CONTENTS

Dissecting the Advertising Effects on Household Milk Purchases

Director’s Column

Next Meeting

Special Cash Opportunity

NEC-63 2002 Spring Meeting March 21-22, 2002 Holiday Inn Albuquerque, New Mexico Allocation Issues in Check-off Programs

Printable pdf version

Dissecting the Advertising Effects on Household Milk Purchases

by Diansheng Dong and Harry M. Kaiser
(Cornell University)

Previous work to analyze the effects of generic advertising has generally been based on aggregate marketing data. However, the availability of panel data (household data over time) on household purchasing behavior allows us to investigate more deeply on how advertising influence household purchase decisions. Particularly, the household data can help us to identify whether the increased milk purchase induced by advertising comes from the new consumers or the old ones. Or whether the increase comes from more frequently purchases, or from more spending per trip. This type of information is valuable for marketing policy makers to craft advertising strategies.

The use of panel data to study household commodity purchases raises, in general, two issues. The first has to do with the temporal linkage of purchasing arising from state dependence caused by the purchase carry over, learning behavior, and other factors. This is a common phenomenon in aggregate time-series models. The temporal linkage of purchasing in panel data models, unlike in aggregate models, arises not only from state dependence, but also from unobservable household heterogeneity. Heterogeneity across households persists over time. It may be caused by different preferences, endowments, or attributes.

Another important issue in the use of panel data is how to control for censoring bias. The censoring problem comes from non-trivial proportion of zero-purchase outcomes made by households. The Tobit-type censored model interprets the household zero-purchase outcomes as being the result of strictly economic decisions, i.e., goods are not purchased when they are too expensive (corner solutions). However, not all zero expenditures reflect corner solutions or rationed behavior. Non-purchase could be the result of short-run consumption behavior (infrequency of purchase), or a social, psychological or ethical distinction unconnected to price and income (double hurdle). As examples, vegetarians do not shun meat because it is expensive and, many non-smokers would not smoke even if tobacco were free. This suggests that zero expenditure may be best modeled by means of discrete variables altering the nature of individual preferences. For household panel data, appropriate temporal aggregation may eliminate the infrequency-of-purchase problem, but the problems of possible corner solutions and double-hurdles will persist.

In this study, we extend the double-hurdle model used in cross-sectional data to a panel data structure. This extended model envisions that households must overcome two hurdles before realizing a positive purchase: (1) entering the market (becoming a potential purchaser), and (2) making the purchase. Therefore, in addition to the Tobit type purchase equation, a discrete equation is defined to determine the participation decisions. While accounting for censoring bias through the double-hurdle structure, we also account for temporal linkage.

We are particularly interested to the factors that influence household milk purchases over time. For each given time unit, we observe whether the household purchases, and if it does, the amount. The model provides information on what variables influence the consumer’s discrete decision of whether or not to participate in the market within a particular shopping period and, if the consumer participates, the continuous decision of how much to purchase. We follow a panel of upstate New York (excluding New York City) households over a four-year period from 1996 through 1999. We focus particularly on the hurdle equation to see if the non-economic barrier exists in milk purchases. We are also interested in the estimation of the parameters that captures the household heterogeneity in preferences and the state dependence, as well as the impacts of price, income, advertising, and other demographic variables on household purchase decisions for fluid milk over time.

Household data are drawn from the ACNielsen Homescan Panel, including household weekly purchase information for fluid milk products and annual demographic information. To eliminate the possible infrequency-of-purchase problem, the weekly purchases are then aggregated to monthly purchases. Monthly generic-fluid-milk advertising expenditures for upstate New York are obtained from Dairy Management, Inc., and the American Dairy Association and Dairy Council (ADADC). The two data sets of purchase and advertising are merged over a 48-months period from January 1996 through December 1999 for 1,320 households. Generic advertising expenditures vary over time, but not across households. The total number of observations in this sample is 63,360 (48 x 1320).

In this application, we concerned with monthly purchases of fluid milk for home consumption only. The monthly household purchase quantities and expenditures are defined as the sum of quantities and expenditures on all types of fluid milk such as whole, reduced fat, and skim milk purchased within that month. The dependent variables in our model are household fluid milk purchase quantities. Among the 1,320 households, 16 did not purchase any fluid milk in the whole time period. Among the purchase households, on average 30 of the 39 months are purchase occasions with a mean purchase of 3.32 gallons over all months and 3.67 gallons for purchase months.

Generic advertising used in this analysis includes monthly national and upstate New York milk advertising expenditures aggregated over all media types. The effect of advertising on consumers’ behavior could last as long as 9 months. In this analysis, the advertising expenditures are lagged 9 months and a polynomial lag model is adopted to capture the long run effect of advertising.

Prices are not observed directly in the household scanner panel data. An estimate of price can be obtained by dividing reported expenditures by quantity for the purchase months. However, no price information is available for those non-purchase months. A number of alternative approaches can be used to obtain estimates of the missing prices. In this analysis, we assume a zero-order correction for the missing prices. For each household the imputed prices for non-purchase months are set equal to the mean price of the purchase months for that household. If the household did not purchase over the whole period, the monthly mean prices over all the households are used. A number of annual household characteristics are also incorporated as explanatory variables, such as household income and household size.

