Online Appendix for The Interplay between Online Reviews and Physician Demand: An Empirical Investigation

Transcription

1 Online Appendix for The Interplay between Online Reviews and Physician Demand: An Empirical Investigation Appendix A: Screen Shots of Original Data A typical interaction of a patient with our focal platform starts along the lines of Figure 1, and normally when a patient enters the system, he starts the search with specialty, location, and insurance. After entering this information, a list of physicians will show up based on patients requirements; see, for example, Figure 2 1. In our work, we first collected a list of physicians that cover all the specialties and locations, and then we collected information for each physician on the list individually. For instance, see Figures 3 and 4 for typical background information from an individual page. From a typical physician page, we collect three types of information: (i) background, (ii) appointment book, and (iii) patient reviews. First, for the background information, we have education, language spoken, insurance, specialties, and the average rating calculated by the website, see Figure 4. Here, only the rating information is dynamic and would change every day, so our crawler only fetched the rating information on a daily basis, and all the other information we only collected once. Next, for the appointment book information, we have the physicians corresponding office address, and the detailed appointment book slots, which include the date and the time of each slot, see Figure 5. Based on the dynamic appointment book information, we can infer the supply and demand of each physician. To make the demand information as accurate as possible, we collected the appointment book data on an hourly basis, because some slots may disappear just because their time has passed. The supply here is measured by the total number of slots offered in a 30-day window starting from the current day and is updated every day at 6am EST. The demand is then measured by the total number of slots booked every day which is computed from the difference between the offered slots of two subsequent days. Note here we acknowledge a potential limitation of our data set is that some bookings that disappear might due to the doctor s change of schedule, however as long as this is systematically uncorrelated with doctor and patient attributes our results should still hold. Note that in our data set, we also measure the average duration of each slot. To compute this average duration, we first compute the difference between all the subsequent slots, and then we take the average after deleting the extreme values. Finally, we also collect the detailed patient reviews, including both text and numeric rating information, see Figure 6. Note that the rating we see in Figure 3 is the average of the individual overall ratings. 1 Note that when we collected our data set between November 27, 2014 and April 12, 2015, patients could only see a list of physicians with their overall ratings. However, the platform has since changed the system, and patients can now see both overall ratings and the physician s calendar. 1

2 2 Figure 1 Online appointment booking system. Figure 2 physician list.

3 3 Figure 3 Individual physician page. Figure 4 physician background information.

4 4 Figure 5 physician appointment book.

5 5 Figure 6 physician reviews. Appendix B: Physician Profile Data Out of the 872 physicians, we have around 55.8% male, around 23% have more than one hospital affiliation, and 57.7% speak more than one language. Figure 7 shows the frequency distribution of physicians by specialty, state, and total number of languages spoken. We can see from Figure 7 that primary care physicians and dentists have the largest population by far in our data. Figure 8 shows New York State has the largest population in our data set; this finding is consistent with this company initially only covering physicians in NYC. From Figure 9, we can see that indeed more than half of the physicians speak more than one language. Of the variety of languages spoken by the physicians, English and Spanish are the most popular ones, and Russian, French, and Chinese are also notable.

6 6 Figure 7 Frequency distribution by specialty. Figure 8 Frequency distribution by state.

7 7 Figure 9 Frequency distribution by number of languages spoken. Appendix C: Review Examples We here show some examples of each service feature. 1. Bedside Manner: He explained everything in a really caring manner! 2. Diagnosis: She didn t provide a good diagnosis. 3. Waiting Time: Waiting time is too long. 4. Service Time: She took her time with me. 5. Insurance Process: Dr. Lane was prompt in taking care of my insurance coverage. 6. Knowledge: Very knowledgeable man and staff. 7. Office Environment: Beautiful office. Open Saturdays. Appendix D: SMOG Index McLaughlin (1969) developed the formula for calculating the SMOG grade, which can be computed with the following steps that we quoted from his paper: 1. Count the number of sentences (at least 30). 2. In those sentences, count the polysyllables (words of 3 or more syllables). 3. Calculate using grade = number of polysyllables 30 numer of sentences

