Biology, images, analysis, design...
|"It has long been an axiom of mine that the little things are infinitely the most important" |
Types of studies
The degree to which we can infer that an association between two variables means that one variable actually causes changes in the other variable depends partly on the type of study. We can divide scientific studies on relationships into three main types: survey-type, observational and experimental.
A descriptive survey is where you sample the population of interest, and record one or more characteristics of that population - known as response variables. The only aim is to estimate (for example) the prevalence of a disease, or the population density of a pest. If you also record information on possible explanatory variables, the study becomes an analytical survey. Here you would look for relationships between the response and (possible) explanatory variables. The study can be done at a single point in time (a cross-sectional study) or over a period of time (a longitudinal study).
You may find there is a strong relationship between two variables. But with this type of study you do not have a strong case for arguing that changes in the (possible) explanatory variable really do cause changes in the response variable. This is because
An observational study is where you examine the effects of selected (possible) explanatory variables upon your chosen response variable, by selecting particular contrasting groups to study. Often one of these groups will be a control. You can select those groups in one of two ways:
Observational studies are better suited for exploring relationships than survey-type studies, because you are often testing a specific hypothesis, and because there is usually a control group. Moreover you have some control over confounding factors, at least those that you know about. However, you still have no control over confounding factors that you don't know about - since there is no random allocation to treatment. There are also still many sources of bias - especially in case-control studies. Hence observational studies still only permit weak inference for causality, albeit stronger than survey-type studies.
In an experimental study one or more of the explanatory variables are under the control of the experimenter. In addition, the different levels of the explanatory variable are randomly allocated as treatments to the different experimental units. In an observational study the level of the explanatory variable is usually self-selected by the individual concerned. Sometimes there may be manipulation, but if there is no random allocation we still refer to this as an observational study - although some use the term quasi-experiment for this type of study. The importance of random allocation in experiments is that we can have more confidence that a strong relationship indicates a causal link - in other words, there is a fairy strong inference for causality.
If one is trying to demonstrate a causal relationship between an explanatory and a response variable, then experimental studies are definitely the way to go. This is why randomized controlled trials now form the basis for assessment of everything from the efficacy of drug treatments, to the choice of fertilizers or insecticides. However, the experimental approach is not always possible for ethical and/or practical reasons. For example, it would be unethical to randomly allocate individuals to experience passive smoking for many years to assess whether it increased the chance of lung cancer. In conservation there has long been a debate over the effect of the size and shape of reserve areas on biodiversity. But it would be impractical to set up an experiment to test this, because the processes which determine diversity operate over very long time periods.
Although there are often very good reasons for doing observational rather than experimental studies, it remains true that it is nearly always better to manipulate and randomly allocate if you can! That way you reduce the chance that something other than your treatment is causing the results you observe. Also much less is learned about systems at equilibrium - manipulation shows how systems respond to changes.
Sometimes data are obtained from a multitude of sources, including studies designed for quite different purposes, and then analysed with a view to revealing any apparent trends and relationships within the data. This approach to research has been described as data trawling. One big danger with this sort of approach (as we show in