Handbook of Research Methods in Consumer Psychology

Frank Kardes, Paul M. Herr, Norbert Schwarz, Frank R. Kardes, Paul M. Herr, Norbert Schwarz

• Publisher: Routledge
• Year: 2019
• ISBN:
  9780429687594
• Edition: 1
• Topic: Psychologie
• Subtopic: Angewandte Psychologie

What impact can various research methods have on consumer psychology? How can they help us understand the workings of the consumer mind? And how can the field of consumer psychology best utilize these methods? In the Handbook of Research Methods in Consumer Psychology, leading consumer psychologists summarize key aspects of the research process and explain how different methods enrich understanding of how consumers process information to form judgments and opinions and to make consumption-related decisions.

Kardes, Herr, and Schwarz provide an in-depth analysis of the scientific research methods needed to understand consumption-related judgments and decisions. The book is split into five parts, demonstrating the breadth of the volume: classic approaches, contemporary approaches, online research methods, data analysis, and philosophy of science. A variety of leading researchers give insight into a wide range of topics, reflecting both long-standing debate and more recent developments in the field to encourage discussion and the advancement of consumer research.

The Handbook of Research Methods in Consumer Psychology is essential reading for researchers, students, and professionals interested in consumer psychology and behavior.

PART I

Classic Approaches

1

EXPERIMENTAL RESEARCH METHODS IN CONSUMER PSYCHOLOGY

Frank R. Kardes and Paul M. Herr

Experimental methodology is the cornerstone of scientific research in consumer psychology. Unlike other research methods, random assignment of subjects to conditions controls for all extraneous variables, including ones the researcher has not considered. True experiments also ensure that the independent variable (or presumed cause) precedes the dependent variable (or presumed effect), unlike correlational research methods which beg questions about which came first, the proverbial chicken or the egg. The ability to control for all extraneous variables permits researchers to draw stronger inferences on the basis of experimental methods than on the basis of correlational methods (Aronson, Ellsworth, Carlsmith, & Gonzales, 1990).

Types of Variables

Although correlational methods continue to improve and to increase in sophistication, correlational methods should not be considered as a reasonable substitute for experimental methods for three reasons. First, it is typically not possible to generate a complete list of potentially influential extraneous variables. Second, even if a researcher could generate a complete list, it is typically not possible to measure every potentially influential extraneous variable. Third, if a researcher tried to measure every potentially influential extraneous variable, several different types of measurement effects would result—including measurement order effects (Sudman, Bradburn, & Schwarz, 1996; see also, Schwarz, this volume), measurement interference effects (Quattrone, 1985), and measurement-induced judgment construction (Cronley, Mantel, & Kardes, 2010; Feldman & Lynch, 1988; see also, Kardes, Escoe, & Wu, this volume).

The experimental method controls for extraneous variables without any need to measure them. Random assignment of subjects to conditions creates groups of subjects that are equivalent, except for error variance, on all dimensions except for the level of the independent variable. For example, imagine that one wanted to test the effectiveness of a new persuasion technique. One would randomly assign subjects to a treatment condition (i.e., exposure to the new persuasion technique) and a control condition (i.e., no exposure to the new persuasion technique). In this simple experiment, the independent variable (or presumed cause) would have two levels: treatment versus control. The dependent variable (or presumed effect) could be assessed on an attitude scale ranging from 0 (very bad) to 10 (very good). Suppose that this study was conducted and the results showed that more favorable attitudes were formed in the treatment condition than in the control condition. We could safely conclude that the new technique was effective, and that this effect was not influenced by any extraneous variable.

If I were to describe this study to a neighbor, however, they would invariably say that some individual difference variable was responsible for the observed results. For example, they might say that the subjects in the treatment condition might have been lower in intelligence and, therefore, more gullible. Or they may say that subjects in the treatment condition were simply more gullible. Or they might say that subjects in the treatment condition were tired, hungry, thirsty, in a good mood, were already favorable toward the target object or issue, were less skeptical and more open-minded, or were different from subjects in the control condition in some way that could be systematically related to persuasion. All of these variables are extraneous variables that contribute to random error, but not to systematic error because the average influence of these variables is held constant across conditions. That is, randomization ensures that the average intelligence of subjects is the same in both conditions, except for error variance (which decreases as sample size increases). Randomization ensures that the average gullibility of subjects is the same in both conditions, and so on, down the list. Simply put, randomization creates equivalent groups, except for error variance. “By randomly assigning subjects to conditions, the experimenter can be sure that no subject variable is more likely to occur in one condition than in the other and thus that no subject variable is a source of systematic error (Aronson et al., 1990, p. 18; original emphasis). Aronson et al. (1990) refer to random assignment as the “great equalizer.” “Because random assignment ensures that all extraneous factors that might influence the subject’s behavior in the experiment are approximately equal in the two (or more) conditions, we would expect that if we left out the experimental treatments and ran both groups of subjects as control groups, the average scores of these two groups on the dependent variable measure would be the same” (Aronson et al., 1990, p. 18; original emphasis).

