Covariate shift adjustment
Carlos Gonzalez Poses & Thom Benjamin Volker
Introduction
The densityratio package offers user-friendly and
efficient software for density ratio estimation. The package contains
multiple methods for estimating density ratios, aiming to estimate the
ratio of the probability density functions underlying two groups of
observations. This density ratio is estimated using a one-shot
procedure, without estimating the separate densities. The package
includes different estimation techniques with built-in hyper-parameter
tuning and easy-to-use helper functions to summarize and visualize the
results.
In this getting-started vignette, we guide the user
through the main functionalities of the package. We start by explaining
the concept of density ratio estimation, and the different methods
available in the package. We then show how to estimate density ratios
using the densityratio package for an example case of
distribution shift (i.e., sample selection bias) adaptation.
Density ratio estimation
The goal of density ratio estimation is to estimate the ratio of two
probability density functions, \[
r(x) = \frac{p_\text{nu}(x)}{p_\text{de}(x)},
\] based on samples from the numerator (\(p_\text{nu}\)) and denominator (\(p_\text{de}\)) distributions. If the
numerator and denominator samples are similar (i.e., they are samples
from a common distribution), the density ratio will be close to one over
the (multivariate) space of the data. If the numerator and denominator
samples are different, the density ratio will be different from one in
those regions where either numerator or denominator samples are
overrepresented. That is, the density ratio will be large in regions
where the numerator samples have high density but denominator samples
have not, and will be small in regions where the denominator samples
have high density but numerator samples have not. Hence, the estimated
density ratio directly shows where and how samples from two
distributions differ.
Data
Throughout the vignette, we illustrate the functionality using the
insurance dataset. The dataset is publicly available on
Kaggle (e.g., here)
and is included as example data in the densityratio
package. The data contain 1338 observations on the following seven
variables:
age: age of the insured (continuous)sex: sex of the insured (old-fashioned binary
variable)bmi: body mass index of the insured (continuous)children: number of children/dependents covered by the
insurance (integer)smoker: whether the insured is a smoker (binary
variable)region: the region of the insured (categorical variable
with categories northwest, northeast,
southwest and southeast)charges: the medical costs billed by the insurance
(continuous)
In this vignette, we first discuss the mechanics of the density ratio
package for distribution comparison, and then show how it can be used in
covariate shift problems. For these examples, we create training and
test data under two settings in which we train a prediction model on the
training data to predict insurance charges in the test data. In the
first setting, the test data is randomly drawn from the insurance data,
and the remaining cases are used for training (this setting is denoted
rs throughout, which is short for random sample). In the
second setting (denoted cs for
covariate shift), we split the data based on the region
variable, such that the training data contains only cases from the
southern regions, while the test data contains cases from the northern
regions. In the second setting, covariate shift may be present, as the
different regions may have different distributions on the covariates
used in the prediction models. Importantly, we assume that the
conditional distribution of the insurance charges given the covariates
is the same in both regions. This assumption may not hold if the regions
also differ on unobserved confounders, but to keep the vignette focused,
we ignore this issue here. We then train a prediction model on the
southern regions and use it to predict insurance charges in the northern
regions.
library(densityratio)
set.seed(123)
south <- insurance$region %in% c("southwest", "southeast")
data <- subset(insurance, select = -region)
trainidx_rs <- sample(south)
train_rs <- subset(data, subset = trainidx_rs)
test_rs <- subset(data, subset = !trainidx_rs)
train_cs <- subset(data, subset = south)
test_cs <- subset(data, subset = !south)
Estimating the density ratio
The densityratio package includes multiple methods for
estimating the density ratio that each has its own estimation function
(e.g., unconstrained least-squares importance fitting
ulsif(), Kullback-Leibler importance estimation procedure
kliep(), spectral density ratio estimation
spectral(), and so on). Each method corresponds to a
slightly different model or estimation routine, although all methods use
non-parametric estimation techniques based on Gaussian kernel
transformations of the input data (with bandwidth parameter \(\sigma\) that is automatically tuned
through cross-validation). Moreover, all functions operate similarly.
The user provides two data sets, containing the numerator and
denominator samples, and the function returns a
densityratio object, which can be used and examined. These
densityratio objects can always be further inspected with
corresponding summary() and plot()
methods.
