Physician review sites are relatively new and were initially greeted with concern by some in the medical community. In particular, some physicians were critical of the lack of transparency in composite statistics [1] and were concerned that online reviews could harm their careers [2]—perhaps unfairly [3,4]. Although ratings are generally high [2,3,5,6], negative ratings undoubtedly influence patient behavior [7] and impact doctors [1,8]. Some doctors have attempted to gag patients by contractually prohibiting them from writing online reviews [9].

Most studies of online physician reviews have focused on portals such as HealthGrades [10], RateMDs [11], Vitals [12], and Yelp [13]. Studies tend to have a small sample size, analyzing approximately 5400 reviews [6]. Many studies aim to understand the factors that influence quantitative physician ratings. The qualitative analysis of 712 reviews by López et al [5] established thematic categories that tended to appear in reviews. Paul et al [14] replicated and expanded this work with a natural language processing (NLP) approach, which they applied to the text of 50,000 online reviews downloaded from RateMDs [11]. Their novel joint topic–sentiment modeling approach found that certain textual accounts of interpersonal skills such as rude, arrogant, and condescending are strongly associated with negative reviews, and drew attention to the role of patient experience of bureaucratic process in reviews, noting that these experiences were often reflected in reviews. Wallace et al [15] expanded on the work by López et al [5] and Paul et al [14] by analyzing 60,000 reviews to identify relationships between overall rating, health outcomes, and cost of care. To date, the only study to investigate the relationship between physician reviews and gender is that by Nwachukwu et al [16] on surgeon quality in sports medicine. They found that communication style influenced the valence of ratings for top- and bottom-tier surgeons and that female surgeons typically had higher ratings [16]. These and other studies of online reviews endeavor to understand how clinical experiences influence patient satisfaction and health outcomes. However, they tend to overlook or minimize questions about whether online review data reflect real experiences of medical care. Reviews may not be representative of the public or reflect demographic variation in health care utilization; indeed, doctor reviews are typically written by educated, younger, affluent, and healthier people [5]. However, a study comparing ratings of over 3000 physicians with licensing data showed a clear relationship between doctor quality and ratings [3].

Little research has studied the impact of gender or other demographic factors on the content and ratings of online physician reviews. Although qualitative studies of doctor-patient relationships have considered both negative and positive experiences [5], including the impact of demographics [17-19], the nature of these studies makes it difficult to estimate the size or scope of gender and other demographic variation in online physician reviews. Furthermore, gender differences, in particular biased interpretations of clinical experiences based on gender stereotypes, may impact online review content, which in turn may negatively impact both patients and physicians and perpetuate false gender stereotypes. The large-scale systematic study we present here documents gender differences in patient reviews with respect to both patients and doctors and proposes improvements for online review systems that could help reduce these disparities, thus improving information quality.

Although we know little about gender in the context of online reviews, gender has been studied extensively in the social sciences for over half a century. Much of this work investigates the role of gender in medical care and health systems more generally.

Gender is a cultural construct that affects people’s expectations and actions [20]. In social contexts and practices [21], gender is assessed independently of one’s identity [22]. Thus, any name appearing in online text is likely to be interpreted in terms of the man/woman binary, which is reflected in the use of gender in the current doctor review literature.

The expectations of one’s behavior differ depending on one’s assessed gender. Indeed, leadership traits praised in men are penalized in women, while traditional feminine behavior is seen as ineffective [23]. When writing references, men are described with more standout and ability-based words and fewer grindstone words (eg, hardworking, conscientious) [24], and women are described with more communal words [25]. Even when all factors are controlled, people rated teachers differently on hard skills (eg, promptness, fairness) and soft skills depending on the gender portrayed by the instructor [26]. Thus, bias may influence review content even when performance is identical.

