It uses a unique data set from one of the largest mail order companies in the United States. Prior to building the models, a brief history of the mail order industry is included to alert the reader to the level and type of information gathered over the past century.
Introduction
- What is Direct Marketing?
- Current Retail Trends
- Current Issues in Mail Order and Direct Mail
- Marketing Issues
- Management Issues
- Court-Mandated Issues
- Firm Background
- Base Business Practice Areas
- Promotional Practices
- Products
- Target Customers
- What’s in the Dissertation
This dissertation focuses on the basic mail order business functions of a company and its customer base.4 The company is unique in many ways. If the customer eventually repays the amount, even after a long period of time, the company will reactivate it.
A Brief History of Mail Order
- The Late Nineteenth Century
- Turn of the Century
- The Depression Years
- The 1960’s
This Midwestern rivalry set the stage for the growth of the large mail order business.2. Another major advance that benefited the mail order business was the addition of rural routes.
Linear Probability Models versus Discrete Choice Models
- Simulation
- Set-Up and Execution
- Results
- Company Data
- Set-Up
- Results
- Modifying the Simulation Dataset
- Set-Up and Execution
- Results
- Correcting for Differences in Underlying Prob
For each replicate, the decile in which the observation (ie, client) occurred was saved for each client. First, it is necessary to examine the reason for the differences between the linear probability model and the logit model.
Set-Up and Execution
Results
Correcting for Differences in Underlying Prob
- Set-Up and Execution
- Results
- Future Research
- Discrete/Continuous Model of Purchase Behavior
- Introduction
- Discrete/Continuous Model of Mail Order De
Furthermore, comparing the false negatives and false positives for each model should be counterproductive. This model uses data from a specific period (December 1993) and estimates the dollar amount purchased based on the consumer's classification.1 2 The results of the preliminary estimate.
Results of the Three-Choice Model
The first tables below give the independent variables used in estimating a three-choice model.4 The number of promotions received (ie, catalogs, 8-prods, and solos) is included in the model. The results of estimating a three-choice model for the period beginning December 1993 are presented in the following tables.
Comparison to Other Estimation Techniques
Comparison of Three Choice and Nested Discrete/Continuous Models
- Three-Choice Model Results
- Two-Level Binary Choice Results
Note also that in the cross-logit model, five of the nine estimated quarters have coefficients in the positive range of unity for the inclusive value (refer to Table 5.2). The three-choice model includes two correction terms—one for the “buy once” and “buy multiple” options. Except for the second and third quarters (June 1992 and September 1992), the estimates for the three-choice model turn out to be statistically significant.
Statistically significant coefficients for the nested model are observed in only four of the nine cases. Again, as in the previous chapter, this appears to support the hypothesis that there is bias in estimating the dollar amount purchased given that a purchase has been made. It was also demonstrated that for the three-choice model for the quarter beginning December 1993, the dollar amount attributable to the selection correction term is significant in dollar terms - 35% of the amount purchased.
At the point when each logit and regression has a better “fit,” the log-likelihood could be compared to determine whether the models' fits are distinguishable. The discrete choice models are presented first, followed by the continuous model estimation of the amount purchased. For each of the nine quarters analyzed, the second step or "buy once" versus "buy multiple times" is given first.
Finally, for each of the nine quarters, the estimate of the dollar amount purchased is presented.
Discrete Time/Discrete Choice Duration Model
- Introduction
- A Model of Consumer Behavior
- Duration Models for Inter-Purchase Times
- Duration Model Notation
- Time Independent Duration Model
- Reservation Value Duration Model
- Comparison to One-Factor Models
- Nested Duration Model
- Estimation Procedures
With the receipt of each piece of "junk mail", the consumer decides whether to purge. It is based on the assumption that each period within an individual is independent of the choices made in every other period for the same individual. Note that the conditional probability of not exiting by period s, given not exiting at period s — 1 is given by.
The exit probability in period s, given that period s — 1 does not exit, can be given by z. The unconditional probability is derived from the integration of the conditional probability with respect to the distribution of the latent factor. The complexity of this last integral limited the utility of the one-factor model.
To estimate the probability of exit in the first phase of the model, period s — 1, the inclusive value must be calculated. The following table summarizes the independent variables used in the estimation of the duration models.
Duration Model Hypotheses
3Note that for the first two hypotheses we only need to optimize one variable at a time. Similarly, the second hypothesis reduces to a problem in one dimension, time since last order. Although not interesting, it would be possible to optimize ∕(c,r) for both and r. . for a different number of catalogs received in each period.
The third hypothesis can be tested by inspecting the coefficients of the variables Spring * Number of Catalogs, Fall * Number of Catalogs, and Winter * Number of Catalogs.
