MANAGEMENT SCIENCE doi /mnsc ec pp. ec1 ec15

Transcription

1 MANAGEMENT SCIENCE doi /mnsc ec pp. ec1 ec15 e-companion ONLY AVAILABLE IN ELECTRONIC FORM informs 2008 INFORMS Electronic Companion Competition Between Local and Electronic Markets: How the Benefit of Buying Online Depends on Where You Live by Chris Forman, Anindya Ghose, and Avi Goldfarb, Management Science, doi /mnsc

2 ec2 Appendix A: Further description on construction of Purchase Circle variables In this section we provide further details on the construction of our data set from Amazon Purchase Circles. Raw Data from Amazon: Our initial data set consisted of 712,359 observations that we collected from the Amazon Purchase Circles. These data were collected once a month using the spider program described in the body of the paper. Each location in our data set had between 10 and 20 products listed as the top products sold when we collected the data for that month. To maintain a consistent sample across locations, we kept only product-location-time observations that appeared in the local top 10. In these raw data, a given product might appear multiple times: once for each location in which it appeared in the local top 10. Appendix Table A.1 describes the variables that appeared in these raw data. Appendix Table A.1: Variables in Raw Purchase Circle Data Variable Name Top 10 Rank Location State ASIN Date List Price Retail Price Sales Rank Release Date Average Rating Number of reviewers Description Number between 1 and 10 that indicates a product s local top 10 ranking. String that indicates the name of the location. State of the location. Book identification string. Date when the data were collected. List price of the book (typically, the price that is indicated on the cover of the book). Price at which Amazon sold the book on the date of data collection. Sales rank of the book indicated on Amazon. Data when the book was initially released. Average valence of reviews for the book. Number of reviewers who had reviewed the book by the date of data collection. Derived Data: Our raw data set contained thousands of ASINs. It was computationally infeasible for us to examine whether each of these thousands of ASINs appeared within the local top 10 list in a particular location in a particular month; moreover, many of these ASINs appear only one or twice in our data so the use of product-location fixed effects was infeasible because of the very small number of location-times in which such unpopular ASINS appear. As described in the text, for each time period we identified the 300 ASINs that appeared in the largest number of locations and made this our choice set. We aggregated all other ASINs into an outside option choice. Details on this procedure are provided below. Thus, the final data set is a panel where the unit of observations is a product-location-time.

3 ec3 As noted in the text, our primary dependent variable, LocalTop10 ijt, is a binary variable that is equal to one if book i is present in the local top 10 in location j in month t, and zero otherwise. In cases where a location purchases one of the products outside of the top 300, LocalTop10 ijt =1 for the outside option. We use a number of characteristics describing Amazon products in our analyses, including relative price, log(number of days since launch), average rating, log(number of reviewers), and a popularity spline. Further details on each of the variables are included in Appendix Table A.2. Most products (ASINs) in our data do not appear in a location-time (as there are over 300 products and only 10 that appear). Since product characteristics in our data do not vary across locations, we compute average values for each of our product characteristics across locations in which the ASINs appear and use these values (there are minor variations in some of the characteristics due to small changes that occur while our Spider is collecting the data; for example an extra review may be posted). For the outside option product, we compute the average of all products outside the top 300 that appear in our data.

