Consumer Impulse Purchase and Credit Card Usage: an Empirical Examination Using the Log Linear Model

ABSTRACT - Most of the work in impulse purchase behavior has investigated the association of socioeconomic variables and unplanned purchases with equivocal results. This paper examines the interrelationship between impulse purchases, credit card usage, cost of items bought, and purchasers' income. A log linear approach is used to test goodness-of-fit of several hierarchical models, and rather surprising results are found. These include the lack of interaction between impulse purchase and credit card possession usage even when consumers' income and cost of items purchased are controlled for.

Citation:

Rohit Deshpande and S. Krishnan (1980) ,"Consumer Impulse Purchase and Credit Card Usage: an Empirical Examination Using the Log Linear Model", in NA - Advances in Consumer Research Volume 07, eds. Jerry C. Olson, Ann Abor, MI : Association for Consumer Research, Pages: 792-795.

Advances in Consumer Research Volume 7, 1980     Pages 792-795

CONSUMER IMPULSE PURCHASE AND CREDIT CARD USAGE: AN EMPIRICAL EXAMINATION USING THE LOG LINEAR MODEL

Rohit Deshpande, University of Texas at Austin

S. Krishnan, University of Pittsburgh

ABSTRACT -

Most of the work in impulse purchase behavior has investigated the association of socioeconomic variables and unplanned purchases with equivocal results. This paper examines the interrelationship between impulse purchases, credit card usage, cost of items bought, and purchasers' income. A log linear approach is used to test goodness-of-fit of several hierarchical models, and rather surprising results are found. These include the lack of interaction between impulse purchase and credit card possession usage even when consumers' income and cost of items purchased are controlled for.

INTRODUCTION

Although unplanned purchases constitute a substantial proportion of consumer buying, little empirical research is available on this subject in marketing literature. This dearth of material is not a reflection on the significance of the research topic. A product category as ubiquitous as brown colas, for instance, depends for its sales almost entirely on unplanned consumer purchases. Limited conceptual and empirical studies of impulse buying have focused largely on demographics as predictors of purchase. Literature in retailing indicates that the level of unplanned purchases increases with the total dollar amount spent on a particular department store shopping trip (Prasad 1975). This is a fairly intuitive finding, yet this and other studies do not differentiate between items purchased along some unit-cost continuum, although this would seem to be a relevant variable to control for. Similarly, the mode in which the payment is made has not received empirical attention, yet it is important to distinguish between purchases made by cash and those using a store or bank card.

This last comment brings up the growing literature on credit card usage. Similar to early work on impulse purchasing, credit and store card owners' characteristics have been carefully studied. These characteristics encompass both traditional demographics and life style measures. In particular, bank card holders appear to be upscale in income and education and more socially active than non-card holders. More recent work has added mobility and benefit predictors to explain greater variance in the use of credit cards (Hirschman, Srivastava, and Alpert 1978). Yet not much light has been shed on the extent of planning prior to purchase. This is surprising since both supporters and detractors of credit instruments seem to hold strong views on whether or not the acquisition of credit cards increases the likelihood of impulse purchases (Hawes, Talarzyk, and Blackwell 1976).

The research reported here attempts to address these issues. The growth of the companies issuing bank cards, travel and entertainment cards, and store cards merits some understanding of their effect on purchases. Additionally, as mentioned earlier, it is necessary to distinguish between costs of the items purchased and also between incomes of the purchasers. The conceptual framework below ties together the elements of credit cards possession and usage, impulse purchase, cost of item, and income of purchaser with a set of propositional hypotheses that are later tested.

CONCEPTUAL FRAMEWORK

The hypotheses described below relate both to earlier empirical work in impulse purchasing and credit card usage areas, and to hitherto unresearched propositions that stern out of the lacunae in retailing literature mentioned briefly above. This permits a grounding of the research discussed in this paper in other work and allows also a venture beyond what is already known. Accordingly, each hypotheses is presented with a brief rationale including, where relevant, citations to previous literature.

Most empirical studies investigating demographic characteristics of credit card possessors have found them to have higher incomes than individuals who do not possess cards (Wiley and Richard 1975, Hirschman and Goldstucker 1978). This is almost tautological since having an income above a certain threshold is one of the prerequisites for card possession eligibility. Yet there are persons with upscale incomes who do not possess cards, and hence it is important to determine if this is true for the consumer sample in this study. Thus the first hypothesis to be tested is:  H1:  The higher the income of a purchaser, the more likely possession and usage of a card.

