2. MARCO REFERENCIAL TEÓRICO CONCEPTUAL 7
2.2. MARCO TEÓRICO 9
2.2.6. LA MÚSICA Y SUS INFLUENCIAS 25
The RDD is essentially a variant of the IV approach and was first established by Thistlethwaite and Campbell (1960) and Campbell (1969). The RD approach assesses causal effects by introducing a cut‐off point or discontinuity that determines programme participation. In more general terms, RDDs follow a deterministic rule, i.e. D = 1 if Z < Ź and D = 0 if Z > Ź where Z is an observed variable which shapes the decision of individuals to select into a programme or not (Heckman, LaLonde and Smith, 1999). This is the case of a ‘sharp’ design, which is the ideal case but cannot often be observed in practice (Blundell and Costa Dias, 2008). The more common case is the ‘fuzzy’ design where the decision to participate in a programme is not an entirely deterministic function of Z (Heckman, LaLonde and Smith, 1999). In a ‘fuzzy’ design, participants and non‐participants exist on either side of the threshold of variable Z and discontinuity cannot be observed (Blundell and Costa Dias, 2008). The variable Z is central since it has an effect on outcome variable Y directly as well as indirectly through D. The indirect effect through D is the causal effect that is of interest and shall be assessed. The publication by Hahn, Todd and van der Klaauw (2001) is widely cited and discusses RDDs in the context of LATE, while Heckman and Vytlacil (2007b) illuminate the RDD from the perspective of MTE. For a more practical guide, see Imbens and Lemieux (2008).
PnK use landownership as the variable Z to determine whether a household is eligible to participate in microcredit. If Z < Ź, then D = 1 and D = 0 if not. In other words,
households are eligible to participate in the programme if they own less than 0.5 acres of land and are not eligible if they own more than 0.5 acres. However, there is a debate surrounding the study of PnK based on that argument that their RDD was not strictly enforced and that the underlying design was ‘fuzzy’ (Chemin, 2008). Furthermore, Ravallion (2008) argues that the study by PnK followed a DID approach (discussed in more depth in chapter 5).
It is commonly argued that the ‘sharp’ design is able to control for selection on observables (Blundell and Costa Dias, 2008). This does not apply in the event of a ‘fuzzy’ design where the problem becomes one of selection on unobservables since participation occurs for various levels of the eligibility variable whose values reflect the unobservables. In the study by PnK land cultivated plays this role, but the ‘unobservables’ are inferred to be quality of land, so that the true variable is value of land equivalent to the value of 0.5 acres of land of average value (this point is further developed and explained in chapter 5).
Much of the appeal and simplicity of RDDs disappear when the criterion is ‘fuzzy’ (Heckman, LaLonde and Smith, 1999; Blundell and Costa Dias, 2008), and other methods that deal with selection on unobservables such as IV, selection models or DID will have to be considered. Heckman, LaLonde and Smith (1999) point out that non‐ participation by individuals who are in fact eligible, i.e. their values of Z satisfy the eligibility rules, is a serious concern in RDDs since unbiased estimates of the mean impact of treatment of eligible individuals can no longer be obtained (Heckman, LaLonde and Smith, 1999).
3.6.3.1.
Applications of regression discontinuity designs
Furthermore, earlier sections mentioned the studies conducted by Dehejia and Wahba (1999 and 2002) that re‐investigated LaLonde’s (1986) study. They employed PSM claiming to illustrate that it can approximate the results obtained from an experimental study. Cook and Wong (2008) did something similar with RDDs applying within‐study comparisons commonly associated with LaLonde (1986) to analyse how well the RD results compare to those obtained from a randomised experiment. They found “considerable correspondence between the experimental and RD results” (Cook and Wong, 2008, p. 32). The RD estimates Cook and Wong (2008) obtained, seemingly using
94 a ‘sharp’ design, were generally robust and a good reproduction of the results of the experimental study (p. 33). A related inquiry was conducted by Green et al (2009).
In addition, further examples that are widely used in the literature to illustrate the application of RDDs are provided by Angrist and Lavy (1999) on schooling and Hahn, Todd and van der Klaauw (1999) on anti‐discrimination laws of minority workers. In brief, Angrist and Lavy (1999) in a widely quoted paper, investigate the impact of class size on students’ test scores in Israeli public schools. Israeli schools typically restrict their class size to 40 students. The authors argue that this cap is based on the so‐called ‘Maimonides’ Rule’ which was established by a rabbinic scholar called Maimonides in the 12th century. This rule states that a class is split into two if the number of its
students exceeds 40, i.e. with the 41st student the class size drops to an average of 20.5
students. Angrist and Lavy (1999) exploit the ‘Maimonides’ Rule’ as an eligibility criterion to estimate the impact of class size on test scores. Students that have been in a class that has been split into two qualify for treatment. As expected, the students’ test scores are directly affected by class size, i.e. the smaller the size of the class, the higher the test scores of the students, in particular in their 4th and 5th grades. In addition, an
indirect effect of class size through D can also be observed (Caliendo, 2006; Caliendo and Hujer, 2005). Hahn, Todd and van der Klaauw (1999) present another well‐known example that assesses the impact of anti‐discrimination laws of minority workers. These laws are only applicable to firms with more than 15 employees. This threshold functions as the discontinuity rule and hence allows assessing the causal effects of this intervention.
Finally, Heckman, LaLonde and Smith (1999) and Caliendo (2006) conclude that RDDs are adequate for controlling for selection on observables as long as a deterministic rule can strictly be observed. In other words, van der Klaauw (2002) argues that
“the validity of the RD approach relies on an appropriate specification of the relationships between the selection and treatment variable, and between the selection and outcome variable“(p. 1284).
If this specification is not done correctly, RDDs fail to provide unbiased impact estimates of programme participation and selection on unobservables becomes an issue. This then requires a different evaluation strategy, e.g. matching or IVs. However,
as mentioned earlier, matching requires common support for treated and non‐treated individuals and if this requirement cannot be fulfilled, then matching is also not an appropriate strategy (Caliendo, 2006; Caliendo and Hujer, 2005). Moreover, an earlier section has shown that matching does not work well when there is selection on unobservables. Hence, the next section takes a closer look at selection models which claim to provide solutions to selection on unobservables.