Meta Analysis of Involvement Research

Carolyn L. Costley, University of North Carolina
ABSTRACT - In recent years the involvement construct has received quite a bit of attention from consumer researchers. Only minimal agreement has been reached on how the construct should be defined. This paper attempts to integrate the research on involvement by applying a meta analysis to studies which have used involvement as an independent variable. The analysis does not identify a "right" definition of involvement. Rather, it highlights that the term cannot be used in a global sense because effects differ depending on how the involvement construct is defined.
[ to cite ]:
Carolyn L. Costley (1988) ,"Meta Analysis of Involvement Research", in NA - Advances in Consumer Research Volume 15, eds. Micheal J. Houston, Provo, UT : Association for Consumer Research, Pages: 554-562.

Advances in Consumer Research Volume 15, 1988      Pages 554-562


Carolyn L. Costley, University of North Carolina


In recent years the involvement construct has received quite a bit of attention from consumer researchers. Only minimal agreement has been reached on how the construct should be defined. This paper attempts to integrate the research on involvement by applying a meta analysis to studies which have used involvement as an independent variable. The analysis does not identify a "right" definition of involvement. Rather, it highlights that the term cannot be used in a global sense because effects differ depending on how the involvement construct is defined.


While scholars claim consensus among researchers that involvement means personal relevance or importance (Greenwald and Leavitt 1985), there are wide variations between the definitions of involvement espoused in the literature. An intuitive agreement that involvement is a unique construct has not alleviated researchers' difficulties in deriving a unique conceptual and, hence, operational definition. Because of thiS, it is difficult to compare one study with another.

This review raises issues about the involvement construct. One of the issues is whether research results have differed based on the definition of involvement used The review does not seek a "right" definition.


Approaches to the involvement construct differ on several dimensions. This author identifies four such dimensions: content, object, nature, and intensity. The three approaches distinguished by the content dimension pretty much correspond to Houston and Rothschild's three types of involvement (1978). Each type may vary on the object, nature, and intensity dimensions. The current formulation provides progressively finer aspects by which to characterize definitions. The following discussion describes these dimensions.

Content Dimension

The content dimension differentiates the way researchers have used the term "involvement" according to its position along an antecedent-consequent continuum. The literature has identified three major approaches: the cognitive approach, the individual state approach, and the response approach.

Cognitive approach: According to the cognitive approach, involvement is a permanent relationship, and thus it can only be measured, not manipulated. This approach suggests that involvement should be included as a covariate in many studies, especially studies of advertising effectiveness. Of course, one must account for involvement with more than one object the product, the medium, and the spokesperson. In this case, it may be the interaction which is enduring (involvement with the product in the ad medium and spokesperson situation) .

State approach: Individual state approaches treat involvement as a mental state at a particular point in time (Mitchell 1979). The state approach separates involvement from its antecedents and consequents. It suffers, however, from the problem of being defined in terms of other concepts. The involvement state is frequently defined in terms of arousal, motivation, attention, or interest. If involvement is any of these, it is not a unique construct and is therefore unnecessary.

Response approach: Response-based approaches take the consequent position and measure involvement in terms of response patterns (Ray 1973; Batra & Ray 1983). Responses don't seem to need a new label such as "involvement" so this conceptualization seems weak. While constructs are typically measured by items that are expected to correlate with each other and with the construct, construct validity is evaluated in terms of the correlation between the measured construct and other constructs to which it theoretically relates. Involvement should certainly be expected to correlate with responses, but involvement is not responses per se. They are theoretically related but not one and the same. If they are the same, then the term "involvement" is unnecessary.

Object Dimension

Mitchell (1979) has drawn attention to the notion that involvement does not exist in the individual independent of an object. It is involvement with something. It has a direction. The object of involvement may be a product, an ad, or a situation.

Product: The attributes of a product as well as the individual's need and experience are thought to influence involvement with the product.

Ad: Attributes of an ad, source credibility, and humor may influence involvement with the ad. Studies of cognitive responses to persuasive messages often deal with involvement with the ad.