8 8 Appendix E: Causal Impact with Regression Discontinuity As stated at the beginning of this paper, we consider ratings to be the primary driver of the online competition among physicians. Therefore, we focus on both the qualitative (causal) impact and quantitative (elasticity) impact of rating with respect to demand. In this section, we use an RD design as a robustness check of the causal impact of numeric rating on demand. The online website displays physicians ratings along two different dimensions as we previously mentioned: (i) the rounded average rating, and (ii) the individual ratings for three different categories: overall, bedside manner, and waiting time. The first type of rating is the average over all the individual ratings, and the website rounds it to the nearest half star. This change in rating is exogenous to physician quality but is correlated with the displayed ratings, and thus provides us an ideal exogenous shock for an RD analysis. The rounding threshold here is defined as the number that below and above which gives different rounded stars. For example, 3.75 is a rounding threshold in the online platform, in that if the average rating for a physician falls to between 3.5 and 3.74, the physician will have a rounded rating of 3.5; however, if the average rating falls to between 3.75 and 4, the physician will have a rounded rating of 4. We start our analysis with observations for a distance of less than 0.1 stars from both sides of the rounding threshold. Therefore, the rating regimes we are considering are (3.70, 3.74) and (3.75, 3.80). We define our binary variable T as T = { 0 if rating is below the rounding threshold 3.75, 1 if rating is above the rounding threshold Therefore, T = 0 if the rating belongs to (3.70, 3.74), and T = 1 if the rating falls into (3.75, 3.80). We consider this T as our exogenous treatment. Denote by D it the total demand of physician i on day t. We first estimate the following model with threshold T it : D it = α + θt it + β 1 r it + β 2 X it + ɛ it, where r it is physician i s average rating on day t over all the reviewers, and X it are the physician specific variables. We found consistent results with Luca and Vats (2014), i.e., a significant positive treatment effect of the overall rating, see Table 1. We then check the results for alternative bandwidth sizes 2, including 0.2 and 0.3, and show similar patterns, see Table 2 and Table 3. Finally, to eliminate the possible bias in RD results (see Hartmann et al. (2011)), that is if the physicians game the system, following Luca and Vats (2014), we conduct McCrary (2008) test. Consistent with Luca and Vats (2014), we did not find any evidence on the physicians gaming behavior. 2 The bandwidth here refers to the smoothing parameter for the regression discontinuity, i.e., the distance to the threshold. For example, if the bandwidth size is 0.1, then we are considering the range of (3.70, 3.80) for the rounding threshold 3.75.

9 9 Table 1 Bandwidth=0.1, Rounding Threshold=3.75 (Signif. codes: 0 *** ** 0.01 * ) Estimate Std. Err t-value Pr(> t ) T OVERALL SPECIALTY CITY GENDER EXT OFF LAN NUM AVE GAP DEATH RATE LIFE EXP PHY ACT REVIEW CNT RATING SQ ID SMOG WORD CNT CHAR CNT SYLLABLE CNT R Squared N Table 2 Bandwidth=0.2, Rounding Threshold=3.75 (Signif. codes: 0 *** ** 0.01 * ) Estimate Std. Err t-value Pr(> t ) T OVERALL SPECIALTY CITY GENDER EXT OFF LAN NUM AVE GAP DEATH RATE LIFE EXP PHY ACT REVIEW CNT RATING SQ ID SMOG WORD CNT CHAR CNT SYLLABLE CNT R Squared N 17440

10 10 Table 3 Bandwidth=0.3, Rounding Threshold=3.75 (Signif. codes: 0 *** ** 0.01 * ) Estimate Std. Err t-value Pr(> t ) T OVERALL SPECIALTY CITY GENDER EXT OFF LAN NUM AVE GAP DEATH RATE LIFE EXP PHY ACT REVIEW CNT RATING SQ ID SMOG WORD CNT CHAR CNT SYLLABLE CNT R Squared N Appendix F: Robustness Checks and Subsets Comparison In this section, we conduct subsets analysis and robustness checks for our main estimation results. We start our discussion with subset analysis. Note in what follows, we check whether our main results hold in a sense that all the signs are the same for the statistically significant variables. We first conduct subset analysis of physicians with different rating categories, namely 3.5, 4, 4.5, and 5 star. We show our results in Table 4. In general, we find our main results still hold. However, in addition to our main estimation results, we find a rating saturation effect, which means the text information is more important for physicians close to the full rating (i.e., 4.5 or 5 star). For physicians with a 3.5 star rating, text information plays a very limited role. Intuitively, for lower-rated physicians, it is likely that patients may not care too much about their detailed reviews. However, for higher-rated physicians, the rating itself does not provide too much additional information to the patients. Therefore, when choosing highly rated physicians, patients tend to read the details of reviews to find the part of the information or some specific points of service quality, such as the waiting time and service time, to base their decisions on. Next, we conduct subset analysis for physicians in the top and bottom 30% of review amount. From Table 5, we first find our main results still hold. Moreover, for the top 30% reviewed physicians, the text review information is more important. We then conduct a third subset analysis for the top 50% reviewed physicians with part of their reviews, that is, reviews contained in the first page (on average in our data set is 23 reviews), the first 10 reviews, the first 5 reviews, and the first 3 reviews. Note here that the first page of reviews are the reviews one can see on a physician s personal page without clicking the more button

