|Home | About | Journals | Submit | Contact Us | Français|
The probability of selection into a treatment, also called the propensity score, plays a central role in classical selection models and in matching models (see, e.g., Heckman, 1980; Heckman and Navarro, 2004; Heckman and Vytlacil, 2007; Hirano et al., 2003; Rosenbaum and Rubin, 1983). 1 Heckman and Robb (1986, reprinted 2000), Heckman and Navarro (2004) and Heckman and Vytlacil (2007) show how the propensity score is used differently in matching and selection models. They also show that, given the propensity score, both matching and selection models are robust to choice-based sampling, which occurs when treatment group members are over- or under-represented relative to their frequency in the population. Choice-based sampling designs are frequently chosen in evaluation studies to reduce the costs of data collection and to obtain more observations on treated individuals. Given a consistent estimate of the propensity score, matching and classical selection methods are robust to choice-based sampling, because both are defined conditional on treatment and comparison group status.
This note extends the analysis of Heckman and Robb (1985),Heckman and Robb (1986, reprinted 2000) to consider the case where population weights are unknown so that the propensity score cannot be consistently estimated. In evaluation settings, the population weights are often unknown or cannot easily be estimated.2 For example, for the National Supported Work training program studied in LaLonde (1986), Dehejia and Wahba (1999, 2002) and in Smith and Todd (2005), the population consists of all persons eligible for the program, which was targeted at drug addicts, ex-convicts, and welfare recipients. Few datasets have the information necessary to determine whether a person is eligible for the program, so it would be difficult to estimate the population weights needed to consistently estimate propensity scores.
In this note, we establish that matching and selection procedures can still be applied when the propensity score is estimated on unweighted choice based samples. The idea is simple. To implement both matching and classical selection models, only a monotonic transformation of the propensity score is required. In choice based samples, the odds ratio of the propensity score estimated using misspecified weights is monotonically related to the odds ratio of the true propensity scores. Thus, selection and matching procedures can identify population treatment effects using misspecified estimates of propensity scores fit on choice-based samples.
Let D = 1 if a person is a treatment group member; D = 0 if the person is a member of the comparison group. X = x is a realization of X. In the population generated from random sampling, the joint density is
where g is the density of the data. By Bayes's theorem, we have, letting Pr(D = 1) = P,
Assume 0 < Pr(D = 1 | x) < 1. From knowledge of the densities of the data in the two samples, g(x | D = 1) and g(x | D = 0), one can form a scalar multiple of the ratio of the propensity score without knowing P. The odds ratio is a monotonic function of the propensity score that does not require knowledge of the true sample weights. In a choice-based sample, both the numerator and denominator of the first term in (2) can be consistently estimated. This monotonic function can replace P(x) in implementing both matching and nonparametric selection models.
However, estimating g(x | D = d) is demanding of the data when X is of high dimension. Instead of estimating these densities, we can substitute for the left hand side of (2) the odds ratio of the estimated conditional probabilities obtained using the choice-based sample with the wrong weights. (i.e. for example, ignoring the fact that the data are a choice based sample). The odds ratio of the estimated probabilities is a scalar multiple of the true odds ratio. It can therefore be used instead of Pr(D = 1 | X) to match or construct nonparametric control functions in selection bias models.
In the choice-based sample, let r(D = 1 | x) be the conditional probability that D = 1 and P* be the unconditional probability of sampling D = 1, where P* ≠ P, the true population proportion. The joint density of the data from the sampled population is
Under random sampling, the right-hand sides of (3a) and (3b) are the limits to which the choice-based probabilities converge. Taking the ratio of (3a) to (3b), assuming the latter is not zero, one obtains
Thus, one can estimate the ratio of the propensity score up to scale (the scale is the product of the two terms on the right-hand side of (4)). Instead of estimating matching or semiparametric selection models using Pr(D = 1 | x) (as in, for example, Ahn and Powell (1993); Heckman (1980); Heckman and Hotz (1989); Heckman et al. (1998); Heckman and Robb (1986); Powell (2001), one can, instead, use the odds ratio of the estimate r(D = 1 | x), which is monotonically related to the true Pr(D = 1 | x). In the case of a logit P(x), P(x) = exp(xβ)/(1 + exp(xβ)), the log of this ratio becomes
where the slope coefficients are the true values and the intercept 0 = β0 + n(P*/(1 − P*)) + n((1 − P)/P)), where β0 is the true value.3
In implementing nearest-neighbor matching estimators, matching on the log odds ratio gives identical estimates to matching on the (unknown) Pr(D = 1 | x), because the odds ratio preserves the ranking of the neighbors. In application of either matching or classical selection bias correction methods, one must account for the usual problems of using estimated log odds ratios instead of true values.4
*This research was supported by NSF SBR 93-21-048 and NSF 97-09-873 and NICHD 40-4043-000-85-261.
1It also plays a key role in instrumental variables models (see Heckman et al., 2006). Heckman and Vytlacil (2007) discuss the different role played by the propensity score in matching IV and selection models.
2The methods of Manski and Lerman (1977) and Manski (1986) for adjusting for choice-based sampling in estimating the discrete choice probabilities cannot be applied when the weights are unknown and cannot be identified from the data.
3See Manski and McFadden (1981, p. 26).
James J. Heckman, University of Chicago, University College Dublin, Cowles Foundation, Yale University and American Bar Foundation.
Petra E. Todd, University of Pennsylvania and NBER.