Situation: Situational involvement encompasses such aspects of the situation as the task and the media. This seems to be related to the amount of effort required of the individual. Tasks which require an evaluation of a product are thought to be more involving than rote tasks such as proof-reading. As for media, print is thought to be more involving than video because the individual must take a more active role in order to process the information.

Nature Dimension

Park (Park & Young 1983; Park & McClung 1986) has suggested that involvement can be affective or cognitive in nature. Affective involvement is the expressive, emotional type of involvement. Nelson, Duncan, and Frontczak (1985) used this type when they measured involvement in terms of whether a message was "interesting," "boring," or "exciting." Cognitive involvement comes from functional motives. Laurent and Kapferer (1985) call it rational. The functional nature of involvement is represented by Petty, Cacioppo, and Schumann (1983). Their subjects who expected to receive a product were highly involved with that product.

The notion of an emotional or functional nature of involvement further serves to make involvement a multidimensional construct. Many authors do not specify this aspect of involvement in their working definitions.



Intensity Dimension

Intensity of involvement is usually referred to in terms of high and low. Some authors use more than two categories, (Zaichkowsky 1985b). In some cases, involvement is conceptualized as a continuous variable and measured in scale scores. This also contributes to the difficulty of comparing studies.


Nearly all definitions of involvement can be categorized along the content dimension. This dimension offers the greatest potential for grouping definitions of involvement. Conceptual definitions of involvement have differed regarding inclusion of the object and nature dimensions. Some address these aspects while others do not. Definitions may be distinguished hierarchically as in Figure 1.

It is apparent that definitions of involvement differ from study to study. Comparisons of involvement studies therefore are difficult. However, these differences suggest that a meta analysis of involvement studies should investigate whether the operational definition of involvement affects the results.


Meta analysis is a quantitative method for reviewing literature. It uses statistical procedures to synthesize the results of investigations which have already been conducted. Meta analysis may address the following questions:

(1) Is there a relationship between two variables?

(2) What is the average strength of that relationship?

(3) Is it consistent across studies?

(4) If not, what explanatory variables account for the variance?

Of particular interest is whether definitional differences account for variation in results across studies (question 4). Questions 1 and 3 will have to be answered ("yes" and "no," respectively) before we can address this issue.

Criteria for Inclusion and Coding of the Studies

We conducted a meta analysis on studies which investigated the effects of involvement. The sampling frame was limited to research published in four marketing publications between 1976 and 1986. Eighteen articles reported usable empirical results. Of these, 1 came from Journal of Marketing, 2 were from Journal of Marketing Research, S were from Journal of Consumer Research, and 10 came from Advances in Consumer Research.

Many of these studies provided useful information for more than one dependent variable. Rather than waste all of this information, these studies became two observations in the meta analysis. No study was allowed more than 2 results in the meta analysis. Table 4 lists the articles and shows the number of results that they provided for the meta analysis.

One rater coded a variety of characteristics for each study. Of particular interest is the definition of involvement. Involvement was coded along the four dimensions discussed above. Along the content dimension a study was coded "cognitive" if it defined involvement as an enduring relationship between individual and object. Perceived personal importance of an object or a predisposition to respond fit this definition. Thirteen of the observations (43%) fell into this category.

A study was coded as taking a "state" approach to involvement if it operationalized the construct in terms of the mental state or goals of an individual at a particular point in time. This state might be evoked by the stimulus or brought to the situation by the individual (such as mood). Thirteen of the meta analysis observations (43%) were coded into this category.

The "response" category of the content dimension was reserved for those studies which operationalized involvement in terms of the extensiveness and pattern of cognitive and behavioral processes. Only 2 of the observations (1%) fit this category. Two observations left the content aspect of involvement unidentified.

Each study's working definition of involvement was also coded as to the object or direction of involvement. This could be a product, an ad, or the situation. Coding "product" was indicated when involvement was defined as interest in a product or importance of a product.

Involvement was directed toward an ad when the message itself was personally relevant to an individual. This was the case in Petty and Cacioppo's studies (1981; Petty, Cacioppo, & Schumann 1983) which used messages about campus policies that would be implemented at the subjects' university (high involvement) or at another university (low involvement).