One can now begin to narrow the inquiry to specific purchases made using a card. Much earlier literature takes a largely macroscopic focus to assess attitudes and opinions of or about, card users (Mathews and Slocum 1969, Plummer 1971). But unless particular purchase situations where cards are actually used are studied, knowledge about the usage phenomenon will lack specificity.

One of the more important parameters when studying particular buying situations is the cost of the item purchased. Following from the first hypotheses, it would also appear that if credit cards are possessed by individuals, they would be used in purchase situations where cost of the item was relatively high. Accordingly, the second hypotheses is:   H2:  The more likely card possession, the greater the likelihood of card usage for higher cost items.

Turning now to the extent of planning that occurs before a purchase is made, it would seem this too is affected by cost of the item being purchased. Although factors such as time of shopping (i.e. morning, day, or evening), nature of the shopping trip (major or fill-in), existence of prepared shopping lists, and so on have not been found to affect the level of unplanned buying (Prasad 1975, Kollat and Willett 1967, Williams and Dardis 1972), earlier work on consumer risk-taking (Ross 1975) would indicate that there is an extended hierarchy-of-effects for a major, high cost of item, purchase. The third hypothesis to be tested, therefore, is:   H3:  The higher the cost of the item purchased, the more likely prior planning of the purchase.

Also, the issue of cost of a purchased item cannot be considered in isolation of purchaser's characteristics. The most relevant such characteristic is the purchasing power of the buyer, generally measured in terms of income. The hypothesis stemming from this reasoning is:  H4:  The higher the purchaser's income, the greater the likelihood of higher cost items being purchased.

Relating the favorite hypothesis to the earlier discussion on pre-purchase planning, it would seem that more affluent consumers can afford to spend less time in deciding what to buy since they are not hampered by budgetary stringency. This is expressed in the fifth hypothesis as:   H5:  The higher the purchasers' income, the greater the likelihood of unplanned purchases. (each of these terms is defined in the sections that follow).

Finally, since availability of credit-instruments such as bank or store cards increases the flexibility of consumers to go beyond (cash) budget limits, it would appear that card possessors are more likely than non card holders to make impulse purchases. This can be stated formally as:   H6:  The more likely card possession, the greater the likelihood of card usage for unplanned purchases.

Each of these six hypotheses were tested using data gathered from individuals in a retail shopping setting. The sample, data, definition of variables, and analysis are discussed below.

METHODOLOGY

Before describing the sample used in this study, it is necessary to make an important distinction. Work in the philosophy of science by Bunge (1963) and others (Kaplan 1964, Stinchcombe 1968) speaks to the major difference between associative and causal models. The former can be described as those where variables under investigation are tested for covariation. Causal models, on the other hand, are tested for the level of explanation provided by a set of one or more predictor variables impacting upon one or more criterion variables. These tests are based upon assumptions that a chronological sequence is theoretically determined or empirically observable, and that independent variables can be controlled so that researchers know that they are the only variables causing a change in the dependent variable.

The model described below is an associative one. Rather than make the assumptions of chronological independence, variables are assumed to be interdependent. Hence, testing of the model involves an empirical investigation of patterns of covariation of variables without implications of causality.

The necessity of an associative model becomes clear when one realizes that the model includes variables such as the unit costs of items purchased, as well as whether purchases were preplanned. The research reported here does not attempt to determine whether the availability of a credit card led to unplanned purchases of high cost item, or whether the cost of the item necessitated the use of a credit card. This study focuses on the specific interactions between card possession/usage, extent of prepurchase planning, cost of items purchased, and purchasers' income that have been suggested in the six hypotheses described earlier.

Data used in this study were taken from a survey of department store customers in tow metropolitan areas of population 2,000,000 and 500,000. 3,000 small intercept interviews were conducted in the first city and 1,225 in the second city. All interviews were conducted on a random intercept basis at five branches (of the same store) in the larger city and at one branch in the other city. Questions asked of respondents included the type and cost of items purchased, the mode of payment, whether the items were planned purchases, and attitudes regarding store image and various credit instruments. This study describes data from a subset of questions from the larger survey.

Respondents existing from a branch of the store were asked if they had purchased anything at the store. If they had made a purchase, the item was recorded by the interviewer as a product category. The cost of the item purchased (c) was coded into two categories: below or equal to $25, and greater than $25. Respondents were also asked whether they had made the trip to the store to buy the item purchased. The question was "Did you come to (this store) today for the specific purpose of buying (the item) or did you just happen to see it while you were here? Answers in the former category were coded as 'planned purchase', and in the latter as 'impulse purchase.' Additionally, respondents' incomes were ascertained and coded (for the purpose of this analysis) into two categories: below or equal to $15,000, and more than $15,000.