In health care, gender differences influence doctors’ communication with patients [27-29], doctor and patient trust [30], and even diagnosis error rates [27]. Female doctors are seen as partners and more involved in the patient-doctor relationship, whereas female patients are treated with more condescension [27,29,31], have their concerns dismissed [29] and credibility doubted [27]. Conversely, patient satisfaction is dependent on more caring communication styles for women than for men [32]. However, many studies documenting these trends are small in scale or have weak evidence [33]. One exception is a study of over 10,000 people experiencing long-term illness in Sweden, where women reported being blamed, interrupted, disbelieved, doubted, and regarded as stupid [19]. These gender differences are likely to impact review scores, as lower patient satisfaction is correlated with high physician dominance, which can manifest itself in gendered actions (eg, poor information sharing and use of medical jargon) [34]. In this study, we investigate how these gender biases are represented in online reviews, which affect patients, physicians, and people using the reviews.

The goal of this study was to broaden and deepen our understanding of the impact of gender bias and other gender differences on online physician reviews. We leveraged reports of patient sentiments about their doctors through a large-scale analysis of online reviews. The Google Place review data analyzed here allowed us to identify patient and reviewer gender and characterize patient sentiment or experience in terms of both overall quality (a reviewer-entered Likert-type scale) and thematic content. Specifically, we formulated the following hypotheses (H):

Hypotheses H1a and H1b relate to physician gender and reflect the findings of prior work on gender inequalities in reviewing in other fields [24-26]. Hypotheses H2a, H2b, and H3 are based on prior work documenting gender differences in clinical encounters [27,28].

Our approach comprised both qualitative and quantitative perspectives on review content that mutually informed one another throughout the research process. We collected and analyzed a corpus of 154,305 reviews of doctors that constitutes a large nationally representative sample of physician reviews across all medical subfields and clinical contexts. Our focus was specifically on characterizing the differences in experience quality and content as they relate to reviewer and doctor gender.

This study contributes to the larger body of work on the impact of gender on clinical interactions and provides insight into what patients value in their doctors. Furthermore, we add to the small but growing body of literature that seeks to develop a general understanding of online reviews and the systems that collect and display them. Our results must be interpreted with caution due to the unstructured and often short nature of the patient narratives in reviews, the relative crudeness of the NLP techniques employed (as compared with human interpretation), and selection biases introduced by nonrepresentative variation in demographic characteristics of reviews and the types of experiences that motivate patients to write reviews. Such selection biases are almost unknown although we make a novel contribution here.

Online reviews provide an opportunity to learn, at scale, about patients’ perceptions of their doctors. Our findings have direct implications for the design of review sites and the presentation of search results. We argued that there are a number of useful ways in which gender differences can be reflected and potentially corrected in the presentation of information on the internet.

To examine how gender influences patient experience at scale, we sought a representative sample of reviews of US physicians. As gender and other demographic variables are rare in social media, we sought data that contained physician and reviewer names, which we used as a proxy for gender.

After exploring possible sources of physician reviews considering various application programming interface (API) features and use in prior work, we selected the Google Places API [36]. The API provides access to patient and physician names, which we leveraged to infer gender, as well as a broad range of areas and specialties. The Google My Business API has a 5-review limit for any particular doctor. Unfortunately, the API documentation does not provide information on how these reviews are selected. We can be fairly confident, however, that reviewer gender is not a factor. Thus, it is unlikely that this introduces bias into the sample with regard to the variables of interest. Furthermore, we took measures to ensure that our personal search histories did not influence review collection.

Reviews, physicians, and practices are likely to vary by location. Differences in locale, such as ruralness or urbanity, can influence health outcomes and care options, as can regional differences. For instance, in the United States, mortality rates of particular conditions have been shown to differ greatly from state to state [37]. Samples were taken across states from multiple regions of the United States using the Google Places API to control for the effect of locale. We steadily increased latitude and longitude intervals throughout each state with a 10,000-m radius to capture both urban and rural regions.

