Strategy for testing a proposition that expresses a probabilistic relation

researcher to observe how changes in the dependent concept follow in time upon changes in the independent concept. If a survey is not feasible, a comparative case study is the next best option see Box 9. In this type of case study the principles of a good survey are followed as closely as possible “quasi-survey case study”. This implies that a population is specified in which the proposition is tested, and that the sample is representative for that population and should be selected randomly. Box 9 How the survey can become a case study An essential characteristic of any survey is probability sampling, e.g. random sampling of instances from the population in which each instance of the population has an equal chance of being selected. This is the only guarantee that a co-variation that is observed in the group of observed instances in the sample also exists in other instances than those included in the sample. Probability sampling is only possible if the sampling frame is specified, i.e. if there is a list of members of a population or a set of directions for identifying each of them. Because there is never or very rarely a sampling frame for all members of an entire theoretical domain, a theory-testing survey is always conducted in a specified population of instances from within that domain. The propo- sition is tested in that population and this test will be followed by other tests in other populations in a replication strategy, in order to achieve generalizability to other parts of the domain. If no population of instances can be identified in the domain no sampling frame is available, it is not possible to test the proposition with a survey. However, this problem can be solved by specifying a smaller population within a domain for which a frame for probability sampling can be defined. It is, for instance, not likely that there is a sampling frame list of innovation projects in general, or of such projects in Europe, or in an economic sector in a country, but it is likely that there is a list of projects for which an EU subsidy was requested or a list of projects within a large company. Such often small populations are not “representative” of the domain, but no population ever is. A consumer behaviour theory, for instance, is always tested in a specific population of consumers say Rotterdam housewives or Toronto students and then replicated in other populations see Chapter 3.2.3. Another problem may then arise with such strategy: the number of available instances from the domain is too small for conducting a statistical analysis of the data, which is the main characteristic of a survey. This problem exists, for instance, in the field of compar- ative politics research when propositions about nations with specific characteristics

4.4.5 Testing more complex conceptual models

Our book focuses on relatively simple causal relations in which one concept causes another concept, which is the effect, as shown in Figure 3.1 of section 3.2 “Principles of theory-testing research”. More com- plex models can be tested as well, as shown in Box 10. One more com- plex and frequently used conceptual model is a model with several independent concepts causes, and one dependent concept effect. Such a model could represent a theory that all causes have, separately, a probabilistic relation with the effect, for example as shown in previous survey research. A further exploration of theory and practice could result in a belief that some factors, when present together, are more important for having the effect, than other combinations of factors. 84 e.g. “Islamic states” or oil-producing states need to be tested. The number of cases may also be small for practical reasons, e.g. if the measurement of the concepts is so time- and labour-intensive that measurements can be conducted in only a limited number of instances. In such situations, in which the number of instances is too small to conduct a statistical analysis, a qualitative analysis of the few instances available can be conducted. The survey has become a quasi-survey case study. Case selection is different in the two conditions mentioned. On the one hand, if statistical analysis is not possible because the population is too small, case selection is not necessary. The quasi-survey comparative case study will include all instances in that small population. The outcome of the study concerns that small population and gener- alization will be sought by replication in other possibly equally small populations from the domain. On the other hand, if a quasi-survey comparative case study is con- ducted because of limitations caused by intensive measurement procedures, instances must be selected from the population. In principle this should be done with probabil- ity sampling, which in practice usually is very difficult to achieve. A test in a quasi-survey, however, is not useful if an outcome in the sample cannot be generalized to the popu- lation. This is the main reason why we advise for all quasi-survey comparative case stud- ies i.e. for all tests of probabilistic propositions to select very small populations in each of which a census can be conducted. The outcome of a test of a probabilistic proposition in a very small population is useful as one test in a series of replications in other small and large populations in the domain, whereas the outcome of a test in a non-probability sample has no significance at all. Then the proposition could be formulated, for example, as a deter- ministic condition e.g. a necessary condition, and the combination of factors is then the independent concept, which is a necessary condi- tion for the effect. The complex model is reduced to again the simple model and testing the propositions is straightforward. Chapter 4 Box 10 More complex conceptual models Our book focuses on relatively simple causal relations in which one concept causes another concept, as shown in Figure A below. However, more complex models are possible as well. For example, it is possible that concept A has an effect on B via another “intervening” or “mediating” concept. A medi- ating concept is a concept that links the independent and the dependent concept in a proposition and which is necessary for the causal relation between the independent and the dependent concept to exist. This is shown in Figure B. First A affects C and then C affects B. Separate propositions can be formulated and tested about the relation between A and C, C and B, and A and B. It is also possible that a concept C has a moderating effect on the relation between A and B. A moderating concept is a concept that qualifies the relation between the independent and the dependent concept in a proposition. For example, the relation between A and B only exists or is stronger if C has a certain value. This is shown in Concept A cause Concept B effect Concept A cause Concept C mediator Concept B effect Figure A Simple causal relation between concept A and concept B Figure B A causal relation with the mediating also called intervening effect of concept C