Logic, Sequences, and Explanation James Mahoney Northwestern University Starting assumption: an individual cause is normally one of five types: (1) necessary but not sufficient; (2) sufficient but not necessary; (3) necessary and sufficient; (4) INUS; or (5) SUIN. Other assumptions: (1) Causation at the population level exists because causes are operating in the individual cases. You can think of this as the “methodological individualism” of causation. (2) Causes that raise the probability of an outcome in a population may still contribute to the absence of that outcome in any one individual case. (3) We first need to understand what “cause” can mean at the level of the individual case. Then we need to explore whether and how this understanding can be extended to the population level. Set-theoretic definition of necessary cause: X is a necessary cause of Y if Y is a subset of X. Set-Theoretic Conceptualization of a Necessary Cause X Y Set-Theoretic Conceptualization of a Necessary Cause: Example Set of all females X Y Set of all pregnant people Set of Females Set of pregnant people Set of Females Set of pregnant people X Not X Y Set of not females Set of females X Not X Y Set of pregnant people Set of females Set of not females Set of pregnant people Set of not pregnant people Method of Agreement (as conventionally used): Eliminates necessary causes. Case Peasant Worker Social Revolts Revolts Revolution --------------------------------------------------------France Yes Yes Yes Russia Yes Yes Yes China Yes No Yes The Relative Importance of Necessary Causes: An Example A B C Y1 Y A = Daytime B = Sunshine C = Sun shower Y = Rainbow Figure 2.2 Set of all females Z X Y Set of pregnant individuals Set of females age 10-60 Framework for Assessing the Relative Importance of Necessary Causes X Y As X approaches Y, it becomes less trivial and more important. Set-theoretic definition of a sufficient cause: X is a sufficient cause of Y if X is a subset of Y. Set-Theoretic Conceptualization of a Sufficient Cause Y X X Y Set of males Set of not pregnant people Set of males Set of Not Males Set of not pregnant people A basic rule of logic: If X is necessary for Y, then Not X is sufficient for Not Y. If X is sufficient for Y, then Not X is necessary for Not Y. Relationship between Sufficient/Necessary Causes Females Not pregnant people Not females Not Y X Y Not X Pregnant people Framework for Assessing the Relative Importance of Sufficient Causes Y X As X approaches Y, it becomes less trivial and more important. Children Males Not pregnant people Method of Difference (as conventionally used): Eliminates individually sufficient causes. Case Relative Social Deprivation Revolution --------------------------------------------------------France Yes Yes Prussia Yes No Set-Theoretic Conception of a Necessary and Sufficient Cause X Y X Not X Y Not Y INUS Cause: A cause that it is one part of a combination of causes that are jointly sufficient for an outcome. An INUS cause by itself is neither necessary nor sufficient. In Mackie’s words: An INUS cause is “an insufficient but necessary part of a condition which is itself unnecessary but sufficient for the result” (Mackie 1965: 246). Logical AND (&/*): INUS causes are connected together with the Logical AND. Again, there are various notations in vogue: Y = A AND B AND C Y=A*B*C Y=A&B&C Y = ABC Here is what I like best: A&B&C→Y ABC → Y Multiple Causation/Equifinality: Different combinations that are sufficient for the same outcome. Here the Logical OR (v/+) is used. For example: Y = (A AND B) OR (Not A AND C AND D) Y = (A & B) v (~A & C & D) Y = (A B) + (~A C D) (A & B) v (~A & C & D) → Y A B v ~A C D → Y A version of Moore’s argument: X & (A v B) → Y where X = strong bourgeoisie; A = alliance between bourgeoisie and aristocracy; B = weak aristocracy; and Y = democratic pathway. Question: Which variables are INUS causes? Set-theoretic definition of an INUS cause: X is an INUS cause of Y if the overlapping set created by X and one or more other factors is a subset of Y. Set-Theoretic Conception of INUS Causes (X & Z) v . . . . → Y Z X Y X Y Z We can also ask about the relative importance of INUS causes. For example, look at the two INUS causes in the original diagram again. Which one is more important? Jim Mahoney’s Rule: Any cause becomes more important as it becomes closer to being a necessary and sufficient cause. Necessary and sufficient cause = the gold standard. Thus, we can think of this rule as the “golden rule” of causality. SUIN Cause: A factor that is sufficient (but not necessary) to constitute a necessary (but not sufficient) cause. Example: X & Z → Y; A v B → Y. where Y = democratic pathway; X = strong bourgeoisie; Z = politically subordinate aristocracy; A = alliance between bourgeoisie and aristocracy; and B = weak aristocracy. Here A and B are SUIN causes of a democratic path. Set theoretical definition of SUIN cause: X is a SUIN cause of Y if Y