Reviews returned by the Google My Business API were either for a place (such as a practice with multiple doctors) or a physician. The API provided additional review-specific data for each review: a 5-point Likert-type rating assigned by the reviewer to the doctor or practice, doctor name, reviewer name, the location of the practice, and the text content of the review. The reviewer and physician gender were determined automatically using third-party software described in the following section.

The collection strategy described in this section yielded 154,305 reviews of physicians across the United States. These reviews spanned 2007 to 2017. Doctor ratings were highly polarized, exhibiting a U-shaped distribution (Figure 1). Of the reviews collected, 46,605 were rated 1 star or 2 stars (negative reviews) and 107,700 were rated 4 stars or 5 stars (positive reviews). Another 3,208 reviews were rated 3 stars and were omitted from our analyses. We did not screen for particular specialties.

The mean length of positive reviews was 50 words, while the mean length of negative reviews was 100 word (both follow heavy-tailed distributions). The distribution of inferred gender for both doctors and reviewers is shown in Figure 1. Note that we expected to see a fairly high rate of unknown gender in these data because some reviews are for medical practices that include multiple physicians. In addition, some physicians’ gender could not be identified. Given the presence of nonperson entities among physicians, it is perhaps surprising that reviewers have a greater rate of gender ambiguity. This likely reflects typos and pseudonyms among reviewers, who have a weaker incentive to use the correct name. However, the high rate of gender detection suggests that this should not be a great concern, as almost all reviewers use real names, not screen names. We were less certain about the rate of pseudonymous users but proceeded on the assumption that even when pseudonyms are used, they accurately reflect the reviewer’s gender.

A total of 3 sets of logistic regression models were fitted to these data. Reviews for which either reviewer or physician gender could not be identified were removed from the analysis in all of the models presented in the Results section. This reduced the data set by 28% from 154,305 reviews to 111,036.

The set of models we present investigated the association between negative reviews and reviewer/doctor gender. The second set comprised models examining the likelihood that a review mentions a soft skill. As our primary variables of interest are binary, and we are interested in interactions among those binary variables, the interpretation of the logistic regression variables is complicated. All interaction terms disrupt the interpretation of their component variables, but this interpretation is even more difficult when estimating all pairwise interactions of a set of variables (in this case, 3). Effectively, this decomposes each main effect coefficient into different components, which must be carefully interpreted and summed to construct odds ratios (ORs) for various conditions. We present one set of regressions on each review theme for each interaction rather than estimating all 3 interactions in a single model for readability. Through these sets, we tested physician gender, reviewer gender, and physician rating for pairwise interactions. The intercept and noninteracting effects are only marginally altered between these models, if at all. An additional set of models estimates the main effects of each dependent variable. When reporting these results, models are grouped by these model classes rather than by the dependent variable, as the relation between the independent variables takes priority over the particulars of the review themes. Finally, a pair of models estimates how administrative process correlates with gender, physician ratings, and soft skills.

Owing to the possibility that reviews could be influenced by clinical processes outside of physician control, we first fit a pair of models with and without mentions of bureaucratic process. The logit models summarized in Table 5 indicate that physician ratings are extensively influenced by gender, irrespective of mentions of process. Four models examine the correlation between sets of independent variables and the probability of a negative review. The A models were fit on all reviews for which we were able to infer both doctor and reviewer gender (N=111,036). In the data set on which we regressed the B models, we stringently controlled for mentions of process, which refer not to the physician but instead to the clinical aspects beyond the patient-physician relationship. The data set for the B models is the result of filtering sentences that mention terms associated with process from the reviews and then removing any reviews that were left without text. This process yielded a slightly smaller data set (N=106,325). Models 1A and 1B present a base model that includes only physician and reviewer gender and an interaction term. Models 2A and 2B control for soft skills, with Model 2A controlling for process. The overlapping coefficient estimates do not differ substantially between the 4 models, suggesting that mentions of process do not substantially alter the correlations captured by the variables of interest in these data. However, we conservatively controlled for them in the remainder of this section, except in models that consider process explicitly.

