Selection Bias Corrections in Julia, Part 1

Selection bias arises when a data sample is not a random draw from the population that it is intended to represent. This is especially problematic when the probability that a particular individual appears in the sample depends on variables that also affect the relationships we wish to study. Selection bias corrections based on models of economic behavior were pioneered by the economist James J. Heckman in his seminal 1979 paper.

For an example of selection bias, suppose we wish to study the effectiveness of a treatment (a new medicine for patients with a particular disease, a preschool curriculum for children facing particular disadvantages, etc.). A random sample is drawn from the population of interest, and the treatment is randomly assigned to a subset of this sample, with the remaining subset serving as the untreated (“control”) group. If the subsets followed instructions, then the resulting data would serve as a random draw from the data generating process that we wish to study.

However, suppose the treatment and control groups do not comply with their assignments. In particular, if only some of the treated benefit from treatment while the others are in fact harmed by treatment, then we might expect the harmed individuals to leave the study. If we accepted the resulting data as a random draw from the data generating process, it would appear that the treatment was much more successful than it actually was; an individual who benefits is more likely to be present in the data than one who does not benefit.

Conversely, if treatment were very beneficial, then some individuals in the control group may find a way to obtain treatment, possibly without our knowledge. The benefits received by the control group would make it appear that the treatment was less beneficial than it actually was; the receipt of treatment is no longer random.

In this tutorial, I present some parameterized examples of selection bias. Then, I present examples of parametric selection bias corrections, evaluating their effectiveness in recovering the data generating processes. Along the way, I demonstrate the use of the GLM package in Julia. A future tutorial demonstrates non-parametric selection bias corrections.

Example 1: Selection on a Normally-Distributed Unobservable

Suppose that we wish to study of the effect of the observable variable X_{i} on Y_i. The data generating process is given by,


where X_i and \epsilon_i are independent in the population. Because of this independence condition, the ordinary least squares estimator would be unbiased if the data (Y,X) were drawn randomly from the population. However, suppose that the probability that individual i were in the data set \mathcal{S} were a function of X and \epsilon. For example, suppose that,

\Pr\left( i \in \mathcal{S}\Big| X_i,\epsilon_i\right)=1 if X_i > \epsilon_i,

and the probability is zero otherwise.  This selection rule ensures that, among the individuals in the data (the i in \mathcal{S}), the covariance between X and \epsilon will be positive, even though the covariance is zero in the population. When X covaries positively with \epsilon, then the OLS estimate of \beta_1 is biased upward, i.e., the OLS estimator converges to a value that is greater than \beta_1.

To see the problem, consider the following simulation of the above process in which X and \epsilon are drawn as independent standard normal random variables:

N = 1000
X = randn(N)
epsilon = randn(N)
beta_0 = 0.
beta_1 = 1.
Y = beta_0 + beta_1.*X + epsilon
populationData = DataFrame(Y=Y,X=X,epsilon=epsilon)
selected = X.>epsilon
sampleData = DataFrame(Y=Y[selected],X=X[selected],epsilon=epsilon[selected])

There are 1,000 individuals in the population data, but 492 of them are selected to be included in the sample data. The covariance between X and epsilon in the population data is,


which is approximately zero, but in the sample data, it is,


which is approximately 0.32.

Now, we regress Y on X with the two data sets to obtain,

2-element Array{Float64,1}:

2-element Array{Float64,1}:

where, in Julia 0.3.0, the command array() replaces the old command matrix() in converting a DataFrame into a numerical Array. This simulation demonstrates severe selection bias associated with using the sample data to estimate the data generating process instead of the population data, as the true parameters, \beta_0=0,\beta_1=1, are not recovered by the sample estimator.

