Institute for Economic Stanford Policy Research (SIEPR) A TASTE OF THEIR OWN MEDICINE: GUIDELINE ADHERENCE AND ACCESS TO EXPERTISE* Amy Finkelstein Maria Polyakova MIT & NBER Stanford University & NBER Petra Persson Jesse Shapiro Stanford University, Brown University Research Institute of Industrial & NBER Economics, & NBER September, 2021 Working Paper No. 21-048 SIEPR | John A. and Cynthia Fry Gunn Building | siepr.stanford.edu | @siepr A Taste of Their Own Medicine: Guideline Adherence and Access to Expertise* Amy Finkelsteint Petra Persson? Maria Polyakova' Jesse Shapiro! September 30, 2021 Abstract We use population administrative data from Sweden to study adherence to 63 medication- related guidelines. We compare the adherence of patients without personal access to medical expertise to the adherence of those with access, namely doctors and their close relatives. We estimate that, among observably similar patients, access to expertise is associated with 3.8 percentage points lower adherence, relative to a baseline adherence rate of 54.4 percent among those without access. This association is larger for recommendations with a weaker clinical motivation. Our findings suggest an important role in non-adherence for factors other than those, such as ignorance, complexity, or failures of patient-provider communication, that would be expected to diminish with access to expertise. Keywords: information, medical decision-making, experts JEL Codes: 112, 118, D83 *E-mail: afinkQ@mit.edu, perssonp@stanford.edu, maria.polyakovaQstanford.edu, jesse_shapiro_1@brown.edu. We acknowledge funding from the Population Studies and Training Center and the Eastman Professorship at Brown University (Shapiro), and the National Institute on Aging under grant KO1AG05984301 (Polyakova). We appreciate comments from Jon Gruber, Eric Patashnik, and seminar participants at the Research Institute of Industrial Eco- nomics and the Wharton School. We are grateful to Aron Malatinszky, Gabriel Swagel, and Yuci Zhou, as well as to Sarah Bogl and Katja Hofmann at the Research Institute of Industrial Economics, for excellent research assistance. 'MIT and NBER +Stanford University and Research Institute of Industrial Economics and NBER. §Stanford University and NBER Brown University and NBER 1 Introduction Widespread non-acherence to medical guidelines is believed to contribute to a large amount of hos- pitalizations, deaths, and health care spending each year (Fonarow et al. 2011; Neiman 2017).! The causes of non-adherence are the subject of significant academic and policy interest (Neiman 2017; Hyun 2017; Lopez-Vazquez et al. 2012). Prominent explanations include patient or provider igno- rance, guideline complexity, and lack of trust or communication in the patient-provider relationship (Alpert 2010; Aslani et al. 2019; Bosworth 2012; Neiman 2017). Such explanations motivate specific policy interventions, such as attempts to simplify treatment regimes or disseminate information to patients or practitioners (Brown and Bussell 2011; Fischer et al. 2016; Irwin et al. 2014; Krueger et al. 2005; McDonald et al. 2002; Nieuwlaat et al. 2014; van Driel et al. 2016). They also suggest the testable implication that patients with greater access to medical expertise will tend to be more adherent, since they are likely to be better informed, better able to make decisions, and better able to communicate with their medical providers. We find that the opposite holds. Specifically, we study the relationship between a patient's ad- herence to medication guidelines and whether the patient has personal access to medical expertise, defined as being a doctor or having one in the close family. To do so, we assemble administrative data on the entire population of Sweden from 2005 through 2016 and use it to measure adherence to 63 government-issued prescription drug guidelines. These include 6 guidelines related to an- tibiotics (e.g., children should start with a narrower- rather than a broader-spectrum antibiotic to combat a respiratory tract infection), 20 guidelines specific to the elderly (e.g., avoid certain sleep medications), 20 guidelines related to specific diagnoses (e.g., take statins after a heart attack), and 17 guidelines on medication use during pregnancy (e.g., avoid certain antidepressants). Adherence to these guidelines requires the compliance of both the provider (to prescribe or not prescribe some medication) and the patient (to take or not take some medication). We use information on a person's completed education to determine if they are a doctor, and we link doctors to their relatives using a population register. In our baseline analysis, we classify 'For example, only a minority of patients adhere to the recommendation to take high-intensity statins following a heart attack (Neiman 2017; Colantonio et al. 2017), and many physicians depart from expert guidance on antibiotic prescribing (Lopez- Vazquez et al. 2012; Fleming-Dutra et al. 2016). a patient as having access to medical expertise if the patient, the patient's partner, or any of the patient's parents or children is a doctor. For each of the 63 guidelines, we narrow in on the set of patients who, based on their health conditions, prior prescription claiming, and demographics, are covered by the guideline; within this risk set, we examine differences in adherence between those with and without access to expertise. We find that access to expertise is generally associated with less adherence to guidelines. Among the 63 guidelines that we study, and controlling for demographics, income, and education, the association between access to expertise and adherence is negative in 41 cases and statistically significant in 20 of those. Since the share of the population covered by any given guideline can vary by up to three orders of magnitude, we summarize these findings by averaging them across guidelines, weighting each guideline by the prevalence of its risk set in the population. We estimate that while the average patient without access to expertise adheres to guidelines 54.4 percent of the time, one with access to expertise adheres only 50.6 percent, a 3.8 percentage point lower adherence rate. The 95 percent confidence interval includes a 4.1 to 3.5 percentage point lower adherence rate for those with access to expertise. The education and income controls in our baseline specification strengthen the negative association between adherence and access to expertise. Relative to our baseline, the association also becomes more negative if we narrow the definition of access to include only being a doctor oneself. It becomes less negative-but remains negative and statistically significant-if we broaden the definition of access to expertise to include having a doctor in the extended rather than just close family, or to include having nurses and pharmacists in the close family. We consider several explanations for the negative association between guideline adherence and access to expertise. We first consider a potential role for unobserved socioeconomic or health differences between those with and without access to expertise, but find little evidence to suggest these differences explain the association. We next consider the possibility that access to expertise is associated with greater comfort with pharmaceutical solutions to medical problems, or, relatedly, greater access to pharmaceuticals. We find some evidence consistent with this hypothesis: access to expertise is associated with greater use of prescription medications. However, we find that the negative association between adherence and access to expertise is similar for guidelines to take a particular medication and guidelines to avoid one, which is not consistent with an explanation based solely on relative comfort with, or access to, pharmaceuticals. The last possibility we consider is that access to expertise gives patients information or confi- dence that prompts them to disregard guidelines that they do not perceive to be in their clinical interest. Several pieces of evidence suggest that this mechanism is at play. One is that the adherence gap between those with and without access to expertise is greater for guidelines with weaker clinical support, although it remains negative and statistically significant even for those with stronger sup- port. Among pregnant women, access to expertise is less negatively associated with adherence for guidelines recommending against drugs classified in category D (which are contraindicated in preg- nancy), than guidelines recommending against drugs in category C (which are to be used only when clearly needed). Likewise, the association between adherence and access to expertise is marginally more negative for guidelines rated by the evidence-based clinical resource UpToDate as based on weaker evidence compared to those rated as having stronger evidence, although here the difference is not statistically distinguishable. Another piece of evidence is that, while the association between access to expertise and adherence is negative for each of the major categories of guidelines we consider, it is most negative for guidelines regarding the appropriate use of antibiotics, which are designed to promote public health rather than the narrow interest of the patient. Our findings contribute to a large