In what follows, we use unconstrained least-squares importance
fitting (ulsif) to estimate the density ratio, as it is fast and
efficient. We remove the outcome variable charges from the
data before passing it to the ulsif() function, because we
typically don’t have access to the outcome variable in the testing
setting. Hence, we do not want to reweigh samples based on the outcome
variable.
dr_rs <- ulsif(test_rs |> subset(select = -charges),
train_rs |> subset(select = -charges))
dr_cs <- ulsif(test_cs |> subset(select = -charges),
train_cs |> subset(select = -charges))
The resulting ulsif objects contain the estimated
density ratio, but also the optimal parameters that can be used to
predict the density ratio for new data.
dr_rs
#>
#> Call:
#> ulsif(df_numerator = subset(test_rs, select = -charges), df_denominator = subset(train_rs, select = -charges))
#>
#> Kernel Information:
#> Kernel type: Gaussian with L2 norm distances
#> Number of kernels: 200
#> sigma: num [1:10] 0.813 1.185 1.366 1.528 1.678 ...
#>
#> Regularization parameter (lambda): num [1:20] 1000 483.3 233.6 112.9 54.6 ...
#>
#> Optimal sigma (loocv): 2.802938
#> Optimal lambda (loocv): 2.976351
#> Optimal kernel weights (loocv): num [1:201] 0.01462 0.01017 0.00757 0.00554 0.00745 ...
#>
The output shows the number of inducing points used in the estimation
(number of kernels), and the grid of bandwidth (sigma) and
regularization parameters (lambda) used in the cross-validation.
Moreover, the output shows the optimal bandwidth parameter, the optimal
regularization parameter lambda and the optimal kernel weights, all
optimized through leave-one-out cross-validation.
dr_cs
#>
#> Call:
#> ulsif(df_numerator = subset(test_cs, select = -charges), df_denominator = subset(train_cs, select = -charges))
#>
#> Kernel Information:
#> Kernel type: Gaussian with L2 norm distances
#> Number of kernels: 200
#> sigma: num [1:10] 0.856 1.225 1.416 1.59 1.753 ...
#>
#> Regularization parameter (lambda): num [1:20] 1000 483.3 233.6 112.9 54.6 ...
#>
#> Optimal sigma (loocv): 1.224742
#> Optimal lambda (loocv): 0.3359818
#> Optimal kernel weights (loocv): num [1:201] 0.27465 0.0706 0.03061 -0.00883 0.01246 ...
#>
Comparing the two density ratio objects shows that under covariate
shift, the optimal bandwidth parameter and regularization parameter are
somewhat smaller than under random sampling, which implies that the
model is more sensitive to smaller deviations between the two
distributions.
One can obtain additional information on the density ratio object by
calling the summary() function. This function again shows
the optimal hyper-parameters, and adds information on the divergence
between the two distributions. Note, here, that different estimation
functions typically use different divergence measures. However, for each
estimation function and corresponding divergence measure, one may
formally test whether the two distributions are equal using a
permutation test. This can be done by setting the
test = TRUE argument in the summary()
function.
summary(dr_rs, test = TRUE)
#>
#> Call:
#> ulsif(df_numerator = subset(test_rs, select = -charges), df_denominator = subset(train_rs, select = -charges))
#>
#> Kernel Information:
#> Kernel type: Gaussian with L2 norm distances
#> Number of kernels: 200
#>
#> Optimal sigma: 2.802938
#> Optimal lambda: 2.976351
#> Optimal kernel weights: num [1:201] 0.01462 0.01017 0.00757 0.00554 0.00745 ...
#>
#> Pearson divergence between P(nu) and P(de): 0.02298
#> Pr(P(nu)=P(de)) = 0.28
#> Bonferroni-corrected for testing with r(x) = P(nu)/P(de) AND r*(x) = P(de)/P(nu).
summary(dr_cs, test = TRUE)
#>
#> Call:
#> ulsif(df_numerator = subset(test_cs, select = -charges), df_denominator = subset(train_cs, select = -charges))
#>
#> Kernel Information:
#> Kernel type: Gaussian with L2 norm distances
#> Number of kernels: 200
#>
#> Optimal sigma: 1.224742
#> Optimal lambda: 0.3359818
#> Optimal kernel weights: num [1:201] 0.27465 0.0706 0.03061 -0.00883 0.01246 ...
#>
#> Pearson divergence between P(nu) and P(de): 0.1221
#> Pr(P(nu)=P(de)) < .001
#> Bonferroni-corrected for testing with r(x) = P(nu)/P(de) AND r*(x) = P(de)/P(nu).
These summaries indeed show that the two distributions are not
significantly different under random sampling, but differ significantly
under our covariate shift setting. Performing the permutation test might
take some time, but it can be sped up by parallelizing the computation
(by setting parallel = TRUE).