Correction 1: Heckman (1979)

The key insight of Heckman (1979) is that the correlation between X and \epsilon can be represented as an omitted variable in the OLS moment condition,

\mathbb{E}\left[Y_i\Big|X_i,i\in\mathcal{S}\right]=\mathbb{E}\left[\beta_0 + \beta_1X_i+\epsilon_i\Big|X_i,i\in\mathcal{S}\right]=\beta_0 + \beta_1X_i+\mathbb{E}\left[\epsilon_i\Big|X_i,i\in\mathcal{S}\right],

Furthermore, using the conditional density of the standard Normally distributed \epsilon,


which is called the inverse Mills ratio, where \phi and \Phi are the probability and cumulative density functions of the standard normal distribution. As a result, the moment condition that holds in the sample is,

\mathbb{E}\left[Y_i\Big|X_i,i\in\mathcal{S}\right]=\beta_0 + \beta_1X_i+\frac{-\phi\left(X_i\right)}{\Phi\left(X_i\right)}

Returning to our simulation, the inverse Mills ratio is added to the sample data as,

sampleData[:invMills] = -pdf(Normal(),sampleData[:X])./cdf(Normal(),sampleData[:X])

Then, we run the regression corresponding to the sample moment condition,

3-element Array{Float64,1}:

We see that the estimate for \beta_1 is now 1.056, which is close to the true value of 1, compared to the non-corrected estimate of 1.452 above. Similarly, the estimate for \beta_0 has improved from -0.859 to -0.167, when the true value is 0. To see that the Heckman (1979) correction is consistent, we can increase the sample size to N=100,000, which yields the estimates,

3-element Array{Float64,1}:

which are very close to the true parameter values.

Note that this analysis generalizes to the case in which X contains K variables and the selection rule is,

\delta_0 + \delta_1 X_1 + \delta_2 X_2 + \ldots + \delta_K X_k > \epsilon,

which is the case considered by Heckman (1979). The only difference is that the coefficients \delta_k must first be estimated by regressing an indicator for i \in \mathcal{S} on X, then using the fitted equation within the inverse Mills ratio. This requires that we observe X for i \notin \mathcal{S}. Probit regression is covered in a slightly different context below.

As a matter of terminology, the process of estimating \delta is called the “first stage”, and estimating \beta conditional on the estimates of \delta is called the “second stage”. When the coefficient on the inverse Mills ratio is positive, it is said that “positive selection” has occurred, with “negative selection” otherwise. Positive selection means that, without the correction, the estimate of \beta_1 would have been upward-biased, while negative selection results in a downward-biased estimate. Finally, because the selection rule is driven by an unobservable variables \epsilon, this is a case of “selection on unobservables”. In the next section we consider a case of “selection on observables”.

Example 2: Probit Selection on Observables

Suppose that we wish to know the mean and variance of Y in the population. However, our sample of Y suffers from selection bias. In particular, there is some X such that the probability of observing Y depends on X according to,

\Pr\left(i\in\mathcal{S}\Big| X_i\right)=F\left(X_i\right),

where F is some function with range [0,1]. Notice that, if Y and X were independent, then the resulting sample distribution of Y would be a random draw from the population (marginal distribution) of Y. Instead, we suppose \mathrm{Cov}\left(X,Y\right)\neq 0. For example,

N = 10000
populationData = DataFrame(rand(MvNormal([0,0.],[1 .5;.5 1]),N)')

1x2 Array{Float64,2}:
 -0.0281916  -0.022319

2x2 Array{Float64,2}:
 0.98665   0.500912
 0.500912  1.00195 

In this simulated population, the estimated mean and variance of Y are -0.022 and 1.002, and the covariance between X and Y is 0.501. Now, suppose the probability that Y_i is observed is a probit regression of X_i,

\Pr\left( i\in\mathcal{S}\Big| X_i \right) = \Phi\left( \beta_0 + \beta_1 X_i \right),

where \Phi is the CDF of the standard normal distribution. Letting D_i=1 indicate that i \in \mathcal{S}, we can generate the sample selection rule D as,

beta_0 = 0
beta_1 = 1
index = (beta_0 + beta_1*data[:X])
probability = cdf(Normal(0,1),index)
D = zeros(N)
for i=1:N
    D[i] = rand(Bernoulli(probability[i]))
populationData[:D] = D
sampleData = populationData
sampleData[D.==0,:Y] = NA

The sample data has missing values in place of Y_i if D_i=0. The estimated mean and variance of Y in the sample data are 0.275 (which is too large) and 0.862 (which is too small).

