v2.1

The Experiment

RATIONALE AND BASIC DESIGN OF THE EXPERIMENT

from: Foundations of Social Research by Nan Lin, Chapter 14, pp. 245- 260. McGraw-Hill, 1976.
The main reason for conducting an experiment is to try to determine the potential effect of one variable (the independent variable) on another (the dependent variable), while eliminating or controlling all other variables which may confound such a relationship. Thus. a "good" experiment should have three basic elements: random assignment of respondents to experimental groups control over extraneous variables, and manipulation of the independent variable. The researcher uses randomization to ensure that the effects of other variables which may affect the independent and the dependent variables are evenly distributed in each and every experimental group, so that if they have any effect on the respondents they will affect all experimental groups similarly.

Other extraneous variables which are known or expected to have an effect on the independent variable, the dependent variable. or both, may be used as control variables and incorporated into the experimental design. For example, in a study of the effect of housing arrangement (integrated versus segregated) on racial prejudice, if the researcher expected that females tend to be less prejudiced than males, he could further analyze his findings for females and males separately in each housing arrangement. As most of the social activities and behaviors have multiple interlocking relationships. the more control variables brought into the design, the more chances the researcher has to demonstrate the effect of the independent variable on the dependent variable. For if in spite of the effects of the control variables on the dependent variable, either directly and independently or jointly with the independent variable, the independent variable still shows a significant effect on the dependent variable, the researcher then has greater confidence in the causal relation between the independent and the dependent variables.

Manipulation of the independent variable represents the most distinctive characteristic of the experiment. as compared to other methods of data collection. In the above mentioned example, the housing arrangement was the independent variable which was manipulated by the researcher-the residents were assigned to different housing developments. In reality, such manipulation may present many problems, but other variables are much easier to manipulate. For example, to assess the effect of source credibility on the believability of a message, the researcher can manipulate the identification of the sources who present the same message to two groups of subjects (as respondents are usually called in an experiment, because they are subjected to manipulation by the researcher). In delivering a message about how to feed young children, to one group of young mothers, the source person may identify herself as a nutritionist affiliated with a medical college and, to another, the same source may identify herself as a housewife with two young children. Many other variables can be manipulated in a similar fashion. The manipulation of the independent variable provides the researcher with an opportunity to vary the kind and the intensity of the values of the independent variable and to assess the effect of such variations on the dependent variable.

These three characteristics. randomization, control and manipulation, make the experiment a desirable method to use in testing specific hypotheses, because of its ability to single out the independent and dependent variables while eliminating and controlling the effects of other variables. The manipulation aspect becomes more attractive to a researcher for testing a causal relation between two variables.

The basic design of an experiment involves the following steps : 1) selection and random assignment of subjects to experimental groups, 2) pretest measurement of the dependent variable, 3) differential treatment (manipulation), and 4) posttest measurement of the dependent variable.

Strictly speaking, the selection of the subjects should also use a representative sampling plan. But in most experiments conducted by social scientists, voluntary subjects are recruited. The researcher then screens the volunteers for competence in receiving, understanding, and responding to the type of manipulation to be performed. Hopefully, any bias introduced by this voluntary recruitment procedure will be eliminated by randomly assigning the subjects to different experimental groups.

Then the subjects are interviewed or required to complete a questionnaire in which measurement of the dependent variable is made. In order to not sensitize the subjects to the purpose of the study, the actual instrument used contains many more questions than those dealing with the dependent variable. In fact, successful manipulation depends to a large extent on the ability of the researcher to camouflage the measurement of the dependent variable in the pretest measurement.

Usually, there is a time lag between the pretest measurement and the treatment (manipulation of the independent variable); should there be any lingering suspicion about the purpose of the study among some subjects, the time lag hopefully will diminish it to a minimum. The variations of the treatment can be either different values or categories of the independent variable, such as the integrated housing versus the segregated housing in our illustration. Or, they could represent the presence and the absence of the independent variable (e.g., one group is exposed to a message while the other is not).