For these reasons, the randomized control experiment is the gold standard in scientific research (Nisbett, 2015; Wilson, 2011). When studies using experimental methods versus correlational methods produce different results, the results of studies using experimental methods are far more likely to be accurate (Aronson et al., 1990; Nisbett, 2015; Wilson, 2011; Wilson, Aronson, & Carlsmith, 2010).

Moderating Variables

We could design a more complex experiment with more than two levels of an independent variable, or with multiple independent variables (Aronson et al., 1990). Manipulating multiple independent variables enables the researchers to test for possible interactions among the independent variables. That is, the effect of one independent variable depends on or is moderated by the level of another independent variable. Two types of interactions are possible: An interaction due to a change in magnitude or an interaction due to a change in direction. If an interaction due to a change in magnitude is observed, the effect of one independent variable on the dependent variable is greater under one level of an independent variable than under another level of this variable. For example, our new persuasion technique might have a larger effect when subjects are in a good mood than when subjects are in a neutral mood. Another possibility is an interaction due to a change in direction. In this case, we see a complete reversal. For example, the new persuasion technique may produce more favorable attitudes when subjects are in a good mood, but less favorable attitudes when subjects are in a neutral mood. The two types of interactions are shown in Figures 1.1a and 1.1b. In Figure 1.1a, the new persuasion technique is effective under conditions of moderator level 1, but not under conditions of moderator level 2. In Figure 1.1b, the new persuasion technique is effective under conditions of moderator level 1, but backfires under conditions of moderator level 2.

The presence of either type of interaction implies that the effect of one independent variable is moderated by another. A moderator variable is a variable that influences the magnitude or the direction of an effect produced by an independent variable on a dependent variable. William McGuire once said that a field is only as advanced as its highest-order interaction. Two-way interactions are common in consumer psychology. Some three-way interactions are also reported in the consumer psychology literature. The record is a four-way interaction predicted and found by Srull and Wyer (1979). The magnitude of the priming effect is moderated by delay, the number of primes, the proportion of critical primes, and the ambiguity of the target.

Moderating variables can be situationally manipulated (e.g., involvement, time pressure, cognitive load, mood; see, e.g., Chaiken & Trope, 1999; Sherman, Gawronski, & Trope, 2014) or measured individual difference variables (e.g., need for cognition, need for cognitive closure, self-monitoring; see, e.g., Chaiken & Trope, 1999; Sherman et al., 2014). It is often desirable to conduct multiple studies showing that both types of moderating variables produce conceptually similar effects when they are related (e.g., involvement and need for cognition often produce conceptually similar effects; time pressure, cognitive load, and need for cognitive closure often produce conceptually similar effects). Confidence in a specific interpretation of a pattern of results increases when many different manipulations and measures of the same conceptual variable produce a similar pattern of results. This is the principle of converging operations (Aronson et al., 1990; Campbell & Fiske, 1959).

When a very specific prediction about the precise pattern of an interaction is possible, contrast analysis should be performed (Rosenthal & Rosnow, 1985). ANOVA addresses the question of whether there are any differences among the experimental groups, whereas contrast analysis provides “(a) very much greater statistical power and (b) very much greater clarity of substantive interpretation of research results” (Rosenthal & Rosnow, 1985, p. 4).

Figure 1.1a Attitude as a function of the treatment and the moderator.

Figure 1.1b Attitude as a function of the treatment and the moderator.

In the case of measured individual difference variables, many reviewers and editors recommend mean centering and treating this variable as a continuous variable in a regression analysis using interaction terms. However, recent research using multiple simulations shows that regression analysis and median splits of continuous variables used in ANOVA typically support the same substantive conclusions (Iacobucci, Posavac, Kardes, Schneider, & Popovich, 2015a). Consistent with the recommendations of Lehmann, McAlister, and Staelin (2011), when a complex analysis and a simple analysis provide the same results, the simpler analysis is preferred. A simpler analysis facilitates communication, enables researchers to reach a broader audience, and is more likely to be used in meta-analyses (Iacobucci et al., 2015a, 2015b; Lehmann et al., 2011). Furthermore, when the relation between the moderating variable and the dependent variable is non-monotonic, analyses using median splits are more accurate than analyses using regression (we thank Robert S. Wyer, Jr., for this astute observation). For these reasons, we recommend the use of median splits, except when independent variables are correlated.