Finally, store and/or bank credit card possession was determined in a series of questions that preceded the ones mentioned above. Following Russell (1975), the analysis reported here categories consumers as possessor or non-possessors of cards without distinguishing between types of cards. After asking about particular items purchased at the store, respondents were asked what mode of payment was used: card or cash. The categories of card possession and mode of payment were collapsed into one variable, called card-possession/ purchase mode (P). The reason for this collapsing procedure is that consumers not possessing cards can only pay in cash and hence the category not possess/ card payment is meaningless. As such separating variables of card-possession and mode of payment will result in what is referred to as "structural zones" (Bishop et al 1975: Ch. 5) in the combined category of not possess/card payment. Accordingly, the variable card-possession/purchase-mode (F) can take 3 values: not possess/cash payment, possess/cash payment, and possess/card payment. The other three variables in the analysis are prepurchase planning (M) in the categories: planned and impulse; cost of item purchased (c) with values: less than or equal to $25 and greater than $27; and purchaser income (I) with values: less than or equal to $15,000 and more than $15,000.

Since the model to be tested is one of association and variables are categorically scaled, analysis used to fit the data utilized a log linear approach. As Green et al (1977) have indicated in their excellent seminal article, the log linear model is eminently suited to the analysis of qualitative (nominal) data to test interdependence structures involving unordered variable categories.

ANALYSIS

The four variables of card possession/purchase mode, prepurchase planning, consumer income, and item cost were cross-tabulated to form a multiway frequency table (Table 1). The total sample size (n=3756) refers to the number of purchases reported. These for the units of analysis.

Following the procedure suggested by Bishop et al (1975), and by Dixon et al (1977), a saturated log linear model can be fitted to the data in Table 1. This model includes all possible interactions (up to the fourth order maximum) for the four variables under consideration. Table 2 shows chi-square statistics for increasing orders of interaction. It can be seen that only the first two orders are significant. This indicates that the log linear model to be used does not require to be more complex than the second order to fit the original data adequately. This can be further seen in Table 3.

TABLE 1

MULTIWAY FREQUENCY TABLE FOR REPORTED PURCHASES

Table 3 describes specific interactions at second and lower orders that are (or are not) significant. As can be seen all main effects are significant as well as interactions of income with purchase mode, and cost with purchase mode and prepurchase planning. Both marginal and partial associations are described following Brown (1976) since a single test may not be sufficient to determine the relative importance of an effect.

The results from Table 3 are extremely interesting. The first three hypotheses posited (IXP, CXP, CXM) are confirmed. However, the next three hypotheses (IXC, IXM, PXM) are rejected. This means that there is

TABLE 2

TESTS OF SIGNIFICANCE FOR INCREASING ORDERS OF INTERACTION

no association (as for as this sample is concerned) between income and cost of item, or income and impulse purchases, or the mode of purchase and impulse purchases. These rather surprising and counterintuitive results need further explanation. Table 4 shows statistics for goodness of fit when the unsaturated model of IP, CP, CM is tested. This is the model suggested by the significant interactions in Table 3. Besides having chi squares that are statistically significant, adding more variables does not substantially improve the fit (i.e. this model is sufficiently complex to explain the hypothesized interactions). As Bishop et al (1975: 327-332) suggest, Pearson's chi squares of 18.01 and likelihood ratio chi squares of 19.02 with 13 degrees of freedom is more than a satisfactory fit.

Since a log linear model is fitted to cell frequencies of the original data (Table 1), it can be expressed as the natural logarithm additive function of main effects and interactions (similar to ANOVA). The unsaturated model in Table 4 therefore describes the function:

LN F_ijkl = q + l^I + l^C + l^P + l^M + l^IP + l^CP + l^CM

where q is the mean and the lambdas are parameter values for each main effect and specified interactions. Estimates of these parameters are shown in Table 5. In terms of main effects, the greatest impact on the logarithm of the expected cell frequency comes from the cost of the item (absolute value 0.924), followed by card possession/purchase mode (primarily not possess/ cash payment of 0.802 and possess/card payment of 0.788), and then preplanning of purchase (0.478) and income (0.274). However, lambda parameters of the iterations are even more interesting.

TABLE 3

TESTS OF PARTIAL AND MARGINAL ASSOCIATION FOR SPECIFIC INTERACTIONS

TABLE 4

GOODNESS-OF-FIT STATISTICS FOR UNSATURATED MODEL

It appears for instance (looking now at both directions and magnitude) that for low income respondents (< $15,000) the not possess/cash payment is used over the possess/cash payment and possess/card payment. This seems intuitively sensible as does the fact that higher income respondents (> $15,000) use cards when they possess them (0.303) over cash (0.190), or cash when they do not possess cards (-0.493).