Situational involvement could be either involvement with the task in the situation or with the ad's medium in the situation. For example, a product evaluation is commonly used to operationalize a high involvement task. Print ads versus television ads are often used to manipulate high versus low involvement in terms of medium. Most of the studies in this meta analysis investigated product involvement (23 of the 30). Three looked at involvement with an ad and four examined involvement with the situation.

The nature of involvement was evaluated by the content of related attitude structures. If involvement was defined by emotional motives, liking and other expressions of affect, it was classified as affective in nature. Four of the studies fell into this category. If involvement was defined as related to rational, cognitive motives, such as a problem to be solved or personal relevance of consequences, it was classified as functional in nature. Twenty of the studies fit this classification. Six of the studies (20%) did not provide information to enable evaluation on this dimension.

Intensity of involvement distinguished the studies as to whether they used a continuous or a dichotomous measure of involvement. Eleven of the studies used continuous measures and 16 used a dichotomy to distinguish levels. Three studies used a categorical measure with more than two levels. These were coded into a third category for intensity.

This meta analysis combines studies which examined different dependent variables in relation to involvement. Effects of involvement may differ depending on the dependent variables. Four categories for dependent variables were generated and used in the analysis. Nine of the studies dealt with attitudes, either attitude formation or change. Nine dealt with some kind of behavior such as gift selection effort or acquisition behavior. Five studies examined brand or ad recall. Seven of the studies used another form of involvement as the dependent variable [Primarily, these followed Houston and Rothschild's framework (1978) and investigated the effects of situational and enduring involvement on response involvement.]. A table of codes for each of the studies is available from the author.

Results of the Mesa Analysis

The first stage in the meta analysis was to compare significance levels. This procedure was implemented to address the question of whether involvement had a significant effect in any of the context in which it was studied. Since most of the articles simply reported significance or nonsignificance, exact p-values associated with test statistics were obtained using SAS functions. Analysis proceeded on the one-tailed p-values. All of the results of the meta analysis are displayed in Table 1.

When comparing significance levels, the first step is to test the null hypothesis that all effects are insignificant. If we reject the null hypothesis then we conclude that there is a significant effect of involvement in at least one study. Two combination procedures have been suggested as most popular-. Fisher's inverse chi-square method (Hedges & Olkin 1985) and Stouffer's inverse normal method (Brinberg & Jaccard 1986; Rosenthal & Rubin 1979). Both will be used here.

For a comparison of significance levels by Fisher's inverse chi-square method, we compute P:

P = -2 S log pi

and compare P to C. C is obtained from the upper tail of the chi-square distribution with 2k degrees of freedom (k is the number of studies) [In this case k is 27 because 3 observations did not provide all of the information needed to compute p.].



We reject the null hypothesis if P 2 C. In this case:

k = 27

71.42 # C # 83.30 (for X2(.95,54))

P = -2 S log pi = 400.8882

P > C so we reject the null hypothesis and conclude that involvement has a statistically significant effect in at least one study.

When significance levels are compared by Stouffer's inverse normal method, we first find the standard normal z's associated with the p-values. This was done using the PROBIT function in SAS. Then we compute


We will reject the null hypothesis of no significance if Zt,Z.05 from the standard normal z-table. We find


Z.05 = 1.645

Zs>Z.05 so we again reject the null hypothesis.

Adjustment for Nonindependence

Since some of the observations in the meta analysis come from the same studies, they are not independent as assumed by the above analyses. Strube (1985) notes that when there is positive correlation between significance tests within studies, the Type-I error rate will be inflated if adjustments are not made. In other words, we will be more likely to reject a null hypothesis which is true. Strube offers a method to account for dependence by including the appropriate covariance terms:


where n is the number of findings per study and r is the correlation among the significance tests within studies. A direct measure of this correlation is hard to come by but it can be estimated by the correlation between the dependent variables. Lacking this information, as we do, upper and lower bounds of the correlation can be estimated using 0 as the lower bound and the maximum reliability of the scales as the upper bound. In this instance, the reliability information is absent and .80 has been used as the upper bound estimate for r. This is noted by Nunnally (1978) as a "good" scale reliability. Since the scales used in the studies probably were not perfect measures and since the correlations between dependent variables are not likely to be perfect either, .80 seems like a liberal estimate for maximum r. And it is a conservative estimate for the lower bound of z. Using this, we arrive at:

Zupper bound = 13.5222


Since both of these values are greater than 1.645, we still reject the null hypothesis. Both Fisher's method and Strube's adjustment to Stouffer's method lead us to reject the hypothesis that there are no significant effects of involvement.