4 ec4 Appendix Table A.2: Description of Derived Variables Variable Name Dummy for Top 10 in location Relative Price (rank 1-150) (rank ) Moderately (rank ) Somewhat Less (rank ) Less (rank ) Unpopular Products (rank over 15000) Dummy for missing price information Average rating Log(says since launch) Broadband Dummy for missing elapsed date information Log(Number of reviews) Discount Store Entry within 5.4 miles Large Bookstore Entry within 5.4 miles Description =1 when ASIN appears in the local 10 top list =(Amazon retail price undiscounted list price)/undiscounted list price =1 if national sales rank is in top 150. =1 if national sales rank is between 150 and 500. =1 if national sales rank is between 500 and =1 if national sales rank is between 1500 and =1 if national sales rank is between 5000 and 15,000. =1 if national sales rank is above 15,000. =1 if there is no information on Amazon retail price or list price for the ASIN. Average valence for the ASIN for all reviews from release data to date of data collection. =log(data collection date release date) Number of broadband provides for each zip code in the location, averaged across zip codes (from FCC Form 477 data). Collected semiannually. Data for April 2005 though May 2005 from FCC report from December 2004; data from June 2005 through November 2005 from FCC report from June 2005; data from December 2005 and January 2006 from FCC report from December =1 if release data is missing =log(number reviews from release date to data collection data) =1 if average distance between location zip codes and zip codes from location of store entry is less than or equal to 5.4 miles =1 if average distance between location zip codes and zip codes from location of store entry is less than or equal to 20 miles

5 ec5 Appendix B: Robustness This appendix includes additional checks of the robustness of our core results. Appendix Table B.1 shows that our results are robust to the use of different splines and measures of product popularity. 6 Appendix Table B.2 shows that our results are robust to using absolute distance rather than the distance dummies (5.4 miles and 20 miles) that we use in our core results; in particular, stores that enter closer to the location will have a stronger effect on online behavior than those that enter farther away. Appendix Tables B.3 and B.4 show that our results are robust to redefining entry as occurring one month after opening date, two months after opening date, and one month before opening date. These results demonstrate that noise in our measurement of the timing of entry would not influence our results. Moreover, they demonstrate that our results do not reflect simply a short-run effect of entry; the effect of entry remains even two months after store opening; they are also complementary to the results in column 5 of Table 3 in the paper than show our results continue to hold when we examine the effects of entry that occurred five months ago. Appendix Table B.5 shows that our results are robust to a broader choice set (1000 products) and to an alternative method of estimating broadband penetration in months that do not coincide with the FCC s collection of Form 477 data. Appendix Table B.6 shows that our results are robust to both high population growth and low population growth locations, and are not capturing a transition of an area being small and rural to becoming larger and more urban. Appendix Table B.7 includes the results of regressions that include regional time trends and regional controls for changes in preferences that may occur over the holiday season. They demonstrate that our results do not reflect changes in preferences in regions where entry occurs. Appendix Table B.8 shows that our results are robust to different methodologies for treating missing observations. Appendix B.9 shows that our results are robust to controlling for interactions between the time since a book was launched, entry and the book s price and national rank. They show that our results do not reflect systematic discounting of new books among locations that experience entry. 6 Note that there is substantial overlap in these different measures of popularity. For example, 95% of the books on the New York Times bestsellers list in November 2005 (and 82% of the books on USA Today s list) were in Amazon s national top 150.