Since the participation equation captures only the non-economic factors that influence the household decision to join the market, as stated earlier, price and income are excluded from the participation equation. However, advertising is assumed to alter household preferences, and therefore impacts the participation equation. As expected, generic advertising has a positive and statistically significant impact on participation. Other variables with positive and significant effects on participation include: percentage of teenagers in the household, percentage of persons over 65 years of age in the household, household size, education level of household head, households living in metropolitan areas, and middle-aged couples with no children. Significant household characteristics negatively associated with participation included: African American households and households where the mother is employed. The direction of impacts of all household characteristics was consistent with our a priori expectations.

Since the double-hurdle model nests the Tobit model, we perform a likelihood ratio test between the two. The test rejects the hypothesis that the first hurdle does not exist, which indicates that some of the zero milk purchases are due to non-economic reasons. However, the expected probability of overcoming the first hurdle is 0.91, which implies that most of the zero purchases are due to the second hurdle, i.e., the economic reason.

In contrast to the participation equation, the purchase equation captures the economic factors that affect household purchases. We found that both household income and generic advertising have positive and significant effects on household milk purchase, as we had expected. We also found household size to be positively related to milk purchase, whereas milk price is negatively related. The percentages of teenage girls and elderly persons in the household have a positive impact on household milk purchases. Consistent with the findings of previous studies, single-person households purchase more, while middle-aged couples without children purchase less. Surprisingly, the percentage of children under 12-years-old in the household had a negative effect on household milk purchases. However, this variable was positively related to participation, as discussed above. This may indicate that the households with more children were likely to participate in the milk market, but given their participation, adults would consume more than children. Relative to white households, Hispanic households have higher, African Americans have lower, and Asian households have the same level of milk purchases. As with participation, the employment status of the female head of household is negatively related to milk purchase.

Habit persistence is found in both the purchase and participation equations from the statistically significant estimates of related parameters. In fact, the correlation coefficient between current purchase and last purchase is 0.7619, and that between the current and last participations is 0.1168. This result means that lagged purchases and participation are positively related to current purchase and participation, respectively. However, more temporal dependence is found in the purchase equation than in the participation equation. This difference indicates that purchase relies more on previous behavior than does participation. Further, for the purchase equation, the component of temporal correlation associated with serial state dependence is 0.0716, and the component of this correlation associated with the household heterogeneity is 0.6903. The positive values imply that if household A purchased more than household B at time t-1, then household A will still purchase more than household B at time t, ceteris paribus. We see that, in this purchase equation, most of the correlation comes from the household heterogeneity. This results from the difference in household preferences for milk: household A prefers to drink more fluid milk than household B does.

To better understand the economic effects and to interpret the dynamic results of the model, we calculate elasticities of some key variables based on the expected values derived from the model. The elasticities of the last month in the sample evaluated at the household sample mean with respect to various expectations are presented in Table 1. The elasticities of the second and the twentieth month are also computed, with results quite close to the last month results.

The long-run elasticity of generic milk advertising is 0.149. In other words, a 1% increase in generic advertising would increase household milk purchases by 0.149%, on average. The 0.149% increase in household purchase counts as 0.058% (38.9%) from the increase of household milk purchase probability and 0.091% (61.1%) from the increase of household conditional milk purchase. An increase in purchase probability implies an increase in purchase incidence or number of purchasers. Thus, of the total impact of advertising on household milk demand, about 40% of the effect comes from purchase incidence. The elasticity of advertising on participation is 0.0065, which contributes 11.2% to the elasticity of the positive purchase probability (0.058). This implies that advertising increases the purchase probability mostly (88.8%) by overcoming the second hurdle. This finding allows us to interpret the effects of advertising as follows. If milk is not in the household’s preference function, advertising may convince them to include it (the first hurdle). Also, if milk is already in the household’s preference function, advertising may increase the weight the household places on it (the second hurdle).

Table 1 - Estimated Elasticities

* The t-test based on the standard errors derived from the Delta Method (Rao) showed that these elasticities are significant at the 0.05 level or higher.

As expected, the price elasticity is negative and inelastic at -0.078. The income effects are relatively low, while household size has a much more prominent effect. Compared to all the households, positive purchase households appeared less sensitive to price changes, given that the total price effect is composed of the purchase probability effect. Interestingly, the effects of all the variables in increasing unconditional purchase quantities through the increase in the conditional purchase quantities are weighted more than through the increase in the probability of purchase.

The last two columns in Table 1 indicate that the elasticities of current purchase probability vary depending upon whether a purchase occurred during the last period: the results were more elastic when there was no purchase occasion than when there was a purchase occasion. The positive value will increase the purchase probability that would increase purchase incidence, or reduce the inter-purchase time.For example, a 1% decrease in price would increase the current purchase probability by 0.1128%, given a non-purchase occasion, and by 0.0203%, given a purchase occasion, during the last period. In both cases, the inter-purchase time tends to shrink.

In conclusion, we found that generic dairy advertising could increase the probability of market participation; that is, advertising attracts new participants into the dairy market. Temporal dependence was found to be statistically significant in both purchase and participation equations. However, purchases are much more dependent on previous behavior than is participation. Generic advertising was also found to increase simultaneously the purchase quantity and purchase incidence. In addition, advertising increases the purchase probability more given non-purchase in the prior time period, than if a purchase occasion occurred, which is an intuitively appealing result.

[ top ]