Finally, after another time gap, the instrument administered in the pretest measurement is again administered to the subjects.

To measure the effect of the independent variable on the dependent variable, a difference score is computed between the pretest measurement and the posttest measurement of the dependent variable for each subject in each group. For example, in an experiment to test the differential effects of source credibility on attitude change toward racial prejudice, if a subject in the high-credibility source group registered a "2" response on a racial prejudice scale of 7 points ("1," most prejudice; "7," least prejudice) in the pretest measurement and a "5" response on the same scale in the posttest measurement, he receives a difference score of positive 3 (5 - 2 = +3). We may then compute the averaged difference score for each group, say, the high-credibility source group and the low-credibility source group. From now on, the term difference score is used to indicate the averaged difference score for each group. Ideally, the difference score should be interpreted as the change in the value of the dependent variable (e.g., attitude change toward racial prejudice) resulting from the treatment (manipulation of the values of the independent variable, e.g., high- and low-credibility sources). But, in reality, such an interpretation is faulty, for the difference score may actually represent the effects of a number of other factors involved in the experiment, in addition to the effect of the independent variable. ln the following sections, the various factors affecting the difference score, and elaborations of the basic design, will be discussed.

FACTORS AFFECTING THE DIFFERENCE SCORE

There are at least seven factors which may contribute to the difference score:

1. The treatment effect T. The effect of the treatment (the independent variable) on the difference score.

2 The pretest measurement effect X. The pretest measurement may sensitize the subjects to the manipulation and the subsequent posttest measurement, thus affecting the difference score.

3 The time effect U. Since the experiment is conducted over a period of time, time-related events and maturation may influence a subject's exposure to the treatment and his responses on the posttest measurement.Time-related factors cannot be controlled by the researcher: thus, they are also known as uncontrolled factors.

4 The interaction effect between the pretest measurement and the uncontrolled factors Ixu. Sensitization to the pretest measurement and factors related to time may jointly affect the posttest measurement scores of a respondent. Here, the symbol I is used to indicate interaction effects, X the pretest measurement effect, and U the time effect. For example, sensitization to the pretest measurement may not by itself affect the posttest measurement. But should some event occur between the pretest measurement and the posttest measurement, it may cause the respondent to recall the pretest measurement, resulting in certain pattern of response which otherwise would not have been formulated. An illustration might be a study focusing on attitudes toward blacks among whites. After the pretest measurement of such attitudes, along with a number of other attitude measurement, and preceding the posttest measurement, civil rights legislation has been passed into law by Congress. This event triggers the respondents' recall of the pretest measurement items and may result in more positive attitudes toward blacks in the posttest measurement than would be the case if the legislation had not occurred.

5 The interaction effect between the pretest measurement and the treatment IXT Sensitization to the pretest measurement and the exposure to the manipulation may jointly affect a subject's posttest measurement score.

6 The interaction effect between the treatment and the uncontrolled factors ITU. Exposure to the treatment and the time-related uncontrolled factors may jointly affect the posttest score of a subject.

7 The interaction effect among the pretest measurement, the treatment, and the uncontrolled factors IXTU. Finally, activities associated with the pretest measurement, the treatment, and the time-related uncontrolled factors may all jointly affect a subject's posttest score.

All the subjects in the basic experimental design discussed in the last section are subject to the influence of these factors. Further, these seven factors may independently affect the difference score. Thus, the difference score may represent the added effect of all seven factors. Using the symbol d to represent the difference score, a simple equation may be constructed to represent the relationship between d and the seven factors: namely, d is equal to the sum of the seven factors:

d = X + T + U + IXT + ITU + IXU + IXTU (14.1)

Equation 14.1 shows that, for the basic experimental design in Figure 14.1, the averaged difference score for all the subjects in each experimental group in fact represents the consequences of the effects of the seven factors on the subjects over the period during which the experiment takes place.