6 ec6 Appendix Table B.1: Difference in Difference on store entry: additional splines Log Linear Spline 1 # Spline 2 # Spline 3 # USA Today # NY Times # (1) (2) (3) (4) (5) (6) (0.0014)** (0.0013)** (0.0010)** (0.0018)** (0.0022)** (0.0008)** (0.0008)** (0.0007)** Moderately (0.0006)** (0.0006)** Less (0.0006)** (0.0009)** Even Less (0.0007)** (0.0007)* Somewhat Unpopular Products (0.0006)** Relative Price (0.0022)** (0.0021)** (0.0022)** (0.0022)** (0.0021)** (0.0021)** Log(Sales Rank) (0.0002)** (0.0027)** (0.0026)** (0.0024)** (0.0039)** (0.0047)** (0.0020) (0.0020)* (0.0016)** Moderately (0.0016) (0.0016)* Less (0.0018)* (0.0026)** Even Less (0.0020)+ (0.0017) Somewhat Unpopular Products (0.0017) Relative Price (0.0063)** (0.0061)** (0.0061)** (0.0061)** (0.0060) (0.0059)** Log(Sales Rank) (0.0004)** Relative price (0.0007)** (0.0007)** (0.0008)** (0.0008)** (0.0007)** (0.0007)** Observations 4,051,254 4,051,254 4,051,254 4,051,254 4,062,326 4,052,722 Number of Fixed Effects 978, , , , , ,611 Controls -Dummy for missing price info -Average rating -Log(days since launch) -Time dummies -Log(Number of reviews) -Broadband competition -Log(Elapsed days since launch) -Dummy for missing elapsed date information -Discount Store Entry within 5.4 miles -Large Bookstore Entry within 5.4 miles -Product-location fixed effects (differenced out) -Book popularity spline (cols 2, 3, & 4 only) -Log(Sales Rank) (Column 1 only) # Spline in column (2) is top 100, , , and ,000; in column (3) is top 150, , , , ,000, and 15,001-50,000; in column (4) is top 250, , and ,000; in column (5) is top 150 books as according to the USA Today list of bestsellers; in column (6) is appearing in a New York Times bestsellers list. Robust standard errors are in parentheses and are clustered by location-time. Regressions include location-product fixed effects. + significant at 10%; * significant at 5%; ** significant at 1%

7 ec7 Appendix Table B.2: Absolute Distance Results Discount Store Entry & Distance to Store (double interactions) Large Bookstore Store Entry & Distance to Store (double interactions) Large Bookstore Entry (interactions) (1) 20 Miles (2) Log Linear # (0.0001)** (0.0001) Moderately ( ) Somewhat Less (0.0001)* Less ( ) Relative Price (0.0002)** (0.0002)** Log(Sales Rank) (0.0000)** Main Interaction: Entry * Distance to Store (0.0001)** (0.0002) (0.0002)** (0.0001) Moderately (0.0001)+ Somewhat Less (0.0002) Less (0.0001) Relative Price ( ) (0.0004) Log(Sales Rank) (0.0000)** Main Interaction: Entry * Distance to Store ( ) (0.0003) (0.0012)** (0.0008)** Moderately (0.0005)** Somewhat Less (0.0007)** Less (0.0005)** Relative Price (0.0017)** ( ) Log(Sales Rank) (0.0002)** (0.0018)** Moderately (0.0013)** (0.0009)**

8 ec8 Somewhat Less (0.0014)** Less (0.0011)** Relative Price (0.0033)** (0.0035)** Log(Sales Rank) (0.0002)** Relative price (0.0008)** (0.0008)** Observations 4,051,254 4,051,254 Number of Fixed Effects 978, ,611 Controls: as above Absolute distance is measured as the great circle distance using the latitude and longitude of the store and that of the location under observation. When then transform this to 20 absolute distance (or in column 2, 5.4 absolute distance) so that bigger numbers correspond to shorter entry distances. To compute radii, we use the average longitude and latitude across zip codes within the location. Robust standard errors are in parentheses and are clustered by location-time. Regressions include location-product fixed effects. + significant at 10%; * significant at 5%; ** significant at 1%. For sales rank regression, we use entry in a 5.4 miles radius.