Although the dictionaries that contain terms related to review themes were developed in conjunction with our qualitative analysis and were thus thoroughly vetted, there remained concerns that these term lists do not adequately capture the themes they purportedly represent. As discussed in the Methods section, dictionary-based or bag of words (presence or absence of terms) approaches to natural language understanding often struggle to overcome or capture nuance in word use, notably suffering an inability to distinguish positive use from negation. We fit 2 models (2A and 2B) to verify that the soft skill dictionaries are correlated with negative reviews, as expected. The coefficients for the themes indicate that our dictionaries capture the basic tendency we anticipated: 3 of 4 positive soft skills, trust, investment, and amicability, are correlated with positive reviews (negative coefficients; P<.001), and 2 negative soft skills, indifference and disrespect, are correlated with negative reviews (P<.001). Notably, candor is not associated with either positive or negative reviews. This suggests the possibility that candor was mischaracterized by its dictionary. However, the intercoder agreement and classifier performance in Table 1 imply that the dictionary for candor captures the theme equally well as the other positive soft skills. Rather, it seems that candor was misclassified as a positive soft skill and, as defined by its dictionary, is perhaps better understood as a neutral soft skill, appearing equally in positive and negative reviews. The models estimate large absolute effects for the three other positive soft skills. A negative soft skill, indifference, fits a considerably greater effect than any positive soft skill, and disrespect fits an even greater effect. This is consistent with the descriptive statistics in Tables 2 and 3, which suggest that although positive soft skills are more strongly associated with positive reviews, they also appear often in negative reviews. Negative soft skills by contrast occur overwhelmingly in negative reviews. Generally, these findings further suggest that our dictionaries accurately represent the themes they attempt to capture.

In our qualitative analysis, we observed that reviews mentioning positive soft skills were primarily associated with high scores for male and female doctors. Reviewers wrote positively about physicians who were candid and direct, “Ladies this doctor listens and responds with respect, she does not talk down to you either” (Female; 5-star; U940); who were trustworthy and supportive, “The delivery would’ve been terrifying without him” (Male; 4-star; O7390); invested, “She asks questions and listens. She makes me feel like I am important” (Female; 5-star; U933); and amicable, for example, “She is an amazing doctor. Kind, caring, empathetic, warm, knowledgeable, quick thinking, funny, and honest” (Female; 5-star; O679). As stated in the discussion of the logit models in Table 5, positive soft skills are more likely to appear in positive reviews (P<.001). These models also estimate large effects for positive soft skills except candor, being at least 2.4 times as likely to appear in positive reviews than negative ones.

We also coded for 2 negative soft skills, indifference and disrespect. Indifference was relatively rare, appearing in only 2% of the reviews. Disrespect was more common, occurring in 8% of all reviews. Unlike positive soft skills, which appeared in both positive and negative reviews, negative soft skills were almost exclusively found in negative reviews. When they did appear in positive reviews, it almost always referred to bureaucratic process, not the physician. Typical reviewer comments coded for negative soft skills relate experiences with doctors who lack courtesy, patience, and warmth toward their patients, for example, “I could not believe how condescending and snippy she was!” (Female; 1-star; O8376); “He was very rude, condescending, arrogant, and appeared angry” (Male; 1-star; U1122). Reviewers also described feeling ignored, “I felt passed around and ignored” (Female; 1-star; O10061), or that their concerns were dismissed, “...brushed it off” (Female; 1-star; O10100) and “I was in tears because he was too stubborn to listen” (Male; 1-star; U1047). Other complaints included ignoring patients’ understanding of their own medical condition and lack of inclusion in decision-making. As reviewers mentioned, “The doctor does not listen to you and forces his opinion down your throat without considering your view. Do not visit here” (Male; 1-star; U398) and “He did not want to listen to anything I had to say and he definitely didn’t want me getting a second opinion. He got defensive and standoffish at the mention of any other opinion which says SHADY all over it” (Male; 1-star; U198).

