Consumer learning and evolution of consumer brand preferences

Another example of prior use experience being only partially relevant when consumer migrate to different versions of products is the camera category, where some consumers switch to more advanced cameras over time and where product usage requires consumer skills (learning-by-doing). Huang (2015) measures the returns to experience on a sample of users of digital cameras, via a measure of their picture quality.

In Section 3 (Data and Identification), we talk about these and other descriptive statistics in detail.

Our application is in the disposable diapers category. Heilman et al. (2000) modeled state dependence in disposable diapers category and their descriptive model results indicated that price sensitivities themselves change as a function of use experience in disposable diapers category. To allow for this possibility, we will also allow diaper-size specific price sensitivities in our application.

For example, Li et al. (2005) investigate customer purchase patterns for products that are marketed by a large bank. To do so, they estimate a multivariate probit model to investigate how customer demand for multiple products evolves over time and its implications for the sequential acquisition patterns of naturally ordered products (e.g., open a credit card account first when young and then applying for a mortgage). They do not model learning explicitly but our modeling approach can be adopted and adapted to model learning in such settings as well.

We thank to an anonymous reviewer for the insight that our framework is equally applicable to any repeated product choices from the same (umbrella) brand over the customer’s life-cycle.

Focusing on first-time parents also alleviates greatly the initial conditions problem that all dynamic models are subject to.

Previous papers that incorporated spill-over learning effects across products or products attributes assumed myopic agents and did not model duration dependence in utility (e.g., Erdem 1998; Coscelli and Shum 2004 and Chan et al. 2013). One exception is Dickstein (2011), who, like us, allows for forward-looking behavior. Dickstein considers a model with forward-looking physicians facing a multi-armed bandit problem, where a physician is uncertain about his patients’ intrinsic preference for drugs’ characteristics, and he makes use of patients’ total utility of consuming a drug in time t to update his belief about their preferences. The proposed model does not allow for risk-averse behavior or evolving needs.

Here “size” refers to the size of the individual diaper, such as the newborn size, and there are 5 sizes. It does not refer to the package size.

When a household does not make a purchase during week t, we assume his choice set included the last purchased size and its adjacent sizes as discussed in the above text.

Please note that we tried data on features as a control variable (we do not have display variable). However, the feature variable was statistically insignificant in the structural models we estimated so we did not use it as a control.

Indeed, Heilman et al. (2000) found that price sensitivities are time-varying and a function of cumulative use experience in the diapers category. We should also note that descriptive models (such as varying parameter models) have shown evidence of changing price sensitivities in few other frequently purchased product categories as well (e.g., Mela et al. 1997).

When babies first grow into a size, the fit to the new size may be not perfect, so utility may first increase as time passes and the fit gets better. Then, when the baby is about to grow out of a size, the fit may again diminish. The duration dependence term is set up so that after a consumer has been in size k for a while, she is more likely to move to k + 1 than to k-1. We thank an anonymous reviewer for this insight.

We also tried a utility specification where Equation 1 has a last brand purchase dummy to capture any one-lag state dependence effects not related to learning (e.g., switching costs (Dubé et al. 2009) or preference inertia (Che et al. 2007; Shin et al. 2012)). Indeed, Osborne (2011) found that in frequently purchased product categories there are both learning and switching costs. The coefficient of the last purchase dummy is identified in such contexts as discussed on Osborne (2011). We found that lagged purchase dummy has statistically significant but (size-wise) very small effect, and the results were very similar between the two models. We turned that component off for three reasons. First, we turned it off for parsimony since our model has already quite a few “moving-parts” as it focuses on evolution of needs (the need to switch to a different size), learning across sizes and changing price sensitivities. Second, a learning model fits our data better than a model with no learning but a lagged dependent variable or with a weighted average of past purchases variable. Third and most importantly, these lagged purchase variables added to learning models are behaviorally difficult to interpret.

The estimate of γ ₀ was not statistically significant so we did not report them in the result tables.

We should also note that we model brand choice and purchase incidence but do not model quantity choice (rather we model the impact of inventories on the probability of purchase incidence in a descriptive way). Previous papers on forward-looking dynamic structural models focused either on quality expectations, learning and strategic sampling in the context of brand choice or on both brand and quantity choice and price expectations but assumed away quality learning and strategic sampling since it is not feasible to model both processes in one structural model that explicitly allows for both quality and price expectations (Erdem et al. (2008). Furthermore, Ching et al. (2014) do so in a semi-structural model and find that in the diapers category the quality learning effects are significant whereas the price expectation effects are not for first-time parents.