Correction 2: Inverse Probability Weighting

The reason for the biased estimates of the mean and variance of Y in Example 2 is sample selection on the observable X. In particular, certain values of Y are over-represented due to their relationship with X. Inverse probability weighting is a way to correct for the over-representation of certain types of individuals, where the “type” is captured by the probability of being included in the sample.

In the above simulation, conditional on appearing in the population, the probability that an individual of type X_i=1 is included in the sample is 0.841. By contrast, the probability that an individual of type X_i=0 is included in the sample is 0.5, so type X_i is over-represented by a factor of 0.841/0.5 = 1.682. If we could reduce the impact that type X_i=1 has in the computation of the mean and variance of Y by a factor of 1.682, we would alter the balance of types in the sample to match the balance of types in the population. Inverse probability weighting generalizes this logic by weighting each individual’s impact by the inverse of the probability that this individual appears in the sample.

Before we can make the correct, we must first estimate the probability of sample inclusion. This can be done by fitting the probit regression above by least-squares. For this, we use the GLM package in Julia, which can be installed the usual way with the command Pkg.add(“GLM”).

using GLM
Probit = glm(D ~ X, sampleData, Binomial(), ProbitLink())

             Estimate Std.Error  z value Pr(>|z|)
(Intercept)  0.114665  0.148809 0.770554   0.4410
X             1.14826   0.21813  5.26414    1e-6

estProb = predict(Probit)
weights = 1./estProb[D.==1]/sum(1./estProb[D.==1])

which are the inverse probability weights needed to match the sample distribution to the population distribution.

Now, we use the inverse probability weights to correct the mean and variance estimates of Y,

correctedMean = sum(sampleData[D.==1,:Y].*weight)

correctedVariance = (N/(N-1))*sum((sampleData[D.==1,:Y]-correctedMean).^2.*weight)

which are very close to the population values of -0.022319 and 1.00195. The logic here extends to the case of multivariate X, as more coefficients are added to the Probit regression. The logic also extends to other functional forms of F, for example, switching from Probit to Logit is achieved by replacing the ProbitLink() with LogitLink() in the glm() estimation above.

Example 3: Generalized Roy Model

For the final example of this tutorial, we consider a model which allows for rich, realistic economic behavior. In words, the Generalized Roy Model is the economic representation of a world in which each individual must choose between two options, where each option has its own benefits, and one of the options costs more than the other. In math notation, the first alternative, denoted D_i=1, relates the outcomes Y_i to the individual’s observable characteristics, X_i, by,

Y_{1,i} = \mu_1\left(X_i\right)+U_{1,i}.

Similarly, the second alternative, denoted D_i=0, relates Y_i to X_i, by,

Y_{0,i} = \mu_0\left(X_i\right)+U_{0,i}.

The value of Y_i that appears in our sample is thus given by,

Y_i = D_iY_{1,i} + (1-D_i)Y_{0,i}.

Finally, the value of D_i is chosen by individual i according to,

D_i=1 if Y_{1,i}-Y_{0,i}-C_i>0,

where C_i is the cost of choosing the alternative D_i=1 and is given by,

C_i = \mu_C\left(X_i,Z_i\right)+U_{C,i},

where Z contains additional characteristics of i that are not included in X.

We assume that the data only contains Y_i,D_i,X_i,Z_i; it does not contain the variables Y_{i,1},Y_{i,0},U_{i,1},U_{i,0},U_{i,C} or the functions \mu_1,\mu_0,\mu_C. Assuming that the three \mu functions follow the linear form and that the unobservables U are independent and Normally distributed, we can simulate the data generating process as,

N = 1000
sampleData = DataFrame(rand(MvNormal([0,0.],[1 .5; .5 1]),N)')
U1 = rand(Normal(0,.5),N)
U0 = rand(Normal(0,.7),N)
UC = rand(Normal(0,.9),N)
betas1 = [0,1]
betas0 = [.3,.2]
betasC = [.1,.1,.1]
Y1 = betas1[1] + betas1[2].*sampleData[:X] + U1
Y0 = betas0[1] + betas0[2].*sampleData[:X] + U0
C = betasC[1] + betasC[2].*sampleData[:X] + betasC[3].*sampleData[:Z] + UC
D = Y1-Y0-C.>0
Y = D.*Y1 + (1-D).*Y0
sampleData[:D] = D
sampleData[:Y] = Y

In this simulation, about 38% of individuals choose the alternative D_i=1. About 10% of individuals choose D_i=0 even though they receive greater benefits under D_i=1 due to the high cost C_i associated with D_i=1.