11 11 under the review section. From Table 6, we first find our main results still hold with these four sets of data. Moreover, we find when the number of reviews increases, the detailed text review information becomes more important. Our fourth subset analysis considers patient choice behavior with income differences. For this analysis, we need the data from an area that contains different income populations in adjacent locations. We thus select the data of NYC s five boroughs, namely, Manhattan, Brooklyn, Queens, the Bronx, and Staten Island. Based on the statistics from the Bureau of Labor Statistics in 2015, 3 we know the average weekly wage for Manhattan was $2,847, while the average weekly wage for Bronx, Brooklyn, Queens, and Staten Island was $901, $818, $936, and $825, respectively. Therefore, we further separate NYC data into two parts: Manhattan and other boroughs. Although geographically adjacent, these two areas would have significantly different income levels. We show our results in Table 7. We first find the factors that drive the choices of patients in NYC are overall rating, total number of reviews, review identity disclosure, review readability, review complexity, bedside manner, diagnosis accuracy, waiting time, service time, ease of insurance process, and physician knowledge. Among these factors, we find ease of insurance and physician knowledge only affect patients in Manhattan and not other areas in NYC, whereas all the other factors are statistically significant for both subsets separately. This finding shows patients in Manhattan seem to make their choices based on more dimensions of online information. Moreover, the review information has a larger effect on patients in Manhattan. Based on this difference, physicians may want to adjust their operational strategies in different NYC areas to increase patient demand. Now we turn to robustness checks. First, we replace the number of text reviews with the number of total reviews, compute the new percentage of identity disclosure, and then conduct the robustness check as shown in Table 8. We find results consistent with our main estimation, and the positive effect of the number of reviews as well as the percentage of identity disclosure is even more significant. Second, we conduct a robustness check without the location flexibility variable, and show in Table 9 that all of our main results hold. Third, we conduct a robustness check without the identity variable, and show our results in Table 10. We find our main estimation results still hold here. Fourth, we conduct a robustness check without the review-quality variables, namely, SMOG, WORD CNT, CHAR CNT, and SYLLABLE CNT. We show our results in Table 11, and find our main estimation results hold. Our fifth robustness check considers the inclusion of lagged demand. The lagged demand helps solve the problem of autocorrelation between the residuals and estimated coefficients. We show the results in Table 12, and we can see our main estimation results hold. In our sixth robustness check, we remove the text review related variables and keep the bedside manner and waiting time rating variables in our model and conduct the robustness check. From Table 13, we can see our main results hold. In our seventh robustness check, we delete the lagged ratings in our set of IVs, and conduct the robustness check. We show the estimation results in Table 14. From this table, we can see our main estimation results hold. In our eighth robustness check, we add Hausman-type IVs on all seven 3 newyorkcity.htm

12 12 sentiment variables as well. We show our estimation results in Table 15. From this table, we can see our main results hold. In our last robustness check, we consider groups of similar specialty physicians in one market, and we find consistent results as shown in the Table 16. To summarize, both our subset analysis and the nine robustness checks show our main results hold. Table 4 Robustness check with different rating categories. Variable 3.5 Star 4 Star 4.5 Star 5 Star GENDER (0.1683) (0.1899) (0.1467) (0.1557) EXT OFF (0.3388) (0.2964) (0.3009) (0.3126) LAN NUM (0.1764) (0.1997) (0.1963) (0.2789) AVE GAP (0.1257) (0.1267) (0.1088) (0.1569) DEATH RATE (0.1364) (0.1543) (0.1126) (0.1084) LIFE EXP (0.1297) (0.1297) (0.1267) (0.1169) PHY ACT (0.0887) (0.0748) (0.1068) (0.1002) REVIEW CNT (0.0098) (0.0067) (0.0088) (0.0089) ID (0.0012) (0.0064) (0.0079) (0.0051) SMOG (0.0653) (0.0094) (0.0067) (0.0053) WORD CNT (0.0438) (0.0489) (0.0072) (0.0014) CHAR CNT (0.0225) (0.0286) (0.0304) (0.0222) SYLLABLE CNT (0.0215) (0.0897) (0.0267) (0.0679) BED MANNER (0.4487) (0.0144) (0.0256) (0.0225) DIAGNOSIS (0.3009) (0.0542) (0.0376) (0.0057) WAITING TIME (0.2555) (0.0099) (0.0094) (0.0059) SERVICE TIME (0.2113) (0.0103) (0.0122) (0.0069) EASE INSUR (0.1997) (0.3326) (0.0073) (0.0094) KNOWLEDGE (0.1958) (0.3021) (0.0855) (0.0822) OFFICE ENVIR (0.1497) (0.1552) (0.1689) (0.0312) BED MANNER (0.0888) (0.0456) (0.0385) (0.0091) ACC DIAG (0.0912) (0.1064) (0.0371) (0.0474) WAITING TIME (0.0884) (0.0501) (0.0262) (0.0336) GENDER (0.0477) (0.0412) (0.1234) (0.1174) BED MANNER (0.0306) (0.0561) (0.0449) (0.0429) ACC DIAG (0.0303) (0.1042) (0.0485) (0.0464) WAITING TIME (0.0499) (0.0284) (0.0253) (0.0049) SERVICE TIME (0.0508) (0.0238) (0.0556) (0.0162) GMM Objective 1.745e e e e-6 N

