Randomization

Randomization (also random allocation , word origin of randomize from English randomize , too random for "random, aimless, random, arbitrary") is a procedure in which the test subjects (e.g. participating patients) are assigned to different groups using a random mechanism . In this way, known and unknown person-related disturbance variables should be distributed equally across experimental and control groupsbe distributed. On average, the same prerequisites or test conditions should be created in order to achieve a higher statistical certainty despite the disturbance variables. The method is used, among other things, in psychological experiments ( field and laboratory experiments ). The aim of the procedure is to exclude alternative explanations and to reduce the probability that the effect proven in an effectiveness test is subject to systematic bias .

General

RA Fisher developed the principle of "randomization" as a consequence of the " ceteris paribus clause ". The experimental treatment conditions are assigned to the test groups, which in turn are assigned to the test persons at random (“randomized”). This excludes bogus declarations after which z. For example, behavior is referred to as an effect of the experimental treatment that has actually already existed pre-experimentally - it was not the new teaching method that led to the better results, the subjects in this test group had a learning advantage even before the investigation. The degree to which it is actually randomized is a characteristic used to distinguish the types of experiment .

Types of randomization

In experimental planning, randomization is often implicitly used to describe the random allocation of test subjects to different test groups, although randomization can be differentiated in a much more differentiated manner:

Randomization ( random allocation )
refers to the random allocation of test subjects to different groups.
Drawing of random samples ( random sampling )
describes, on the other hand, the random drawing of the test persons from the population, see also random sample .
Random assignment ( random assignment )
means the random assignment of the test persons to different test conditions (e.g. to control or one of the experimental conditions). Just in case studies, or studies with very small sample sizes, the subjects are often not groups, but only experimental conditions assigned because they are often run by only one Family (z. B. in an experimental design with multiple baseline ( English multiple baseline design ) ).

Proof of effectiveness

Clinical studies are carried out in order to generalize a statement of the study result to the population while considering the effectiveness of a treatment in the sample . Before the start of the study, various information must be provided in the test plan , such as proof of effectiveness .

Statistical aspect

An effectiveness test (e.g. significance test ) is carried out on the basis of the data obtained in the study to determine an effect. If an effect has now been found, it may have occurred by chance, correspond to reality (i.e. also found in the population ), or it may have been influenced by a systematic bias.

A systematic bias may exist, for example, if the treating investigator gets along particularly well with a treatment method. This falsifies the effect, as this relationship is not representative. The attending physician is therefore an influencing factor for the effect. Another influencing factor could be a patient's age. If, by chance, only patients who are older than a certain age are included in the study for treatment A and who are below a certain age for B, this can significantly influence the effect. A distinction is made between the following types of systematic bias:

• Selection bias :The next treatment allocation can be determined and influences the effect. Improvement: unpredictable randomization process
• Accidental bias : The balance between the various influencing factors is not taken into account. Influencing factors can be prognostic factors or risk factors . Improvement: Stratifying Procedure.
• Information bias : Knowledge of the method used has an indirect influence on the result, as those involved (patient, investigator, statistician) tend to prefer the treatment that they think is better due to the information. Improvement: blinding of all those involved.
• Concealment bias : Theinvestigator'sknowledge of the procedure to be used prior to treatment can influence the outcome of the treatment itself. Improvement: blinding the investigator.

In order to be able to validate that an effect proven with the test would (most likely) also be found in the population , all influencing factors that could cause systematic bias must be taken into account. There are various randomization methods for this. With a randomization process, a randomization list can be available prior to the study (block randomization, stratified randomization) or it can be based on properties of the randomized samples. The latter is calculated dynamically during the study (adaptive randomization, stratified randomization).

Organizational aspect

Even if the type of randomization were ideal from a statistical point of view, there may be a systematic bias. With block randomization , randomization boxes are often created for each participating center. In such a box there are consecutively numbered envelopes that have to be sealed so that nobody can see them and thereby influence the patient's choice, for example.

There are two methods of randomization:

• centralized randomization: allocations via the internet or by telephone for each center.
• Decentralized randomization: Allocation with the help of the mentioned randomization boxes.

Randomization process

There are several ways in which treatments are assigned to patients to perform a randomization.

Evaluation factors

The evaluation of a procedure is based on the following factors:

• Predictability : A treating investigator cannot foresee the next assignment of a treatment type (otherwise he could influence the result).
• Poised awareness : Bzgl. different factors such as the type of treatment itself should be assigned equally strong. In order to be able to carry out an interim evaluation during the study, the balance should be present at all times if possible.

Quasi-randomization

In the case of quasi-randomization , the allocation is determined by a mechanism that cannot be ascribed to any real randomness, e.g. B. by an alternating allocation. The all-important arrival of a patient at the center is not a valid random mechanism. The balance is guaranteed, but the next allocation is easy to see.

Simple, unrestricted randomization

Allocation of treatment to a new patient is random without any restrictions. With two treatments this corresponds to a coin toss for each patient without paying attention to balance (this is only statistically guaranteed from approx. 1000 samples). Accordingly, poor predictability is guaranteed.

Block randomization, balanced randomization

In order to be able to guarantee a better balance compared to simple randomization , it is ensured that in N patients there is a certain predefined ratio between the assigned treatments (e.g. 1: 1 for two types of treatment). The N (and unknown) patients to be included are divided into blocks before the start of the study. The ratio of the permuted treatment types is then taken into account in each block. The randomization list then consists of the compositions of the individual blocks.

