Confounding

Nuisance (also third variable) is a term from the empiricism experiments. Are all the factors that may affect both the dependent variable and the independent variable and not be manipulated. These may be characteristics of subjects or external factors. Third variables are seen as alternative or competing explanations for the initial hypothesis of the research problem. For control of confounders, there are special techniques.

Under a confounder (English for, ' disruptive ', confundere from Latin: confuse, mix, pour together ) or Germanized Konfundierungseffekt ( see also " confused " ) is understood within epidemiological studies a problem, namely the two factors with the observation of exposure and the endpoint is related. A confounder is a variable that influenced the occurrence of a risk factor and observed at the same end point.

The observed exposure is not the sole cause of the observed effect - this is at least partly caused by a confounder.

• 6.1 Hawthorne effect

Species

In particular, in the social sciences, it is necessary to investigate the disturbances in more detail. This will make use of the variance. The variance is a measure of dispersion for formal representation. The variance is the square of the standard deviation. Mathematically, the variance of the average of the squared deviations of individual values ​​from the grand mean.

Conditions for confounding

In an experiment to test whether a random variable having impact on a random variable. But influenced not only by the well-known variable, an unknown confounding the random variable, then it is called confounding. If a third variable influences two random variables, then the causal interpretation of the effects will be distorted. Confounding is a possibility that distortion. It depends on two conditions:

Expressed mathematically, a regression confounded if:

Sources of confounding factors

Sources that affect the internal validity and external validity:

• Between temporal events ( events that additionally influence the next stimulus, the dependent variable, such as " Black Friday " )
• Maturation processes ( " intrapersonal " processes that are independent of the stimulus, such as development of a small child )
• Subjects - motivation, for example in the form of the effect of social desirability
• Effects of certain test conditions and methods ( engl. testing effects), see reactivity (Social Sciences)
• Aids ( change in meter ), which may also be the involuntary change of gestures when the experimenter
• Distorted selections and failures ( difference between control and experimental groups not only in relation to the stimulus, but also in other characteristics that influence the dependent variable)
• Experimenter effects.

Are confounding factors and stimulus mixed, one speaks of a confounding.

Common risk factor

An examination of the relationship between tobacco smoking and liver cirrhosis or hepatocellular carcinoma, one can notice a significant association. However, there is no biological connection: Smoking does not lead to liver cirrhosis. Rather, many drinkers are also smokers ( statistical association due to a higher-level common cause ( addictive personality) ) and alcohol consumption is an independent risk factor for liver cirrhosis. In this example would be alcohol addiction and personality confounders in the context of the measurement of the effect that smoking has on the outcome liver cirrhosis.

Addiction alcohol personality →          ↓ ↓        Smoking → ​​cirrhosis Testing of confounding

Measurement results are characterized by internal validity in the ideal case. This means that the dependent variable is actually declared by the research approach, including the independent variables. If the measurement is influenced by a confounding and distorted, so confounding is present, then the internal validity is not (any longer ) where.

It must be expected that the independent variable is affected by other variables, as it also influence the dependent variable. Through this overlay a detailed breakdown of the influences of the dependent variables is difficult or impossible.

To find out whether confounding is present and where necessary mitigate this, an examination of the model is necessary. However, there is no specific test for confounding, as test problems are tested asymptotically normally. For this large sample sizes necessary and statistical inaccuracies regarding the significance levels are to be expected. However, the empirically often hardly to be met prerequisite is that the disturbing influences defined, demarcated and can be measured reliably and validly (see the methodological hardly possible differentiation between different response tendencies in psychological assessment ).

Instead, the conditions for confounding be used. First you have a potential confounding that may be responsible for the confounding be found. Then it is important to test the two conditions for confounding. This is partly the stochastic dependence between the independent variable and the disturbance variable. The events and must be stochastically dependent. This can be checked, for example, with a test. The first condition is satisfied, in addition, the diversity can be checked with regard to the expected values ​​of the model. If the relationship between the independent variable and the dependent variable changes when the model is a potential confounding is added, the second condition for confounding is satisfied.