literature comparing the medical decisions of practitioners and their families to those of the general population. Comparisons have included preventive health behavior (Glanz et al. 1982), treatment decisions and outcomes (Bunker and Brown 1974; Chen et al. 2020; Ubel et al. 2011), end-of-life care (Gramelspacher et al. 1997; Weissman et al. 2016; Wunsch et al. 2019), and use of Cesarean sections (Chou et al. 2006; Grytten et al. 2011; Johnson and Rehavi 2016). A related literature compares the health outcomes of doctors and their families to those of the general population (Chen et al. 2019; Artmann et al. forthcoming; Leuven et al. 2013). Most closely related to our paper, Frakes et al. (2021) compare the propensity of military physicians and that of other U.S. military personnel to use several types of low-value and high- value medical care (as defined by the health policy community). They find that physicians are only slightly more likely to avoid low-value care or engage in high-value care than non-physicians, and conclude that policies aimed at improving patients' information and medical knowledge would therefore do little to affect adherence." Our findings reinforce this conclusion in a different setting (and focusing on different types of guidelines) by showing that doctors and their families tend to be less adherent to guidelines, even those backed by strong evidence. More broadly, our paper contributes to a literature comparing expert and non-expert behavior in contexts such as consumer purchases (Bronnenberg et al. 2015), real estate (Levitt and Syverson 2008; Rutherford et al. 2005), household finance (Bodnaruk and Simonov 2015), and health insur- ance (Handel and Kolstad 2015). Some of this literature treats the behavior of informed individuals as a normative benchmark of optimal behavior. It is unclear whether this perspective is appropriate in our context. Medication guidelines represent broad "rules of thumb" that may not apply in all circumstances. It is possible that the care of more informed patients is guided by clinically relevant knowledge that is not used in the care of less informed patients, and that the more informed may benefit from their greater departures from guidelines. However, consistent with recent evidence that practitioners' departures from prescribing guidelines lead to worse patient outcomes (Abaluck et al. 2020; Cuddy and Currie 2020; Currie and MacLeod 2020), it is also possible that more in- formed patients are overconfident or otherwise mistaken in deviating from guidelines.* We return to this issue in the conclusion. 2 Data 2.1 Population and Characteristics The backbone of our data is an extract from the Total Population Register consisting of all in- dividuals residing in Sweden from 2005 through 2016 (Swedish Research Council n.d.). For each ?Relatedly, Abaluck et al. (2020) use data from clinical notes to argue that most departures from clinical guidelines do not stem from practitioners' lack of awareness of these guidelines. 3 Alpert (2010), for example, argues that evidence-based "guideline-directed therapy for a particular condition has been shown to lead to better clinical outcomes compared with 'eminence-based,' personally derived, therapeutic strategies." individual, we observe information about their biological parents, which allows us to link any given individual to their grandparents, parents, siblings, children, and cousins, from which we also infer links to aunts, uncles, nieces, and nephews. We are also able to link individuals to their spouses using marital records and to cohabiting partners using information about addresses and shared biological children. Hereafter, we refer to a person's spouse or cohabiting partner as the person's partner. We merge these data to Statistics Sweden's longitudinal database of individuals (LISA) from 1990 through 2016, which contains information drawn from various administrative records (Statis- tics Sweden n.d.). From the education records we obtain information on each individual's highest completed degree in each calendar year. We define an individual as a doctor if their highest degree is a medical degree, taking the most recent degree in cases of multiple higher-level degrees. We define other specialized occupations analogously. We define an individual as having access to medical expertise in a given year if the individual, the individual's partner, or any of the individual's parents or children is a doctor. We sometimes use a broader definition of access which also includes having a doctor in the extended family, i.e., among the individual's siblings, cousins, aunts, uncles, nieces, nephews, grandchildren, or grandparents. We define the presence of other occupations (e.g., nurses, pharmacists) in the family analogously. We also use LISA to define control variables. From the location records, we construct a cate- gorical variable indicating the individual's municipality (kommun) of residence as of the preceding year, using the mother's municipality for those age 17 and under. From the education records, we construct a categorical variable indicating the individual's highest level of completed schooling (i.e., no college, some college, completed college) as of the preceding year, using the mother's schooling for those age 26 and under. Finally, from the tax records, we define a measure of pre-transfer in- come for each individual and year, using the average of parents' nonmissing income for individuals age 26 and under and using income as of age 60 for those over 60. We compute the percentile rank of each individual's income in the preceding year, among those with strictly positive income of the same gender and birth cohort. 2.2. Medical Records Sweden has universal health insurance. Patients pay at most a small co-pay for medical treatments or prescription drugs. We link the data from Statistics Sweden to health records from the National Board of Health and Welfare (Socialstyrelsen 2019). For each individual, we observe the universe of prescription drug purchases made in outpatient pharmacies from July 2005 through 2017. For each purchase, we observe the name of the drug and the drug's seven-digit Anatomical Therapeutic Chemical (ATC) classification code. We also observe the universe of inpatient hospital visits, outpatient visits (excluding those for primary care and pre-natal care), and births from 2005 through 2016. For each visit, we observe the date of the visit and the diagnosis codes (ICD-10) attached to the visit. For each birth, we infer the date of conception (by subtracting 280 days from the due date), and the date the pregnancy ended (by adding the gestational age at birth to the date of conception). To form control variables for sensitivity analysis, we follow Chen et al. (2019) and define separate indicators for whether an individual had a heart attack, heart failure, lung cancer, Type II diabetes, or asthma diagnosis in any preceding year. 2.3 Prescription Drug Guidelines Multiple government agencies promulgate medical guidelines in Sweden. We focus on the subset of guidelines that are for prescription drugs because we are best able to measure adherence for these guidelines. We consider two types of guidelines. The first type are guidelines issued by the Swedish Board of Health and Welfare (Social- styrelsen), a government organization that issues national guidelines for treatment of various dis- eases. These guidelines are written by panels of physicians appointed by the Board. The government tracks adherence to the guidelines as a mechanism for improving quality of care, but does not insist that all practitioners should follow the guidelines in all circumstances. In October of 2019, we identified 93 active guidelines pertaining to prescription drugs that had been issued by the the Socialstyrelsen.* We analyze the 46 of these for which we can measure adherence in our data.® The second type are guidelines covering the use of prescription drugs in pregnancy. The potential for a drug to harm fetal development is reflected in a letter grade classification (A, B, C, or D) (Danielsson and Dencker 2019). These classifications are in turn based on text selected by the drug manufacturer from a standardized set of options provided by the EU.