13 13 Table 5 Robustness Check Based on Reviews. Variable Top 30% Bottom 30% OVERALL (0.0678) (0.0069) GENDER (0.1197) (0.5879) EXT OFF (0.6681) (0.6675) LAN NUM (0.5897) (0.6897) AVE GAP (0.0778) (0.1579) DEATH RATE (0.3254) (0.1987) LIFE EXP (0.4161) (0.1658) PHY ACT (0.0987) (0.1974) REVIEW CNT (0.0093) (0.0068) RATING SQ (0.0254) (0.0015) ID (0.0178) (0.0054) SMOG (0.0987) (0.0098) WORD CNT (0.0715) (0.0014) CHAR CNT (0.0157) (0.1021) SYLLABLE CNT (0.0125) (0.1011) BED MANNER (0.0069) (0.0897) DIAGNOSIS (0.0087) (0.0446) WAITING TIME (0.0098) (0.0395) SERVICE TIME (0.0086) (0.7963) EASE INSUR (0.0501) (0.4990) KNOWLEDGE (0.0323) (0.6554) OFFICE ENVIR (0.0799) (0.0867) OVERALL (0.0097) (0.0057) BED MANNER (0.1075) (0.0438) ACC DIAG (0.0727) (0.2356) WAITING TIME (0.0345) (0.1469) OVERALL (0.0067) (0.0042) GENDER (0.0456) (0.0336) BED MANNER (0.0515) (0.2298) ACC DIAG (0.0485) (0.0500) WAITING TIME (0.0411) (0.0338) SERVICE TIME (0.0352) (0.0337) GMM Objective 5.647e e-5 N

14 14 Table 6 Top 50% reviewed physicians with part of their reviews. Variable First page First 10 First 5 First 3 OVERALL (0.0096) (0.0059) (0.0088) (0.0044) GENDER (0.1200) (0.1887) (0.0826) (0.1112) EXT OFF (0.5547) (0.5546) (0.7714) (0.3446) LAN NUM (0.3124) (0.4012) (0.3218) (0.4129) AVE GAP (0.0557) (0.1138) (0.1006) (0.1008) DEATH RATE (0.1558) (0.2008) (0.2016) (0.2054) LIFE EXP (0.3779) (0.3067) (0.3331) (0.3007) PHY ACT (0.0774) (0.1099) (0.1989) (0.1446) REVIEW CNT (0.0088) (0.0007) (0.0056) (0.0077) RATING SQ (0.0061) (0.0098) (0.0081) (0.0097) ID (0.0040) (0.0048) (0.0063) (0.0051) SMOG (0.0822) (0.0153) (0.0072) (0.0046) WORD CNT (0.0621) (0.0482) (0.0358) (0.0076) CHAR CNT (0.0336) (0.1067) (0.0772) (0.0899) SYLLABLE CNT (0.0332) (0.1167) (0.0226) (0.1067) BED MANNER (0.0055) (0.0056) (0.0067) (0.1404) DIAGNOSIS (0.0094) (0.0038) (0.0576) (0.0985) WAITING TIME (0.0091) (0.0078) (0.0450) (0.1559) SERVICE TIME (0.0082) (0.0053) (0.0462) (0.1269) EASE INSUR (0.0563) (0.0509) (0.0885) (0.1118) KNOWLEDGE (0.0304) (0.0604) (0.0774) (0.1001) OFFICE ENVIR (0.0293) (0.0452) (0.0996) (0.0964) OVERALL (0.0087) (0.0078) (0.0051) (0.0106) BED MANNER (0.0697) (0.0483) (0.0428) (0.0321) ACC DIAG (0.1465) (0.1964) (0.0811) (0.0446) WAITING TIME (0.0333) (0.0438) (0.0732) (0.0333) OVERALL (0.0055) (0.0051) (0.0077) (0.0113) GENDER (0.0430) (0.0422) (0.1444) (0.0384) BED MANNER (0.0499) (0.0485) (0.0511) (0.0776) ACC DIAG (0.0448) (0.0460) (0.1367) (0.0723) WAITING TIME (0.0392) (0.0407) (0.0439) (0.0469) SERVICE TIME (0.0277) (0.0251) (0.0351) (0.0228) GMM Objective 3.154e e e e-5 N