File Drawer Problem

This analysis may be biased due to using only published research. It is possible that work that has been done but is still in file drawers and work that is yet to be done show no effects for involvement and that meta analysis results could differ if these studies were included. To address this issue, we can estimate the tolerance for null results using Rosenthal's formula for "Fail-Safe N" (1979). This provides an estimate of the number of studies which find null results but which are unavailable to this analysis (remaining in file drawers or yet to be conducted) that must exist for the meta analysis to fail to reject the null hypothesis.

The formula, with Strube's (1985) adjustment for nonindependence is


NS is the number of studies.

n is the number of results per study. The number fixed at 2 as it was in this meta analysis.

r is the within study correlation. An interval (low bound and upper bound) will be calculated as discussed above.

s2 is the variance of the linear combination of z's (i.e., the sum of the variance of the z's plus 2 times their covariances).

Since they are z's (mean 0, variance 1), each variance = 1 and the sum is k (the number of studies). The covariances are the correlations between results from the same study. Again, we use .80 as the maximum expected correlation. In this meta analysis there are: dependent observations, observations which come from the same study. The covariances between studies are assumed to be 0. The variance-covariance matrix in Table 2 makes this clear. Therefore, the computation Fail-Safe N adjusted for 80% correlation is:


The upper bound, computed for r = 0 is:


We need at least 494 studies with null results to lead us to fail to reject the null hypothesis of no significant effect of involvement. Thus, it looks like we can be fairly confident that there is a significant effect of involvement. Analysis of significance levels, however, does not tell us much about the direction, magnitude, or consistency of effects. The second stage of the meta analysis examines effect sizes from the pool of studies.

Analysis of Effect Sizes

The first step toward analysis of effect sizes is to obtain a common metric for comparison across studies. All test statistics were converted to r's as suggested by Wolf (1986). The guidelines for conversion are shown in Table 3.

Rather than comparing r's, Hedges & Olkin (1985) recommend transforming them to z's by the Fisher r-to-z transformation. This serves to normalize the distribution and to make the variance independent of the population correlation. The formula for transformation is:

z / z(r) = 1/2 log (1 + r)/(1 - r)

Now, effect magnitudes can be compared across studies by comparing z's. When the sample size of r's is large, z is approximately normally distributed with mean Z and variance 1/n. For moderate size samples, 1/(n-3) is a more accurate approximation to the variance of z (Hedges & Olkin 1985). The correlations and z-transforms are shown in Table 4.

The first thing we want to know is whether the effects of involvement are homogeneous across studies. If they are, then we can pool estimates to obtain an estimate of the common effect of involvement. If effect magnitudes are heterogeneous then it would be misleading to pool the estimates. In this case, the source of the differences should be traced. Given that involvement has been inconsistently defined across studies, it is reasonable to expect to find heterogeneity of effects. It is also reasonable to expect that the differences may be related to the definitions used.

We will reject the hypothesis of homogeneity if:

Q = Sk (ni - 3)(zi - z+)2

exceeds a critical value from the chi-square distribution with k-1 degrees of freedom.

k = number of studies = 30

ni=sample size in the ith study.

z+=weighted average effect size


z+ = w1z1 + ... + wkzk

wi = (ni - 3)/k(nj - 3)

(ni - 3) = the inverse of the covariance of the z's.

If Q is less than chi square with k-l degrees of freedom, then the effect sizes are considered homogeneous and z+is the z-est mate of the common correlation. If this is the case, the next step would be to test whether z+=0. If Q > chi square with k-l degrees of freedom, then the effect sizes are heterogeneous and we will proceed to test possible explanations.