The simplest variant consists of just one block. Balance is then only guaranteed at the end of the study, but predictability is low.

• Permuted blocks of the same length : If several generated blocks are used, these are arranged randomly within the randomization list. By using several blocks of the same length, there is a better balance over the entire recruiting period, but at the expense of predictability. If all blocks have a length of 4, this can be determined by an observer. With the knowledge of the block length, it is then easy to determine which treatment assignment is being made when using the next block. For example, if the first two assignments already specified treatment type A, B must now be assigned twice.
• Permuted blocks of variable length : In order that the length of the blocks cannot be recognized and the treatment allocation cannot be easily predicted, permuted blocks of variable length can be used. I.e. Blocks of different lengths are generated, the assignment definitions of which are determined randomly and which are randomly strung together within the randomization list.
• Stratified randomization : The methods described so far only take into account the balance with regard to the type of treatment. In order to be able to balance the influencing factors, a randomization list can be created for each combination of factors (before the start of the study). I.e. with two influencing factors (each with two characteristics) there would be 4 lists for each type of treatment and thus eight lists for two types of treatment. The influencing factors and the type of treatment itself are referred to as strata . The individual randomization lists can be calculated by any of the above methods. Stratified randomization should be used for small studies (less than 50 samples) and multicenter studies. A central randomization is necessary for the implementation.

With adaptive randomization , the treatment is allocated with a probability that depends on the previous distribution of the treatments in the individual strata . The data from the randomized samples are taken into account. The randomization thus takes place dynamically during the study.

• Baseline adaptive randomization : With this adaptive randomization, the number of previous allocations is taken into account and thus the probability of the next allocation is influenced. If the treatment methods have been assigned the same number of times, the probability for the method is the same. If a method has been assigned less often, the assignment to the current patient is higher. There are more complex methods for calculating both the probability and the decision size (instead of the difference in the number of assignments). An example of this is the biased coin method . This can improve balance. Predictability is limited by the likelihood of allocation. However, this method can only be used if no further strata have to be included. Very important for open studies (without blinding) so that the predictability is minimized.
• Minimization Method: The Minimization Method is a stratifying method for Baseline Adaptive Randomization . For a patient with fixed values ​​for each stratum, the sum of the respective values ​​for the treatment types is calculated. If the patient in center 1 is male and over 50 years old, then for treatment A it is determined how many treatments were carried out by A in center 1, how many were male and how many over 50 years. This calculation is carried out in the same way for treatment type B. The patient is then assigned the type of treatment (with a higher probability) that has the lower sum. Thus, a balance with regard to all strata is taken into account. The good balance is offset by the equally good predictability.
• Response adaptive randomization : In contrast to baseline adaptive randomization , the allocation is clearly determined by the strata that have already been measured. So no probability is defined depending on the strata.
• Play-the-winner : Successful treatment with one method means that it is also applied to the next patient. A change only takes place if the application was not successful. The poor balance and good predictability contrasts with the fact that the better method was probably used more often (patient benefit).

Choice of procedure

For large studies, block randomization can be selected if it is not a multi-center study. Then a stratified randomization is certainly necessary. If there are several strata, an adaptive procedure should be selected, otherwise a large number of randomization lists would have to be managed (product of the strata characteristics). Overall, however, the number of strata should be kept low.

Blinding

A distinction is made between blind, double-blind and triple-blind studies. In a blind study , only the patient does not know which treatment alternative he is getting. However, it is also important that the treating doctor does not know which patient is being treated with which drug. This procedure is known as a double-blind study. In order to preserve the objectivity of the data evaluation, this can also be carried out without knowledge of the treatment that has taken place, in which case a triple blind study is available.

However, emergencies - for example severe side effects - can make it necessary for individual test subjects to identify their assignment to the study groups prematurely (so-called unblinding ).

Treatment diffusion

Treatment diffusion, when planning a psychological experiment, describes the blurring of the boundary between the control and the experimental group. The treatment - the treatment whose effectiveness is to be demonstrated - is not only applied to the experimental group, but also sometimes unintentionally to the control group. Thus, the resulting effect can no longer be clearly traced back to the treatment during the evaluation and can therefore be traced back to a faulty randomization.

Examples

• A control group is treated according to the usual psychotherapy method, while the experimental group is treated using a new method. Both groups are cared for by the same psychotherapist. As he is convinced of the new (yet to be checked) method, he unconsciously also applies some of the new method's procedures to the control group.
• In an orthopedic department of a clinic, the patients are divided into a control group and an experimental group. The control group receives the usual therapy, the experimental group is instructed to independently do additional exercises (e.g. consciously tensing and loosening muscle groups) several times a day. The patients talk to each other and exchange ideas about the forms of therapy, whereupon some patients in the control group also start doing these exercises because they expect a better course of therapy.

literature

• International Committee on Harmonization, ICH guidelines (E8 General Considerations for Clinical Trials)
• LM Friedman, CD Furberg, DeMets DL (1999). Fundamentals of Clinical Trials. Springer, Heidelberg.
• Musahl H.-P. & Schwennen C. (2000) Design of experiments in: Lexicon of the Red.: Gerd Wenninger - Heidelberg: Spectrum Akad. Verl.
• Markus Pospeschill: Empirical Methods in Psychology . tape 4010 . UTB, Munich 2013, ISBN 978-3-8252-4010-3 .
• Jürgen Bortz, Nicola Dörig: Research methods and evaluation for human and social scientists. 4th edition. Springer, Heidelberg 2006, ISBN 3-540-33305-3 .