The logit models for negative soft skills indicate that negative soft skills are far more likely to appear in negative reviews than positive ones. The log-likelihood coefficients are considerably stronger for negative than for positive soft skills. This is because positive soft skills can be negated to note the absence of a positive quality, whereas negative soft skills are rarely negated to indicate a positive quality.

Although we made no predictions about administrative process, it is worth noting several patterns that emerged from the logistic models on mentions of process. We fit 2 models of process reported in Table 6 in the reviews for which gender could be inferred and which did not filter out mentions of process (N=111,036). Model 1 parallels the models of themes in Table 4 and accordingly reports interaction effects as separate models. Model 2 estimates correlations between each theme and process, controlling for rating and gender.

Process is much more likely to be mentioned in negative reviews (P<.001). This is consistent across all models. This may be because when process is smooth, it is more likely to go unnoticed, whereas poor experiences with process are more likely to color the overall experience. Parallel to the trend observed in soft skills, process more often occurs in reviews written by and about women (P<.001). The Doctor_F × Rating_Neg model estimates that negative reviews of female physicians are 2.61 (95% CI 2.41-2.88) times as likely to mention process than negative reviews of male physicians (log OR 0.96, 95% CI 0.88-1.06; P<.001). By contrast, female and male reviewers mention process in negative reviews at equal rates. Women who see female doctors are less likely to mention process, which produces an equalizing effect that offsets the greater rate of reports of process for both female reviewers and physicians.

We also examined the association of process with soft skills. Model 2 demonstrates a positive correlation between mentions of process and soft skills. The correlations with amicability and disrespect are sizable (log OR 1.01, 95% CI 0.98-1.04 and 0.70, 95% CI 0.65-0.76, respectively; P<.001), indicating that patients value the ease of interpersonal interactions with staff, that is, whether they are friendly or rude, and likely interpret bureaucratic competence through the framework of how nice the staff are. In our qualitative analysis, we found that positive reviews expressing disrespect overwhelmingly do so with regard to process, indicating that reviewers are able to separate relationships with their doctor and the overall clinical experience as reflected in the final review. This compartmentalization is well illustrated by a patient who reported a positive experience with a doctor but faced problems with poor administration such that the patient ultimately severed their relationship with the clinic:

Our results provide compelling evidence for a number of effects of gender on patient experience, as reported in physician reviews. These findings may be interpreted through 2 distinct frames. First, a patient experience frame attempts to interpret gender dynamics in the context of the patient-physician relationship. This frame should be familiar to readers versed in the literature on gender and clinical experience. The second frame, the online review system frame, seeks not to understand or improve the clinical aspects of health care but rather considers how gender differences may subvert or be leveraged to improve reviews as a valuable public resource that informs decisions about care-seeking.

The following discussion of the results in context of the hypotheses of this study assumes the patient experience perspective. The Summary and Recommendations section, however, largely reflects the online review system perspective, which is less concerned with controlled statistical inference than it is how the descriptive statistical patterns in Tables 2 and 3 might affect public perception of physician quality and how we might design online review systems to improve physician-patient matching and offset bias.

We hypothesized that female physicians would be more likely to receive lower ratings. This hypothesis was supported by our study. These data indicate that there is considerable reviewer bias against female physicians. This is consistent with well-documented patterns of bias against women in other fields, notably when reviewing instructor performance in a controlled online classroom [26].

We hypothesized that female patients would report overall worse clinical experiences than men. The results here support the opposite hypothesis. We found that men are slightly more likely to report negative experiences than women.