is a subset of the joint space created by X when combined with one or more other causal factors. Set-Theoretic Conception of SUIN Causes A = Weak landed elites Y = Democratic regime Y A Set-Theoretic Conception of SUIN Causes X X A Z A = Weak Landed Elites X = Small Landed Class Z = Politically Marginal Landed Class Set-Theoretic Conception of SUIN Causes X Z Y X = Small Landed Class Z = Politically Marginal Landed Class Y = Democratic Regime It is worth asking which SUIN cause in the diagram is the more important one. Remember the golden rule. Why do I believe that this list of five causes is exhaustive from a logic and set theory perspective? Because, from a set theory perspective, a cause or a combination of causes must be necessary and/or sufficient for an outcome. And these five seem to express the range of possibilities. What is the relationship between logical types of causes and the understanding of causation used in statistical analysis (e.g., possible outcomes model)? Possible answer: Variables that show up as exerting effects in correctly specified statistical models have those effects because they picking up important INUS causes (i.e., INUS causes that come close to being necessary and/or sufficient). Z Y Y V x x xx xxx Y x xx x xx x x x xx x Z Y It is possible to imagine the Venn diagram that corresponds to different scatter plots. x x x x xx x x x x x x x x x x V x x x x x The main point here is that we can think about correlations in terms of logic. I think we can, in a sense, “reduce” them to their logical status. We learn something important when we ask about the extent to which a causal variable approximates a necessary and/or sufficient cause. Example: economic development is associated with democracy is because it approximates a sufficient cause. If a cause is close to being a necessary, then its effect will be more like an “enabler” – i.e., it is a variable that allows some outcome to happen. You can think about it this way: the cause exists in most of the various different combinations of factors that are each sufficient for the outcome. If a cause is close to being sufficient, then it is more like a factor that does not need much help in order to generate the outcome. It does not need to combine with many other factors to produce the outcome. It is almost enough by itself. Now I want to discuss a method for assessing the relative importance of causes in case study research. I want to do so by building on the notions of causal importance introduced earlier, including especially the golden rule. Normally, we evaluate causal importance using populations of cases: X Z Y Y X = Female, Z = Female age 10-60, Y = pregnant individual In the example, the size of the circles corresponds to the number of cases that are members of the categories represented by the circles. But how could we assess causal importance when N is small, or when N = 1? Let me make three points to start . . . (1) Small-N scholars routinely debate the importance of causes associated with different time periods. For example: Yashar and Mahoney on Central America; Precolonial, colonial, and postcolonial arguments about the NICs; and “Critical juncture” arguments more generally. (2) Scholars have trouble deciding if and when an intervening cause should be treated as more important than the original cause. For example: Colonialism => Developmental State => Sustained High Growth in Korea and Taiwan (3) Are proximate or historical causes more important in general? Elster: The best explanation is the one that identifies the causes most proximate to the outcome. Diamond: An ultimate explanation is the one that identifies the original historical causes. The Method Sequence Elaboration (Mahoney, Koivu, and Kimball 2009) Sequence Elaboration: Like Lazarsfeld’s “elaboration model” but applies to the kinds of causes used in historical explanation. Basic idea: Start with an initial two factor relationship (e.g., X =n=> Y) and elaborate it by introducing new antecedent and/or intervening causes. With correlations, we know that controlling for a third variable can influence how we think about the original relationship. The timing of that third variable is crucial. For example: imagine that strong positive correlation drops to zero when we control for a third variable. It makes a big difference if that control variable is antecedent or intervening. What happens when we start with a two variable relationship that is expressed in terms of the logical types of causes (e.g., X is necessary for Y), and then introduce a third variable (either antecedent or intervening) that is also a logical type of cause? For example: Start with X =n=> Y. Then add an antecedent cause: Z –?–> X –n–> Y ║--------?