We find Q = 203.9178 which is greater than chi square (.95, 29) = 4256. We conclude that the effect sizes are not homogeneous and we will proceed to analyze for explanations.




Explanations for Effect Sizes

Heterogeneous effect sizes are consistent with the notion that different definitions of involvement may lead to different findings. The next step is to construct and test a model of effect sizes using the dimensions of involvement definitions as explanatory variables. Hedges and Olkin (1983) extended the application of generalized least squares to handle correlations and their approach is applied here. It was suggested earlier that the definitions of involvement could be represented hierarchically with the content dimension at the highest level (see Figure 1). Therefore, a nested analysis of variance model was constructed to account for the differences in effect sizes. The object dimension was nested within levels of the content dimension and Nature within those levels and Intensity within each of those. Because the sample sizes were unequal across studies, a weighted analysis was performed under the recommendation of Hedges and Olkin (1985). The estimation procedure was modified using the inverse of the covariance of the z's as the weighting factor (ni-3). Analysis was executed under the SAS General Linear Models procedure. The model can be depicted:

z = b0 + b1C + b2O + b3N + b4I


C = the content dimension,

O = the object dimension nested within levels of the content dimension,

N = the nature dimension nested within levels of object within content,

I = the intensity dimension nested within levels of nature within object within content.

The first step in this stage of the meta analysis is to test the model specification. This provides a basis for deciding whether the variation in effect size is accounted for by the explanatory variables. When the model is correctly specified and when the sample size is large, the error sum of squares has an approximate chi-square distribution with k - p degrees of freedom (p is the number of parameters in the model) [This model contains four levels for the content dimension (cognitive, state, response, and unidentified), three levels for the object dimension (product, ad, and situation), three levels for the nature dimension (affective, functional, and unidentified), and three levels for the intensity dimension (continuous, high/low, and multi-categories). The nesting procedure results in 16 distinct definitions in this selection of studies.]. So, to test that the model

Z = Xb

adequately fits the data, we compare the weighted error sum of squares to the critical value for chi-square with k p degrees of freedom. The weighted error sum of squares for the model is 69.72. Comparison to X2(.95,14)=23-68 leads us to reject the hypothesis that the model is correctly specified. This model does not adequately account for the variation in z. Because of this, the parameter estimates do not necessarily converge on the true values. Therefore, their interpretation is inappropriate.



Specification of the same model was also tested for each level of the dependent variables studied. That is, for each category of dependent variable investigated by studies in the pool, a separate model test was performed. For studies which investigated the relationship between involvement and attitude, the error sum of squares is 3.93. When this is compared to chi square (.95, 1) = 3.84, we conclude that the model does not adequately explain the effect sizes for this group. For studies which investigated behavior or involvement as a dependent variable, the model is accurately specified. Error sum of squares equal 4.80 and 1.62, respectively and are compared to chi square (.95, 2) = 5.99. Not enough studies in the sample used recall as a dependent variable to enable evaluation of model specification (the degrees of freedom were used up by the explanatory variables in the model).

Sources of Poor Fit

Potential sources of poor fit are outlier studies, multicollinearity of the independent variables, omission of important independent variables, and measurement error. The effect sizes listed in Table 4 do not show any extreme outliers. lt could be that some of the involvement dimension levels are not discriminators and could be combined. Combining categories could reduce multicollinearity and increase the available degrees of freedom.

Measures of other characteristics of the studies are available for use as independent variables. Differences in results were not accounted for by the type of dependent variable studied. Unfortunately, the sample size prohibits inclusion of very many explanatory variables. Degrees of freedom are used up rapidly by these categorical variables.

It is likely that the model doesn't fit because a great deal of variation is due to measurement error, both in the meta analysis and in the individual studies. Integration of studies that contain measurement error by a procedure that itself contains measurement error just magnifies the problem. We did not attempt to adjust for measurement error in the studies because the necessary information was not available. We also do not have any measure of the error in the meta analysis procedure because only one rater coded all of the studies.


