The Effects of Consumer Education on Consumer Product Safety Behavior

In a similar manner let TB_B denote a person's true safety behavior prior to the educational program and TB_" the behavior after the program. Since we believe changes in knowledge affect behavior we hypothesize the following relationship,

TB_" =  TB_B + Changes in Behavior due to Changes in Knowledge + Exposure to Other Aspects of the Program.   (3)

Using the notation of (2),

Equations (2) and (4) model the basic hypothesized effects of the educational process. However, it is still necessary to specify each consumer's initial level of knowledge and safety behavior. The following models are postulated;

and

MODEL OF THE MEASUREMENT PROCESS

The preceding section was concerned with a person's true level of knowledge and behavior. However, it is next to impossible to measure these true levels without error. In other words, one never observes TK or TB but instead some fallible measure of these unobserved variables. Denote the fallib1e measure of the true, but unobservable constructs as MK and MB respectively. Then

Equations (2) and (4)-(6) represent our conceptualization of how the educational process ultimately affects behavior while equations (7)-(10) document our belief that the instruments used to measure the process are fallible. The next step is to specify more precisely the functional forms of the relationships (i.e., the g_i(.)'s) and the distributional assumptions of the error terms. Table 1 lists the variables used to measure the general constructs previously mentioned. These variables are assumed to affect the variable of interest in a linear and additive fashion. Thus, for example g₁(.) is operationalized to be a linear additive function of the student's home environment (measured by the occupation of the head of household), the school environment (measured by a dummy variable for the identity of the school), the student's intellectual ability (measured by IQ and grade point) and exposure to basic safety principles (measured by whether or not the student took specific academic courses, had experience with specific types of consumer products and the identity of the student's sex).

The above discussion can now be summarized in more precise mathematical form. Let Xij be the row vector of demographic variables (i.e., those specified in Table 1) which are hypothesized to be part of g_i(.) for student j. Also let Bi be a row vector of the weights of specific demographic variables. Then equations (2), (4)-(6) can be restated as follows for any student j (the j subscript has been suppressed for clarity),

EQUATIONS  (5'),   (2'),  (6'),  (4')

where b₃and b₄ indicate the effects of knowledge and changes in knowledge respectively on the measures of interest.

TABLE 1

SPECIFICATION OF VARIABLES USED TO MEASURE . HYPOTHESIZED INFLUENCES ON KNOWLEDGE AND BEHAVIOR

ESTIMATION

Thus far we have specified our conceptualization of the educational and measurement processes. The next step is to determine how to estimate the anticipated effects. Of particular importance is the coefficient b₄ of equation (4') which represents the effect of a change in knowledge on the student's behavior. Before we can estimate these effects, however, we must first "marginal out" (substitute for) the unobserved variables TK_B, TK_", TB_B, and TB_". This can be accomplished by substituting equations (7)-(1 0) in equations (2'), (4')-(6') yielding,

EQUATION (5"),   (2"),  (6"),  (4")

It should be noted that the error terms in the above four equations are not independent, since at the very least u₁ appears in each equation. In other words measurement error affects each of the equations, causing them to be linked through the covariance matrix. This linkage should be considered in estimating the parameters of the four equations. [The issue of this linkage is efficiency not consistency.] For a more detailed discussion of the problems of estimation and possible models see Kadane et al. (1976).

Also of interest in terms of estimation is the occurrence of the endogenous elements MK_B and (MK_" - MK_B) on the right side of equations (6") and (4") respectively. These elements are not independent of the error structure, consequently avoiding biased estimates requires that this dependency be taken into account during the estimation process (Johnston, 1972).

The estimation procedure used here was originally proposed by Zellner and Theil (1962) and is known as three-stage least-squares. In the first stage, separate reduced-form equations are estimated using ordinary least squares estimates for MK_B and (MK_" - MK_B). In the second stage, each equatio5 (i.e., -(2") and (4")-(6") is estimated separately using OLS, however, the estimates ^MK_B and (^MK_"-^MK_B) from the first stage estimates are used in place of MK_B and (MK_" - MK_B). Then the residuals from these four regression equations are correlated to obtain an estimate of the covariance matrix. Finally, the four equations are solved using a generalized least square formulation based upon the estimated covariance matrix (Johnston, 1972).