® Sweden's pharmaceutical database (FASS) classifies a drug as D-class if the selected text says that the drug is "contraindicated during pregnancy" and as C-class if the text says the drug "has harmful pharmacological effects on pregnancy and/or the fetus/newborn" and "should be used during pregnancy only when clearly needed."' We obtained information on which drugs had C and D classifications in Sweden in March 2019. We define one guideline for each of the 10 categories of C-class drugs (e.g., C-class opioids) and one for each of the 5 categories of D-class drugs (e.g., D-class tetracyclines, a type of antibiotic) most frequently purchased by women in our data during the six months before conception. We also define one guideline for all other C-class drugs, and one for all other D-class drugs, yielding a total of 17 pregnancy-related guidelines. We group guidelines into one of four mutually exclusive and exhaustive categories: 6 guidelines covering the use of antibiotics, 20 covering medication use specifically by the elderly (defined as those age 75 and older), 20 covering medication use following specific diagnoses, and 17 covering medication use in pregnancy. Appendix Tables 1-4 provide additional details on the guidelines in each of these four groups, as well as their classification along two additional dimensions. Specifically, we classify guidelines according to whether they recommend against taking a par- ticular drug or class of drugs ("don't take" guidelines) or in favor of taking a particular drug or class of drugs ("do take" guidelines). We do not include antibiotics guidelines in this classification, as they advocate taking one drug over another. "We obtained this information from www.vardenisiffror.se, Indikatorer (indicators), Lakemedelsbehandling (drug treatment), and from Socialstyrelsen (2010). ®We exclude those that rely on special drug registries that are not available to us (38 guidelines), do not have a clear direction or target value (5 guidelines), or track dosage rather than type of medication (4 guidelines). ®See https: //emcsupport .medicines.org.uk/support/solutions/articles/7000007888-what-is-an-smpc-. 7B-class drugs include those whose text indicates that prescribing during pregnancy should be done "only when clearly needed." A-class drugs include those whose text indicates that prescribing during pregnancy should be done "with caution" or that the drug "can be used during pregnancy." See https://www.fass.se/LIF/menydokument? userType=O&'menyrubrikId=124. Translations via GoogleTranslate. We also classify a subset of guidelines according to the strength of the evidence underlying them. For guidelines covering medication use in pregnancy, we do this by distinguishing between C-class and D-class drugs. For other guidelines, we do this by determining whether UpToDate gives the guideline a 1A rating-its strongest recommendation based on the highest possible quality of evidence-or not.® 2.4 Measuring Adherence To measure adherence to each prescription drug guideline, we first define the circumstance under which the guideline applies. We consider each patient-year that meets this circumstance to be in the risk set for the corresponding guideline. We then use the prescription drug purchase data to define a binary indicator for adherence for each case in the risk set. Appendix Tables 1-4 provide additional information on the definition of the risk set and of adherence for each guideline. For 39 of our 46 guidelines issued by the Board, we follow the Board's definition as closely as possible in defining the risk set and adherence. For example, one guideline recommends that indi- viduals should use statins 12-18 months after a myocardial infarction (ie., heart attack) diagnosis. We define the risk set to include each individual's first observed inpatient diagnosis for myocar- dial infarction, and we define adherence by whether the individual purchases a statin within 12-18 months after discharge from that inpatient episode. The remaining 7 of our guidelines issued by the Board recommend against certain prescriptions or combinations of prescriptions for those 75 and older. For these guidelines, we define the risk set. to be person-years who are 75 years and older and where the person purchased the given prescription or combination of prescriptions when she was 74. For example, one of these guidelines recommends that individuals 75 and older should avoid a particular set of potentially risky drugs including some tranquilizers and opioids. We therefore define the risk set to be the intersection of person-years who are 75 years or older and people who purchased at least one of the potentially risky drugs at age 74. We define adherence for a given person-year as an indicator for not purchasing any of these drugs. 8UpToDate is a US-based publisher of clinical decision support tools for practitioners. It uses the system of ratings developed by the Grading of Recommendations Assessment, Development and Evaluation working group; see www. gradeworkinggroup.org. For 13 guidelines we are unable to obtain a rating from UpToDate. For the 17 guidelines recommending against the use of certain drugs during pregnancy, we define the risk set to be the set of pregnancies in which the mother purchased the drug(s) during the 24 months prior to conception, and we define adherence as not purchasing the drug(s) during the pregnancy. 2.5 Averaging across Guidelines We analyze results separately for each guideline, but also average results across guidelines. When we average, we account for the fact that different guidelines affect different shares of the population by weighting each guideline by its prevalence in the population. Specifically, for each gender and for each age from 0 through 85, we weight each guideline by the fraction of people in our sample, of the given gender and age, who are in the risk set in a reference year.? We then take an unweighted average across ages for each gender and across both males and females as our measure of prevalence. The resulting summary statistics on guideline prevalence and adherence should be interpreted as reflecting the average experience of a (hypothetical) person who lives each age of life, from 0 through 85, during our sample period. 2.6 Descriptive Statistics Our analysis sample consists of 5,887,471 individuals aged 85 or younger for whom we have valid information on completed education and who fall into the risk set for at least one guideline over the 2005-2017 period during which we measure prescription drug purchases. Of these individuals, 149,399 have access to expertise at some point during the sample period, of whom over 95 percent have access to expertise throughout the entire sample period. The share of the population in the risk set ranges from 7.6 percent for the guideline that adults should use penicillin V for their first antibiotic treatment (as opposed to starting with a broader- spectrum antibiotic), to 0.003 percent for the recommendation against using anti-epileptics during °For guidelines recommending against certain prescriptions for those 75 and older, the reference year is 2017, the most recent year in which we measure prescription drug purchases. Otherwise, the reference year is 2016, the most recent year in which we measure inpatient and outpatient hospital visits. 10Of those who fall into the risk set for at least one guideline, 2.4 percent are excluded from the analysis sample because of missing or invalid information on completed education. 10 pregnancy (Appendix Figure 1, Panel A). On average, over their life cycle, an individual is exposed to 36.32 guidelines. The average woman is subject to 43.23 guidelines and the average man to 28.85 guidelines, with the difference driven primarily by the pregnancy guidelines. Guidelines are substantially more prevalent for the elderly. Appendix Figure 2 shows the age and gender-specific patterns of guideline prevalence in more detail. Rates of adherence vary considerably across guidelines (Appendix Figure 1, Panel B). Among those without access to expertise, adherence ranges from 20.4 percent for the recommendation that individuals age 50 and older take osteoporosis medications in the 12 months after a fracture diagnosis, to 98.8 percent for the guidelines against taking D-class tetracyclines (antibiotics) and progestogens (hormones) during pregnancy. On average, over their life cycle, an individual without access to expertise adheres to guidelines 54.4 percent of the time. Appendix Tables 1-4 report the share of the population covered by each guideline, as well as the average adherence rate for each guideline among those without access to expertise. 3 Results 3.1 Individual Guidelines Figure 1 presents estimated differences in adherence between otherwise similar individuals with and without access to expertise for each of the 63 guidelines we study. Specifically each row reports the coefficient and the 95 percent pointwise confidence interval from a linear regression of an indicator for adherence on an indicator for access to expertise and a set of baseline controls; the sample is the set of patient-years in the risk set for the given guideline. The baseline controls are indicators for: income percentile, calendar year, month, age in years, gender, highest level of education, municipality of residence, and the number of children previously born to the person (zero for males). Figure 1 orders guidelines by the size of the coefficient on access to expertise. Darker colors indicate guidelines that affect a larger share of the population, which is also reflected in generally 11 smaller confidence intervals. Out of 63 guidelines, we find a negative point estimate for 41 of them, indicating that access to expertise is associated with lower adherence. For these 41 guidelines, 20 of the estimates are statistically significantly different from zero. For example, for the guideline that individuals aged 75 and older should avoid a particular set of potentially risky drugs including some tranquilizers and opioids, we find that access to expertise is associated with a statistically significant 4.1 percentage point lower adherence (95 percent confidence interval 5.7 to 2.6), relative to a 49.4 percent adherence rate among those without access to expertise. Likewise, for the guideline advising that pregnant women not take C-class opioids, access to expertise is associated with a statistically significant 1.9 percentage point lower adherence (95 percent confidence interval 3.4 to 0.3), relative to an 85.1 percent adherence rate among those without access to expertise. For the remaining 22 guidelines, the point estimate is positive, with 3 of these estimates sta- tistically significantly different from zero. For example, for the guideline recommending the use of statins 12-18 months after a myocardial infarction diagnosis, access to expertise is associated with a statistically insignificant 0.5 percentage point greater adherence (95 percent confidence interval -2.1 to 3.0) relative to a 53.2 percent adherence rate among those without access to expertise. 