Mediating Variables

It is also possible (and often desirable) to investigate mediating variables (Hayes, 2018; Rucker & Preacher, in this volume). Mediating variables are intervening variables that occur between the independent variable and the dependent variable. For example, our new persuasion technique might influence attention, comprehension, memory retrieval, or inference processes occurring following exposure to the treatment, but prior to attitude formation. Multiple mediating variables are also possible. With serial mediation, the independent variable influences a mediating variable, which influences another mediating variable, and so on (e.g., the independent variable influences attention, which influences comprehension, which influences retrieval, which influences inference). With parallel mediation, the independent variable influences multiple mediating variables simultaneously (e.g., the Brunswik lens model; Brunswik, Hammond, & Stewart, 2000; see also, Hirt and Guevara, in this volume).

Multiple mediating variables could also produce mediated moderation or moderated mediation (Muller, Judd, & Yzerbyt, 2005). In the former case, an interaction is observed on the dependent variable because different processes are occurring at different levels of the moderator variable. This is the key assumption of all dual-process models (Chaiken & Trope, 1999; Sherman et al., 2014). In the case of moderated mediation, however, an interaction is observed on a mediating variable, but not on the dependent variable. That is, the same effect is observed at different levels of moderator variable, but for different reasons. For example, asking subjects to write a counterattitudinal essay leads subjects change their attitudes in the direction of the essay. However, this could occur because subjects were motivated to reduce cognitive dissonance (i.e., a motivational process; Festinger & Carlsmith, 1959) or because subjects observed their own behavior and drew inferences about their attitudes (i.e., a non-motivational, cognitive process; Bem, 1965).

Types of Validity

Internal validity refers to the extent to which the design, methods, and procedures of a study allow one to conclude that the independent variable (and no extraneous variable) influenced the dependent variable (Aronson et al., 1990; Campbell, 1957; Wilson et al., 2010). External validity refers to the extent to which the results of a study generalize to other people and to other situations. For many decades, it was assumed that laboratory experiments are high in internal validity because they control for all extraneous variables, but low in external validity because undergraduate subjects are different from other types of people and because laboratory settings are more artificial than field settings. It was also assumed that internal validity is more important than external validity, because, if internal validity is low, why should anyone care about external validity?

Fortunately, research by Anderson, Lindsay, and Bushman (1999) shows that it is possible to design studies that are high in both types of validity. They computed effect sizes for published laboratory studies on a wide range of topics (e.g., aggression, helping, leadership style, social loafing, self-efficacy, depression, memory, and others). They also computed effect sizes for published field studies on the same topics. Even though the laboratory studies and the field studies differed on many dimensions (e.g., subjects, designs, procedures, measures, and others), the correlation between the effect sizes of the laboratory studies and the field studies was positive and very large, r = .73. This result suggests that the results of published laboratory studies are likely higher in external validity than previously assumed, and that the results of published field studies are, in turn, likely higher in internal validity than previously assumed.

When internal and external validity are high, the results of a study should replicate readily to other people and other situations. Recently, however, a replication crisis was declared by the Open Science Collaboration (2015), a group of researchers who attempted to replicate 100 published studies. Depending on the criterion that was used, 36–47% of the original studies were replicated. However, Gilbert, King, Pettigrew, and Wilson (2016) reanalyzed the data from the Open Science Collaboration (2015) and showed that replication rates were much higher when error, power, and bias were controlled for. In addition, many of the studies conducted in the Open Science Collaboration (2015) replicated manipulated variables rather than conceptual variables, and many were conducted in large groups of ten or more studies, unlike the original studies (Fabrigar, Wegener, Vaughan-Johnston, Wallace, & Petty, this volume; Schwarz & Clore, 2016).

When randomization is not possible owing to ethical or practical concerns, the greatest threat to internal validity is selection (Cook & Campbell, 1979). Different types of subjects may select or choose to participate in different...

Table of contents

• Cover Page
• Half Title
• Title Page
• Copyright Page
• Table of Contents
• Notes on Contributors
• Preface
• Part I Classic Approaches
• Part II Contemporary Approaches
• Part III Online Research Methods
• Part IV Data Analysis
• Part V Philosophy of Science
• Index