DATA AND MEASUREMENT INSTRUMENTS

The data were obtained in the spring of 1974 from three high schools in the Pittsburgh public school system. The sample consists of 494 juniors, of which 239 were assigned to receive instruction on product safety, the assignment procedure was accomplished by randomly assigning predetermined classes to the control and experimental groups.)

Both before and after the educational program, each student was administered a paper and pencil test designed to measure the student's a) knowledge of safety principles and b) safety behavior. Safety knowledge was measured by a scalar index formed by equally weighting 26 multiple choice questions concerning electricity, insulation, materials and fasteners, stability, heat and flammability. Safety behavior was measured by a scalar index formed by equally weighting 11 multiple choice questions which asked how each respondent behaves in given situations. Typical situations included the use of a three-prong electric plug, pouring or handling volatile liquids, and use of a power mower. Responses were scored with respect to safe use behavior as determined by a panel of engineers.

It should be noted that the unobserved construct TB was measured by asking the students about their behavior patterns instead of their injury history. A major reason for concentrating on reported use behavior rather than injury behavior is that consumer product related injuries for a given individual are rare events and are a function of exposure to the hazard; this latter point being extremely difficult to measure and control for. Thus, efforts were centered on a type of behavior which is easier to measure and duplicate. Also, even though injury history and reported safety use behavior are different measures, a previous study (Staelin and Weinstein, 1974) reported they are positively related. More specifically, the results indicate that respondents with better reported use behavior live in households which tend to have fewer consumer product related injuries. For example, adults in the upper 25 percentile with respect to reported safety behavior also reported on the average 1/4 the number of injuries in their household compared to those in the lowest 25 percentile.

In addition to the above questions on knowledge and behaviors t-he respondents a) were asked how one should behave in the 11 behavior situations (the index formed from these questions is referred to as normative behavior versus actual behavior), b) were given five situationally-based questions involving risk and c) were asked a few questions about their family and school experiences (i.e., grades, courses, father's and mother's occupations, etc.). Finally, this information was supplemented by school records concerning the student's grade point and IQ.

Since there was a possibility that students might not report their true behavior in the above mentioned behavioral situations but instead give responses which reflect their belief of what they should do, correlations were calculated between the two forms of the behavior questions (i.e., normative and actual). The correlations were low (ranging from .3 to .6) implying that the respondents when asked about how they acted in a given situation reported their actual behavior instead of the behavior they felt was appropriate,

This degree of non-correspondence between reported actual behavior and normative behavior points out one of the major potential weaknesses of an educational program, namely that even though you can inform a consumer about how to safely use a consumer product, there is no assurance that this new knowledge will lead to changes in behavior. It is clear from the above results that, at least for our sample, there was not a one-to-one correspondence between beliefs about how one should behave and how they actually behaved. Consequently, changes in the belief structure with respect to correct behavior may not have a significant impact on changes in actual behavior, since facts such as the perceived cost (effort) to behave "correctly" also enter in the behavior decision.

RESULTS

Two different measures of true behavior were available for analysis; the first measuring normative behavior the second, the actual behavior. Each was analyzed separately using the estimation procedure described previously.

The normal approach for displaying estimates of the influence of each exogenous variable in "regression" type studies is to display the estimated coefficients (i.e., the Bj's) along with the coefficients' estimated standard errors. We did not use this approach for two reasons. First, the magnitude of the coefficients is a function of the units of the associated variables. Since most of the variables used to estimate the proposed model have an arbitrary scale (up to a linear transformation) the magnitude of the coefficients is also arbitrary. Second, and more importantly, we believe the real issue is the magnitude of the effect of specific exogenous variables on the measure of interest, where magnitude is measured in policy terms not statistical terms (i.e., t-tests). Thus we want to highlight the results in terms of the effects of each variable on the measure of interest (i.e., knowledge or behavior). The unit of analysis used in the presentation is the fraction of the standard deviation of the population distribution of the measure of interest prior to the educational program. For instance, the standard deviation of the knowledge measure (as determined from the pretest) was .42 units. Thus the effects on a student's knowledge score for a variable having an estimated coefficient of say .19 is .45 (=.19/.42) standard deviations for a one unit increase in the independent variable. It is this fraction (i.e., .45) that we present. Finally, for readers who may be interested in the "statistical significance" of the estimate from zero, we present the ratio of the estimated coefficient to the estimated standard error of the coefficient for continuous variables and one level dummy variables. [Many of the variables used in the analysis are discrete and thus lend themselves to a multi-level dummy variable formulation. Since the restriction used to estimate these multi-level dummy variables is arbitrary the above mentioned ratio is not useful for testing whether the specific level is different from zero, and thus not presented in these instances.]