3.2 Aggregate Patterns On average across all of these guidelines, individuals with access to expertise are 3.8 percentage points less likely to adhere to guidelines (Figure 2, top row); this estimate is highly statistically significant, with a 95 percent confidence interval spanning 4.1 to 3.5. The point estimate repre- sents a 7.0 percent lower adherence rate among those with access to expertise, compared to the baseline adherence rate of 54.4 percent for those without access to expertise. In other words, while the average patient without access to expertise adheres to guidelines 54.4 percent of the time, a demographically similar patient with access to expertise adheres only 50.6 percent of the time. Figure 2 also reports the average adherence gap for particular subgroups of patients or guide- lines. Panel A shows that adherence among those with access to expertise is about 3.3 percentage points lower for women (95 percent confidence interval 3.7 to 2.9) and 4.4 percentage points lower 12 for men (95 percent confidence interval 4.9 to 3.9); these estimates-which are statistically distin- guishable (p-value 0.0003)-represent an 6.0 and 8.3 percent lower adherence rate for women and men, respectively, relative to those without access to expertise. Panel B shows that adherence is statistically significantly lower for those with access to expertise in each of the four, mutually exclusive and exhaustive categories of guidelines we created. The adherence gap is most pronounced for antibiotic guidelines, where those with access to expertise are on average 5.2 percentage points (about 9.8 percent) less likely to adhere (95 percent confidence interval 5.6 to 4.9 percentage points). The adherence gap for antibiotics is statistically different from the adherence gap for each of the other three guideline categories (p-values are 0.0009, < 0.0001, and < 0.0001 for tests of equality with elderly, diagnosis-specific, and pregnancy guidelines, respectively). The adherence gap is least pronounced (but still statistically significantly different from zero) for the pregnancy guidelines, where those with access to expertise are on average 2.1 percentage points (about 2.4 percent) less likely to adhere (95 percent confidence interval 2.6 to 1.6 percentage points). We also find that guidelines that have a higher adherence rate among those without access to expertise tend to have a more pronounced adherence gap, although this relationship is not statistically distinguishable from zero (Appendix Figure 3). Panel C shows how the relationship between access to expertise and adherence changes as we narrow or broaden the definition of either "access" or "expertise." Our baseline definition of access to expertise defines doctors as experts and access based on being a doctor, partnering with one, or having one in the close family. When we narrow the definition of access to expertise to include only being a doctor oneself, access to expertise is now associated with a more negative, 8.4 percentage point lower adherence rate (95 percent confidence interval: 9.0 to 7.8), compared to our baseline estimate of 3.8.11 When we broaden the definition of access to include having a physician in one's extended family, access to expertise is associated with only a 1.6 percentage point lower adherence rate (95 percent confidence interval: 1.9 to 1.3). Likewise, if we leave the definition of access unchanged but broaden the definition of experts to include nurses and pharmacists, access to expertise is associated with only a 0.9 percentage point lower adherence rate (95 percent confidence '1When we exclude those who are doctors from the sample entirely (not shown), access to expertise is associated with a less negative 1.8 percentage point lower adherence rate (95 percent confidence interval: 2.2 to 1.4). 13 interval 1.0 to 0.7). 3.3 Interpretation 3.3.1 Socioeconomic Status One explanation for the lower adherence to medication guidelines among those with access to expertise is that the negative relationship between access to expertise and adherence is driven by unobserved socioeconomic differences between those with and without access to expertise. Recall that we control for income percentile and education in our main analysis. Since doctors are a relatively high-SES occupation, and prior evidence indicates that adherence is positively associated with SES in both the US (e.g., Kennedy and Erb 2002; Mojtabai and Olfson 2003; and Madden et al. 2008) and Sweden (Wamala et al. 2007), we expect any remaining, unmeasured differences in socioeconomic status (SES) between those with and without access to expertise to bias against our findings, toward a more positive association between adherence and access to expertise. Consistent with this expectation, Panel D of Figure 2 shows that removing income percentile and education from our set of controls produces a less negative association between adherence and access to expertise. The scatterplot in Figure 3 evaluates the role of income more directly. The y-axis variable is a measure of the association between adherence and access to each of a broad set of specialized occupations, obtained by augmenting the models underlying Figure 1 to include indicators for access to each occupation. The z-axis variable is the average income percentile of those with access to the given occupation. Not surprisingly, we find a positive association between the average income of people with access to a given occupation and their adherence rate. However, doctors are a major outlier; although those with access to doctors have very high incomes, access to doctors is associated with markedly lower adherence.!? Figure 3 thus suggests that access to doctors is associated with lower adherence despite, rather than because of, the high socioeconomic status of those with access to doctors. Consistent with Panel C of Figure 2, those with access to nurses and pharmacists are close to the line of best fit in Figure 3. 14 3.3.2 Health Our finding could also be driven by health differences between those with and without access to expertise. Existing evidence, including prior work in our setting, indicates that doctors and their families tend to have better health and health behaviors (e.g., Chen et al. 2019; Artmann et al. forthcoming; and Leuven et al. 2013).18 The sample for each of our regression analyses of adherence is restricted to those who fall within the risk set for a given guideline. We expect that this reduces the scope for differences in health between those with and without access to expertise. But it is likely that some unmeasured variation in health remains among those in the risk set, both because the construction of the risk set considers only a limited number of health factors, and because selection into the risk set may depend on non-health factors such as willingness to seek out diagnosis or treatment. Whether people who are in better health are more or less likely to follow guidelines is a priori unclear. In cases where adherence to the guideline trades off the health of the patient against other considerations-such as antibiotic guidelines which recommend starting with a less aggressive treatment for public health reasons, or recommendations against medication in pregnancy which trade off the health of the mother against potential risks to the fetus-we might expect patients in poorer health to be less likely to follow the guideline. If those in the risk set with access to expertise are in better health, this would bias the estimates against our findings, toward a positive association between adherence and access to expertise. In practice, adding controls for the five health conditions described in section 2.2 makes little difference to the estimated relationship between adherence and access to expertise (Figure 2, Panel D). 3.3.3. Comfort with or Access to Pharmaceuticals Another possible explanation for the adherence gap is that access to expertise is associated with greater familiarity and comfort with pharmaceutical solutions to medical problems, or greater ease of filling prescriptions, and thus a greater propensity to take medications even in contradiction of guidelines. Consistent with this explanation, for guidelines whose risk set is based on taking a Consistent with this evidence, Panel A of Appendix Figure 4 shows that those with access to expertise are less likely to be in the risk set for some guidelines where the risk set is based on a diagnosis. 