9 9 Appendix Table B.3: Alternate Entry (with a One or Two Month Lag) Results One Month Lag Two Month Lag (1) (2) (3) (4) (5) (6) 5.4 Miles 20 Miles Log Linear 5.4 Miles 20 Miles Log Linear # (0.0012)** (0.0007)** (0.0012)** (0.0007)** ( ) (0.0004)** (0.0008)* (0.0005)** Moderately (0.0005)** (0.0003)** (0.0006)** (0.0003)** Somewhat Less (0.0008)** (0.0004)** (0.0011)** (0.0005)** Less (0.0007)* (0.0003)** (0.0009) (0.0004) Relative Price (0.0020)+ (0.0010)** (0.0020)** (0.0022) (0.0011)** (0.0023)* Log(Sales Rank) (0.0002)** (0.0002)** (0.0026)** (0.0011)** (0.0033)** (0.0014)** ( ) ( ) (0.0022) (0.0011) Moderately ( ) (0.0006)** (0.0021) (0.0009) Somewhat Less ( ) (0.0009)** (0.0033) (0.0014) Less ( ) (0.0008)+ (0.0027) (0.0011)+ Relative Price (0.0057) (0.0023) (0.0059) (0.0071) (0.0033)* (0.0072) Log(Sales Rank) (0.0004)** (0.0005)** Relative price (not interacted) (0.0007)** (0.0007)** (0.0007)** (0.0007)** (0.0007)** (0.0007)** Observations 4,051,254 4,051,254 4,051,254 4,051,254 4,051,254 4,051,254 Number of Fixed Effects 978, , , , , ,611 Controls: as above Robust standard errors are in parentheses and are clustered by location-time. Regressions include location-product fixed effects. + significant at 10%; * significant at 5%; ** significant at 1%. For log linear regression, we use entry in a 5.4 miles radius.

10 ec10 Appendix Table B.4: Alternate Entry (One Month Lead) Results (1) (2) (3) Row 5.4 Miles 20 Miles Log linear 1 # (0.0011)** (0.0007)** (0.0007)** (0.0004)** 3 Moderately (0.0005)** (0.0003)** 4 Somewhat Less (0.0007)** (0.0004)** 5 Less (0.0005)** (0.0002)** 6 Relative Price (0.0016)** (0.0009)** (0.0017)** 7 Log(Sales Rank) (0.0002)** (0.0021)** (0.0009)** (0.0015)** (0.0007)** 10 Moderately (0.0011)** (0.0004)** 11 Somewhat Less (0.0017)** (0.0006)** 12 Less (0.0011)* (0.0004)** 13 Relative Price (0.0041)** (0.0014)** (0.0043)** 14 Log(Sales Rank) (0.0003)** 29 Relative price (0.0008)** (0.0008)** (0.0007)** Observations 4,051,254 4,051,254 4,051,254 Number of Fixed Effects 978, , ,611 Controls: as above Robust standard errors are in parentheses and are clustered by location-time. Regressions include location-product fixed effects. + significant at 10%; * significant at 5%; ** significant at 1%. For log linear regression, we use entry in a 5.4 miles radius.

11 ec11 Appendix Table B.5: Larger Choice Set and Alternative Broadband Definition: linear interpolation Row (1) (2) (3) (4) Larger Alternative Broadband Definition Choice set 5.4 Miles 20 Miles Sales Rank 1 # (0.0011)** (0.0012)** (0.0008)** (0.0005)** (0.0008)** (0.0005)** 3 Moderately (0.0003)** (0.0006)** (0.0003)** 4 Somewhat Less (0.0003)+ (0.0009)** (0.0004)** 5 Less (0.0002)* (0.0007)** (0.0003)** 6 Relative Price (0.0012)** (0.0022)** (0.0010)** (0.0022)** 7 Log(Sales Rank) (0.0002)** (0.0023)** (0.0025)** (0.0011)** (0.0012) (0.0020) (0.0008)** 10 Moderately (0.0007)** (0.0016) (0.0006)** 11 Somewhat Less (0.0009) (0.0025)** (0.0009)** 12 Less (0.0005)+ (0.0018) (0.0006)** 13 Relative Price (0.0031)** (0.0061)** (0.0021)** (0.0063)** 14 Log(Sales Rank) (0.0004)** 15 Relative price (0.0004)** (0.0007)** (0.0008)** (0.0007)** Observations 9,420,562 4,051,254 4,051,254 4,051,254 Number of Fixed Effects 2,933, , , ,611 Controls: as above Robust standard errors are in parentheses and are clustered by location-time. Regressions include location-product fixed effects. + significant at 10%; * significant at 5%; ** significant at 1%. For sales rank regression, we use entry in a 5.4 miles radius.