The individual estimates are presented in Tables 2-5. The 2nd and 3rd columns of each table pertain to the analysis using the normative behavior measures, the 4th and 5th columns pertain to the analysis using the actual behavior measures.

The estimated effects for knowledge before the program (see Table 2) are approximately the same for both analyses. The estimates for identity of students' sex indicates that males tend to have higher knowledge levels of safety principles than females (the average estimated effect is .48 (=(.42 + .54)/2) standard deviations). Also, as might be expected, knowledge is positively influenced by native intelligence. To get a better feel for the influence of IQ we compare two students, the former having an IQ score one standard deviation above the population average (for our sample this was 11.4 IQ points), the other having a score equal to the population average. We would estimate that the difference in knowledge of safety measures due to the IQ difference to be approximately .34 (=.30 x 11.4/10) standard deviations. More simply, consumers with more native intelligence tend to know more about the principles of safety.

There seems to be a consistent negative influence of "theoretical" classroom experience which a priori was felt to be related to safety knowledge. In contrast there is a positive influence of "practical" exposure to consumer products. In both cases most of the effects are small, the major influences being the course in Auto Lab (a negative influence) and practical training in how to repair a TV or automobile (positive influences).

The home environment as measured by occupation of the head of the household has little influence on safety knowledge level. However, the estimated effect for students whose parents are skilled or unskilled workers is positive relative to the other occupational categories, suggesting again that consumers exposed to practical training tend to have better knowledge of basic safety principles. This feeling is reinforce by the estimated effects for the school dummies. School 2 (which is made up of students whose parents are primarily blue collar workers) had the most positive influence in contrast to the more professionally oriented school (school 1).

TABLE 2

ESTIMATED INFLUENCES ON TRUE KNOWLEDGE PRIOR TO EDUCATION PROGRAM

In summary, knowledge of safety concepts is positively affected by exposure to repairing or using consumer products and general intelligence. However, exposure to the more theoretical concepts taught in Physics or Chemistry or the practical school courses of Shop, Home Economics, or Auto Lab do not seem to help the student understand concepts associated with consumer product safety.

Table 3 displays the results for the two analyses of behavior prior to the program. Females, after adjusting for knowledge levels of general safety principles (where they tend to score lower), tend to report safer behavior. They also have a better understanding of how to behave (i.e., normative behavior). Students who indicated a general tendency to take risks tended to report less safe behavior; the strongest effects being observed in the normative behavior analysis. For example, we would estimate a person one standard deviation above the mean with respect to a propensity to take risks would report actual behavior..06 standard deviations below a person who was average with respect to risk taking. This relationship between safe behavior and propensity to risk is compatible with previous re-suits reported by Staelin and Weinstein [1974], who used an adult population for estimation.

TABLE 3

ESTIMATED INFLUENCES ON TRUE BEHAVIOR PRIOR TO THE EDUCATION PROGRAM

The relationship between knowledge of the concepts of safety principles and the two types of behavior differed depending on the measure of interest. For actual behavior, increased knowledge of safety concepts leads to significantly safer behavior. For example, the estimated effect for a student one standard deviation above the population average in knowledge is .85 standard deviations when contrasted to the behavior of a student having exactly the average knowledge level. This positive relationship is in accord with previous cross-sectional results (Staelin and Weinstein, 1974). However, there seems to be no significant relationship between the level of safety knowledge of a student and the student's ability to determine how to behave. We find this result to be counter-intuitive since our priors were that knowledge of safety concepts had a more positive influence on normative behavior than actual behavior. (Note that normative behavior is, in effect, a measure of knowledge of how to behave, and thus should be highly related to knowledge of safety principles.) We are unable to explain the discrepancy between our prior opinion about the relative magnitude of the coefficients and the estimates of these coefficients.