15 15 Table 7 Subset analysis in NYC. Variable NYC Manhattan Others OVERALL (0.0323) (0.0256) (0.1485) GENDER (0.2248) (0.1146) (0.1249) EXT OFF (0.3564) (0.3326) (0.3069) LAN NUM (0.1179) (0.1674) (0.1084) AVE GAP (0.1238) (0.1347) (0.0886) DEATH RATE (0.1143) (0.1894) (0.1789) LIFE EXP (0.3027) (0.1884) (0.7746) PHY ACT (0.3394) (0.1215) (0.4456) REVIEW CNT (0.0034) (0.0057) (0.0044) RATING SQ (0.0037) (0.0038) (0.0079) ID (0.0044) (0.0047) (0.0451) SMOG (0.0799) (0.0097) (0.0455) WORD CNT (0.0293) (0.0097) (0.0116) CHAR CNT (0.1287) (0.1444) (0.1213) SYLLABLE CNT (0.0881) (0.0931) (0.0441) BED MANNER (0.0043) (0.0091) (0.0084) ACC DIAG (0.0079) (0.0033) (0.0046) WAITING TIME (0.0066) (0.0071) (0.0058) SERVICE TIME (0.0081) (0.0059) (0.0287) EASE INSUR (0.0721) (0.0668) (0.0585) KNOWLEDGE (0.0361) (0.0403) (0.0137) OFFICE ENVIR (0.0449) (0.0557) (0.0105) OVERALL (0.0056) (0.0057) (0.0098) BED MANNER (0.0101) (0.0082) (0.0033) ACC DIAG (0.0451) (0.0999) (0.0418) WAITING TIME (0.0241) (0.0306) (0.0336) OVERALL (0.0077) (0.0082) (0.0030) GENDER (0.0975) (0.0777) (0.1003) BED MANNER (0.0440) (0.0481) (0.0886) ACC DIAG (0.0884) (0.1029) (0.0772) WAITING TIME (0.0265) (0.0079) (0.0272) SERVICE TIME (0.0333) (0.0374) (0.0505) GMM Objective 2.497e e e-5 N

16 16 Table 8 Robustness Check with Total Number of Reviews. Variable I II OVERALL (0.0067) (0.0059) GENDER (0.0272) (0.0267) EXT OFF (0.0089) (0.0066) LAN NUM (0.0599) (0.0612) AVE GAP (0.0536) (0.0881) DEATH RATE (0.0499) (0.0592) LIFE EXP (0.0913) (0.0936) PHY ACT (0.1111) (0.1184) REVIEW CNT (0.0064) (0.0051) RATING SQ (0.0145) (0.0161) ID (0.0046) (0.0051) SMOG (0.0163) (0.0336) WORD CNT (0.0033) (0.0077) CHAR CNT (0.0076) (0.0055) SYLLABLE CNT (0.0082) (0.0093) BED MANNER (0.0099) (0.0074) ACC DIAG (0.0072) (0.0045) WAITING TIME (0.0815) (0.8026) SERVICE TIME (0.1589) (0.1268) EASE INSUR (0.6635) (0.5498) KNOWLEDGE (0.3264) (0.3367) OFFICE ENVIR (0.1109) (0.1137) OVERALL (0.0057) (0.0061) BED MANNER (0.0352) (0.0357) ACC DIAG (0.0401) (0.0405) WAITING TIME (0.0350) (0.0448) OVERALL (0.0049) (0.0053) GENDER (0.0316) (0.0355) BED MANNER (0.0389) (0.0384) ACC DIAG (0.0499) (0.0489) WAITING TIME (0.0433) (0.0425) SERVICE TIME (0.0355) (0.0346) GMM Objective e e-6 N I - Our focal platform data only II - Our focal platform together with another platform data

