There are two dominant approaches to research design in the political and social sciences: statistical-oriented and case-oriented. The former is an extremely useful tool for identifying correlations, and the latter for teasing out causal mechanisms.
Imagine you were interested in trying to assess the effect of changes in diet on general health, and you had access to an enormous dataset collected by the world health organization. What methodological approach would you use?
Statistics are a powerful tool for population oriented studies.
You can slice and dice all the relevant information into variables and capture what is relevant with a few indicators. You could do this with a series of long interviews, but how much more relevant information would you receive?
Statistics are most useful when researchers are trying to explain marginal changes and marginal effects. You can find out how a change in lifestyle (more exercise and less alcohol) might effect a change in a given outcome (longevity).
Causes of effects
Now imagine you want to explain why some countries revert to being authoritarian regimes after a period of stable democracy. There are not very many cases of this in history. But Weimer Germany is the obvious example worth considering.
The USA today might be a case in the future.
Small N case studies contribute to social science because they can explain the complex configuration of factors that interact to produce a given outcome. This is what’s described as a “causes of effect” approach. Population studies are interested in “effects of causes”.
In the case of Germany, many different factors have been identified: the first world war, the Versailles treaty, unemployment, inflation, conflict between communists and social democrats, a polarizing class structure, culture of anti-semitism, rise of the Nazi party.
The complexity and uniqueness of case studies is what makes them so interesting for researchers. There are multiple pathways for a democracy to collapse (equifinality), and the integrity of the configuration of factors that make up the case is crucial.
Case studies, therefore, are not very useful for generalizations.
However, if you want to explain why mass genocides occur, then the number cases (N) increases – Rwanda, Balkans. You can then subtract the specific conditions that apply to Germany and identify more general trends and characteristics that all of these cases share.
Statistical and configurational case studies differ along 4 key lines:
- the type of explanation (marginal versus holistic),
- the unit of observation (variables versus cases),
- the method of answering (additive versus conditions),
- and the type of answers they provide (general versus specific).
Sometimes both methods can be combined (identifying a strong statistical relationship or correlation via regression analysis) and then using paired case studies to tease out the causal mechanism. But they cannot always be combined.
This course is qualitative methods in political science , so we are more interested in the type of causal arguments that the case study approach can provide.
What is it a case of?
There is one fundamental rule or criteria that case study researchers tend to follow in the process of research design: ensuring that the case sheds light on a broader question, or universe of cases. To put it another way, always ask yourself – what is this a case of?
For example, if you are interested in examining why Mumbai has the largest red light district in Asia, frame and justify the case selection of Mumbai against a wider universe of cases such that it speaks to a broader theoretical literature on prostitution.
It is also crucial to consider the dimensions of your case.
If your research question can be captured in one dimension then two well chosen cases is enough. If it requires two dimensions, you can use four cases.
Political scientists usually capture these in a 2 x 2 table. For example, see table 1
Table 1: Modes of Wage-Setting and Political Economy
Cases do not necessarily have to be countries or nation-states. They can be economic sectors, time periods, institutions, regions or even individuals.
Time and history
Another distinguishing feature of causal inference in the case study method is sensitivity to time. As Paul Pierson (1996) argues, all politics takes place in time.
Some statistical analyses are timeless, in that they attempt to identify the marginal effects of a individual variable across time and space.
Pierson outlines four core points to consider when thinking about causality in time:
- Sequence (does it matter if property rights are constitutionally enshrined before the state privatizes public assets?)
- Timing (does it matter when a country democratizes? For example southern Europe in the 1970’s and central Europe in the 1990’s)
- Asymmetry (if social democrats created the welfare state does that then imply that a decline in social democracy leads to retrenchment of the welfare state, or does the politics of retrenchment have its own causal dynamic)?
- Change (is change in form the same thing as change in function, how do we consider inconsequential change from critical change?)
Statistical and case-study oriented research designs have different understandings of causation, sometimes complementary but not always.
Goldthorpe (2001) identifies three different understandings of causation in the social sciences (robust dependence, consequential manipulation and generative processes).
“Causation is not correlation”. This states an obvious fact, namely that an association between the variables x and y does not imply X caused y.
Causation as robust dependence attempts to solve this problem through the use of various statistical inference to detect spurious causation. X is a cause of Y to the extent that dependence of Y on X can be shown to be robust.
That is, it cannot be eliminated through introducing other variables into the analysis.
A lot of political science has critiqued this approach by arguing that such techniques can show relations among variables but not necessarily that these relations are actually produced. They can forecast but they cannot explain.
Statistical inference is not causal inference.
Income is dependent on educational levels, and therefore the dependence is robust. But why? How does this dependence come about? Is it about the supply and demand for skills. To establish a causal link requires specifying the relationship within a theory.
Causation as consequential manipulation attempts to establish causal inference through experimental methods. Causes must serve as treatments that are manipulable. If X is manipulated then it must have a consequential response or effect on Y.
Only a randomized experiment can verify if Y is a consequential effect of X. How useful is this understanding of causation when explaining political and social phenomena?
Case study researchers are usually interested in explaining why X has causal significance for Y, and that the association must be generated through some sort of causal mechanism, even if it is not directly observable. They are interested in the causes of effects.
Case studies and causal mechanisms
Causation as a generative process usually takes place via three steps in research design: establishing the phenomena to be explained, hypothesizing a generative process or causal mechanism, and testing the hypotheses.
For example, in the study of political economy, supply-side factors such as declining interest rates and financial liberalization are associated with housing-asset price booms (explananda). But does this explain the wide variation in house prices in the OECD?
An alternative explanation might be income growth and wage setting institutions (mechanism) explain housing inflation. This might lead to a particular hypothesis on the extent to which sectoral-level interests shape asset-inflationary outcomes (hypothesis).
Necessary and sufficient conditions
In conclusion, political science researchers conceptualize causation in contrasting ways when they pursue explanation in particular cases versus large populations.
Mathematically, both research traditions stem from different understandings of causality. One originates in the study of linear algebra, and the other, boolean logic.
But is there a contradiction in conceptualizing causation as a configurational process that generates particular outcomes within specific cases versus causation as a statistical probability that exists across all populations? Can there be a unified theory of causality?
For example, what causes democracy?
The conditions that caused democracy in India (all encompassing mass party) should probabilistically decrease democracy in general. So what causes democracy in this case?
The all encompassing party turns out to be a necessary causal variable in India.
Causal inference in case studies are grounded in a philosophy of logic rather than the logic of probability. Logic identified necessary and sufficient causes for an outcome.
For example in political science, Barrington Moore (1966) famously quipped “no bourgeoise, no democracy”. The presence of a middle class is considered a necessary cause for democracy. But the presence of a middle class is not a sufficient cause in itself.
Think about this in terms of formal logic (see James Mahoney 2008):
Y1 = X1 & (A1 v B1)
Y1 = democratic pathway; X1 = strong bourgeoisie; A1 = alliance between bourgeoise and aristocracy; and B1 = weak aristocracy.
A1 and B1 is neither individually necessary nor sufficient. Instead they combine with X1 to produce two possible combinations for Y1. A1 and B1 are “causes” that combine with X1 to form two possible combinations that are sufficient to produce Y1.