12 ec12 Appendix Table B.6: Differences in population growth based on census place data Row (1) (2) (3) High Growth Low Growth Intermediate Growth 1 # (0.0028)** (0.0022)** (0.0018)** (0.0020)+ (0.0015)** (0.0012) 3 Moderately (0.0015)** (0.0009)** (0.0009)** 4 Somewhat Less (0.0025)** (0.0012)** (0.0013)** 5 Less (0.0016)** (0.0008) (0.0012)+ 6 Relative Price (0.0052)* (0.0029)** (0.0035)** (0.0053)** (0.0051)** (0.0034)** (0.0050)* (0.0043) (0.0024)* 9 Moderately (0.0044)* (0.0033) (0.0020)* 10 Somewhat Less (0.0061) (0.0055) (0.0031)** 11 Less (0.0036) (0.0043) (0.0024) 12 Relative Price (0.0192)* (0.0155)** (0.0067) 13 Relative price (0.0016)** (0.0016)** (0.0011)** Observations 926, ,896 1,935,702 Number of Fixed Effects 222, , ,509 Controls: as above High growth locations are defined as those in which population change in the Census Place for the location between the 1990 and 2000 decennial Censuses was above the 75 th percentile; low growth locations are defined as those for which the population change was below the 25 th percentile; intermediate growth areas as defined as those between the 25 th and 75 th percentiles. Robust standard errors are in parentheses and are clustered by location-time. Regressions include location-product fixed effects. + significant at 10%; * significant at 5%; ** significant at 1%.

13 ec13 Appendix Table B.7: Inclusion of State and MSA time trends and holiday controls Time Trends 5.4 Miles State Controls Holiday Dummies 5.4 Miles Time Trends + Holiday Dummies 5.4 Miles Time Trends + Holiday Dummies 20 Miles MSA and non-msa State Time Trends Holiday 5.4 Miles Controls 5.4 Miles (1) (2) (3) (4) (5) (6) # (0.0012)** (0.0012)** (0.0012)** (0.0008)** (0.0012)** (0.0012)** (0.0008)** (0.0008)** (0.0008)** (0.0005)** (0.0008)** (0.0008)** Moderately (0.0006)** (0.0006)** (0.0006)** (0.0003)** (0.0006)** (0.0006)** Somewhat Less (0.0009)** (0.0009)** (0.0009)** (0.0004)** (0.0009)** (0.0009)** Less (0.0007)* (0.0007)** (0.0007)* (0.0003)** (0.0007)** (0.0007)** Relative Price (0.0022)** (0.0022)** (0.0022)** (0.0021)** (0.0022)** (0.0022)** (0.0025)** (0.0025)** (0.0025)** (0.0011)** (0.0025)** (0.0025)** (0.0020) (0.0019) (0.0019) (0.0008)** (0.0020) (0.0019) Moderately (0.0016) (0.0016) (0.0016) (0.0006)** (0.0016) (0.0016) Somewhat Less (0.0025)** (0.0025)** (0.0025)** (0.0009)** (0.0025)** (0.0025)** Less (0.0018) (0.0018) (0.0018) (0.0006)** (0.0018) (0.0018) Relative Price (0.0063)** (0.0061)** (0.0061)** (0.0021)** (0.0061)** (0.0061)** Relative price (0.0061)** (0.0007)** (0.0007)** )** (0.0007)** (0.0007)** Observations 4,051,254 4,051,254 4,051,254 4,051,254 4,051,254 4,051,254 Number of Fixed Effects 978, , , , , ,611 Controls: As above plus: - State time trends (column 1) - State dummies interacted with holiday (November-December) dummies (column 2) - State time trends plus state-holiday (November-December) dummies (columns 3 and 4) - State MSA and non-msa time trends (column 5) - State MSA and non-msa holiday (November-December) dummies (column 6) In columns 5 and 6, controls are included for all MSAs within a state plus non-msa state areas (91 dummies interacted with time trends and holiday dummies). Robust standard errors are in parentheses and are clustered by location-time. Regressions include location-product fixed effects. # Base is unpopular books ranked and up. + significant at 10%; * significant at 5%; ** significant at 1%.