Confounding is present, therefore, and can be tested if both mentioned conditions are fulfilled, ie independent variable and confounding are stochastically dependent, and the expectation values ​​of the model with and without confounding each different size.

Control of confounding factors ( third variable control) and avoidance

If a confounding of two variables is found only in hindsight and the confounding variables were not collected in the experiment, the whole experiment is useless because it can not be definitively established by the independent variables on the dependent variable.

An effective way to prevent a confounding in advance is the randomization. Here, the subjects are assigned by a random process the different experimental conditions. This ensures that no systematic relationship between the dependent variable and potential confounders, such as certain persons properties is. However, randomization is only possible with real experiments in which the assignment of persons to the respective treatment groups is under the influence of the experimenter. For all other collection methods, such as quasi- experiments or field or pure observation method, a random allocation of subjects is not possible and the risk of confounding thus in principle available.

The application of a randomized assignment of subjects is purposeful only in large samples, as can be expected only when a sufficiently large sample size from an equal distribution within each group. In reality, the samples are often, however, due to economic considerations or other practical reasons rather small and randomization therefore not useful. In these cases, for example, by keeping the potential confounding variable attempts to confounding, thus preventing a distortion of the result. Another possibility is the balancing, in which the various forms of possible confounding be distributed equally among the experimental groups.

In the experimental forms in which it is not possible to influence the sample composition before, it is important that the experimenter is concerned about possible confounding variables in advance and this rises in the investigation. Only in this way can be afterward checked whether a conflation of two variables is present and confounding can be accounted for by statistical control techniques in the result.

In experiments, there are techniques for the control of confounding factors. These techniques are especially important in the social sciences. In the experiment, one can test and control group (s ) form, which serve to eliminate the influence of trial characteristics that may act as confounders. There are two methods for the formation of the groups:

• Randomization means that the assignment of subjects to experimental is done and control group randomly. This ensures that average out the differences between the experimental groups at a sufficiently large sample. Randomization is possible that it comes from the splitting of the subjects in the experimental and control group to systematic bias in the results.
• Parallelization or matching refers to methods for forming groups that are homogeneous with respect to one or more confounding factors interfering factor. If, for example, a teaching method are evaluated as similar as possible to student groups can be formed by two parallelization in terms of their notes.

In laboratory experiments, environmental factors can be controlled:

• Elimination refers to the elimination of potential confounding variables. Your goal is to ensure that the subjects, in addition to the independent variables as possible is not exposed to other factors. To ensure that the subject is not influenced by external events, experiments can be performed, for example in windowless, soundproof cabins.
• Constant: To ensure that the observed effect is due to the variation of the independent variable, an attempt is made to keep all other factors constant. Since the natural brightness often vary from day to day and throughout the day, eg tests for visual perception should be carried out in an over all experimental procedures across equal -lit laboratory.

Examples

Hawthorne effect

A famous example of the occurrence of confounding is the so-called Hawthorne experiment from the 1920s. The occurred in these group-based observational studies in the United States Hawthorne effect describes the influence of confounding variables on an experiment.

In the Hawthorne Works (Illinois, USA ), a telephone equipment -producing industrial company, was to motivate employees to higher production volumes in several experimental passages specifically altered the environment. Better lighting addition, each wall of the color was in the further steps of changing or increasing the room temperature. Immediately after each change in the short term could be observed an increased production rate, but after a few days went back to the initial level. Thus, no single change in the work environment led to a permanent increase in the production rate. Rather, a mixture of different variables, or the occurrence of a third variable ( confounding ), thus confounding was present. The increased work performance thus could be caused by a short-term increase in work motivation and not by improving lighting, explain the change of wall color, or to increase the room temperature.

Crozby calls this phenomenon as " third- variable problem ". He argues that as no direct correlation between the variables leisure activity and restlessness must exist, but possibly a higher income allows more time for extensive recreational activities. If income is the determining variable, there is no cause-and- effect chain between the studied variables can be established recreational activity and restlessness. The relationship is influenced by a third variable representing an alternative explanation for the observed effects.