Table 4 indicates the degree of success of the educational program in terms of changes in knowledge of safety concepts. The estimated effects differ depending on the particular behavior analyzed, but the general pattern is similar. The effects of attending the eight sessions (denoted Treatment) was estimated to increase the student's knowledge level by .44 or .17 standard deviations depending on the behavioral measure used. The program seemed to be more effective for females. Consequently, the differential between the two sexes was reduced by approximately 38 per cent. Also, some minor increase in learning of safety principles occurred for the average grade point student relative to those who get the highest or lowest grades even though the student's IQ did not seem to be important. Finally, school 2 seemed to learn mere relative to the other two schools although the differences were not major. [It is impossible to determine if the between schools effects are due to the different teachers or the interaction of the characteristics of the student body and the educational program. Since the author taught primarily in school 3, he is pulling for the latter hypothesis.]

TABLE 4

ESTIMATED INFLUENCE ON CHANGES IN TRUE KNOWLEDGE DUE TO EDUCATIONAL PROGRAM

The effects of the educational program on behavior are found in Table 5. Again we note differences between the normative and actual behavior analyses; the estimated effect due to changes in knowledge of safety principles was positive and significant for actual behavior, but minuscule (and negative) for normative behavior.

TABLE 5

ESTIMATED INFLUENCES ON CHANGES IN BEHAVIOR DUE TO EDUCATIONAL PROGRAM

With respect to actual behavior the estimated influence of a shift in knowledge level of every student by one standard deviation is 1.03 standard deviations with respect to the behavior measure. Since the estimate for the effects of our particular program was a shift in the average level of knowledge of .44 standard deviations with respect to knowledge, we estimate the average change in actual behavior due to the increase in knowledge to be an upward change of approximately .45 (=1.03 x .44) standard deviations. However, this shift is substantially reduced by other aspects of our program which caused a mean decrease of .39 standard deviations. (See the Treatment variable in Table 5.) Thus the net gain associated with the educational program is only .06 standard deviations.

In contrast, the analysis of normative behavior indicates almost no influence due to changes in knowledge levels, but a significant positive shift due to other aspects of the program. Consequently the net effect is a mean positive shift of .37 standard deviations.

We find the above results somewhat puzzling. As in the equation for behavior before the educational program, knowledge is not related (significantly) to normative behavior but is to actual behavior. Also we are unsure of how to explain the negative mean shift in actual behavior associated with the program. What aspects of the program would cause the students to have a downward shift in actual behavior but not normative behavior? We can not answer the question but offer the following suggestions.

The first hypothesis, as might be expected, concerns the applicability of the basic model and/or the data used to estimate the model. Perhaps our data are so error-prone that the reported results are due to random fluctuations. Although this is always a possibility, we believe our measures, although fallible, are reasonable measures of the basic constructs. Also the model is designed specifically to handle fallible data. More likely, the effects of the educational process are more complex than our postulated model. For example, the program may have forced students for the first time to explicitly think about the issues of safety and this in turn caused the students to modify their basic risk taking structure. Consequently, changes in the student's propensity to take risk might be included in the model. Also, it is possible to argue that changes in normative behavior (which is really knowledge) must precede any changes in actual behavior. Thus, the educational process (i.e., changes in knowledge of safety principles) alters knowledge of how to behave, which in turn alters the student's actual behavior. Clearly these models are somewhat more complex than the one used in this study, but they are feasible with respect to estimation.