14 ec14 Row Appendix Table B.8: Different ways to treat missing prices (1) (2) (3) (4) Includes Imputing missing prices using price interaction prior/future month prices Drop missing prices Imputing prices using linear regression 1 # (0.0013)** (0.0012)** (0.0012)** (0.0012)** (0.0009)** (0.0008)** (0.0008)** (0.0008)** 3 Moderately (0.0007)** (0.0006)** (0.0006)** (0.0006)** 4 Somewhat Less (0.0011)** (0.0009)** (0.0009)** (0.0009)** 5 Less (0.0007)** (0.0007)** (0.0007)** (0.0007)** 6 Relative Price (0.0028)** (0.0027)** (0.0023)** (0.0022)** 7 Missing Price (0.0012) (0.0028)** (0.0026)** (0.0025)** (0.0025)** (0.0022)** (0.0021) (0.0020) (0.0020) 10 Moderately (0.0019)* (0.0018) (0.0016) (0.0016)** 11 Somewhat Less (0.0032)** (0.0026)* (0.0025)** (0.0025) 12 Less (0.0023)** (0.0019)** (0.0018) (0.0018) 13 Relative Price (0.0076)** (0.0073)** (0.0062)** (0.0061)** 14 Missing Price (0.0028)** 15 Relative price (not interacted) (0.0009)** (0.0008)** (0.0006)* (0.0005)** Observations 3,790,471 4,051,254 4,051,254 4,051,254 Number of Fixed Effects 956, , , ,611 Controls: as above in Appendix Table B1 Robust standard errors are in parentheses and are clustered by location-time. Regressions include location-product fixed effects. # Base is unpopular books ranked and up. + significant at 10%; * significant at 5%; ** significant at 1%. Entry is defined with a 5.4 mile radius.

15 . ec15 Appendix Table B.9: Interactions with log of days sin ce launch (1) (2) (3) (4) Row 5.4 Miles 20 Miles Sales Rank Dropping observations with missing elapsed date 5.4 Miles 1 # (0.0184)** (0.0086)** (0.0223)** (0.0182)** (0.0083)** (0.0225)** 3 Moderately (0.0330)** (0.0081)** (0.0208)** 4 Somewhat Less (0.0179)** (0.0081)** (0.0202)** 5 Less (0.0172)+ (0.0078) (0.0180)** 6 Relative Price (0.0131)** (0.0071)** (0.0135)** (0.0264)** 7 Sales Rank (0.0014)** (0.0341)** (0.0153)** (0.0412)** (0.0334)** (0.0150)** (0.0417)** 10 Moderately (0.0330)* (0.0148)** (0.0392)** 11 Somewhat Less (0.0331)+ (0.0147)** (0.0382)** 12 Less (0.0324) (0.0142)* (0.0347) 13 Relative Price (0.0275)** (0.0123)** (0.0288)** (0.0512)** 14 Sales Rank (0.0029)** 15 Relative price (not interacted) (0.0078)** (0.0083)** (0.0075)** (0.0064)** Observations 4,051,254 4,051,254 4,051,254 3,946,318 Number of Fixed Effects 978, , , ,190 Controls: as above in Appendix Table B1 plus log of days since launch interacted with entry and popularity Results of regressions including interactions with log(days since launch) with entry plus (1) popularity spline (or sales rank) and (2) relative price. Additional controls include interactions of log(days since launch ) with popularity spline (or sales rank) and relative price. Robust standard errors are in parentheses and are clustered by location-time. Regressions include location-product fixed effects. # Base is unpopular books ranked and up. + significant at 10%; * significant at 5%; ** significant at 1%.