Visualization methods
The densityratio package contains several visualization
methods that may aid in interpreting the estimated density ratio. These
plots functions build on the ggplot2 library, and return a
customizable ggplot object. The default plot()
method produces a histogram of the log density ratio values for both
groups of samples (numerator and denominator). The plot()
function takes several arguments that allow to customize the plot, such
as the samples argument, which allows to only plot
numerator or denominator samples, the logscale argument to
indicate whether to plot the log density ratio values, and so on (run
?plot.ulsif for more information). The default
plot() summarizes the similarity of the samples through the
density ratio. When the numerator and denominator samples are drawn from
the same distribution, the density ratio should be close to 1 (and close
to 0 on a log-scale). Moreover, the distribution of the density ratio
values should be very similar for both the numerator and denominator
samples.
plot(dr_rs)
#> `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
Distribution of density ratio values under
random sampling.
In cases where the numerator and denominator samples have different
densities, the density ratio values will be different from 1, and more
importantly, the distributions of density ratio values over the two
groups will overlap to a lesser extent. Hence, numerator samples will
typically have estimated density ratio values larger than 1, while
denominator samples will have density ratio values smaller than 1 (or,
on a logarithmic scale, larger and smaller than zero, respectively).
This is exactly what we see in the figure below: numerator samples
(i.e., the test cases) typically have larger estimated density ratio
values than the denominator (training) samples, although the difference
is not that large here (but note the different scale of the x-axis
compared to the previous plot). ´
plot(dr_cs)
#> `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
Distribution of density ratio values under
covariate shift.
The package comes with two additional plotting functions:
plot_univariate() and plot_bivariate(). The
plot_univariate() plots the density ratio values against
each variable separately, which can help to identify those variables
that are most informative for the density ratio. Below, we plot the
density ratio values against the variables age and
bmi (and set grid = TRUE to add the two
figures in a grid). For age there seems no trend: the density ratio
values are quite randomly distributed over the age range. For
bmi, the pattern is different. People in the numerator
sample (i.e., the test cases), seem to have lower bmi
values in general, whereas the higher bmi values occur more
often for the denominator (training) samples. This results in a higher
density ratio for lower bmi values, and a lower density
ratio for higher bmi values.
plot_univariate(dr_cs, vars = c("age", "bmi"), grid = TRUE)
Univariate plots of age and
bmi against the estimated density ratio values for the two
groups of samples under covariate shift.
The plot_bivariate() function is available to plot the
density ratio values for pairs of variables (with the density ratio
value mapped to the color scale). We can again select the variables for
which we want to create the scatterplots using vars, and
the function will create a scatterplot for each pair of variables.
Again, we solely observe a pattern for bmi, with higher
values corresponding to lower density ratio values.
plot_bivariate(dr_cs, vars = c("age", "bmi", "children"), grid = TRUE)
Bivariate scatter plots for age,
bmi and children for two groups of samples
under covariate shift, with density ratio values mapped to the color
scale.
Covariate shift
We now show how the densityratio package can be used to
deal with covariate shift. Throughout, we attempt to predict the outcome
variable charges using the predictor variables
age, sex, bmi,
children, and smoker. A common assumption of
many statistical analyses is that training and test samples are drawn
from the same distribution. If this assumption is met, least-squares
regression (and many other statistical learning techniques, such as
random forests and support vector regression) yield consistent estimates
of the model parameters, in the sense that the estimates converge to the
optimal model parameters as the sample size increases. We can
obtain estimates of the model parameters by solving the least-squares
problem \[
\hat{\beta} = \arg\min_{\beta} \sum_{i=1}^{n} (y_{\text{tr}, i} -
x_{\text{tr}, i}^T \beta)^2 =
(X_\text{tr}^TX_\text{tr})^{-1}X_\text{tr}^Ty_\text{tr},
\] where \(X_\text{tr}\) is the
design matrix of the training data, \(y_\text{tr}\) is the vector of training
outcomes, and \(\beta\) is the vector
of model parameters.
Under covariate shift, this estimator must no longer be consistent.
In fact, it only is consistent if the analysis model is correctly
specified. This problem can be compensated by reweighing the training
samples using the estimated density ratio, which yields a consistent
estimator under the true density ratio model \[
r(x) = \frac{p_\text{test}(x)}{p_\text{train}(x)}.
\] Since the true density ratio model is not known, all we can do
is to estimate it from the training and test data. Consequently, we have
\[
\hat\beta_\text{R} = \arg\min_{\beta} \sum_{i=1}^{n}
\hat{r}_{\text{tr},i} (y_{\text{tr}, i} - x_{\text{tr}, i}^T \beta)^2 =
(X_\text{tr}^T \hat{R} X_\text{tr})^{-1}X_\text{tr}^T \hat{R}
y_\text{tr},
\] where the density ratio weights \(\hat{r}_{\text{tr},i}\) (collected in the
diagonal matrix \(\hat R\)) are
obtained from the estimated density ratio model.