----------↑ One possibility is: Z –n–> X –n–> Y ║--------n----------↑ Z –n–> X –n–> Y ║--------n----------↑ Downing: Z = Roman Empire, X = medieval constitutionalism, and Y = early democracy. Question: How do the new relationships affect our understanding of the original relationship? What is more important, X or Z? The Roman Empire or medieval constitutionalism? Answer: X is the more important cause (from a logical perspective). Z –n–> X –n–> Y ║--------n----------↑ Z Z X Z X Y Y Z –n–> X X –n–>Y Z –n–> Y X Y Z –n–> X –n–> Y ║--------n---------↑ Z X Y We call this “contextualizing” the original X-Y finding. We learn something about the causes of our original cause. As a Rule: In a sequence of linked necessary causes, the importance a cause increases as it becomes more temporally proximate to the outcome. “For want of a nail, a horseshoe was lost. For want of a horseshoe, a horse was lost. For want of a horse, a rider was lost. For want of a rider, the battle was lost. For want of a battle, the war was lost.” This approach identifies logical or empirical importance. It does not tell us about theoretical or normative importance. Some factors (e.g., slavery; US intervention; ideas) may be important as causes because they raise critical normative issues, or because they are associated with important theoretical traditions. To what does the size of a circle correspond in this example? Possible answer: (1) The real case/s in which the necessary causes and the outcome are present (this determines the size of the outcome circle); (2) Hypothetical/counterfactual cases in which the necessary cause is present but the outcome is not. The number of these cases can be inferred from the logical structure of the argument. Question: What would have happened to Downing’s argument if the Roman Empire was sufficient of medieval constitutionalism (and still necessary for the final outcome of democracy)? That is: Z –s–> X –n–> Y ║---------n---------↑ Z = Roman Empire, X = medieval constitutionalism, and Y = early democracy. Answer: the Roman Empire (Z) would be the more important cause. Z –s–> X –n–> Y ║--------n---------↑ X1 X1 Z1 Z1 –s–> X1 Z1 X1 Z1 Y1 Y1 X1 –n–>Y1 Y1 Z1 –n–> Y1 Z –s–> X –n–> Y ║--------n---------↑ Z X Z = Roman Empire X = medieval constitutionalism Y = early democracy. Y We call this: Diminishment. The importance of the original causal factor is diminished. Great merit of sequence elaboration: gives precise answers! This structure is common. One tries to “diminish” an initial necessary cause by: (a) Finding a new necessary cause; (b) Showing that the new necessary cause comes before the original one; (c) Showing that the new necessary cause is sufficient for the original one. Illogical Relationships: Not all combinations are logically possible. Sequence elaboration tells us when a set of causal claims is impossible. For example: Z –n–> X –n–> Y ║----------s---------↑ Z –n–> X –n–> Y ║----------s---------↑ Z1 Y1 Z1 X1 Y1 X1 Y1 Z1 Z1 –s–> X1 X1 –n–>Y1 Z1 –s–> Y1 X1 Z1 Z –n–> X –n–> Y ║----------s---------↑ This doesn’t work; can’t be done. Several other interesting types exist: (1) When we introduce an intervening cause, the original cause may be less important. We call this discovering the mechanism through which the original cause exerts its effect. Example from Waldner (1999) X–s–>Y X–s–>Z–s–>Y ║-------s-------↑ Z is a more important cause than X. We call this identifying the mechanism through which X does its causal work. Y Z X = high elite conflict Z = nondevelopmental state Y = failed development X X = high elite conflict Y = failed development X Y X = high elite conflict Z = nondevelopmental state Y = failed development X Z Y (2) Of course, logically speaking, it is also possible for the intervening cause to be less important than the original cause. We call this identifying a partial mechanism. For example: X–n–>Y X–s–>Z–n–>Y ║------n--------↑ Z X Y For example: X–n–>Y X–s–>Z–n–>Y ║------n--------↑ X Z Y We also distinguish between: (3) Background Contextualization; and (4) Pathway Contextualization. So let’s do it. Background Contextualization (Example from Mahoney 2001) X–s–>Y Z–n–>X–s–>Y ║------n--------↑ Z Y Y X Z provides a background within which X does its causal work. Z = partially centralized state/incipient export economy X = radical liberalism (critical juncture) Y = military authoritarianism Background Contextualization Z–n–>X–s–>Y ║------n--------↑ Z X Y Pathway Contextualization Luebbert: ~LL =n=>SD ~LL –n–>RG –s–>SD ║--------n----------↑ ~LL SD RG The intervening cause identifies one pathway through which ~LL exerts its necessary causal effect. LL = liberal-labor alliance; RG = red-green alliance; SD = social democracy. ~LL –n–>RG –s–>SD ║--------n----------↑ ~LL RG Pathway Contextualization SD (1) There is a whole world of methodology out there waiting to be developed. It is based in the philosophy of logic, not statistical/probability theory. But most graduate students learn next to nothing about formal logic. (2) This world of methodology could have profound implications for statistical analysis. A unified theory of causality might resolve major methodological problems in the social sciences.