The comparison of significance levels indicates that involvement does indeed have a significant effect in the pool of studies. Bias due to publication of statistically significant results would have to be countered by at least 494 studies with null results.

Comparison of effect sizes indicates that the effect of involvement was not consistent across the studies. This was not surprising given the variety of conceptualizations of the involvement construct. However, the dimensions of involvement that we identified as being sources of variation in definitions did not adequately explain the differences in effect sizes. So, we have learned that studies of involvement do not find consistent results, but we were unable to explain the differences.


Arora, Raj (1982), "Validation of an S-O-R Model for Situation, Enduring, and Response Components of Involvement," Journal of Marketing Research 19, 505-16.

Batra, Rajeev and Michael L. Ray (1983), "Operationalizing Involvement as Depth and Quality of Cognitive Response" in Richard P. Bagozzi and Alice M. Tybout (eds.) Advances in Consumer Research Volume 10, Ann Arbor, MI: Association for Consumer Research, 309-313.

Belk, Russell W. (1982), "Effects of Gift-giving Involvement on Gift Selection Strategies," in Andrew A. Mitchell (ed.) Advances in Consumer Research Volume 9, Ann Arbor, MI: Association for Consumer Research, 408-411.

Brinberg, David and James Jaccard (1986), "A Review of Meta-Analytic Techniques," in Richard J. Lutz (ed.) Advances in Consumer Research Volume 13, Provo, UT: Association for Consumer Research, 606-611.

Clarke, Keith and Russell W. Belk (1979), "The Effects of Product Involvement and Task Definition on Anticipated Consumer Effort," in William Wilkie (ed.) Advances in Consumer Research Volume 6, Ann Arbor, MI: Association for Consumer Research, 313-318.

Gardner, Meryl P., Andrew A. Mitchell, and J. Edward Russo (1978), "Chronometric Analysis: An Introduction and an Application to Low Involvement Perception of Advertisements," in H. Keith Hunt (ed.) Advances in Consumer Research Volume 5, Ann Arbor, MI: Association for Consumer Research, 581-589.

Greenwald, Anthony G. and Clark Leavitt (1985), "Cognitive Theory and Audience Involvement," in Linda Alwitt and Andrew Mitchell (eds.) Psychological Processes and Advertising Effects, Hillsdale, NJ: Lawrence Erlbaum, 221-240.

Hedges, Larry V. and Ingram Olkin (1985), Statistical Methods for Meta-Analysis, Orlando, FL: Academic Press.

Hedges, Larry V. and Ingram Olkin (1983), "Regression Models in Research Synthesis," The American Statistician, Vol. 37, No. 2, 137-140.

Houston, Michael J. and Michael L. Rothschild (1978), "Conceptual and Methodological Perspectives on Involvement," in Subhash C. Jain (ed.) Research Frontiers in Marketing: Dialogues and Directions, Chicago: AMA, 184-187.

Kahle, Lynn R. and Pamela M. Homer (1985), "Physical Attractiveness of the Celebrity Endorser A Social Adaptation Perspective," Journal of Consumer Research, 11, 954-961.

Lastovicka, John L. (1979), "Questioning the Concept of Involvement Defined Product Classes," in William L. Wilkie (ed.) Advances in Consumer Research Volume 6, Ann Arbor, MI: Association for Consumer Research, 174-179.

Lastovicka, John L. and David M. Gardner (1978), "Low Involvement Versus High Involvement Cognitive Structures," in H. Keith Hunt (ed.) Advances in Consumer Research Volume 5, Ann Arbor, MI: Association for Consumer Research, 87-92.

Laurent, Gilles and Jean-Noel Kapferer (1985), "Measuring Consumer Involvement Profiles," Journal of Marketing Research, 22, 41-53.

Mitchell, Andrew A. (1979), "Involvement: A Potentially Important Mediator of Consumer Behavior," in William L. Wilkie (ed.) Advances in Consumer Research Volume 6, Ann Arbor, MI: Association for Consumer Research, 191-196.