17 17 Table 9 Robustness Check with No Location Flexibility. Variable I II OVERALL (0.0044) (0.0043) GENDER (0.0993) (0.0913) LAN NUM (0.0666) (0.0587) AVE GAP (0.0928) (0.0904) DEATH RATE (0.1667) (0.1548) LIFE EXP (0.1256) (0.1057) PHY ACT (0.1996) (0.1987) REVIEW CNT (0.0064) (0.0064) RATING SQ (0.0087) (0.0077) ID (0.0088) (0.0074) SMOG (0.0101) (0.0135) WORD CNT (0.0121) (0.0085) CHAR CNT (0.0335) (0.0123) SYLLABLE CNT (0.0117) (0.0139) BED MANNER (0.0084) (0.0097) ACC DIAG (0.0046) (0.0044) WAITING TIME (0.0099) (0.0937) SERVICE TIME (0.0475) (0.0334) EASE INSUR (0.2257) (0.2106) KNOWLEDGE (0.1579) (0.1357) OFFICE ENVIR (0.1657) (0.1367) OVERALL (0.0052) (0.0057) BED MANNER (0.0334) (0.0458) ACC DIAG (0.1338) (0.1179) WAITING TIME (0.0304) (0.0444) OVERALL (0.0055) (0.0064) GENDER (0.0432) (0.0413) BED MANNER (0.0470) (0.0469) ACC DIAG (0.0410) (0.0408) WAITING TIME (0.0339) (0.0351) SERVICE TIME (0.0447) (0.0471) GMM Objective 1.525e e-6 N I - Our focal platform data only II - Our focal platform together with another platform data

18 18 Table 10 Robustness check without ID disclosure. Variable I II OVERALL (0.0112) (0.0103) GENDER (0.1267) (0.1199) EXT OFF (0.0966) (0.0897) LAN NUM (0.1026) (0.1122) AVE GAP (0.0882) (0.0901) DEATH RATE (0.1056) (0.1021) LIFE EXP (0.1128) (0.1257) PHY ACT (0.1025) (0.0999) REVIEW CNT (0.0044) (0.0054) RATING SQ (0.0079) (0.0078) SMOG (0.0065) (0.0059) WORD CNT (0.0195) (0.0185) CHAR CNT (0.0111) (0.0128) SYLLABLE CNT (0.0099) (0.0097) BED MANNER (0.0079) (0.0093) ACC DIAG (0.0092) (0.0078) WAITING TIME (0.0846) (0.0851) SERVICE TIME (0.0632) (0.0628) EASE INSUR (0.2579) (0.2577) KNOWLEDGE (0.1279) (0.1326) OFFICE ENVIR (0.1054) (0.1046) OVERALL (0.0086) (0.0079) BED MANNER (0.0340) (0.0321) ACC DIAG (0.0805) (0.0799) WAITING TIME (0.0305) (0.0152) OVERALL (0.0348) (0.0366) GENDER (0.0180) (0.0172) BED MANNER (0.0374) (0.0379) ACC DIAG (0.0392) (0.0395) WAITING TIME (0.0777) (0.0754) SERVICE TIME (0.0276) (0.0279) GMM Objective 5.666e e-5 N I - Our focal platform data only II - Our focal platform together with another platform data

19 19 Table 11 Robustness check without review quality variables. Variable I II OVERALL (0.0126) (0.0133) GENDER (0.1168) (0.1157) EXT OFF (0.1055) (0.1034) LAN NUM (0.1364) (0.1458) AVE GAP (0.1848) (0.1755) DEATH RATE (0.1164) (0.1178) LIFE EXP (0.1339) (0.1441) PHY ACT (0.3364) (0.3511) REVIEW CNT (0.0055) (0.0049) RATING SQ (0.0078) (0.0081) BED MANNER (0.0102) (0.0113) ACC DIAG (0.0114) (0.0109) WAITING TIME (0.0098) (0.0096) SERVICE TIME (0.0549) (0.0552) EASE INSUR (0.0529) (0.0522) KNOWLEDGE (0.1964) (0.1888) OFFICE ENVIR (0.1258) (0.1256) OVERALL (0.0128) (0.0124) BED MANNER (0.0526) (0.0509) ACC DIAG (0.1355) (0.1336) WAITING TIME (0.1287) (0.1297) OVERALL (0.0108) (0.0099) GENDER (0.0157) (0.0152) BED MANNER (0.0448) (0.0442) ACC DIAG (0.0402) (0.0399) WAITING TIME (0.0351) (0.0353) SERVICE TIME (0.0360) (0.0357) GMM Objective 4.879e e-5 N I - Our focal platform data only II - Our focal platform together with another platform data

