Awesome

Overview

dbt_linreg is an easy way to perform linear regression and ridge regression in SQL (Snowflake, DuckDB, and more) with OLS using dbt's Jinja2 templating.

Reasons to use dbt_linreg:

• 📈 Linear regression in pure SQL: With the power of Jinja2 templating and some behind-the-scenes math tricks, it is possible to implement multiple and multivariate regression in pure SQL. Most SQL engines (even OLAP engines) do not have a multiple regression implementation of OLS, so this fills a valuable niche. dbt_linreg implements true OLS, not an approximation!
• 📱 Simple interface: Just define a table= (which works with ref(), source(), and CTEs), a y-variable with endog=, your x-variables in a list with exog=..., and you're all set! Note that the API is loosely based on Statsmodels's naming conventions.
• 🤖 Support for ridge regression: Just pass in alpha=scalar or alpha=[scalar1, scalar2, ...] to regularize your regressions. (Note: regressors are not automatically standardized.)
• 🤸‍ Flexibility: Tons of formatting options available to return coefficients the way you want.
• 💪 Durable and tested: The API provides feedback on parsing errors, and everything in this code base has been tested (check the continuous integration).

Installation

dbt-core >=1.2.0 is required to install dbt_linreg.

Add this the packages: list your dbt project's packages.yml:

  - package: "dwreeves/dbt_linreg"
    version: "0.2.5"

The full file will look something like this:

packages:
  # ...
  # Other packages here
  # ...
  - package: "dwreeves/dbt_linreg"
    version: "0.2.5"

Examples

Simple example

The following example runs a linear regression of 3 columns xa + xb + xc on y, using data in the dbt model named simple_matrix. It outputs the data in "long" format, and rounds the coefficients to 5 decimal points:

{{
  config(
    materialized="table"
  )
}}
select * from {{
  dbt_linreg.ols(
    table=ref('simple_matrix'),
    endog='y',
    exog=['xa', 'xb', 'xc'],
    format='long',
    format_options={'round': 5}
  )
}} as linreg

Output:

Note: simple_matrix is one of the test cases, so you can try this yourself! Standard errors are constant across xa, xb, xc, because simple_matrix is orthonormal.

Complex example

The following hypothetical example shows multiple ridge regressions (one per product_id) on a table that is preprocessed substantially. After the fact, predictions are run, and the R-squared of each regression is calculated at the end.

This example shows that, although dbt_linreg does not implement everything for you, the OLS implementation does most of the hard work. This gives you the freedom to do things you've never been able to do before in SQL!

{{
  config(
    materialized="table"
  )
}}
with

preprocessed_data as (

  select
    product_id,
    price,
    log(price) as log_price,
    epoch(time) as t,
    sin(epoch(time)*pi()*2 / (60*60*24*365)) as sin_t,
    cos(epoch(time)*pi()*2 / (60*60*24*365)) as cos_t
  from
    {{ ref('prices') }}

),

preprocessed_and_normalized_data as (

  select
    product_id,
    price,
    log(price) as log_price,
    (time - avg(time) over ()) / (stddev(time) over ()) as t_norm,
    (sin_t - avg(sin_t) over ()) / (stddev(sin_t) over ()) as sin_t_norm,
    (cos_t - avg(cos_t) over ()) / (stddev(cos_t) over ()) as cos_t_norm
  from
    preprocessed_data

),

coefficients as (

    select * from {{
      dbt_linreg.ols(
        table='preprocessed_and_normalized_data',
        endog='log_price',
        exog=['t_norm', 'sin_t_norm', 'cos_t_norm'],
        group_by=['product_id'],
        alpha=0.0001
      )
    }}

),

predict as (

  select
    d.product_id,
    d.time,
    d.price,
    exp(
      c.const
      + d.t_norm * c.t_norm
      + d.sin_t_norm * c.sin_t_norm
      + d.cos_t_norm * sin_t_norm) as predicted_price
  from
    preprocessed_and_normalized_data as d
  join
    coefficients as c
  on
    d.product_id = c.product_id

)

select
  product_id,
  pow(corr(predicted_price, price), 2) as r_squared
from
  predict
group by
  product_id

Supported Databases

dbt_linreg should work with most SQL databases, but so far, testing has been done for the following database tools:

• Snowflake
• DuckDB
• Postgres*

If dbt_linreg does not work in your database tool, please let me know in a bug report and I can make sure it is supported.

* Minimal support. Postgres is syntactically supported, but is not performant under certain circumstances.

API

The only function available in the public API is the dbt_linreg.ols() macro.