15 particular medication, those with access to expertise are on average more likely to be in the risk set (Appendix Figure 4, Panel B). As noted above, differential selection into the risk set could also affect the association between access to expertise and unmeasured factors, such as health, among those in the risk set. However, Panel E of Figure 2 shows that the relationship between adherence and access to ex- pertise is similar between guidelines that recommend against taking a specific drug or class of drugs ("don't take" guidelines) and those that recommend in favor of doing so ("do take" guidelines). For the 30 "don't take" guidelines, we estimate that access to expertise is associated with a 3.4 percentage point (95 percent confidence interval 4.2 to 2.6) lower probability of adherence. This is similar to the 2.9 percentage point lower adherence (95 percent confidence interval 3.3 to 2.6) for the 27 "do take" recommendations. The adherence gaps are not statistically distinguishable between these two groups (p-value = 0.1947). This suggests that comfort with or access to pharmaceuticals does not account for the negative association between adherence and access to expertise. 3.3.4 Superior Information about Guidelines The final explanation we consider is that access to expertise brings with it access to information that contradicts the guidelines in some situations, and/or the confidence (or ability) to act on this information. One testable implication of this hypothesis is that access to expertise will be more negatively associated with adherence to guidelines that are based on weaker clinical evidence. Consistent with this implication, Panel F of Figure 2 shows a larger adherence gap where the evidence is weaker. For guidelines related to medication use in pregnancy, the recommendation against C-class drugs is weaker than that against D-class drugs (see Section 2.3); correspondingly, the adherence gap is -2.3 on average for C-class drugs and -1.2 on average for D-class drugs, and these two values are statistically distinguishable (p-value = 0.0044). For guidelines related to specific diagnoses and to medication use among the elderly, among those for which we are able to find a rating on UpToDate (again see Section 2.3 for details), we find an adherence gap of -3.7 among those with weaker evidence and a gap of -3.4 among those with stronger evidence, though 16 the difference between the two groups is not statistically distinguishable (p-value = 0.1962).'4 Another testable implication is that access to expertise will be most negatively associated with adherence to guidelines whose recommendations are intended to serve goals other than the narrow interest of the patient. The antibiotic guidelines to use narrower- rather than broader-spectrum antibiotics are an example of recommendations motivated by public (rather than private) health considerations (Hyun 2017; Pichichero 2002; Sirota et al. 2017). As reported in section 3.2, the adherence gap is largest-by a considerable and statistically significant margin-for the antibiotic guidelines, among the four groups of guidelines that we consider. 4 Conclusion As of mid-2018, the US National Guidelines Clearinghouse described over 1,400 currently active medical guidelines (Timmermans and Berg 2003; Agency for Healthcare Research and Quality 2018a,b). Guidelines can help move average practice towards evidence-based standards, but can also discourage customizing care to relevant medical circumstances (Basu 2011; Lugtenberg et al. 2011; Boudoulas et al. 2015).1° We find that patients with access to medical expertise are, on average, less adherent to med- ication guidelines. This suggests an important role in non-adherence for factors other than those emphasized in much of the literature-such as ignorance, complexity, or failures of patient-provider communication-which would be expected to diminish with access to expertise. The normative implications of our findings are not clear. It is possible that lower guideline adherence among those with access to expertise may partly reflect these patients' superior under- standing of guidelines. Our finding that the negative relationship between access to expertise and guidelines adherence is more pronounced for guidelines based on weaker clinical evidence, and for guidelines intended to serve interests beyond those of the patient, is consistent with this interpreta- tion, as is other evidence from our setting that those with a health professional in their family are "For the 13 guidelines in these categories for which we were not able to find a rating on UpToDate, we find an adherence gap of -4.3 percentage points (95 percent confidence interval -5.0 to -3.6). 'SGerber et al. (2010) report that more than 80 percent of the US public is somewhat or very convinced by the argument that treatment guidelines prevent customizing care; see also Patashnik et al. (2017). 17 healthier overall (Chen et al. 2019). However, there is also evidence that practitioners' departures from prescribing guidelines lead to worse patient outcomes (Abaluck et al. 2020; Cuddy and Currie 2020; Currie and MacLeod 2020). An important avenue for further research is to identify whether and when non-adherence is in the patient's best interest. 18 References Abaluck, Jason, Leila Agha, David Chan, Daniel Singer, and Diana Zhu, "Fixing misal- location with guidelines: Awareness vs. adherence," Working Paper 27467, National Bureau of Economic Research July 2020. Agency for Healthcare Research and Quality, "About NGC and NQMC: Guidelines and measures," 2018. http://www.ahrq.gov/gam/about/index.html. - , "Web archive of National Guideline Clearinghouse," July 2018. Accessed May 15, 2020. https: //webo.archive.org/web/20180713042326/https: //www. guideline. gov/. Alpert, Joseph S., "Why are we ignoring guideline recommendations?," The American Journal of Medicine, February 2010, 123 (2), 97-98. Artmann, Elisabeth, Hessel Oosterbeek, and Bas van der Klaauw, "Do doctors improve the health care of their parents? Evidence from admission lotteries," American Economic Jour- nal: Applied Economics, forthcoming. Aslani, Parisa, Rana Ahmed, and Filipa Alves da Costa, "The role of adherence in phar- maceutical care," in Filipa Alves da Costa, J. W. Foppe van Mil, and Aldo Alvarez-Risco, eds., The Pharmacist Guide to Implementing Pharmaceutical Care, Cham: Springer International Publishing, 2019, pp. 41-57. Basu, Anirban, "Economics of individualization in comparative effectiveness research and a basis for a patient-centered health care," Journal of Health Economics, May 2011, 30 (3), 549-559. Edition: 2011/04/23. Bodnaruk, Andriy and Andrei Simonov, "Do financial experts make better investment deci- sions?," Journal of Financial Intermediation, October 2015, 24 (4), 514-536. Bosworth, Hayden B., "Causes of medication nonadherence," in Hayden B. Bosworth, ed., Enhancing Medication Adherence: The Public Health Dilemma, Tarporley: Springer Healthcare Ltd., 2012, pp. 9-16. Boudoulas, Konstantinos Dean, Carl V. Leier, Paraschos Geleris, and Harisios Boudoulas, "The shortcomings of clinical practice guidelines," Cardiology, 2015, 130 (3), 187- 200. Bronnenberg, Bart J., Jean-Pierre Dubé, Matthew Gentzkow, and Jesse M. Shapiro, "Do pharmacists buy Bayer? Informed shoppers and the brand premium," The Quarterly Journal of Economics, November 2015, 130 (4), 1669-1726. Brown, Marie T. and Jennifer K. Bussell, "Medication adherence: WHO cares?," Mayo Clinic Proceedings, April 2011, 86 (4), 304-314. Bunker, John P. and Byron Wm. Brown, "The physician-patient as an informed consumer of surgical services," New England Journal of Medicine, May 1974, 290 (19), 1051-1055. 19 Chen, Stacey H., Hongwei Chuang, Jennjou Chen, and Tzu-Hsin Lin, "Physicians Treat- ing Physicians: The Relational and Informational Advantages in Treatment and Survival," SSRN Scholarly Paper, Social Science Research Network December 2020. Chen, Yiqun, Petra Persson, and Maria Polyakova, "The roots of health inequality and the value of intra-family expertise," Working Paper 25618, National Bureau of Economic Research February 2019. Chou, Yiing-Jeng, Nicole Huang, I-Feng Lin, Chung-Yeh Deng, Yi-Wen Tsai, Long- Shen Chen, and Cheng-Hua Lee, "Do physicians and their relatives have a decreased rate of cesarean section? A 4-year population-based study in Taiwan," Birth, September 2006, 33 (3), 195-202. Colantonio, Lisandro D., Lei Huang, Keri L. Monda, Vera Bittner, Maria-Corina Serban, Benjamin Taylor, Todd M. Brown, Stephen P. Glasser, Paul Muntner, and Robert S. Rosenson, "Adherence to high-intensity statins following a myocardial infarction hospitalization among medicare beneficiaries," JAMA Cardiology, August 2017, 2 (8), 890-895. Cuddy, Emily and Janet Currie, "Rules vs. discretion: Treatment of mental illness in U.S. adolescents," Working Paper 27890, National Bureau of Economic Research October 2020. Currie, Janet M. and W. Bentley MacLeod, "Understanding doctor decision making: The case of depression treatment," Econometrica, May 2020, 88 (3), 847-878. Danielsson, Bengt R. and Lennart Dencker, "Graviditets- och amningsinformation," Febru- ary 2019. https: //www.fass.se/LIF/menydokument ?userType=OémenyrubrikId=124. Electronic Medicines Compendium, "What is an SmPC? Help and support." Accessed May 20, 2020. https://emcsupport .medicines.org.uk/support/solutions/articles/7000007888- what-is--an-smpc-. Fischer, Florian, Kerstin Lange, Kristina Klose, Wolfgang Greiner, and Alexander Kraemer, "Barriers and strategies in guideline implementation-a scoping review," Healthcare, 2016, 4 (3). Fleming-Dutra, Katherine E., Adam L. Hersh, Daniel J. Shapiro, Monina Barto- ces, Eva A. Enns, Thomas M. File Jr, Jonathan A. Finkelstein, Jeffrey S. Gerber, David Y. Hyun, Jeffrey A. Linder, Ruth Lynfield, David J. Margolis, Larissa S. May, Daniel Merenstein, Joshua P. Metlay, Jason G. Newland, Jay F. Piccirillo, Re- becca M. Roberts, Guillermo V. Sanchez, Katie J. Suda, Ann Thomas, Teri Moser Woo, Rachel M. Zetts, and Lauri A. Hicks, "Prevalence of inappropriate antibiotic pre- scriptions among US ambulatory care visits, 2010-2011," JAMA, May 2016, 315 (17), 1864-1873. Fonarow, Gregg C., Clyde W. Yancy, Adrian F. Hernandez, Eric D. Peterson, John A. Spertus, and Paul A. Heidenreich, "Potential impact of optimal implementation of evidence- based heart failure therapies on mortality," American Heart Journal, 2011, 161 (6), 1024-1030.