20 20 Table 12 Robustness check with lagged demand control. Variable I II OVERALL (0.0079) (0.0074) GENDER (0.2246) (0.2012) EXT OFF (0.0548) (0.0521) LAN NUM (0.1199) (0.1346) AVE GAP (0.0662) (0.0658) DEATH RATE (0.1064) (0.1076) LIFE EXP (0.1349) (0.1371) PHY ACT (0.1124) (0.1134) REVIEW CNT (0.0046) (0.0092) RATING SQ (0.0042) (0.0041) ID (0.0345) (0.0349) SMOG (0.0101) (0.0109) WORD CNT (0.0148) (0.0136) CHAR CNT (0.0076) (0.0083) SYLLABLE CNT (0.0091) (0.0088) BED MANNER (0.0063) (0.0071) ACC DIAG (0.0081) (0.0077) WAITING TIME (0.0970) (0.0961) SERVICE TIME (0.0358) (0.0361) EASE INSUR (0.1258) (0.1476) KNOWLEDGE (0.0967) (0.0921) OFFICE ENVIR (0.0854) (0.0810) LAG DEMAND (0.0312) (0.0309) OVERALL (0.0352) (0.0359) BED MANNER (0.0345) (0.0339) ACC DIAG (0.0775) (0.0799) WAITING TIME (0.0523) (0.0513) OVERALL (0.0087) (0.0094) GENDER (0.0215) (0.0209) BED MANNER (0.0411) (0.0415) ACC DIAG (0.0481) (0.0487) WAITING TIME (0.0395) (0.0407) SERVICE TIME (0.0333) (0.0364) GMM Objective 8.884e e-7 N I - Our focal platform data only II - Our focal platform together with another platform data

21 21 Table 13 Robustness check with bedside manner and waiting time ratings. Variable I II OVERALL (0.0128) (0.0115) BEDSIDE (0.0135) (0.0138) WAITING (0.0133) (0.0127) GENDER (0.0774) (0.0578) EXT OFF (0.1257) (0.1193) LAN NUM (0.1199) (0.1287) AVE GAP (0.0571) (0.0565) DEATH RATE (0.1137) (0.1233) LIFE EXP (0.1233) (0.1436) PHY ACT (0.0322) (0.0305) REVIEW CNT (0.0103) (0.0115) RATING SQ (0.0078) (0.0084) ID (0.0101) (0.0099) SMOG (0.0348) (0.0341) WORD CNT (0.0096) (0.0097) CHAR CNT (0.0977) (0.0985) SYLLABLE CNT (0.1046) (0.1011) OVERALL (0.0047) (0.0054) BEDSIDE (0.0079) (0.0088) WAITING (0.0112) (0.0104) OVERALL (0.0022) (0.0024) BEDSIDE (0.0013) (0.0011) WAITING (0.0031) (0.0027) GENDER (0.0167) (0.0166) GMM Objective 5.469e e-5 N I - Our focal platform data only II - Our focal platform together with another platform data

22 22 Table 14 Robustness check without lagged rating IV. Variable I II OVERALL (0.0078) (0.0088) GENDER (0.1025) (0.1006) EXT OFF (0.1155) (0.1278) LAN NUM (0.1056) (0.1055) AVE GAP (0.1001) (0.0997) DEATH RATE (0.1034) (0.1251) LIFE EXP (0.1188) (0.1165) PHY ACT (0.1002) (0.1033) REVIEW CNT (0.0101) (0.0113) RATING SQ (0.0082) (0.0073) ID (0.0074) (0.0065) SMOG (0.0160) (0.0151) WORD CNT (0.0293) (0.0196) CHAR CNT (0.0154) (0.0125) SYLLABLE CNT (0.0184) (0.0177) BED MANNER (0.0111) (0.0123) ACC DIAG (0.0101) (0.0107) WAITING TIME (0.0859) (0.0833) SERVICE TIME (0.0362) (0.0383) EASE INSUR (0.4479) (0.3598) KNOWLEDGE (0.2257) (0.2008) OFFICE ENVIR (0.1085) (0.1157) OVERALL (0.0113) (0.0111) BED MANNER (0.0429) (0.0573) ACC DIAG (0.0458) (0.1025) WAITING TIME (0.0445) (0.0429) OVERALL (0.0097) (0.0098) GENDER (0.1005) (0.1087) BED MANNER (0.0525) (0.0529) ACC DIAG (0.0499) (0.0491) WAITING TIME (0.0421) (0.0422) SERVICE TIME (0.0284) (0.0291) GMM Objective 7.548e e-6 N I - Our focal platform data only II - Our focal platform together with another platform data

