fastreg 1.2

Fast sparse regressions

Fast sparse regressions with advanced formula syntax. Good for high-dimensional fixed effects. Installation and usage are described below. Detailed documentation can be found further down.

New: generalized linear models and maximum likelihood estimation with JAX.

Install

To install from PyPI with pip:

pip install fastreg

To install directly from GitHub:

pip install git+https://github.com/iamlemec/fastreg

Alternatively, you can clone this repository locally and run

pip install -e .

Optionally, for the maximum likelihood routines, you'll need jax (and jaxlib) as well. See here for detailed instructions.

Usage

First import the necessary functions

import fastreg as fr
from fastreg import I, R, C

Create some testing data

data = fr.dataset(N=100_000, K1=10, K2=100, models=['linear', 'poisson'])

	y0	y	x1	x2	id1	id2
0	0.140	3.450	-0.260	0.958	E	37
1	-0.552	0.955	0.334	-1.046	I	65
2	-0.683	1.517	0.067	-0.631	I	10
...

We can construct formulas to define our specification. To make a real Factor on x1, use R('x1') or more conveniently R.x1. These can then be combined into Terms with * and then into Formulas with +. Regress y0 on 1, x1, and x2 given pandas DataFrame data:

fr.ols(y=R.y0, x=I+R.x1+R.x2, data=data)

	coeff	stderr	low95	high95	pvalue
I	0.099	0.003	0.093	0.105	0.000
x1	0.304	0.003	0.297	0.310	0.000
x2	0.603	0.003	0.597	0.609	0.000

Regress y on 1, x1, x2, categorical id1, and categorical id2:

fr.ols(y=R.y, x=I+R.x1+R.x2+C.id1+C.id2, data=data)

	coeff	stderr	low95	high95	pvalue
I	0.153	0.033	0.088	0.218	0.000
x1	0.295	0.003	0.289	0.302	0.000
x2	0.594	0.003	0.588	0.600	0.000
id1=B	0.072	0.014	0.044	0.099	0.000
id1=C	0.168	0.014	0.140	0.195	0.000
...

Regress y on 1, x1, x2, and all combinations of categoricals id1 and id2 (Note that * is analogous to : in R-style syntax):

fr.ols(y=R.y, x=I+R.x1+R.x2+C.id1*C.id2, data=data)

	coeff	stderr	low95	high95	pvalue
I	0.158	0.107	-0.051	0.368	0.138
x1	0.295	0.003	0.289	0.301	0.000
x2	0.593	0.003	0.587	0.599	0.000
id1=A,id2=1	-0.068	0.144	-0.350	0.213	0.634
id1=A,id2=2	0.060	0.155	-0.244	0.363	0.700
...

Instead of passing y and x, you can also pass an R-style formula string to formula, as in:

fr.ols(formula='y ~ 1 + x1 + x2 + C(id1):C(id2)', data=data)

There's even a third intermediate option using lists and tuples, which might be more useful when you are defining specifications programmatically:

fr.ols(y=R.y, x=[I, R.x1, R.x2, (C.id1, C.id2)], data=data)

Right now, categorical coding schemes other than treatment are not supported. You can pass a list of column names to cluster to cluster standard errors on those variables.

Categorical coding

For categorical variables, one must avoid collinearity by either not including an intercept term or by dropping one value. The default for categorical variables is to drop the first value in alphabetical/numerical order. You can specify which value to drop by passing that as an argument to the specified variable. For instance, if one wanted to drop B from the factor id1, they would write C.id1('B') or equivalently C('id1', 'B'), or more verbosely C('id1', drop='B'). You can also tell it to not drop any values by passing fr.NONE and explicitly tell it to drop the first value with fr.FIRST.

In the case of interacted categorical variables, you would typically specify the dropped value for each factor and this will be inherited to the term level. For instance, if one wished to drop id1 = B and id2 = 3 from the interaction of these two terms, they would write C.id1('B')*C.id2(3). An alternative method would be to write (C.id1*C.id2).drop('B', 3). When creating compound categorical terms, an attempt is made to find the correct drop strategy. In the case of ambiguity or when no information is given, the default is again FIRST. When interacting categorical and real variables, the default is NONE, as this source of collinearity is no longer an issue.

High dimensional

Point estimates are obtained efficiently by using a sparse array representation of categorical variables. However, computing standard errors can be costly due to the need for large, dense matrix inversion. It is possible to make clever use of block diagonal properties to quickly compute standard errors for the case of a single (possibly interacted) categorical variable. In this case, we can recover the individual standard errors, but not the full covariance matrix. To employ this, pass a single Term (such as C.id1 or C.id1*C.id2) with the hdfe flag, as in

fr.ols(y='y', x=I+R.x1+R.x2+C.id1, hdfe=C.id2, data=data)

You can also pass a term to the absorb flag to absorb those variables a la Stata's areg. In this case you do not recover the standard errors for the absorbed categorical, though it may be faster in the case of multiple high-dimensional regressors. This will automatically cluster standard errors on that term as well, as the errors will in fact be correlated, even if the original data was iid.

Generalized linear models

We can do GLM now too! The syntax and usage is identical to that of ols. For instance, to run a properly specified Poisson regression using our test data:

fr.poisson(y=R.p, x=I+R.x1+R.x2+C.id1+C.id2, data=data)

	coeff	stderr	low95	high95	pvalue
I	0.322	0.011	0.300	0.344	0.000
x1	0.294	0.001	0.293	0.296	0.000
x2	0.597	0.001	0.596	0.599	0.000
id1=B	0.072	0.005	0.062	0.081	0.000
id1=C	0.178	0.005	0.169	0.187	0.000
...