¢3. Frakes, Michael, Jonathan Gruber, and Anupam Jena, "Is great information good enough? Evidence from physicians as patients," Journal of Health Economics, January 2021, 75, 102406. 20 Gerber, Alan S., Eric M. Patashnik, David Doherty, and Conor Dowling, "A national survey reveals public skepticism about research-based treatment guidelines," Health Affairs, Oc- tober 2010, 29 (10), 1882-1884. Glanz, K., S.B. Fiel, L.R. Walker, and M.R. Levy, "Preventive health behavior of physi- cians," Academic Medicine, 1982, 57 (8). Gramelspacher, Gregory P., Xiao-Hua Zhou, Mark P. Hanna, and William M. Tierney, "Preferences of physicians and their patients for end-of-life care," Journal of General Internal Medicine, June 1997, 12 (6), 346-351. Grytten, Jostein, Irene Skau, and Rune S¢grensen, "Do expert patients get better treatment than others? Agency discrimination and statistical discrimination in obstetrics," Journal of Health Economics, January 2011, 30 (1), 163-180. Handel, Benjamin R. and Jonathan T. Kolstad, "Health insurance for 'humans': Information frictions, plan choice, and consumer welfare," American Economic Review, 2015, 105 (8), 2449- 2500. Hyun, David, "What drives inappropriate antibiotic use in outpatient care?," Technical Report, Pew Charitable Trusts 2017. Irwin, Amy, Susan Moore, Connie Price, Tim Jenkins, Lauren DeAlleaume, and David West, "Flexible interventions to decrease antibiotic overuse in primary care practice: A report from SNOCAP-USA," Technical Report, Agency for Healthcare Research and Quality 2014. Johnson, Erin M. and M. Marit Rehavi, "Physicians treating physicians: Information and incentives in childbirth," American Economic Journal: Economic Policy, 2016, 8 (1), 115-41. Kennedy, Jae and Christopher Erb, "Prescription noncompliance due to cost among adults with disabilities in the United States," American Journal of Public Health, July 2002, 92 (7), 1120-1124. Krueger, Kem P., Bruce A. Berger, and Bill Felkey, "Medication adherence and persistence: A comprehensive review," Advances in Therapy, July 2005, 22 (4), 313-356. Leuven, Edwin, Hessel Oosterbeek, and Inge de Wolf, "The effects of medical school on health outcomes: Evidence from admission lotteries," Journal of Health Economics, July 2013, 32 (4), 698-707. Levitt, Steven D. and Chad Syverson, "Market distortions when agents are better informed: The value of information in real estate transactions," Review of Economics and Statistics, Novem- ber 2008, 90 (4), 599-611. Lopez- Vazquez, Paula, Juan M. Vazquez-Lago, and Adolfo Figueiras, "Misprescription of antibiotics in primary care: A critical systematic review of its determinants," Journal of Evaluation in Clinical Practice, April 2012, 18 (2), 473-484. Lugtenberg, Marjolein, Jako S. Burgers, Carolyn Clancy, Gert P. Westert, and Eric C. Schneider, "Current guidelines have limited applicability to patients with comorbid conditions: A systematic analysis of evidence-based guidelines," PloS one, 2011, 6 (10), e25987. 21 Madden, Jeanne M., Amy J. Graves, Fang Zhang, Alyce S. Adams, Becky A. Briesacher, Dennis Ross-Degnan, Jerry H. Gurwitz, Marsha Pierre-Jacques, Dana Gelb Safran, Gerald S. Adler, and Stephen B. Soumerai, "Cost-related medi- cation nonadherence and spending on basic needs following implementation of Medicare Part D," JAMA, April 2008, 299 (16), 1922-1928. McDonald, Heather P., Amit X. Garg, and R. Brian Haynes, "Interventions to enhance patient adherence to medication prescriptions: scientific review," JAMA, December 2002, 288 (22), 2868-2879. Mojtabai, Ramin and Mark Olfson, "Medication costs, adherence, and health outcomes among medicare beneficiaries," Health Affairs, July 2003, 22 (4), 220-229. Neiman, Andrea B., "CDC Grand Rounds: Improving medication adherence for chronic dis- ease management - Innovations and opportunities," MMWR. Morbidity and Mortality Weekly Report, 2017, 66. Nieuwlaat, Robby, Nancy Wilczynski, Tamara Navarro, Nicholas Hobson, Rebecca Jeffery, Arun Keepanasseril, Thomas Agoritsas, Niraj Mistry, Alfonso Iorio, Su- san Jack, Bhairavi Sivaramalingam, Emma Iserman, Reem A Mustafa, Dawn Jeraszewski, and Chris Cotoi, "Interventions for enhancing medication adherence," Cochrane Database of Systematic Reviews, 2014, (11). Patashnik, Eric M., Alan S. Gerber, and Conor M. Dowling, Unhealthy Politics: The Battle over Evidence-Based Medicine, Princeton, NJ: Princeton University Press, 2017. Pichichero, Michael E., "Dynamics of antibiotic prescribing for children," JAMA, June 2002, 287 (23), 3133-3135. Rutherford, R.C., T.M. Springer, and A. Yavas, "Conflicts between principals and agents: evidence from residential brokerage," Journal of Financial Economics, June 2005, 76 (3), 627- 665. Sirota, Miroslav, Thomas Round, Shyamalee Samaranayaka, and Olga Kostopoulou, "Expectations for antibiotics increase their prescribing: Causal evidence about localized impact," Health Psychology, 2017, 36 (4), 402-409. Socialstyrelsen, "Indikatorer for god lakemedelsterapi hos aldre," 2010. Accessed October 2019. https: //www.socialstyrelsen.se/globalassets/sharepoint-dokument/artikelkatalog/ ovrigt/2010-6-29.pdf. -, "Halsodataregister - réddar liv och ger battre vard," 2019. Accessed August 2021. https: //www.socialstyrelsen.se/globalassets/sharepoint--dokument/dokument- webb/ovrigt/halsodataregister-information-om-nyttan-med-register. pdf. Statistics Sweden, "Longitudinal integrated database for health insurance and labour mar- ket studies (LISA)." Accessed May 2019. https://www.scb.se/en/services/ordering- data-and-statistics/ordering-microdata/vilka-mikrodata-finns/longitudinella- register/longitudinal-integrated-database-for-health-insurance-and-labour- market-studies-lisa/. 22 Sveriges Kommuner och Regioner, "Om varden i siffror.". Accessed May 20, 2020. https: //vardenisiffror.se/about. Swedish Research Council, "The Total Population Register (RTB)." Accessed September 27, 2021. https: //www.registerforskning.se/en/register/population-statistics/. The GRADE Working Group, "GRADE home." Accessed June 10, 2020. https://www. gradeworkinggroup.org/. Timmermans, Stefan and Marc Berg, The Gold Standard, Temple University Press, 2003. Ubel, Peter A., Andrea M. Angott, and Brian J. Zikmund-Fisher, "Physicians recommend different treatments for patients than they would choose for themselves," Archives of Internal Medicine, April 2011, 171 (7), 630-634. UpToDate, Inc. and/or its affiliates, "Evidence-based clinical decision support at the point of care." Accessed June 10, 2020. https: //www.uptodate.com/home. van Driel, Mieke L., Michael D. Morledge, Robert Ulep, Johnathon P. Shaffer, Philippa Davies, and Richard Deichmann, "Interventions to improve adherence to lipid- lowering medication," Cochrane Database of Systematic Reviews, 2016, (12). Wamala, Sarah, Juan Merlo, Gunnel Bostrom, Christer Hogstedt, and Gunner Agren, "Socioeconomic disadvantage and primary non-adherence with medication in Sweden," Interna- tional Journal for Quality in Health Care, June 2007, 19 (3), 134-140. Weissman, Joel S., Zara Cooper, Joseph A. Hyder, Stuart Lipsitz, Wei Jiang, Michael J. Zinner, and Holly G. Prigerson, "End-of-life care intensity for physicians, lawyers, and the general population," JAMA, January 2016, 315 (3), 303-305. Wunsch, Hannah, Damon Scales, Hayley B. Gershengorn, May Hua, Andrea D. Hill, Longdi Fu, Therese A. Stukel, Gordon Rubenfeld, and Robert A. Fowler, "End-of-life care received by physicians compared with nonphysicians," JAMA Network Open, July 2019, 2 (7), e197650. 