23 23 Table 15 Robustness check with more IVs. Variable I II OVERALL (0.0046) (0.0049) GENDER (0.1155) (0.1179) EXT OFF (0.1064) (0.1058) LAN NUM (0.0925) (0.0937) AVE GAP (0.0651) (0.0663) DEATH RATE (0.1117) (0.1136) LIFE EXP (0.1389) (0.1277) PHY ACT (0.1158) (0.1034) REVIEW CNT (0.0063) (0.0061) RATING SQ (0.0073) (0.0077) ID (0.0044) (0.0038) SMOG (0.0226) (0.0231) WORD CNT (0.0172) (0.0169) CHAR CNT (0.0177) (0.0146) SYLLABLE CNT (0.0355) (0.0351) BED MANNER (0.0091) (0.0088) ACC DIAG (0.0074) (0.0069) WAITING TIME (0.0066) (0.0069) SERVICE TIME (0.0381) (0.0327) EASE INSUR (0.1079) (0.1033) KNOWLEDGE (0.0879) (0.0812) OFFICE ENVIR (0.0649) (0.0655) OVERALL (0.0043) (0.0037) BED MANNER (0.0454) (0.0446) ACC DIAG (0.0351) (0.0342) WAITING TIME (0.0155) (0.0149) OVERALL (0.0033) (0.0041) GENDER (0.0114) (0.0124) BED MANNER (0.0397) (0.0405) ACC DIAG (0.0350) (0.0346) WAITING TIME (0.0786) (0.0599) SERVICE TIME (0.0398) (0.0386) GMM Objective 1.687e e-6 N I - Our focal platform data only II - Our focal platform together with another platform data

24 24 Table 16 Robustness check with groups of similar specialty physicians. Variable I II OVERALL (0.0115) (0.0109) GENDER (0.0846) (0.0859) EXT OFF (0.0806) (0.0810) LAN NUM (0.0876) (0.0846) AVE GAP (0.0885) (0.0832) DEATH RATE (0.1134) (0.1129) LIFE EXP (0.1346) (0.1332) PHY ACT (0.1008) (0.1012) REVIEW CNT (0.0066) (0.0063) RATING SQ (0.0072) (0.0069) ID (0.0074) (0.0079) SMOG (0.0321) (0.0323) WORD CNT (0.0385) (0.0390) CHAR CNT (0.0158) (0.0162) SYLLABLE CNT (0.0202) (0.0210) BED MANNER (0.0071) (0.0075) ACC DIAG (0.0084) (0.0087) WAITING TIME (0.0062) (0.0063) SERVICE TIME (0.0771) (0.0769) EASE INSUR (0.2006) (0.2205) KNOWLEDGE (0.1177) (0.1164) OFFICE ENVIR (0.1012) (0.1011) OVERALL (0.0034) (0.0041) BED MANNER (0.0071) (0.0073) ACC DIAG (0.0898) (0.0908) WAITING TIME (0.0297) (0.0291) OVERALL (0.0042) (0.0048) GENDER (0.0258) (0.0241) BED MANNER (0.1222) (0.1235) ACC DIAG (0.0503) (0.0514) WAITING TIME (0.0419) (0.0421) SERVICE TIME (0.0356) (0.0351) GMM Objective 1.746e e-6 N I - Our focal platform data only II - Our focal platform together with another platform data

25 25 Appendix G: Other Tables Table 17 Results from Nested Logit Model Variable OVERALL (0.0610) GENDER (0.0102) EXT OFF (0.1134) LAN NUM (0.2458) AVE GAP (0.2357) DEATH RATE (0.0611) LIFE EXP (0.2469) PHY ACT (0.4254) REVIEW CNT (0.0466) RATING SQ (0.0442) ID (0.0613) SMOG (0.0428) WORD CNT (0.0842) CHAR CNT (0.0261) SYLLABLE CNT (0.0093) N Table 18 Correlation Matrix of Reveiw Variables OVER.N BED.N WAIT.N BED. ACC DI. WAIT SERV INSUR KNOW. OFFICE OVER.N 1 BED.N WAIT.N BED ACC DI WAIT SERV INSUR KNOW OFFICE References Hartmann, W., H. S. Nair, S. Narayanan Identifying causal marketing mix effects using a regression discontinuity design. Marketing Science 30(6) Luca, M., S. Vats Digitizing doctor demand: The impact of online reviews on doctor choice. Health & Healthcare in America: From Economics to Policy. Ashecon. McCrary, J Manipulation of the running variable in the regression discontinuity design: A density test. Journal of Econometrics 142(2) McLaughlin, G. H Smog grading-a new readability formula. Journal of reading 12(8)