Phenomena in nature are often complex and are influenced by many different factors. Systematic investigations in form of scientific studies are conducted to understand and describe their interrelationships which lead to their expression. For this purpose, scientists use specific strategies and study designs so that they exclude perturbating factors and arrive at an evidence-based result.
1. Descriptive Study Designs
Descriptive study designs are scientific observations, that means they include observing, measuring, and describing of phenomena without any kind of intervention. These are often used when a natural behaviour without artificial influences is to be scientifically investigated. In general, you can distinguish between four different descriptive study designs:
Longitudinal designs are usually (but not always) used in dependent studies, i.e. these kind of studies, one sample is measured or observed repeatedly over a longer period of time. The benefit of longitudinal studies is that they can give insight to intraindividual changes over time without genetical influences. Disadvantages are high expenditure of time and money as well as the risk of dropouts between measurements. Furthermore, there is the danger of the learning effect, where the test subjects become better at a task because they get used to the test conditions over time.
Cross-sectional designs are carried out in independent studies. Here, only one measurement per sample at one point in time is carried out. For example, different age groups of subjects are investigated in the same time span. From that, conclusions about typical phenomena in one stage of the development can be made. This type of study requires less effort (time and money) than longitudinal studies. However, correlations between characteristics cannot be interpreted as causal, since even unrecorded variables, such as the cohort effect, can influence the result.
The mixed longitudinal study design is a mixture of longitudinal and cross-sectional. Different groups are measured over a (shorter) period of time. The advantage of this combination is that the duration of the study can be quite short, whereas the effective age span investigated is not reduced. However, individual effects in the development are investigated over a shorter period.
In time-lag studies, identical groups are measured in different points in time. So, this type of study will investigate how variables changes over time in the same group (e.g. Did the movement behaviour of 14-year-olds change between 2000 and 2020?). Similar to longitudinal studies, the expenditure of time and money is very high. However, there is no danger of dropouts, since each sample is only examined once.
2. Experimental Study Designs
The goal of an experiment is to determine the influence of one of the independent variables by changing it while keeping the other independent variables as controlled and constant as possible. The change in the dependent variable is then measured and a statement about the influence of the independent variable on the overall system can be obtained.
With scientific experiments, a few basic principles must be followed:
Three main quality requirements for scientific research needs to be met. These are:
1. Validity (Does the test measures what it is supposed to measure?)
2. Objectivity (Is the measurement outcome independent from the test executer?)
3. Reliability (Is the outcome of the measurement replicable under identical conditions?)
When conducting experiments, homogeneous cohorts should be selected in order to achieve a high degree of comparability. Therefore, inclusion and exclusion criteria for subjects are usually introduced in scientific studies.
Although a high degree of comparability within a study should be made possible, generalizability must not suffer. If the inclusion and exclusion criteria are too narrowly defined, the study results can only be related to this one sample and projection to the entire population becomes more difficult.
By replicating the measurements of an experiment, the experimental error can be estimated, and the effect of non-systematic errors can be reduced.
In scientific experiments, different conditions should be measured in a randomized or in a blocked order. This allows for an unbiased estimation of the effect as well as an independency of the results from perturbing factors. Also, the estimated error can be obtained unbiased and the normalization of the data is generally higher. By that, unwanted and unknown systems of correlations are prevented which leads to an elimination of systematic and non-systematic errors.
Often in scientific investigations, not only the independent variable influences the dependent variable but also uncontrollable perturbating factors. If these are known and measurable, an analysis of covariance can be carried out. If these are known but not measurable, the groups of the experiments should be divided into blocks that should agree as much as possible with the perturbing factor. If the factors are unknown and not measurable, randomization of the group allocation and replication of the experiment lead to an elimination of systematic and non-systematic errors.
There are many different types of experimental setups and study designs in science. These differ in their applicability and quality. Here, we want to give you a quick overview over them. We will explain the different designs to you by means of an example study, so that the differences become clear. Starting with the simplest design, we will then move over to more complex, but higher quality ones.