23 Figure 1: Adherence and Access to Expertise, by Guideline Don't initiate UTI treatment with quinolones, women Initiate RTI treatment with penicillin V, children Don't take hazardous drugs after heart failure diagnosis, elderly Don't take NSAIDs, elderly Initiate antibiotics treatment with penicillin V, children Take zopiclone conditional on taking hypnotics and sedatives, elderly Don't take other D-drugs, pregnancy No polypharmacy w.r.t. drugs in the same therapeutic group, elderly Don't take C-class NSAIDs, pregnancy Don'y take specific drugs, elderly Initiate UTI treatment with recommended antibiotics, elderly Don't take hazardous sleep medication, elderly Don't take hazardous drugs after dementia diagnosis, elderly Don't take other C-drugs, pregnancy Initiate antibiotics treatment with penicillin V, adults Take oxazepam conditional on anxiolytics, elderly Continue with antidepressant treatment Don't take C-class corticosteroids, pregnancy Don't initiate UTI treatment with quinolones, elderly Take antipsychotics 0-6m after schizophrenia diagnosis Take antipsychotics 12-18m after schizophrenia diagnosis Don't take C-class antidepressants, pregnancy Don't take C-class opioids, pregnancy Take beta blockers & ACE inhibitors or ARBs after heart failure diagnosis, elderly Don't take C-class antibiotics (other), pregnancy Don't take antiplatelet drugs after Afib diagnosis, elderly Take antiplatelet drugs 0-1m after TIA w/o Afib diagnoses Don't take D-class tetracyclines, pregnancy Take statins 0-1m after ischemic stroke diagnosis Take ACE inhibitors or ARBs after heart failure diagnosis, elderly Don't take hazardous drugs after ischemic heart disease & heart failure diagnoses, elderly Take RAAS inhibitors & beta blockers 0-24m after heart failure diagnosis Don't take C-class anxiolytics, pregnancy Take anticoagulants 0-6m after Afib diagnosis with high stroke risk Don't take D-class immunosuppressants, pregnancy Don't take D-class progestogens, pregnancy Don't take C-class hypnotics and sedatives, pregnancy Don't take D-class macrolides, lincosamides or streptogramins, pregnancy Take antiplatelet drugs 0-1m after ischemic stroke w/o Afib diagnoses Take statins 12-18m after cerebral infarction diagnosis Avoid drugs with certain interactions, elderly Take ADP receptor blocker 0-1m after myocardial infarction diagnosis Take digoxin after heart failure & Afib diagnoses, elderly Take anticoagulants after Afib diagnosis, elderly Don't take C-class corticosteroids (systemic use), pregnancy Don't take hazardous drugs after ischemic heart disease & Afib diagnoses, elderly Take antihypertensives 0-1m after acute stroke diagnosis Take antihypertensives 0-1m after TIA diagnosis Take statins 0-1m after TIA diagnosis Take statins 12-18m after myocardial infarction diagnosis Take anticoagulants 12-18m after ischemic stroke & Afib diagnoses Take osteoporosis drugs 0-12m after fracture diagnosis No polypharmacy w.r.t. all drugs, elderly Don't take C-class hypothalamic hormones, pregnancy Take RAAS inhibitors or beta blockers 12-18m after heart failure diagnosis No polypharmacy w.r.t. psychotropic drugs, elderly Take RAAS inhibitors & beta blockers 0-6m after heart failure diagnosis Take anticoagulants 0-1m after ischemic stroke & Afib diagnoses Don't take C-class psychostimulants, pregnancy Take recommended drugs after dementia diagnosis, elderly Take anticoagulants 0-1m after TIA & Afib diagnoses Don't take antipsychotic drugs, elderly Don't take D-class antiepileptics, pregnancy 24 Tbh ies tite spae 5 = a0 Fiedsspasijancncncstse® wont { -.2 0 2 A Difference in adherence Note: For each guideline, we run an OLS regression of an indicator variable for adherence on an indicator for access to expertise, controlling for indicators for the calendar year and seasonal month at which we start measuring adherence, indicators for age in years, gender, highest level of education (or of mother's education for those under 27), municipality of residence (or of mother's residence for those under 18), income percentiles, and number of children previously born to the person (zero for males), all measured in the previous year. We include separate indicators for whether income is missing, zero, or negative (these represent, respectively, 7.3, 3.0, and 3.5 percent of the analysis sample) and for whether municipality of residence is missing (2.7 percent of the analysis sample). The sample for each regression is the set of patient-years in the corresponding risk set. Depicted 95% pointwise confidence intervals are based on standard errors which are clustered at the patient level for any guidelines for which a patient can appear in the risk set more than once, and which are heteroskedasticity-robust otherwise. The color code represents eight equally-sized bins of guideline prevalence, with darker colors representing higher prevalence. Figure 2: Adherence and Access to Expertise, Averages Baseline (63) > Panel A. By gender Men +e Women +o Panel B. By guideline category Antibiotics (6) +o! Elderly (20) re Diagnosis-specific (20) to) Pregnancy (17) He Panel C. Varying definition of access to expertise Including self only (61) Fe Including extended family to Including nurse or pharmacist in family o Panel D. Varying controls Dropping education and income controls tet Controlling for richer health measures +e Panel E. Dos vs. Don'ts Don't take (30) eH to Do take (27) Panel F. Strength of evidence Weaker evidence: pregnancy C-drugs (11) re Stronger evidence: pregnancy D-drugs (6) FH Weaker evidence: UpToDate Grade below 1A (18) te Stronger evidence: UpToDate Grade 1A (15) te I I I I | I I I I I | I | I I I I | I I | I I I I I I I | I I I | I I | 0 -08 -06 -04 -.02 Difference in adherence 02 Note: The plot shows the prevalence-weighted average coefficient on access to expertise from the regressions described in Figure 1. Spikes indicate upper and lower bounds of the 95% confidence interval. Row labels describe the analysis, with the number of included guidelines in parentheses if different from baseline. We bootstrap the estimation with 50 replicates drawn at the patient level, and construct confidence intervals based on the bootstrap standard errors. The prevalence weights are the guideline- and age-specific empirical probabilities of being in the risk set in the reference year. Baseline includes all patients and guidelines. In Panel A, we estimate the regression described in Figure 1 separately by guideline and gender of patient and use gender-specific prevalence weights. In Panels B, E and F, we average the coefficients for each set of guidelines. In Panel C, we re-estimate the regressions described in Figure 1 modifying the definition of access to expertise in three ways, first by excluding patients who are not themselves doctors (and thus excluding two guidelines that apply only to children), second by including patients with a doctor in their extended family, and third by including patients with access to a nurse or pharmacist. In the first row of Panel D, we re-estimate the regressions described in Figure 1 excluding the education and income variables. In the second row of Panel D, we re-estimate the regressions described in Figure 1 including as controls separate indicators for whether an individual had a heart attack, heart failure, lung cancer, type II diabetes, or asthma diagnosis in any preceding year. Panel E excludes the 6 antibiotic recommendations while Panel F excludes 13 guidelines for which there was no UptoDate grade available. 25 Figure 3: Adherence By Access to Different Occupations Teachers .05 Pharmacists ~~ Nursgs ° § awyers ankers 0 e@ Doctors -.05 30 40 50 60 70 Average income percentile of individuals with access to the occupation Note: The plot shows the prevalence-weighted average relationship between adherence and access to a given oc- cupation, plotted against the average income percentile of those with access to the occupation. To construct the y-axis variable, we augment the models underlying Figure 1 to include indicators for access to each occupation; the regressions include all of our baseline controls (including own income percentile), and a given case in the risk set may have access to multiple occupations. We then average the coefficient on the access indicator using the method underlying Figure 2. We exclude occupations with fewer than 150 individuals. We normalize the average coefficient for each occupation by subtracting the average coefficient for access to a doctor. To construct the x-axis variable, we calculate the average income percentile of those with access to the given occupation in 2016. We plot the line of best fit excluding doctors, nurses, and pharmacists, weighted by the number of individuals in the occupation, which is proportional to circle areas. 