Using Python typing notation, the full API for dbt_linreg.ols() looks like this:

def ols(
    table: str,
    endog: str,
    exog: Union[str, list[str]],
    add_constant: bool = True,
    format: Literal['wide', 'long'] = 'wide',
    format_options: Optional[dict[str, Any]] = None,
    group_by: Optional[Union[str, list[str]]] = None,
    alpha: Optional[Union[float, list[float]]] = None,
    method: Literal['chol', 'fwl'] = 'chol',
    method_options: Optional[dict[str, Any]] = None
):
    ...

Where:

• table: Name of table or CTE to pull the data from. You can use ref() or source() here if you'd like.
• endog: The endogenous variable / y variable / target variable of the regression. (You can also specify y=... instead of endog=... if you prefer.)
• exog: The exogenous variables / X variables / features of the regression. (You can also specify x=... instead of exog=... if you prefer.)
• add_constant: If true, a constant term is added automatically to the regression.
• format: Either "wide" or "long" format for coefficients. See Formats and format options for more.
  • If wide, the variables span the columns with their original variable names, and the coefficients fill a single row.
  • If long, the coefficients are in a single column called coefficient, and the variable names are in a single column called variable_name.
• format_options: See Formats and format options section for more.
• group_by: If specified, the regression will be grouped by these variables, and individual regressions will run on each group.
• alpha: If not null, the regression will be run as a ridge regression with a penalty of alpha. See Notes section for more information.
• method: The method used to calculate the regression. See Methods and method options for more.
• method_options: Options specific to the estimation method. See Methods and method options for more.

Formats and format options

Outputs can be returned either in format='long' or format='wide'.

(In the future, I might add one or two more formats, notably a summary table format.)

All formats have their own format options, which can be passed into the format_options= arg as a dict, e.g. format_options={'foo': 'bar'}.

Options for format='long'

• round (default = None): If not None, round all coefficients to round number of digits.
• constant_name (default = 'const'): String name that refers to constant term.
• variable_column_name (default = 'variable_name'): Column name storing strings of variable names.
• coefficient_column_name (default = 'coefficient'): Column name storing model coefficients.
• strip_quotes (default = True): If true, strip outer quotes from column names if provided; if false, always use string literals.

These options are available for format='long' only when method='chol':

• calculate_standard_error (default = True if not alpha else False): If true, provide the standard error in the output.
• standard_error_column_name (default = 'standard_error'): Column name storing the standard error for the parameter.
• t_statistic_column_name (default = 't_statistic'): Column name storing the t-statistic for the parameter.

Options for format='wide'

• round (default = None): If not None, round all coefficients to round number of digits.
• constant_name (default = 'const'): String name that refers to constant term.
• variable_column_prefix (default = None): If not None, prefix all variable columns with this. (Does NOT delimit, so make sure to include your own underscore if you'd want that.)
• variable_column_suffix (default = None): If not None, suffix all variable columns with this. (Does NOT delimit, so make sure to include your own underscore if you'd want that.)

Methods and method options

There are currently two valid methods for calculating regression coefficients:

• chol: Uses Cholesky decomposition to calculate the pseudo-inverse.
• fwl: Uses a "Frisch-Waugh-Lovell" approach, which consists of calculating univariate regressions to get multiple regression coefficients.

chol method

👍 This is the suggested method (and the default) for calculating regressions!

This method calculates regression coefficients using the Moore-Penrose pseudo-inverse, and the inverse of X'X is calculated using Cholesky decomposition, hence it is referred to as chol.

Options for method='chol'

Specify these in a dict using the method_options= kwarg:

• safe (default = True): If True, returns null coefficients instead of an error when X is perfectly multicollinear. If False, a negative value will be passed into a SQRT(), and most SQL engines will raise an error when this happens.
• subquery_optimization (default: True): If True, nested subqueries are used during some of the steps to optimize the query speed. If false, the query is flattened.

fwl method

This method is generally not recommended.

Simple univariate regression coefficients are simply covar_pop(y, x) / var_pop(x).

The multiple regression implementation uses a technique described in section 3.2.3 Multiple Regression from Simple Univariate Regression of TEoSL (source). Econometricians know this as the Frisch-Waugh-Lovell theorem, hence the method is referred to as fwl internally in the code base.

Ridge regression is implemented using the augmentation technique described in Exercise 12 of Chapter 3 of TEoSL (source).

There are a few reasons why this method is discouraged over the chol method:

• 🐌 It tends to be much slower in OLAP systems, and struggles to efficiently calculate large number of columns.
• 📊 It does not calculate standard errors.
• 😕 For ridge regression, coefficients are not accurate; they tend to be off by a magnitude of ~0.01%.
• ⚠️ It does not work in all databases because it relies on COVAR_POP.

