Covariate

Research Design

Although covariates are neither the design variable (i.e., the independent variable) nor the primary outcome (e.g., the dependent variable) in research, they are still explanatory variables that may be manipulated through experiment design so that their effect can be eliminated or minimized. Manipulation of covariates is particularly popular in controlled experiments. Many techniques can be used for this purpose. An example is to fix the covariates as constants across all experimental treatments so that their effects are exerted uniformly and can be canceled out. Another technique is through randomization of experimental units when assigning them to the different experimental treatments. Key advantages of randomization are

(a) to control for important known and unknown factors (the control for unknown factors is especially significant) so that all covariate effects are minimized and all experimental units are statistically comparable on the mean across treatments, (b) to reduce or eliminate both intentional and unintentional human biases during the experiment, and (c) to properly evaluate error effects on the experiment because of the sound probabilistic theory that underlies the randomization. Randomization can be done to all experimental units at once or done to experimental units within a block. Blocking is a technique used in experimental design to further reduce the variability in experimental conditions or experimental units. Experimental units are divided into groups called blocks, and within a group, experimental units (or conditions) are assumed to be homogeneous, although they differ between groups.

However ideal, there is no guarantee that randomization eliminates all covariate effects. Even if it could remove all covariate effects, randomization may not always be feasible due to various constraints in an experiment. In most circumstances, covariates, by their nature, are not controllable through experiment designs. They are therefore not manipulated and allowed to vary naturally among experimental units across treatments. Under such circumstances, their value is often observed, together with the value of the dependent variables. The observation can be made either before, after, or during the experiment, depending on the nature of the covariates and their influence on the dependent variables. The value of a covariate may be measured prior to the administration of experimental treatments if the status of the covariate before entering into the experiment is important or if its value changes during the experiment. If the covariate is not affected by the experimental treatments, it may be measured after the experiment. The researcher, however, should be mindful that measuring a covariate after an experiment is done carries substantial risks unless there is strong evidence to support such an assumption. In the hypothetical nutrition study example given below, the initial height and weight of pupils are not covariates that can be measured after the experiment is carried out. The reason is that both height and weight are the response variables of the experiment, and they are influenced by the experimental treatments. In other circumstances, the value of the covariate is continuously monitored, along with the dependent variable, during an experiment. An example may be the yearly mean of ocean temperatures in a long-term study by R. J. Beamish and D. R. Bouillon of the relationship between the quotas of salmon fish harvested in the prior year and the number of salmon fish returned to the spawning ground of the rivers the following year, as prior research has shown that ocean temperature changes bear considerable influence on the life of salmon fish.

...