The problem then becomes whether it is possible to identify each of the effects so that the difference score d can be decomposed and interpreted adequately and the effect of the treatment T on the difference score identified. For the researcher, equation 14.1 contains only one known score- the difference score computed from the change between the pretest measurement and posttest measurement. In general, this type of equation (called a nonhomogenous equation) can be solved for only one unknown. But equation 14.1 has seven unknowns, and therefore we cannot determine the effects individually.

The strategy, then, is to construct different experimental groups, so that different equations can be formulated containing these effects. Theoretically, if seven different equations, or seven different experimental groups, could be constructed, then all seven effects could be identified. This is possible because seven equations would provide solutions for seven unknowns. Unfortunately, it is impossible to construct seven different experimental groups. In other words, there is no perfect experimental design.

Available Experimental Groups and Alternative Experimental Designs

Experimental Groups

There are four different experimental groups a researcher can utilize in an experiment. These groups are presented in Table 14.4. Group 1 is the basic experimental group discussed previously. The subjects in this group are exposed to the pretest measurement, the treatment, the uncontrolled factors, and the posttest measurement. In group 2, the subjects are administered the pretest measurement but do not participate in the treatment. They are, however, subjected to the effects of the uncontrolled factors.

Group 3 subjects are not administered the pretest measurement, but are exposed to the treatment ; therefore they are affected by T, U, and P. Finally, group 4 subjects do not participate in the pretest measurement and do not receive the treatment ; they are administered only the posttest measurement.

There are other possible but useless groups which can be constructed. For example, groups could be constructed which participate in the pretest measurement only, the treatment only, or the pretest measurement and the treatment only. But lack of posttest measurement would prevent the computation of the difference scores for the subjects in these groups. Thus, they cannot be considered.

For each of the four available groups, we may discuss the computation of the difference score. For clarity d1 will represent the difference score computed for group 1, d2 for group 2 etc.

Group 1

As represented in equation 14.1, the group 1 difference score d is composed of: (1) the effect of the pretest measurement X, (2) the effect of the treatment T, (3) the effect of uncontrolled events U, (4) the effect of the interaction of the pretest measurement and the treatment IXT, (5) the effect of the interaction of the pretest measurement and the uncontrolled events IXU. (6) the effect of the interaction of the treatment and the uncontrolled events ITU, and (7) the effect of the interaction among the pretest measurement, the treatment, and the uncontrolled events IXTU.

Thus,

D1 = X + T + U + IXT + IXU + ITU + IXTU

This equation has one known score (d,) and seven unknowns. Thus, there is no way to identify each individual effect.

Group2

ln the same way, the difference score d2 between the posttest and the pretest measurements can be shown as follows:

D2 = X + U + IXU

There is one equation but three unknowns.

Group 3

The equation for the effects in group 3 is :

D3 = T + U + ITU

Since this group does not have a pretset measurement, d3 is not known. Therefore, equation 14.3 (group 3) is meaningful only if the pretest score can be estimated from another group (either group 1 or group 2)). In that case, and if it can be assumed that the groups are relatively large (therefore that the estimated pretest score is stable) and randomized (therefore that there are no prearranged differences between the groups), then the averaged pretest score of another group can be used to find d3. This estimation method will be demonstrated later in the chapter.

Equation 14.3 then, also has three unknowns (T,U, and ITU).

Group 4

Group 4 does not have a pretest measurement. The discussion for group 3 also applies here to obtain d4 for the following equation :

D4 = U

In summary, there can be up to four groups, and their equations are :

Since four equations can be solved only for four unknowns, no matter which groups are used, no experiment design can identify all the effects.

One solution to this problem of lack of information from the data is for the researcher to make necessary assumptions about certain unknowns. For example, if a researcher chooses a one-group design (group 1), he has one equation and seven unknowns. In order for him to solve for the effect of one factor, say the effect of the treatment T, he would have to make assumptions about the other six unknowns in equation 14.1. Three types of assumptions can be made. An easy way out is simply to assume that all the other unknowns are equal to zero. In other words, all other factors have no effect on the difference score. Or, a researcher can assume that these factors have effects on the difference score but that effects canceled each other out. Third, he can assign numbers to the various factors to represent the extent of the effects they have on the difference score.