So when should you use fwl? The main use case is in OLTP systems (e.g. Postgres) for unregularized coefficient estimation. Long story short, the chol method relies on subquery optimization to be more performant than fwl; however, OLTP systems do not benefit at all from subquery optimization. This means that fwl is slightly more performant in this context.

Notes

• ⚠️ If your coefficients are null, it does not mean dbt_linreg is broken, it most likely means your feature columns are perfectly multicollinear. If you are 100% sure that is not the issue, please file a bug report with a minimally reproducible example.

• Regularization is implemented using nearly the same approach as Statsmodels; the only difference is that the constant term can never be regularized. This means:

  • A scalar input (e.g. alpha=0.01) will apply an alpha of 0.01 to all features.
  • An array input (e.g. alpha=[0.01, 0.02, 0.03, 0.04, 0.05]) will apply an alpha of 0.01 to the first column, 0.02 to the second column, etc.
  • alpha is equivalent to what TEoSL refers to as "lambda," times the sample size N. That is to say: α ≡ λ * N.

• Regularization as currently implemented for the chol method tends to be very slow in OLTP systems (e.g. Postgres), but is very performant in OLAP systems (e.g. Snowflake, DuckDB, BigQuery, Redshift). As dbt is more commonly used in OLAP contexts, the code base is optimized for the OLAP use case.

  • That said, it may be possible to make regularization in OLTP more performant (e.g. with augmentation of the design matrix), so PRs are welcome.

• Regression coefficients in Postgres are always numeric types.

Possible future features

Some things I am thinking about working on down the line:

• Optimization: Given access to Jinja2 templating, there may be more efficient ways to calculate the get a closed form OLS solution than the approach taken in this code base.

• Standard errors and t-stats: For the format='long' output (or perhaps a new format?), there is space to sensibly add t-stats and standard errors. The main challenge is that this necessitates inverting a covariance matrix, although this is theoretically doable using Jinja2 templating.

FAQ

How does this work?

See Methods and method options section for a full breakdown of each linear regression implementation.

All approaches were validated using Statsmodels sm.OLS(). Note that the ridge regression coefficients differ very slightly from Statsmodels's outputs for currently unknown reasons, but the coefficients are very close (I enforce a <0.01% deviation from Statsmodels's ridge regression coefficients in my integration tests).

BigQuery (or other database) has linear regression implemented natively. Why should I use dbt_linreg over that?

You don't have to use this. Most warehouses don't support multiple regression out of the box, so this satisfies a niche for those database tools which don't.

That said, even in BigQuery, it may be very useful to extract coefficients within a query instead of generating a separate MODEL object through a DDL statement, for a few reasons. Even in more black box predictive contexts, being able to predict in the same SELECT statement as training can be convenient. Additionally, BigQuery does not expose model coefficients to users, and this can be a dealbreaker in many contexts where you care about your coefficients as measurements, not as predictive model parameters. Lastly, group_by is akin to estimating parameters for multiple linear regressions at once.

Overall, I would say this is pretty different from what BigQuery's CREATE MODEL is doing; use whatever makes sense for your use case! But keep in mind that for large numbers of variables, a native implementation of linear regression will be noticeably more efficient than this implementation.

Why is L2 regularization / ridge regression supported, but not L1 regularization / LASSO supported?

There is no closed-form solution to L1 regularization, which makes it very very hard to add through raw SQL. L2 regularization has a closed-form solution and can be implemented using a pre-processing trick.

Is the group_by=[...] argument like categorical variables / one-hot encodings?

No. You should think of the group by more as a seemingly unrelated regressions implementation than as a categorical variable implementation. It's running multiple regressions and each individual partition is its own y vector and X matrix. This is not a replacement for dummy variables.

Why aren't categorical variables / one-hot encodings supported?

I opt to leave out dummy variable support because it's tricky, and I want to keep the API clean and mull on how to best implement that at the highest level.

Note that you couldn't simply add categorical variables in the same list as numeric variables because Jinja2 templating is not natively aware of the types you're feeding through it, nor does Jinja2 know the values that a string variable can take on. The way you would actually implement categorical variables is with group by trickery (i.e. center both y and X by categorical variable group means), although I am not sure how to do that efficiently for more than one categorical variable column.

If you'd like to regress on a categorical variable, for now you'll need to do your own feature engineering, e.g. (foo = 'bar')::int as foo_bar

Why are there no p-values?

This is planned for the future, so stay tuned! P-values would require a lookup on a dimension table, which is a significant amount of work to manage nicely, but I hope to get to it soon.

In the meanwhile, you can implement this yourself-- just create a dimension table that left joins a t-statistic on a half-open interval to lookup a p-value.

Trademark & Copyright

dbt is a trademark of dbt Labs.

This package is unaffiliated with dbt Labs.