You can use the hdfe flag here as well, for instance:

fr.poisson(y=R.p, x=I+R.x1+R.x2+C.id1, hdfe=C.id2, data=data)

Under the hood, this is all powered by a maximum likelihood estimation routine in general.py called maxlike_panel. Just give this a function that computes the mean log likelihood and it'll take care of the rest, computing standard errors from the inverse of the Fisher information matrix. This is then specialized into a generalized linear model routine called glm, which accepts a loss function along with data. I've provided implementations for logit, poisson, negbin, zinf_poisson, zinf_negbin, and gols.

Custom factors

The algebraic system used to define specifications is highly customizable. First, there are the core factors I (identity), R (real), and C (categorical). Then there are the provided factors D (demean) and B (binned). You can also create your own custom column types. The simplest way is using the factor function decorator. For instance, we might want to standardize variables:

@fr.factor
def Z(x):
    return (x-np.mean(x))/np.std(x)

The we can using this in a regression with either Z('x1') or Z.x1, as in:

fr.ols(y=R.y0, x=I+Z.x1+Z.x2, data=data)

We may also want factors that use data from multiple columns. In this case we need to use eval_args to tell it what expressions to map, as it defaults to only the first argument (0). For example, to implement conditional demean (which is also included by default as fr.D), we would do:

@fr.factor(eval_args=(0, 1))
def CD(x, i):
    datf = pd.DataFrame({'vals': x, 'cond': i})
    cmean = datf.groupby('cond')['vals'].mean().rename('mean')
    datf = datf.join(cmean, on='cond')
    return datf['vals'] - datf['mean']

and then use it in a regression, though we can't use the convenience syntax with multiple arguments

fr.ols(y=R.y0, x=I+CD('x1','id1')+CD('x2','id2'), data=data)

The factor decorator also accepts a categ flag that you can set to True for categorical variables. Finally, it may be useful to inject functions into the evaluation namespace rather than create a whole new factor type. To do this, you can pass a dict to the extern flag and prefix the desired variable or function with @, as in:

extern = {'logit': lambda x: 1/(1+np.exp(-x))}
fr.ols(y='y0', x=I+R('@logit(x1)')+R.x2, data=data, extern=extern)

Documentation

The core functionality of this library lies in creating well-structured data matrices (often called "design matrices") from actual data in the for of Pandas DataFrames and a regression specification, either Fastreg-style or R-style. For that, we have the following function defined in formula.py. You must always pass data as well as either y/x or formula.

fastreg.design_matrices(
    y=None, x=None, formula=None, data=None, dropna=True, prune=True, validate=False,
    flatten=True, extern=None, warn=True
)

• y: specification for the outcome variable, a column name (str) or a single Term, which might be the combination of multiple Factors
• x: specification for the input variables, a Formula or list of Terms
• formula: an R-style specification string, this will override any y or x given above
• data: a DataFrame with the underlying dataset
• dropna: drop any rows containing null data
• prune: prune categories that have no instances
• validate: return binary mask specifying which rows were dropped
• flatten: combine dense and sparse x variables into one matrix
• extern: a dictionary of functions for use in specification
• warn: output info on dropped rows or categories

This returns (data, name) pairs for both y and x variables. In addition, if you only want to deal with the x variables, you can use design_matrix, which has nearly identical syntax but does not accept the y argument. Next is the ols function defined in linear.py that handles regressions.

fastreg.ols(
    y=None, x=None, formula=None, data=None, cluster=None, absorb=None, hdfe=None,
    stderr=True, output='table'
)

• y: specification for the outcome variable, a column name (str) or a single Term, which might be the combination of multiple Factors
• x: specification for the input variables, a Formula or list of Terms
• formula: an R-style specification string, this will override any y or x given above
• data: a DataFrame with the underlying dataset
• cluster: cluster standard errors on the given Term
• absorb: regress on differences within groups specified by given Term
• hdfe: use block inversion to speed up standard error calculation for given Term
• stderr: standard error type, True for basic, and hc0-hc3 for robust types
• output: control output, table gives DataFrame of estimates, dict gives much more info

Other estimators use syntax very similar to that of ols. This includes glm in general.py, which also accepts custom a loss functions. For instance, the built-in poisson uses a Poisson likelihood loss function (with an exponential link). Below only the arguments not common to ols are listed.

fastreg.glm(
    y=None, x=None, formula=None, data=None, hdfe=None, loss=None, model=None,
    extra={}, raw={}, offset=None, epochs=None, display=True, per=None, stderr=True,
    output='table'
)

• loss: the loss (log likelihood) function to use for optimization, can be one of 'logit', 'poisson', 'negbin', 'normal', 'lognorm', 'lstsq', or a custom function that accepts (params, data, yhat, y)
• model: in lieu of a loss function, one can specify a model function mapping from (params, data) to an average log likelihood
• extra: a dict of extra parameter names mapping to initial values that can be accessed by the loss function
• raw: a dict of extra Term specifications that are evaluated and passed to the loss function as part of data
• offset: a Term to evaluate and add to the linear predictor (for instance, R('log(t)'))
• epochs: how many full iterations over the dataset to do during optimization
• display: whether to display updates on objective and parameter values during optimization
• per: how often to display updates during optimization