4.31 Causes: Which variables are causes? (Discussion)

  • Typical causes (X) in the social sciences (e.g. Freese and Kevern 2013)
    • Income, class, education, participation, attitudes, genes, personality, height, ethnic background, regulatory regimes, policies
      • Q: Which of the cursive causes could be outcomes as well? Why or why not?
  • Q: Outcome is exam result: What is the difference between the “causes” in the following examples?27
    1. She did well on the exam because she is a woman.
    2. She did well on the exam because she studied for it.
    3. She did well on the exam because she was coached by her teacher.
  • Q: Which of these three “causes” could we manipulate in an experiment?

References

Freese, Jeremy, and J Alex Kevern. 2013. “Types of Causes.” In Handbook of Causal Analysis for Social Research, edited by Stephen L Morgan, 27–41. Springer Netherlands.


  1. “First, the mechanism that assigns subjects to treatment and control groups is not usually known to be random. Rather, an event occurs in the world that happens to affect some subjects but not others, and the researcher assumes that the naturally occurring intervention was assigned as-if at random (Dunning 2008). The condition of as-if randomness is sometimes referred to as exogeneity. In most studies, researchers go to great lengths to argue that this condition is satisfied. Exogeneity implies that the treatment and control groups created by the natural experiment are similar in terms of all observed and unobserved factors that may affect the outcome of interest, with the exception of the treatment and confounders that the researcher controls for. If the two groups are similar in this way, then the design is said to be valid, and comparing outcomes across the groups identifies the causal effect of treatment. The second distinguishing feature of natural experiments is more subtle. The naturally occurring intervention generates some subjects who receive treatment and other subjects who do not. It is often possible, however, to define a number of different treatment and control groups from the natural intervention. Yet, only some of these groups are similar and thus valid to compare, even when nature intervenes randomly. The problem arises because the researcher does not directly control the design of the experiment. In a randomized controlled experiment, the researcher picks what the treatment and controls groups should be and then randomly assigns subjects to the groups. In a natural experiment, however, the researcher finds some intervention that has been implemented and also finds some subjects. She then constructs treatment and control groups to address a particular hypothesis. But the treatment and control groups constructed post hoc may not be comparable, even if one assumes that the natural intervention was randomly assigned” (Sekhon and Titiunik 2012, 35–36). Sekhon and Titiunik (2012, 36) discuss four examples of natural experiments: redistricting as a natural experiment to study the personal vote; the impact of representatives’ decisions to move from the U.S. House to the U.S. Senate on their roll-call voting scores; The impact of Indian randomized electoral quotas for women on the probability that women will contest and win elections after the quotas are withdrawn; Regression discontinuity designs that have been used, among other things, to estimate the incumbency advantage in U.S. House elections - how time in office affects the size of the incumbency advantage;