These numbers can be based on evidence from past research. How realistic these assumptions are depends on the experimental situation, the nature of the study, the activities which have taken place during the experiment, and past evidence reported in the literature. The fewer assumptions a researcher has to make, the less likely it is that he will distort the data.

In general, two criteria help the researcher to select a particular design: (1) to minimize the number of assumptions. and (2) to assume effects which either have been determined m previous studies or are less consequential in relation to other alternative effects.

Thus, a research design which requires two assumptions is in general preferred to a design requiring three assumptions. A research design assuming effects determined in the past is preferred to a design assuming effects without such previously determined values. A research design assuming effects which intuitively have no important consequences on the crucial variables is preferred to one which has to assume values for effects affecting the crucial variables.

Experimental Designs

In the following, we will discuss several popular experimental designs and assess their relative merits in terms of the two criteria just mentioned.

Two-Group Design 1

In this design, groups I and 2 are used. Thus, equations 14.1 and 14.2 are employed:

D1=X+T+U+IXT+ITU+IXU+IXTU
D2=X+U+IXU

This is the most commonly used experimental design, in which one experimental group receives the treatment and the other does not. Since two equations are involved, and there are seven unknowns, the researcher must make assumptions about five unknowns in order to find solutions for the other two unknowns. To illustrate the solutions, an example is provided in Table 14.5. In this example, the pretest measurement of the dependent variable shows an average of 20 points on the scale for both groups, indicating that randomization was effective in the assignment of subjects to the two groups. The posttest measurement scores are 100 and 60 points, respectively, for the two groups. Thus, the averaged difference scores can be computed for the two groups. The researcher decided that he would assume that all the interaction effects and the uncontrolled factors did not appreciably affect the posttest scores of the subjects. After eliminating the terms assumed to vanish in the equations, the researcher obtained two equations containing only two unknowns. The two unknowns were then solved. The researcher concluded that the treatment induced a change of 40 points from the pretest measurement to the posttest measurement, and that the pretest measurement effected a change of 40 points also.

Table 14.5 Example Solving Two Unknowns in Two-Group Design 1

Instead of assuming that the uncontrolled factors U had no effect on the difference score, the researcher could alternatively make a similar assumption about the pretest measurement effect X or the interaction effect between the pretest measurement and the uncontrolled factors lxu The decision would not affect the solution for the treatment T, but would drastically change the effect of X, U, or lXu, depending on which one was selected for solution.

Thus, it can be concluded that the experimental design utilizing groups 1 and 2 requires the researcher to make many assumptions and that the solutions are extremely unreliable.

Two Group Design 2

An alternative two-group design utilizes groups 3 and 4. The equations involved are:

D3 = T + U + I

(14.3)

D4 = U
(14.4)

Thus, there are two equations and three unknowns. The researcher needs to make only one assumption to solve for two unknowns. However, since neither group received a pretest measurement, how can the difference score be computed? If we denote the pretest measurement score for group 3 as b3, and the posttest measurement score as a3, and likewise for group 4 as b4 and a4, then the difference scores for the two groups are:

D3 = a3 - b3
D4 = a4 - b4

Thus, D3 - D4 = a3 - b3 - (a4 - b4)
= (a3 - a4) - (b3 - b4)

However, if randomization was in effect in the assignment of subjects to the two experimental groups, then the pretest measurement scores for the two groups should be approximately the same. In order words, b3 and b4 should be approximately the same and cancel each other out in the last equation. Then,

D3 - D4 = a3 - a4

Thus, the difference between the two difference scores for the two groups can be computed from the difference between the two posttest measurement scores alone.

From equations 14.3 and 14.4, we know that :

D3 - D4 = a3 - a4 = T + ITU

Thus, we only need to make an assumption about ITU to solve for T. An example of this design is presented in Table 14.6.