There are several valid interpretations of these findings. Men may receive worse care than women, as captured by patient experiences (rather than health outcomes). Alternatively, it could be that men have higher expectations for care than women or are less competent at navigating the clinical setting. Both are plausible given that men less frequently utilize health services [40,41]. Finally, we might attribute the discrepancy to women’s greater propensity for agreeableness [42,43], and for forgiveness [44] and compromise [45] in interpersonal conflict. These interpretations are not mutually exclusive, and further research is warranted to account for this trend.

We hypothesized that reviews of female physicians would be more likely to critique their soft skills. Our results indicate that this is true of all soft skills. All soft skills were more often mentioned in the reviews of female physicians. In negative reviews of female physicians, reviewers were considerably more likely to mention candor and amicability.

The results supporting this hypothesis indicate that female physicians’ soft skills are more likely to be critiqued and that female physicians are much more likely to be associated with disrespect. We also present evidence that women may be penalized for lacking candor and amicability to a much greater degree than men. This may be attributed to failure to live up to a positive stereotype, as women are generally expected to be more open and personable, and female physicians in particular are expected to be more caring [46]. Furthermore, the physician role is one of authority, and research has extensively documented that women are punished for leadership styles that men are rewarded for [47,48].

We hypothesized that female reviewers are more likely to reflect on their doctor’s soft skills. This hypothesis was supported by our study.

In positive reviews, women mentioned all positive soft skills with greater probability. However, the magnitude of these effects was diminished in negative reviews. These patterns suggest that women may be more inclined to value a physician’s soft skills. However, it also indicates that men may be more sensitive to a lack of positive soft skills than their presence.

We found that women were more likely to mention the negative soft skills, indifference and disrespect. This likely reflects the wealth of literature documenting the tendency for physicians to take women’s concerns less seriously and treat them with condescension [27,29,31].

We expected that female patients would report more negative experiences with male doctors. Our results support this hypothesis. Although there was no significant interaction effect of reviewer and physician gender on the probability of a negative experience, we found a significant interaction between reviewer and physician gender on the likelihood of reporting disrespect, which overwhelmingly occurs in negative reviews. Female reviewers apparently benefit from seeing female doctors, as they are less likely to mention disrespect when reviewing female physicians than when they review men. This compensating effect neutralized the overall greater association of disrespect with female reviewers, but not female doctors.

Given the literature and our previous finding that female physicians are subject to biased reviews, these results suggest that even women harbor bias against female physicians. However, this bias is considerably smaller among women than it is among men and is complicated by physician-gender/patient-gender preferences for different communication styles [32].

We made no predictions regarding the relationship between process and soft skills or reviewer/physician gender. However, we found that both amicability and disrespect were highly correlated with process, suggesting that the ease of social interaction with staff is important to reviewers. Importantly, negative reviews of female physicians are considerably more likely to mention aspects of the clinical experience beyond experiences with the doctor.

The strong association between process and negative reviews of female doctors may reflect a tendency for patients to assess male doctors “on their own merits,” whereas women are more likely to be held accountable for poor process. This reflects a bias against women interpretation. A realist account might hypothesize that female doctors are more likely to work in clinics with less competent or accommodating staff.

This work is complementary to previous qualitative studies on the influence of gender on the doctor-patient relationship. We acknowledge that it is unclear whether gender differences reflect patient perception or the reality of physician behavior. For example, given the ample evidence in other contexts on bias against women, the high rate of negative reviews for female physicians likely reflects reviewer bias against physicians rather than genuine differences in treatment. However, our approach does not allow certainty in this regard. The Discussion section provides a more detailed explanation of the interpretation of gender differences in the context of the findings of previous studies. We also recognize that the data do not contain information related to patients’ health outcomes. Although the health outcome of each patient is not represented in our data, other studies have shown that reviews can reflect real health consequences [14]. Similarly, treatment noncompliance, unwarranted recalcitrance, and other patient characteristics beyond the reviewer’s narrative are not captured in the reviews.