Below, we compare the performance of least squares regression with
and without reweighing on the training and test data. For the sake of
illustration, we do this for data under random sampling, and for our
covariate shift setting. We first fit a least squares regression model
to the training data, and then predict the outcome variable for the test
data, and evaluate the root mean squared error (RMSE) and (\(R^2\)).
Performance under covariate shift
We repeat exactly the same procedure as in the previous section, but
now for the setting with covariate shift between training and test
data.
fit_cs <- lm(formula, data = train_cs)
We again predict the outcome variable for the training and test data,
and evaluate the RMSE and \(R^2\).
y_train_cs <- train_cs$charges
y_test_cs <- test_cs$charges
pred_train_cs <- predict(fit_cs, newdata = train_cs)
pred_test_cs <- predict(fit_cs, newdata = test_cs)
c(
RMSE_train = sqrt(mean((y_train_cs - pred_train_cs)^2)),
RMSE_test = sqrt(mean((y_test_cs - pred_test_cs)^2)),
R2_tr = 1 - var(y_train_cs - pred_train_cs) / var(y_train_cs),
R2_te = 1 - var(y_test_cs - pred_test_cs) / var(y_test_cs)
)
#> RMSE_train RMSE_test R2_tr R2_te
#> 5860.0592523 6434.1388040 0.7944538 0.6722150
Now, the performance of the prediction model decreases substantially
on the test data, as is expected under covariate shift. Hopefully, we
can account for this by reweighing the training data.
weights_cs <- predict(dr_cs, newdata = train_cs |> subset(select = -charges))
fit_cs_w <- lm(formula, data = train_cs, weights = weights_cs)
pred_train_cs_w <- predict(fit_cs_w, newdata = train_cs)
pred_test_cs_w <- predict(fit_cs_w, newdata = test_cs)
c(
RMSE_train = sqrt(mean((y_train_cs - pred_train_cs_w)^2)),
RMSE_test = sqrt(mean((y_test_cs - pred_test_cs_w)^2)),
R2_tr = 1 - var(y_train_cs - pred_train_cs_w) / var(y_train_cs),
R2_te = 1 - var(y_test_cs - pred_test_cs_w) / var(y_test_cs)
)
#> RMSE_train RMSE_test R2_tr R2_te
#> 6005.6703347 6191.6472513 0.7841192 0.6971867
We see that we get a modest improvement in the RMSE and \(R^2\) values for the test data, compared to
the unweighted solution. The fact that the improvement is not larger can
have multiple reasons. First, of course, the estimated density ratio
model is not perfect, which may lower the performance of the reweighted
regression model. Second, we have no way of knowing whether our
important assumption, namely that only the covariate distribution of
train and test set changed, is actually met. There could very well be
unobserved processes that affect the selection mechanism and the
outcome. Finally, it could very well be that the performance under
random sampling is not even attainable, because, for example, the
distribution of the outcome is different in the two settings.
lm(formula, data = test_cs) |>
summary()
#>
#> Call:
#> lm(formula = formula, data = test_cs)
#>
#> Residuals:
#> Min 1Q Median 3Q Max
#> -8867 -3221 -1166 1187 27132
#>
#> Coefficients:
#> Estimate Std. Error t value Pr(>|t|)
#> (Intercept) -12619.838 1419.984 -8.887 < 2e-16 ***
#> age 242.068 17.155 14.111 < 2e-16 ***
#> sexmale 8.815 477.858 0.018 0.985287
#> bmi 379.957 43.460 8.743 < 2e-16 ***
#> children 759.908 201.638 3.769 0.000179 ***
#> smokeryes 21320.176 605.795 35.194 < 2e-16 ***
#> ---
#> Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#>
#> Residual standard error: 6077 on 643 degrees of freedom
#> Multiple R-squared: 0.7061, Adjusted R-squared: 0.7038
#> F-statistic: 309 on 5 and 643 DF, p-value: < 2.2e-16
Fitting the same model on the test data
shows that, given our model specification, an \(R^2 = 0.71\) is the best we can do, which
suggests that our reweighted model is already quite close to the best
possible performance.
Conclusion
In this vignette, we showed how the densityratio package
can be used to estimate the density ratio between two data sets, and how
the density ratio can be used to reweigh the training data in a
regression model under covariate shift.