Discrete Duration Model Results
- Hypothesis Testing for the Time Independent Duration Model
- Tables
- Hypothetical Catalog Policies
- Possible Extensions to Analysis of Hypotheti
At subsequent periods, the propensity to repurchase decreased to the point where the optimal number of catalogs exceeded twenty per thirteen week period. The results of the test of the second hypothesis indicate that with the passage of time since the last order, given that the number of catalogs received per period is constant, there is a greater propensity to purchase. In this section we analyze the propensity to buy across a variety of policies that determine how many catalogs per quarter a consumer will receive.4 The profile of a typical individual will be used to compare the effectiveness of the current practice of catalog distribution with those of different levels.
The average number of catalogs per quarter, given that the customer reordered during this period, is and 4 for the eight quarters studied.5 This policy will contrast with various alternatives that use a more diverse sequence in terms of the number of catalogs per quarter. The first policy is simply the current average of the catalogs shipped, given that the consumer is not 4This analysis used a three-choice model with NoPurchase, Purchase, and BadDebt/Not Pro. The third and fourth are variations that reduce the number of catalogs in the periods immediately after purchase, increase and then decrease.
In addition, the peak probability is higher in the policies that vary the number of catalogs than it is in the constant number of catalogs strategy. The effect of an additional catalog in the second period is striking when comparing the current strategy with constant number of catalogs.
- Discussion and Conclusion
- Duration/Continuous Model — Theory
- Introduction
- Continuous Model Notation
- Time Independent Duration/Continuous Model
- Reservation Value Duration∕Continuous Model
- Nested Duration/Continuous Model
- Duration/Continuous — Results
- Discussion and Conclusion
- Future Research
An important point is that this analysis assumes that the order size is equal regardless of the inter-purchase time. The effective rate of return on the firm's capital (i.e., the trade-off in sending catalogs this period versus next) can then be solved for using the first-order conditions. Not only have we shown that the relationship between the number of catalogs sent to a consumer and their purchasing behavior is not linear, but also that the timing of the catalogs relative to their last purchase is important.
For the company used in the analysis, a 0.1% increase in attractiveness is equivalent to adding $1 million to operating income.9 The importance in terms of profitability should not be underestimated. This can help identify additional market segments and personalize the catalogs. Because the amount is only observed for individuals who take action, we consider a correlation between ¾ and the unobserved factors that influence the timing of the choice.
The first table below presents the continuous dollar amount results for the time-independent duration model, and the second table presents those for the continuous dollar amount for the reservation duration model. A comparison of the two models reveals that both have a "low" R2, but recognize important factors in deciding the dollar amount.
Appendix A Micro-level Data
- Current Customer Database
- Order Detail
- Promotion Detail
- Customer Service Detail
- Dependent Variable Construction
- Discrete/Continuous Model
- Time Independent Duration Model
- Independent Variables
- Duration Model Quarters
- Data Overview
Credit History (Internal) - records summary information about a customer's payment history, including the number of days since last payment, the amount, their balance and the number of days past due. In addition, statistics such as the maximum, average for the last six months and number of cases of the payments and overdue amounts are tracked. A customer's response statistics, including the number of sales/no sales and total minutes on the phone, by period, by quarter, by year, by day of the week, by time of day, etc.
An overview is also kept of the number of orders placed, canceled orders and time spent as an intermediary customer. Although it does not contain detailed information about the prizes (free gifts) sent with an order, it does record the number of such gifts. It contains the sales project, page and list for each piece of junk mail. The combination.
The total number of back-end promotions sent to the selected customers is given in Table A.3. However, it should be noted the relatively large number of individuals who become non-promotable within six months (two quarters) of an order.
Appendix B Future Research with Additional Data
- Pricing
- Promotion Detail — Additional Information
- Order Detail — Additional Information
- Cohort Analysis
First, Kashyap [28] reported on price changes and found support for the concept of fixed prices in catalogs. This could be related to the merchandising of “kits” and to the practice of dividing a general merchandise catalog into several. As for this company, applying Krishna's model to the analysis of end-of-season electronics purchases can provide new information about price elasticity when promotions are announced.
The data available for this analysis includes the products ordered by the 50,000 customers over a period of thirteen quarters. The post dates and quantities of each promotion for the period examined are available. Each item ordered and the free gifts, along with their prices, order dates and shipping dates, are available.
Because these data are available over a relatively long period of time, an analysis of cus. Cohorts that have become back-end customers at the same time can be analyzed for the length of time they remain customers and compared to other cohorts.
Appendix C Proof of Proposition 2
Bibliography
The validity of state use tax on the distribution of catalogs and other promotional materials.