A comparison between this design and the two-group design 1 discussed earlier (Table 14.5) shows that this design needs only one assumption as contrasted to three assumptions for the other design to solve for the effect of the treatment T. Further, no pretest measurement represents savings for the researcher. Thus, we may conclude that in general the two-group design utilizing groups 3 and 4 is superior to the two-group design utilizing groups 1 and 2 in finding the effect of the treatment. This conclusion is valid only if the assignment of the subjects to the groups is completely random, so that the pretest measurement scores can be assumed to have been similar for the two groups if a pretest measurement was administered.

The only advantage of the two-group design 1 over the two-group design 2 is the fact that the former permits determination of the effect of another factor (for examplea in Table 14.5 the effect of the pretest measurement), whereas the latter does not (it is impossible to find U, because the difference score for each group is unknown). However, this advantage is based on the heavy price a researcher has to pay for making assumptions about the effects of five factors.

Three-group design 1

The best known three-group design uses groups 1, 2, and 3. The system of equations involved is :

Thus, this design has three equations and seven unknowns. To solve for three unknowns, assumptions must be made for the other four unknowns. In general, the researcher is interested in finding T, so no assumption should be made for it. Further, assumptions can be made for only two of the unknowns X, U, and IXU, since the third value would automatically be derived from equation 14.2. Thus, the researcher can make assumptions about the four factors selected from the combination of two or three of the three interactions IXT, IXU, and IXTU, and one or two from among X, U, and IXU. An example of such a design is presented in Table 14.7. In this example, the researcher assumed that the effects of IXT, ITU, IXU and IXTU were minimal, so that solutions were found for T, X, and U.

Table 14.7 Example of solving three unknowns in Three Group design 1

Note that in this example the third group (TUP) did not take the pretest measurement. In order to compute d3, an estimate of the pretest measurement score was computed by taking the average of the pretest measurement scores from the first two groups. This could be done because, again, it was assumed that the random assignment of subjects to the three groups ensured similar pretest scores for all groups. However, when taking the average from the pretest scores from the first two groups, the researcher should take into account the fact that the scores differed slightly from subject to subject. This variation indicates that there might be some random error involved in the estimate of the pretest score for group 3. Thus, the final estimate must include a point estimate (250) and an interval estimate (plus/minus 20). The two estimates must be carried in all subsequent computations.

There are three equations and five unknowns, making it necessary to assume the effects of two factors. An example of this design is shown in Table 14.8. Note that estimates for the pretest scores for groups 3 and 4 were computed from the pretest scores of group 2. The effect of the pretest measurement on the change score X was estimated as -60 + 20, a negative score. It suggested that the subjects exposed to the pretest measurement tended to be affected negatively toward the dependent variable in the posttest measurement.

A comparison between the two three-group designs discussed shows that the three-group design 2 needs fewer assumptions (two versus four) and pretest measurements (one versus two) than the three-group design 1. Thus, in general, the three-group design utilizing groups 2, 3, and 4 is superior to the three-group design utilizing groups 1, 2, and 3. Again, this statement is valid only if complete random assigmnent of the subjects to the groups is assumed, so that the estimates of the pretest measurement scores for the groups not taking the pretest measurement are reliable and within a tolerable range.

Four-Group Design

The design utilizing all four groups is known as the Solomon four group design and involves all four equations and seven unknowns. With assumptions made for the effects of three factors, solutions for the effects of four factors can be found. An example of the design and the solutions appears in Table 14.9.

Note several things in Table 14.9. First, the pretest and posttest measurement scores are not given; only the difference scores and the estimated change scores are presented. Second, the assumptions about the effects of the three factors X, IXT. ITU are actual scores rather than zeroes. This suggests that past research evidence has provided realistic estimates of the effects of the factors. Finally, the effect of the interaction between the pretest measurement and the uncontrolled factors is 15 + 25, an interval including both positive and negative scores. This suggests that the effect of this factor was small, thus making the estimate unreliable.