Nelson, James E., Calvin P. Duncan, and Nancy T. Frontczak (1985), "The Distraction Hypothesis and Radio Advertising," Journal of Marketing, 49, 60-71.

Nunnally, Jum C. (1978), Psychometric Theory (second edition), New York: McGraw Hill.

Park, C. Whan and Gordon W. McClung (1986), "The Effect of TV Program Involvement on Involvement with Commercials," in Richard J. Lutz (ed.), Advances in Consumer Research Volume 13, Provo, UT: Association for Consumer Research, 544-548.

Park, C. Whan and S. Mark Young (1983), "Types and Levels of Involvement and Brand Attitude Formation," in Richard P. Bagozzi and Alice M. Tybout (eds.), Advances in Consumer Research Volume 10, Ann Arbor, MI: Association for Consumer Research, 320-323.

Parkinson, Thomas L. and Carolyn Turner Schenk (1980), "An Empirical Investigation of the S-O-R Paradigm of Consumer Involvement," in Jerry C. Olson (ed.), Advances in Consumer Research Volume 7, Ann Arbor, MI: Association for Consumer Research, 696-699,

Petty, Richard E. and John T. Cacioppo (1981), "Issue Involvement as a Moderator of the Effects on Attitude of Advertising Content and Context," in Kent B. Monroe (ed.), Advances in Consumer Research Volume 8, Ann Arbor, MI: Association for Consumer Research, 20-24.

Petty, Richard E., John T. Cacioppo, and David Schumann (1983), "Central and Peripheral Routes to Advertising Effectiveness: The Moderating Role of Involvement," Journal of Consumer Research, 10, 135-146.

Rajaniemi, Pirjo and Martti Laaksonen (1986), "The Role of Profile Analysis in the Conceptual Development of Involvement," European Marketing Conference, June 1986, Helsinki.

Ray, Michael L. (1973), "Marketing Communications and The Hierarchy-of- Effects," in P. Clarke (ed.), New Models for Mass Communication Research, Beverly Hills, CA: Sage Publications, 147-176.

Rosenthal, Robert (1979), "The 'File Drawer Problem' and Tolerance for Null Results," Psychological Bulletin, 86, 638-641.

Rosenthal, Robert and Donald B. Rubin (1979), "Comparing Significance Levels of Independent Studies," Psychological Bulletin, 86, 1165-1168.

Rothschild, Michael L. and Michael J. Houston (1980)1 "Individual Differences in Voting Behavior: Further Investigations of Involvement," in Jerry C. Olson (cd.), Advances in Consumer Research Volume 7, Ann Arbor, MI: Association for Consumer Research, 655-658.

Strube, Michael J. (1985), "Combining and Comparing Significance Levels From Nonindependent Hypothesis Tests," Psychological Bulletin, 97, 334-341.

Swasy, John L. and James M. Munch (1985), "Examining the Target of Receiver Elaborations: Rhetorical Question Effects on Source Processing and Persuasion," Journal of Consumer Research, 11, 877-886.

Swinyard, William R. and Kenneth A. Coney (1978), "Promotional Effects on a High- Versus Low Involvement Electorate," Journal of Consumer Research, 5, 41-48.

Wolf, Fredric M. (1986), Meta-Analysis: Quantitative Methods for Research Synthesis, Beverly Hills: Sage Publications.

Zaichkowsky, Judith Lynne (1985a), "Familiarity: Product Use, Involvement, or Expertise?" in Elizabeth C. Hirschman and Morris B. Holbrook (eds.), Advances in Consumer Research Volume 12, Provo, UT: Association for Consumer Research, 296-299.

Zaichkowsky, Judith Lynne (1985b), "Measuring the Involvement Construct," Journal of Consumer Research, 12, 341-352.

Zinkhan, George M. and Aydin Muderrisoglu (1985), "Involvement, Familiarity, Cognitive Differentiation, and Advertising Recall: A Test of Convergent and Discriminant Validity," in Elizabeth C. Hirschman and Morris B. Holbrook (eds.), Advances in Consumer Research Volume 12, Provo, UT: Association for Consumer Research, 356-361.