In this paper and similar research, gender representation is reported as binary, which does not capture the full spectrum of gender or gendered interaction. Even though gender is likely to be interpreted in a binary fashion by most review readers [21], the doctor-patient relationship is more complex, and other data could offer more nuanced and richer perspectives. Furthermore, our work does not consider the intersection of gender and other identities, such as race [49,50].

As noted in the Methods section, the Google Places API limits data collection to 5 reviews per physician or practice. Google provides no documentation on how these reviews are chosen from all the reviews written. We acknowledge that the small sample may not be fully representative of a doctor or practice; however, our contribution is more focused on the biases within the reviews and not on the doctors themselves. We also note that the Google Places reviews are subject to selection biases. Demographics undoubtedly play a role in determining who writes physician reviews (eg, consider the high proportion of female reviewers in our data set). However, the data likely suffer selection biases, irrespective of demographic differences. For instance, it is probable that the U-shaped distribution of physician ratings is both a product of overall polarized attitudes and strong experiences providing greater motivation to write a review. It seems likely that other such selection biases were present in these data but were unknown to us as we performed our analysis.

The increasing prevalence of online reviews of physicians affects both medical practices and patient choices. However, little is known about biases that may be present in these reviews or whether they reflect the real biases documented in doctor-patient interactions [27,28]. Conversely, most studies of gender bias in doctor-patient interactions to date have been limited to qualitative analyses, smaller-scale data sets, or specific medical conditions.

This study is the first to provide evidence that gender biases and other gender differences are observable at scale in physician reviews. We provide extensive evidence of differences in physician ratings and review content with respect to both physician and reviewer gender. Our statistical inference indicates that these differences are robust when controlling for possible confounding relationships and therefore are likely to reflect gender differences and biases in the patient-physician relationship.

It is difficult to disentangle which aspects of these gender differences may be attributed to review selection bias, gender bias, or gender behavioral variation. However, these patterns undeniably affect prospective patients as they peruse online reviews to select a doctor. Thus, it is important to consider how we might educate the public about the effects of gender bias on physician ratings and how online review systems could be improved to control for bias. We propose several concrete steps that could be taken to better support patients.

Review systems can draw attention to gender differences in reviews to aid prospective patients in building their own understanding of a physician’s potential gender biases. One way to do this is to organize reviews by gender. This could either be the default presentation or a special gender-separated view. Alternatively, reviewer gender might be indicated explicitly only in automatically generated summaries of physician reviews. Additionally, prospective patients might benefit from a panel that provides a sense of how a particular physician compares with other physicians on gender. For example, if there is a discrepancy between men’s and women’s ratings for a physician that differs greatly from the gender discrepancy of other similar physicians, a prospective patient might benefit from reading their reviews with this information in mind.

An online review system could also help to correct for gender differences that generally affect reviews. As female physicians receive many more negative reviews on average, a prospective patient might find it easier to select among physicians if ratings are adjusted to control for the physician’s gender or if ratings are reported relative to physicians of the same gender. On the other hand, online review systems could implement measures to reduce bias in the reviews. For instance, when writing a review and using a word that is commonly used to critique female physicians, the system could prompt the reviewer about gender stereotyping in word choice or alternative terms that are gender neutral. This might encourage a more balanced approach to review writing and help reviewers recognize their own biases. Alternatively, information can be solicited from the reviewer in such a way that greatly reduces the effects of gender stereotypes on performance evaluation [51].

These approaches would ideally lead to reviews that more accurately reflect the quality of care provided by physicians. Finally, this study draws attention to several important areas for future work. We advocate researchers adopt mixed methods approaches similar to the one presented here when pursuing quantitative analyses of text. Furthermore, this study raises questions specifically related to online review systems as objects of study in their own right. Little is known about how readers interpret online reviews, notably in the context of health care and gender. It also highlights the need to study how review systems can be designed to improve review accuracy and inform review readers and writers on gender bias in online reviews. We propose that experimental studies in review cognition and system design will be most fruitful to these ends.