Hartmann (2006) and (2010) are two exceptions that allow for richer unobserved heterogeneity structures.

In our sample, consumers only buy products across adjacent sizes and do not purchase more than one size up or down. The correlation matrix between adjacent sizes is set up to reflect the fact that there is a natural sequence in consumer’s purchase of different sizes. In a more general setting, to model the correlation between product classes (e.g., sedan, SUV and truck), it would be useful to generalize the correlation matrix to allow more flexible spill-over effects. We thank an anonymous reviewer for this insight.

More details of the learning model are provided in a Technical Appendix available upon request from the authors.

We fixed the weekly discount factor β at 0.995 since the discount factor is often difficult to identify even when certain variables can be found that affect expected payoffs but nor current utility (that is, exclusion restrictions may exist). For example, Erdem and Keane (1996) found, in a similar but simpler model, the likelihood was quite flat over a range of discount factors in the vicinity of 0.995, which was the case for us too. We estimated the model with few different weekly discount factors but the results were not very sensitive to the exact value of the discount factor. Please note that the best way to identify the discount factor is either to find contexts where proper exclusion restrictions and practical identification exist (e.g., Chung et al. (2014)) or use (experimental or field) data that has information on behavior both in static and dynamic contexts/regimes to pin down the discount factor (e.g., Yao et al. (2012)) but we do not have such data. There are indeed very few cases where such data are available.

That is, count all brand switches when a (an adjacent) size switch was made and divide it by the total number of size switches between adjacent sizes irrespective of whether there was brand switching or not.

When consumers are just starting with the newborn size 1, probably they are just learning about diapers and babies’ diaper needs in general and not too motivated to start sampling a lot right away.

We should also note here that most of the size switches occur to the adjacent bigger, rather than the adjacent lower size. Indeed, brand switching probabilities conditional on switching to the next bigger size are 12.5, 23.3, 21.6 % and 19.7 for sizes 1 through 4, respectively. When one counts only switches to the next bigger size, there is of course no such switch for size 5 as this is the biggest size.

The increased number of switches in larger sizes would occur if the impact of increased price sensitivity & elasticity dominates the effect of diminished overall strategic sampling on brand switching. To again check data patterns, we categorized the switching observations into two groups: when a household switched to a different brand when the price of the brand switched to is at least 5 % lower than its mean price, we categorized the brand switch observations as brand switching due to price promotion. Otherwise, we classified the brand switch as a not price promotion related brand switching (which could be due to strategic trial or other reasons). The size-specific brand switching observations categorized as “price promotion related” yielded percentages of brand switching to be 44, 48, 51, 59 and 63 % for Size 1 through Size 5, respectively. That is, while 44 % of all size 1 brand switching was “price promotion related” (and 56 % was “non-promotion-related), 63 % of total size 5 brand switching was “price promotion-related”. Thus, the data patterns suggest non-promotion related brand switches decline over time relative to price promotion related switches. Thus, consumers switch early on more for non-price related reasons (e.g., for strategic trial) while they will switch due to price variation in later periods.

The “true qualities” can be different across segments if there is a baby-diaper match issue. The same issue holds in many other categories as well though and it is not feasible to estimate a dynamic structural model that allows true qualities varying by households. Furthermore, even if the match issue exists, there is no reason to believe that not modeling it would bias the results in a systematic way.

This implication is consistent with data that show that there is indeed more brand switching conditional on size switching.

In the model fit simulations and counterfactual analyses, we first draw the parameter estimates from their joint distribution, then we solve for choice shares for each draw, and lastly calculate the mean choice shares across the draws.

Additionally, we also repeated the price simulations with the parameters obtained from a model without size-specific price coefficients (this model differed from ours only by the fact that price coefficients were not size specific). We obtained similar results and price elasticities are still larger for larger sizes but compared to the model with size-varying price parameters, this effect is dampened a bit as one would expect. Thus, price elasticties are larger for larger sizes even when does not allow size specific price coefficients but the effect is more pronounced when there are size specific price coefficients.