Finally we put forth a technical explanation for the results. The estimated coefficients for Change in Knowledge and Treatment in both analyses have a negative correlation of approximately -.54, indicating a reasonably high degree of multicollinearity between the estimates. Consequently, the individual estimates, although unbiased, could be highly unstable. However it is still possible to estimate the overall influence of the program by dropping either the Change in Knowledge variable or the Treatment variable. The latter was done and the results indicated, as would be expected from the previous discussion, that normative behavior was modified positively by the program, while actual behavior was for all intents and purposes unchanged.

CONCLUDING REMARKS

The above discussion although confirming a number of relationships previously determined from cross-sectional analyses, also raises a number of new questions. Specifically it points out the difficulties of uncovering the relationships between knowledge and behavior. However, the potential payoff associated with determining these relationships is great since policy makers could make better decisions if they knew more about how education works. For example, the results pertaining to the two different effects of education, if correct, offer much encouragement for consumer product safety education provided one can design a program which does not have the general negative influence associated with our program. On the other hand, taken overall, the program was successful only in modifying normative behavior not actual behavior. Only with careful modeling and closely controlled experimental design will it be possible to unravel the complex interrelationship of education, knowledge and behavior. Hopefully, the models and analyses presented above offer a foundation for further work.

REFERENCES

Paul N. Bloom, "How Will Consumer Education Affect Consumer Behavior?" in Advances in Consumer Research; Vol. 3, ed. B.B. Anderson (Cincinnati: Association for Consumer Research, 1976, 208-212).

Consumer Product Safety Commission, "Short Study to Estimate the Fraction of Product Related Injuries Addressable by Mandatory Standards," Office of Program Planning and Evaluation, Internal Consumer Product Safety Commission Document, February, 1976.

Sheila H. Feichtner, "Improving Consumer Safety through Innovative Consumer Education," Speech presented at the Women's National Safety Conference, Chicago, Illinois, October, 1973.

Joseph B. Kadane, Timothy W. McGuire, Richard Staelin, and Peggy Sanday, "Models of Environmental Effects on the Development of IQ," Journal of Educational Statistics, Autumn 1976.

J. Johnston, Econometric Methods, 2nd Editions New York: McGraw Hill, 1972.

Joseph C. Miller and A. Parasuraman, "Advising Consumers on Safer Product Use: The Information Role of the New Consumer Product Safety Commission," 1974 Combined Proceedings (Chicago: A.M.A. 1975).

Richard W. Olshavsky and John O. Summers, "A Study of the Role of Beliefs and Intentions in Consistency Restoration," Journal of Consumer Research, 1(1974), 63-70.

Jagdish H. Sheth and Nicholas J. Mammana, "Recent Failures in Consumer Protection," California Management Review, 1974, 16, No. 3, 64-72.

Richard Staelin and David Pittle, "Consumer Product Safety: A Discussion Paper," A.M.A. Consumerism Workshop Monograph, (Chicago: A.M.A. 1977).

Richard Staelin and Alan G. Weinstein, "Correlates of Consumer Safety Behavior," Advances in Consumer Research, Vol. 1, edited by Scott Ward and Peter Wright, (Urbana: Association for Consumer Research, 1974, 87-100).

Jeffery L. Thompson, "Product Safety: Suggestions for Better Use and Purchase Behavior Through Consumer Education and Information," Advances in Consumer Research; Vol. 1, ed. by Scott Ward and Peter Wright,(Urbana: Association for Consumer Research, 1974, 101-107).

George Tokuhata, "Childhood Injuries Caused by Consumer Products," Pennsylvania Department of Health, Division of Research and Biostatistics, May, 1972.

R. Whitaker, Hearings on H.R. 8110; H.R. 8157; H.R. 260, and H.R. 3813 Before the Subcommittee on Commerce and Finance of the House Committee on Interstate and Foreign Commerce, 92nd Congress, 1st and 2nd Sess., P.T. 3, at 1135 (1971-72).

Arnold Zellner, "An Efficient Method of Estimating Seemingly Unrelated Regressions and Tests for Aggregation Bias," Journal of the American Statistical Association 57, 1962, 348-368.

Arnold Zellner and H. Theil, "Three Stage, Least Squares: Simultaneous Estimation of Simultaneous Equations," Econometrica, 30, 1962, 54-78.

----------------------------------------