval <= .01
    coef_block[name] = val, se, 3

How you choose to do it is mostly a matter of taste and coding style. Note that by default, the number of stars is always set to zero. In other words, TexTables will not assume that it can infer the number of significance stars from the standard errors and the coefficients alone. If you want to annotate your table with significance stars, you must explicitly choose in your model-specific code which entries to annotate and how many stars they should have.