“The maximum power output during cycling is higher with a 60° hip angle than with a 50° hip angle”
With this hypothesis, we get one independent variable with two levels (L1: 60° hip angle and L2: 50° hip angle) as well as one dependent variable (DV: maximum power output).
The sample is not divided into any groups. The whole sample is first tested under condition L1 and then under L2. This design is not really adequate since the order effect may influence the results. This can be either a learning effect, where participants perform better in the second trial because they are more familiar with the test procedure or a fatigue effect, where the participants perform worse due to muscle fatigue. Also, there is no control group which counteracts these effects.
The sample is not split into any groups. Each test person is subjected to a total of four measurements. First L1, then L2 and after that both conditions again in reverse order, i.e. L2 and then L1. This mirroring of the Repeated Measures Design eliminates order effects. However, this method is only applicable, if the order of tests is not relevant (which is not in our example).
The sample is split up into two groups. Group A is tested with condition L1 only and Group B is tested with L2 only. Because two groups are formed here, more test persons than in the Repeated Measures Design are needed. However, we will not see order effects in this design since each subject is only tested once. In this method, group assignments are critical because both groups could be very different from each other, making it impossible to compare them.
This design is similar to the Independent Measures Design, with the difference that the subjects are allocated randomly to groups A and B. This leads to an elimination of uncontrollable perturbation factors and a higher comparability.
Also this design is similar to the Independent Measures Design, with the difference that the subjects are allocated to groups A and B according to the results of a pre-test which is performed before the actual measurements (e.g. according to maximum power at preferred hip angle). The results of the pre-test are ranked, and the groups are divided in such a way that both groups have a similar result distribution (matched pairs). Thus, one of the perturbating factors can be eliminated and the comparability is increased.
The sample is split into two groups. The subjects are allocated randomly to Group A and to Group B. Group A is first tested with condition L1 and then with L2. Group B is tested in the opposite order.
With the randomized allocation, perturbation factors can be eliminated. Furthermore, by the cross over design, a learning effect can be precluded while not reducing the comparability of the results. However, this method is only applicable, if the order of tests is not relevant (which is not in our example).
Similar to the Balanced Repeated Measures Design but the groups are allocated by the matched pairs method of the Parallelized Group Design.
In science, however, there is often not only one independent variable, but several, which do not simply affect the dependent variable in isolation, but also have interaction effects. If we come back to our bicycle example, another independent variable would be the saddle height for example, which of course also influences the hip angle as well as the maximum power output. If you would now perform two isolated experiments on both variables, you would simply ignore the interaction both of them. To avoid this, so-called factorial experiments are carried out.
In our example we have two independent variables, which are also called factors (hip angle and saddle height) with two levels each (hip angle: 60° and 50° hip angle; saddle height: 85% leg length and 95% leg length):
|F1: Hip angle|
|F2: Saddle height|
85% leg length
95% leg length
In factorial experiments, all factors are measured at all levels. The number of individual measurements thus results from the number of factors (n) and the number of levels (k): kn. In our case it is as 22 = 4. Here, usually a Randomized Trial Design is selected, so that the sample is split up randomly into kn groups.
As mentioned several times in this article, subjects are often randomly assigned to individual experimental groups. This minimizes the differences between those groups by distributing persons with certain characteristics equally by use of the probability theory. Thereby, unbiased measurements are possible, and blinding will be facilitated. Also, the statistical power increases by randomization. A general distinction is made between three different types of randomization:
The simple randomization is an intuitive procedure such as coin tossing and is therefore also called coin randomization. Hereby, selection and allocation biases can already be excluded. Especially in small studies, however, it may happen that group sizes differ a lot due to random allocation. The larger the sample, the higher the probability that the groups are of equal size.
The restricted randomization method serves the purpose of keeping the group sizes equal. It is also called permuted-block randomization. Here, the block size and allocation ratio are given, and the randomization happens within these boundaries.
The adaptive randomization serves the purpose of keeping group sizes equal as well. In contrast to the restricted randomization, there are no fixed boundaries. Rather here, the probability of being assigned to a group is constantly adapted. That means, it decreases if the group is overrepresented and increases if the group is under-represented. Therefore, it is called adaptive biased-coin randomization.