Today: (1) Go over process tracing and process tracing tests; (2) Finish up lecture on causality; (3) Hand out assignment #3. Key ideas for assignment #3: Process tracing: a qualitative method of looking at data within a particular case to make inferences about that case. Process tracing is a method that can work with only one case. It has strong parallels to detective work and the kind of reasoning that juries use in trials. Causal-process observations (CPOs): specific pieces of information or data from within a case that are useful for making judgments about causation. You can think of them as similar to “clues” about a case. For example, in the Tannenwald article, specific conversations in which foreign policy decision makers had trouble discussing nuclear use are CPOs. Data-set observation: the scores or values for a case across all systematically measured variables. This is often what we mean by a “case” when we have a rectangular dataset. Eye Color Height Sex You Green 5’ 3’’ Me Blue 5’ 10 Male (almost 5’ 11’’ really) Age Female 21 42 Place of Birth Miami, FL Davenport, IA Process tracing tests: (1) Hoop tests; (2) Smoking gun tests; (3) Straw in the wind tests; (4) Doubly decisive tests. All of these tests involve relating: (1) Specific pieces of evidence (CPOs), to (2) Preexisting generalizations that apply more broadly. Hoop test: Passing a hoop test is necessary but not sufficient for the validity of a given hypothesis. This kind of test can eliminate a given hypothesis but it cannot always provide strong support that the hypothesis is valid. Example Hypothesis: O.J. Simpson intentionally caused the death of Ron Goldman. Hoop test: Was O.J. in the general area at the time that Goldman was killed? Some hoop tests are harder to pass than others: (1) Was O.J. on the planet Earth at the time that Goldman was killed? (2) Was O.J. at the Nicole Brown Simpson home at the time that Goldman was killed? Failing a hoop test always eliminates a hypothesis. Passing a hoop test lends support in favor of a hypothesis only to the degree that it is a hard test. What makes a hoop test easy or hard? The difficulty of a hoop test is related to the frequency at which the CPO is typically or normally present. Hoop tests that make reference to rare CPOs are hard hoop tests. Hoop tests that make reference to common CPOs are easy hoop tests. Other hoop tests: (1) Is O.J. right handed? (2) Did O.J. have motive to carry out a violent murder? (3) Does O.J.’s hand fit the glove? Smoking gun test: With these tests, one has access to evidence that acts like a smoking gun in a murder investigation: the evidence is sufficient but not necessary for the validity of a given hypothesis. Smoking gun tests are used primarily to confirm the validity of a hypothesis. Failing a smoking gun test does not mean that a hypothesis is necessarily wrong. Example from O.J. investigation: Traces discovered at the crime scene included Simpson’s DNA in blood samples from footprints and the glove. Goldman’s DNA was found in Simpson’s Bronco. Nicole Brown Simpson’s DNA was found in the Bronco and on a sock in Simpson’s bedroom. The defense suggested that one of the detectives may have intentionally planted blood in Simpson’s Bronco by using the bloody glove found at Simpson’s estate. They also suggested that the DNA samples were degraded and suffered from crosscontamination. How consequential was the absence of the murder weapon for challenging the validity of the hypothesis that Simpson was the murderer? It might depend on difficult/unusual it would have been to not find the weapon. Straw in the wind test: Provides some support for or against a hypothesis, but it is not decisive. Example: Passing a hard hoop test provides straw in the wind evidence in favor of a hypothesis. Back to causality lecture . . . 4 Criteria of Causality: (1) Time Order (2) Association (3) Not Spurious (4) Mechanism (2) Association: there is a “systematic relationship” between variables or events. One kind of association: a correlation Correlation with dichotomous variables I: present Cases here Dependent Variable absent Cases here absent present Independent Variable Correlation with dichotomous variables I: present Cases here Good Grade absent Cases here absent present Studies a lot Correlation with dichotomous variables II: present Cases here Dependent Variable Cases here absent absent present Independent Variable We can also think about correlations using continuous data. High Dependent Variable Low Low High Independent Variable Example: xxx High xxxx xxxx xx Level of Democracy Low xxxxxx xxxxx xx xxxx x x x xxx xxxx Low x xx High Level of Economic Growth Please notice the relationship between the two-by-two table for dichotomous variable correlations and the scatterplot for continuous variable correlations. (I will help you notice it!) Another kind of association: a necessary condition (actually a type of asymmetrical correlation). present Dependent Variable absent absent present Independent Variable For example: present Pregnancy absent absent (Male) Female present (Female) For example: present Social Revolution absent absent present Authoritarian Government Question: How would a necessary condition look with continuous variables? Dependent Variable Independent Variable Necessary cause: powerful notion of causality because it implies that if the cause had not occurred, then the outcome would not have occurred. In that sense, the cause really seems to have “made a difference.” Counterfactual Statement: “If it had been the case that C (or not C), it would have been the case that E (or not E).” If I had eaten breakfast, I would not be hungry right now. Counterfactual because it makes a claim about the effect of an event in the past that did not happen (me eating breakfast). Always useful to ask about the counterfactual. For example: If there was no nuclear taboo, what would have happened? If systematic vulnerability did not exist, what would have happened? Problem of counterfactuals: We can never directly test them. It would require “re-running history.” Instead, we must imagine a “possible world” in which the cause did not occur. Issue: How close is the “possible world” to the real world? How much do we have to rewrite history even to imagine the possible world? Example: “If the United States did not exist, Iraq would be allied with France.” Useful counterfactuals require us to rewrite history only a little bit. But it is often hard to find such counterfactuals. Example: If McCain had won the election, the economy would be stronger right now. The problem of “over-determination” and necessary causes: Soldier 1 Soldier 2 Prisoner Dies Soldier 3 Soldier 4 The problem of “preemption” and necessary causes: Two assassins shoot 5 seconds apart at a dictator. Are these actions both causes of the dictator’s death? How do we distinguish “trivial” from important necessary causes? Rule of thumb: important necessary causes are rare events. “We may define a cause to be an object followed by objects similar to the second. Or, in other words, where, if the first object had not been, the second never ceased to exist.” --David Hume Another kind of association: a sufficient condition (actually a type of asymmetrical correlation). present Dependent Variable absent absent present Independent Variable Example: Being caught cheating on the final exam is sufficient for failing class. present Failure in Class absent absent present Caught cheating on Final Exam Question: How would a sufficient condition look with continuous variables? Dependent Variable Independent Variable Enough on association!!! Clearly, it is a big but complicated part of causation. Let’s talk about the third criteria: (3) Non-Spurious Relationship Sometimes a correlation (whether symmetrical or asymmetrical) is spurious. That is, the correlation exists, but it is not causal. Spurious correlation: a noncausal correlation. Note: A spurious correlation is a correlation. But it is not a causal association. We need some examples to make this point concrete . . . The number of cracks in the road is correlated with heart attacks (areas with lots of cracks in the road have higher rates of heart attacks). Is this a causal relationship? If not, why not? How a correlation can become spurious: The trick is often to find a new variable that comes before both the original independent variable and dependent variable, and that causes them both. Z X Y X Y Antecedent variable: A variable that comes before other variables. This kind of variable can (but does not always) make an initial correlation spurious. The number of storks living in an area is correlated with the birth rate of that area (i.e., areas with more storks tend to have higher birth rates). Countries where people eat lots of pizza tend to have higher literacy rates than countries where they don’t eat so much pizza. Caffeine consumption is correlated with miscarriage. Watching violence on TV is correlated with carrying out violent behavior. Drinking alcohol in moderation is correlated with fewer heart attacks. Smoking is correlated with lung cancer. Let us always be mindful also of time order. X Y Y X When you see a correlation, ask about: (1) Time order; (2) Association; and (3) Spuriousness. Example: Smoking marijuana and low grades in high school. (4) Causal Mechanism If there is a non-spurious association between variables, then why does the association exist? Answer: there must be some mechanism through which the independent variable affects the dependent variable. X X Y M Y How do we find the causal mechanism? Possible answer: Process Tracing Process Tracing: We try to identify the historical events that link a cause with its effect. Example: nuclear taboo => nonuse nuclear taboo => actual decision makers’ calculations/strategies => non-use Note: We can always potentially identify more causal mechanisms. X Y X X M A M Y B Y When do we stop looking for mechanisms? Possible Answer: When the relationships are so obvious that no further process tracing is needed.