Solution 3: Heckman Correction for Generalized Roy Model

The identification of this model is attributable to Heckman and Honore (1990). Estimation proceeds in steps. The first step is to notice that the left- and right-hand terms in the following moment equation motivate a Probit regression:


where U_{D,i}\equiv -\left(U_{1,i}-U_{0,i}-U_{C,i}\right)/\sigma_D\sim\mathcal{N}\left(0,1\right) is the negative of the total error term arising in the equation that determines D_i above, \sigma_D \equiv \sqrt{\mathrm{Var}\left(U_{1,i}-U_{0,i}-U_{C,i}\right)}, and,

\mu_D\left(X,Z\right) \equiv \left([1,X]\beta_1 -[1,X]\beta_0 -[1,X,Z]\beta_C\right)/\sigma_D \equiv [1,X,Z]\beta_D,

In the simulation above, \beta_D = [-.4,.7,-.1]/\sqrt{.5+.7+.9}\approx[-0.276,0.483,-0.069]. We can estimate \beta_D from the Probit regression of D on X and Z.

betasD = coef(glm(D~X+Z,sampleData,Binomial(),ProbitLink()))
3-element Array{Float64,1}:

Next, notice that,

\mathbb{E}\left[Y_i\Big|D_i=1,X_i,Z_i\right] =[1,X_i]\beta_1+\mathbb{E}\left[U_{1,i}\Big|D_i=1,X_i,Z_i\right],


\mathbb{E}\left[U_{1,i}\Big|D_i=1,X_i,Z_i\right] =\mathbb{E}\left[U_{1,i}\Big|\mu_D\left(X_i,Z_i\right)>U_{i,D},X_i,Z_i\right]=\rho_1 \frac{-\phi\left(\mu_D\left(X_i,Z_i\right)\right)}{\Phi\left(\mu_D\left(X_i,Z_i\right)\right)},

which is the inverse Mills ratio again, where \rho_1 \equiv \frac{\mathrm{Cov}\left(U_1,U_D\right)}{\mathrm{Var}\left(U_D\right)}. Substituting in the estimate for \beta_D, we consistently estimate \beta_1:

fittedVals = hcat(ones(N),array(sampleData[[:X,:Z]]))*betasD

sampleData[:invMills1] = -pdf(Normal(0,1),fittedVals)./cdf(Normal(0,1),fittedVals)

correctedBetas1 = linreg(array(sampleData[D,[:X,:invMills1]]),vec(array(sampleData[D,[:Y]])))
3-element Array{Float64,1}:

To see how well the correction has performed, compare these estimates to the uncorrected estimates of \beta_1,

biasedBetas1 = linreg(array(sampleData[D,[:X]]),vec(array(sampleData[D,[:Y]])))
2-element Array{Float64,1}:

Similar logic allows us to estimate \beta_0:

sampleData[:invMills0] = pdf(Normal(0,1),fittedVals)./(1-cdf(Normal(0,1),fittedVals))

correctedBetas0 = linreg(array(sampleData[D.==0,[:X,:invMills0]]),vec(array(sampleData[D.==0,[:Y]])))
3-element Array{Float64,1}:

biasedBetas0 = linreg(array(sampleData[D.==0,[:X]]),vec(array(sampleData[D.==0,[:Y]])))
2-element Array{Float64,1}:

In summary, we can consistently estimate the benefits associated with each of two alternative choices, even though we only observe each individual in one of the alternatives, subject to heavy selection bias, by extending the logic introduced by Heckman (1979).

Bradley J. Setzler

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s