Reductionist = "bottom-up":
Most of the experiments and observations in science are likely to be reductionist in nature. The approach often is to concentrate on parts of systems, e.g. cells or molecules etc., and describe their properties and thus what they contribute to the overall biological functioning of the system. Unfortunately, if such investigations are done in isolation from the rest of the system, the behaviour of the studied component may be different from when it is interacting with the others. The interactions between the components are equally important and should not be overlooked when taking a reductionist approach. (cf. Neurophysiology at Summer School)
Holistic = "top-down":
In contrast, a holistic investigation examines the behaviour and properties of the intact biological system, e.g. you as an individual organism. Such an approach recognises the importance of the interactions between system components as well as their individual properties; it also recognises that the properties and actions of a component may be very different in the "environment" of the integrated system than when they are studied in isolation, e.g. in vitro. The system is more than just the sum of its parts; it has emergent properties, which are only manifested when the intact system is operating as a whole entity.
It is, however, the case that by devising appropriate observations
and experimental approaches, that the behaviour of the system,
e.g. under different circumstances, can yield data permitting
inferences to be drawn about the behaviour of some of its components.
(cf. Perception at Summer School)
The shortest definition of an experiment = "What if....?"
You need to give operational definitions of your variables; a variable must be (a) able to vary, (b) show mutually exclusive states/values, that (c) you can perceive and record/measure.
Independent variable (Vind): values determined by the experimenter and otherwise invariant.
Dependent variable (Vdep): values recorded by the experimenter and supposed to be potentially affected ONLY by changes in the independent variable.
You are looking for a causal relationship between Vind and Vdep.
In other words, you want to be able to decide whether to accept the null hypothesis (H0), that any variation in Vdep is just random due purely to chance, or the alternative (experimental) hypothesis (H1) that variation in Vdep is because of the variation in Vind.
Unfortunately, extraneous variables may be unavoidable and may affect the dependent variable directly, or the causal relationship of Vind and Vdep, or (if you are not checking) cause variation in Vind B any of which could result in the unwanted influence of a confounding variable. The best way to eliminate these is to keep them constant, or alternative make sure they balance out. These confounding effects may be either random errors or systematic bias.
Within- and Between-subject experimental designs
In the context of holistic physiological/behavioural experiments, there is a marked contrast between -
within-subject designs (each subject is tested under all of the experimental conditions) and
between-subject designs (different subjects under different conditions). This might be forced on you by the nature of the test, but it means that you cannot avoid the criticism that any differences are due to the different subjects rather than the different conditions, i.e. complete confounding of subject and condition effects.
Your entire effort in designing any experiment is to isolate the independent and dependent variables and either eliminate or control for any effect due to confounding variables.
So, in a between subjects design, you would have to (a) have a very large sample size so that any subject-related variation is random (and hopefully does not mask out the effects of Vind), or (b) match the subjects in one group with those in another so that the variation between subjects is minimised.
In contrast, with a within-subjects design there is no subject variation between the conditions as they are the same subjects - so this can't be a confounding variable because it is constant.
The only snag with the within-subjects design is that (usually) you can't test subjects under both conditions at once, so one condition has to be before or after another. This introduces inevitable order effects, which can be controlled for by either randomisation or , preferably, structured counterbalancing.