The pattern for cost of item also shows a trend toward cash payment even when cards are possessed for low cost items (0.179 and 0.097) and a reverse trend for higher cost items (0.275 and -0.179) where cards tend to be used.

Finally, impulse purchases seem to occur for low cost items (0.093) and prior planning is done for items that cost over $25.

Although the analysis thus far is replete with both anticipated and nonintuitive findings, it is incomplete without an examination of the stability of the parameters described. Tables 6 and 7 attempt to remedy this deficiency. Table 6 shows standardized residuals which are calculated as follows.

standardized residuals = observed cell frequency - fitted frequency

                                                                           (fitted frequency)^1/2

                           i.e.   = (f_ijkl - F_ijkl)

                                       (F_ijkl)^1/2

TABLE 5

ESTIMATES OF THE LOG-LINEAR PARAMETERS (LAMBDA)

TABLE 6

STANDARDIZED RESIDUALS

The sum of squares of these residuals is the chi square test. And it can be seen that values of not possess/ cash payment for cost of items above $25 seem to be causing the error (for instance, -1.492 and 1.626 for planned purchases). If we refer back to the multiway contingencies in Table 1, we can see that these high residuals are due to low numbers of cases in the respective cells (for instance, 19 and 24 for the two residuals cited). Clearly, the relative paucity of observations in these cells is leading to the minor instability which keeps our model from providing a better fit, although the overall fit is extremely good (as evidence by consistently low residuals in other cells).

CONCLUSION

Briefly summarizing the findings from the analysis, ownership and usage of credit cards is associated with both income and cost of items purchased. Higher income consumers seem more likely to possess cards, and use them to buy higher cost products and services (i.e. over $25). Additionally, impulse purchases are associated with lower coat items.

However, unplanned purchases appear not to be associated with income (i.e. both higher and lower income consumers make impulse purchases), or with the mode of purchase. This last result means that the popular sentiment that credit card ownership is more likely to stimulate impulse purchases is not borne out by this data. As indicated above, cost of item (rather than income or purchase mode) is the only variable in this analysis that is likely to explain any variance in planned or unplanned purchase behavior.

Finally, consumer income and cost of items purchased do not show an interaction. We caution that this is true for the particular coded categories of income thresholds of $15,000 and cost thresholds of $25.

As indicated in the analysis, third order interactions are absent. So more explanation is unlikely from using additional variables in the above propositions. This leads us to offer some initial statements about causality.

If we were to redefine this study to build a predictive, causal model rather than descriptive, associative one, we might seek to measure variables such as item cost and income on an interval scale in the hope of getting additional explanation. Also, the research design itself may be further refined to look at other variables such as types of product and service categories purchased, and also whether they were on sale at the stores. This last issue could be a major predictor of whether a consumer decided to purchase an item on "impulse." Theoretical discussion in this area is limited, but it is possible to hypothesize that desire to purchase an item is latent in consumers' minds, and is triggered by discounts or heavy point-of-purchase promotional displays. Although this may indeed provide further explanation, the study described in this paper suggests that the presence or absence of credit cards will not directly stimulate an impulse purchase when such a purchase is defined as the acquisition of an item which the consumer just happened to see in a store (without prior planning to make a trip to buy the item). Alternative definitions of impulse purchases that distinguish between latent consumer cognitive demand and physical purchase are possible, but are beyond the scope of this paper.

In conclusion, it appears that the phenomenon of impulse purchasing allows for more theoretical and empirical investigation. The area is rich enough to justify further programmatic research on the cognitive processes a consumer goes through in making an unplanned purchase and also in determining what factors (such as word-of-mouth communication, point of purchase pronunciation, etc.) increase the likelihood that disposable funds will be allocated to one product or service over another in retail settings.

REFERENCES

Bishop, Y. M., Fienberg, S. E., and Holland, P. W. (1975), Discrete Multivariate Analysis: Theory and Practice, Cambridge, MA: MIT Press.

Brown, M. B. (1976), "Screening Effects in Multidimensional Contingency Tables," Applied Statistician, 25, 37-46.

Bunge, Mario. (1963), Causality: The Place of the Causal Principle in Modern Science, New York: Meridian.

Dixon, W. J., Brown, M. B., Engelman, L., Frame, J. W., and Jennich, R. I.(1977), BMDP-77: Biomedical Computer Programs P-Series, Berkeley, CA: University of California Press, 297-332.

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