26 "'quaryed oy} ST sisAyeue Jo yun oY} 'd[qQ~e} STU} UL Pa4si] SOUI[EPINS oy} BULATOAUT SUOISSaIZeI 107 "esijrodxe 04 sseo0e INOYAIM ssoyy SuOWIe sOUOIOYpe Jo oseraAe poyysiom sousTRAId-spoAD9IT] pure SOUT[SpINS pure soSe [ye sso1oe Ieok soUsIEJOI OY} UL OOUETeAoId SUT[EpINS oy1oeds-o8e Jo oSvI9AVe OY} SMOYS MOI JS] IIL], "OJOIQIyUe pojos1ey-IOMOITeU @ YIM Suryerpur Sulpuswuioser Apo dun ere (OMO0IQIJUe poyeSIey-Jopeolq @) YIM SFeIIUL 1,UOP, PepIOM SUOCT}EPUSUTUIODEI 9T[} 'SpIOM JOJO Ul "OTOIqIyUe Ue Zupye} WO [UOT {IPUOD SsI9Y40 IOAO SBNIP VUIOS JO IOAR] UL pusUIUIODEI Ady} OSNeIEq 924 1,UOP,, 10 ,AY2} OP, Se SOUT[APINS ssoy} AjIssepo Jou Op 944 'osTZJodxe 04 ss900e YNOYIIM dSOY} IO] 9YVI VOUSIOYPR 9} 84SI] ,IOUIIOYpPe,, Jopun ,aTeys,, UUINTOS 'read soUDIIJOI OY} Ul 49S HSI SUT[EpMs yoo ul uorye_ndod ystpemg oy Jo oreYs oY} 84ST] 198 FSI, Jopun @reys, UUINOH *200% A[ng Jeqye/ul uoldiioseid 4sI] YIM sTenplAIpUl Japisuoo aa 'Goog A[N¢ Ul ape sUTe[> Snip SulAjasqo 47e\s aM sy "e}ep SUITe[D Snip ay} Ul aArasqo am uonjdiioseid ysiy oy} e1ojoq saeok Z UIGAIM UOT}dIJoseid oTjOIqIyUe Ue SAVY JOU PIP sys JI JWOULyeoT} OLJOIQIYWe SozeIIIUL [ENPIAIPUL Ue Je} OUTOP 9A "UOTJOOTUT 4oe!Iy AIOyerIdser {LIM 'woroeyur yoery Areurin :[ LQ 'sororqrue urnsyoeds-iepeoiq 0} pesoddo se -romoireu YIM yreys syuaTyed yey} SUOTyepUSUTUIODEI OTe 4SI oy} YNq seUTTEpINS [Ty :270A7 Si 92,000" osvioAy A Ulyorued Joy st uory +8T o8e '2002 Ane roqye/ut soto synpe 'A ulpiorued v -diroseid soqjorqijue \siy s.fenprArpuy 920° -Iqiyue Joy uordrosaid ysiy YIM S[eNplAIpuy = YY JUOUT}BI} SOTJOIqIyUe 9yeI}TUT A Uypforued Joy st uory 21-0 288 '2002 Aine Joye /ur soryo ueIpryp 'A uloraed 9° -djioseid sonorqryue ysig syenplATpuy TZO" "iqijue Joy uorjdiroseid ysiy YA STeNpIAIpUr «=yYA Jwoulywer, soNorqime syeryruy souojournb 10j 4OU ST WOT} 6L-8T eB '200 Ane seqe/ur TA ysurese ustIOM 'sauojoumnb vs" -diroseid sonorqijue \sig 8 fenprlArpuy 810° = SOMorqrue Joy uoTdIsoseid ysIy YUM UaMIOM YY yWouryeoT, TL eyeryur 4u0q A Urlyiorued 107 st wor, 9-0 282 '7002 Ane seqye/ur TTY ysure8e o10 wWoIpTIyo 'A 62° -dioseid sororqryue ysiy s,fenpLAlpuy TO" -qiyae roy uoydizoseid ysry qyLM sTenprtaIpuy = = ur_loraed gytMm yuouryeeIy [TAY eyeryTUyT ureUuly -Hoeuraid JO UlOJWeINJOI}IU IOFZ St UOT} +g) a8e '2002 A[ne 109s /uI TL Isurese Ajiopye 'sororqrjyue popusur cq -doseid sorjorqiyue 4s s,[enplAIpuy 8€00° «=: SOMJOTQTIUe Joy UOT|dIIOSeId ysIT UM ApIop[y -Woosal YM JuoUyeETy [Lf seu sauojoumnb 10} 4OU st UOT} +g) 032 '200% Atne Joye /UI TL ysurese Ajrepye 'souojoumb &9° -diioseid sonorqiue 4s s,[enplAalpul 8€00° = SOHOIqTyWe Joy uOdIoseid ysIg YM ALOp( = YM JUaUTywET] [LN eyerTar 4,u0q ereys vorraged ereys uvoruged ssuerTs4ypVy qos 4STY euljepins) SOIOIQUUY JO asf) A} SULIAAOD souTjapM :[ e[qey, xipueddy sooipueddy 27 asreyosip 19qyze goog Aine soye/ur o8reyostp 'aanyrey yVey IOJ stsouserp juaryed ATI9pya 'SISOUSBIP SINTTey Jeey IOEqye VI Mop ayey oc 98° Snip Jeyye Joy woMdioserg 82000° -3no Jo yuolyedul asi yytM ApIep[y «=-- SE W'V Jo s1oxiqryul FOV ee, suoydurds o1joyoAsd @IBAVS OJ you sBnzp 10d -Asdique Jo 'syue8e Suidasys cooe Ame seiye/ut Brey Ajiap 'syoayja OIsJeuljoyoIue YH -SIp 'eIjUeUep JO} sisouseIp juered = =_ -ja 'sIsoUSeIpP eljusWap soIyTe VI Mop aye. 1,u0q ves ssnip soy wormdiioseid oN TS000° -jno JO yuetyedur ysiy YIM AlJop[y ssnip snoprezey eyez 1,u0q esreyosip Joye (SuTyWeUTSUL Jo 'aurmSryseat 'izadauop gooz Ang saqye/ur eS8reyo Ajrap 'guII0e] Jo ou0) Snap pepusul SIP 'eIJUeWIOp Oj sisouserp queryed = -ja_ 'sIsouSeIp eljUeWIep 104 VI Mop ayey og wQ «2s "wooed «6e «Oy ~=UONdIIOSserg Tg000° = -3ho _ Jo quaTyedur ysiy YIM ApJoply Je ssnip pepueurustose oyeVT, 002 ATar Joye /ur egreyosip 'stsouserp qiyy yim 'on [re]. Jaeoy Joy sisouserip juermed Ajsepye 'sesouseip quy 27 am VI Moped oye} Od ge" a8reyosip sayye wordrroserg gg000° = -¥no Jo quaryedur ysigy YM ATIeply -«--[Fe} reo Joye UTXOIP oye], osreyo gooz Aine seqye/ur e8reyo "SIP JoIye sieyxooTq e19q YIM -sIp 'sisouseip quy qjIM 'eseasip Alsop 'sesouserp quy UOTJVUIqUIOD UI TTUredeIoA IO qIVSY OTMLEYDSI 10} sisouseip yuetyed = -«-2Y SBOSIP JIVOY OTWIAYSI J0qFe pepersuy) ayey 4,u0q 96° WoeZzeIy[Ip Joy uodtoseid ON TI000° = -3ho _ Jo quaTyedur ysIy YYM A[Jop[y ssnIp snoprezey oye. 1,u0q gooz Aja seqye/ut o8reyosip Ajrep 'SISOUSBIP SIN[Iey JAvaY WAL 'oseasIp ja 'sasouselp aIn[rey yreoy 2 eSreypsiIp Joye [Turedviea 10 qreo OMIEYOsI 10] sisouseip quetyed = aseaSIP JIVE] OTUIEYOSI Joqye pepersuy, aye} 1,U0q 86° «- WazeIy[Ip Joy uoldt1oseid ON Z80000° -jno Jo yusTyeduI ysIy YIM AlJOp[y ssnIp snoprezey eye, 1,u0q aoUeplAg ores vorTuyed oreys woruyged jo e472} ,u0p yySue13¢ "sa od sousraypy yes 4STY eulepmy (+g, se8y) Arepry oy Aq as-) UOT}eOIpay] BUIIOAOD souTjepIny :z s[quy, xIpueddy 28 [4zedauop pue weidofeyip 'lruredeioa pue WoezelyIp '[Iuederea pue sio -yoo[q B30q 'soljomip Surreds -umissejod pue winissejod 'SGIVSN pue ourreyrem turaid -se pue oulieyrem read USAT @ jo jJoyenb oules oy} Ul ssnip Jo uomoesoyur 9soy} JO} swuirel> ou sey [eNPrATpUl pL uimy Aaya weak ayy Jo Joyrenb @ ur (uum -[O9 Jxeu 94} 99S) SUOT}OBIEYUT UTeY Ajropye 'suo1joR.104 pepersuy) axed 7,U0(, 99° 'JeaAg] Ieak-yuatyed ye pouyecy TIO00° -120 JO SSNIp OJ SUITeTD YIM ApIaply -UI UTey190 [JIM ssnip ploay 00% Ane seyye/ur adreyosiIp 'qy Vy JO} stsouserp yworyed Ajsopya 'stsouserp quy VI ayet od 92° eSreypsip Joye uordiioseig Too" -qno Jo yueTyeduI ysIy WIM AlJop[y =-«Joye_=-s squR[NBeoorue aye, 00% Ajar soqye/ur asreyo eSIVLPSIp 'qyV JO} stsouseip yuaryed Ajsepys 'sisouseip quy Joye VI MOS sxe} ,u0d 99° "sIp Joye uorndrioserd oN TOO" "no Jo yusryedur yszy yYM APopy -«- SBN jopoyerdyue oyeq 4,u0q pl, ode reaX uaals & Ul SNIp o14049 ye Snap syoyDAsdrjue soy uoTdrI0s Ajrepye 'ssnaip pepeisuy aye} 1,u0q ee" -Asdiyue soy uorjdiioseid oN 62000° -eid yyim (GZ eaoge a8e) Ap10ply oyoyoAsdiyue = aye} = 4}, U0, osreyosip Jaye joreyos Io suorepsucIp 'oplureseyy 'guouayedoid 'gprurerAdosip = 'sysTuoseyue goo Ajnc soiye/ut a8reyostp AyI9pya UWINIOTe) BSULIEMOT oye1 {IeOy 'oIN[Tey JTVeY IOJ sisouserp yuerjed 'sisouseIp oN]rey yreey Iq? VI Moped ayey 3,u0q g2; 'SCIVSN Jo} woMdiioseid oN 92000° -3ho Jo quolyedul ysiy y{tM Ajlep[y =: SSHAP snoprezey oyey 4,u0q aSIeYyPSIp Joye sYV gooz Aja seqye/ut e8reyosip Ajr9epys 'sisouserp oanyrey JO siOJIQIYUT YOY jo ouo pue 'ommprey jreoy Joy sisouserp yooryed yreoy Joye sqYYY JO ss0qqry VIL Mopg ayey od 21, - SAOYOOTG BJoq JO] UoTydtIose1g 82000" -qno JO yueTyeduI ysIy WIM ATOp[y = -Ul FOV 2 siaxpolq eyoq oye], aoUuspIAS oreugs uoTTayed ereus wonmyed exe} 1,u0p oO OTLUIYWOD J "SA OG eousTeypVy yes YSTY (Penupquoo) ousfopIn sD q8uer4¢ 29 reak usals @ Ul spoT dBIOUT[OTOIyUE Wu ssnip io 'eprurepueqrs 'soisesfeue proido-uou s0q}0 Jo joureysoered pue sulepoo 'ourzewoidoid 'jopeurely 'uredezeryUn yg 'nredoeiq1u 'uredozerIp 10} uotydiiosoid ou PZ 888 1e@ peploae oq plnoys yey} Ajrop pepersuy) oye} 4,u0q 67" ']eAo] reeA-queryed ye pouyod €10° snap soy uoMdisoseid yyM Ajopry = -J9. 'sSnap oytoeds oye} 4,u0q rea UdAIs & Ul GIVEN JO} uordrioseid ou PL ose ye pepersuy ayey 3,u0q 69° oA] Teoh-quotyed ye pouyoq TIO" CIVSN 40j uordiioseid yyLA ATopLy ATJoppe 'SCIVSN e722 3,u0q zeak UsAIs & jo Joqzenb oures ou} ul ssnip qUsleyip sul wey} s10ul ou yi, un} Aoy) Teak Joy surej> sey [enpraArput ay} jo royrenb @ ul ssnip JuereqIp Ajzepye 'ssnaip pepeisuy aye 7,u0q 8° 'esa, reak-quoryed ye pouyeq g800° a10Ul 10 QT 10J swmrepo YIM ATIOPTA re 'yam Aoeurreyddsjod on reak UdATS @ JO Jaqrenb sures oY} Ul dnois 9,Ly .nedesey) sures 72, win} ASY4 9y} WO Snip suo ueYy} a10UI zeof oy} Jo Joyrenb e ul dnois OLY Ajzapye 'dnois 914 OU JOJ SUITED sey [eNprATput o1nedeisy} oures sy} Woy ssnip -nedeiey} sures oy} Ul ssnip pepeisuy) aye} 1,u0q, [9° ']eAg] eak-juatyed ye pouyeqd G100° 91OU IO OM} JOT SUMTeTD YIM ATIOPTY 'yom = ADeuFeydATod = ON uredeze -xO Joy ore read UsATS & UT 2L10@-9002 'So1)4Torxue Ajsepye 'sorAjorxue uo pepeisuy, aye} og 8@ «=-s- suorydizosoad somAjorxue [[V ¢900° = Joy: «UOTJdiz0seid yyIM savod-A[Jop[y =-«-« [RMOTJIpuoo uredozexo oe], Teak USAIS @ jo Joyrenb oures 94} Ul ssnip oidorjoyoAsd 0M} uwey} 210UI 72, win, Ay} weak ay} Jo 104renb e ou JOJ sUITe[D sey feNprarput ul ssnip sidomoyodsd juosoyip arou Ajrapya 's8nap ordorzyoy9 pepeisuy, aye} 3,u0q v ']eA9| reof-quotyed ye pouyod ZZ00° 10 90149 Joy surep> yy Aplappy -Asd -y-1-m AoeurreydAjod on aousplIAS ereugs uoTyTayed ereus woRnmyed exe} 4,u0p oO ONUIyUOS) sUTTEpIN 3 "SA OM eouerleypy qos ¥STY (penuryuoo) euljepiny) qBuer¢ 30 'UUIN[OO ,20USJOYpe,, oY} Ul IveA-quoTyed 943 se poyloods ssolun yuelyed 943 ST sIshjeue jo Jun 944 'e[qe} SI) Ul poysI, SouTjEpIns oy} SULA[OAUT SUCISSeIZeI IOq "USSTeIA\sTeIDOG 944 Aq pouyeEp jou ore Syos YSII SsOYM SoUTSpINS IOJ ZTOZ Ul peInseour st sousTeacig 'as]piodxe 04 sse00R YNOYPM VsoYyy SUOUIY sdUeIOYpe Jo eseioae poyysiom aouaTeaoid-spoAoojl] pus souljepmMs pue sosev [[e ssolo' IeoA dDUAIJoI 94} Ul soUsTeAeId oulfepms oyioeds-e8e Jo aseiaae oY] SMOYS MOI Ise] OTT, 'oVeqoTdy Aq payers you st Jt papeisun st aurjapins e 'Aue are arayqy jl 'suaping pue systr sysramqno A[reapo aower;duroo jo sqyoueq ey} pue 'souepras omryueros Ayipenb yBry Aq poysoddns ore (,,20ueptae jo yySuer4s,, WUN[OO ut peyeorput) seutjepmS yT oper 'ayeqoydy Aq payer sy "Snap ureyzseo SUIYVY JSUITe SPUSUIUIODEL JI JI oe} 1.UOp, se pue 'Snip ureyieo Sue} SpusUIUIODEL JI JI oye} Op,, Se SUT[epIns wv AjIsse[D oAA 'osTJJodxe 04 ssooov JNOYYIM asoy} [[e IO} ayer aoueleype 9} S4SI[ ,2OUsIOYpe,, JopuN ,aJVys,, UUIN[OO '1eeA aodUdIaJOI OY} Ul JOS YS S,SUl[apIMs yove ul UoTye_ndod ysipamg jo areys 94} S4SI[ 398 YSIL, Jopun ,ereys, UUIN[OD 'ssnip AloyeuIUreyul-tyue [eploieysuou :QTVSN +4999" OIWeYS! yUSTsUeIy :WLL, 'UOTyeT[AGY [eye :quy 'uorjoeyur youry Aroyerdsad :[ LY SWOTOezUI your} ATeULIN :[.LN :270N7 eg" 82000" Belay suIzeyjouloid Jo oUlzewoule 'oulzrxorpAy 'guizemoldoid = '(uredezerym ng 'uredozeru 'uredozerp) souidezeipozueq = Surjoe-Su0] Joy zeof udAts @ UT SsuUOT} -diioseid uoryeorpeu desis ou 2Z10¢-9002 'woreotpeur desjs Ajsepye 'aoryeorpeut pepesuy) ayey 1,u0q aL ']eAo] Teok-quetyed ye pouyeq 910° Jo} uoMdrioseid yyIM sreoX-AjJoply desjs snopiezey eye, 41,u0q auopidoz 10} are Ieok UsAIs B UL suUOTydIIOseId ApIopye 'seat SaATyepes pue sonouddéy [ye L102-9002 'seatyepes pue sorjoudéq -epes pue soroudéy Sune, VI Mopeg ayey Od 12° ']eA9| Feef-queryed ye pouyed €10° Joy uoMdrioseid yim sxeod-A[Jop[y «UO [eUOTyIpUOD auOpIdoz oye, aoUSplAg oreus uorTuyed ereys worfuyod eye} 7,u0p oO eNUIyUOD) SUTTapmM 3 "SA OM eouateypy qos ¥STY (penuryuoo) suyjapimy q78uer4¢ 31 osreyosIp qeye syquoU gT--ZT UTM +02 e88 '910% oun erojeq/ut pue ¢ooZ Ajme¢ seye/ur osTeYOSIp 'ore TVVY IOJ sisouseip sIsouseIp oIn[rey qreoy i9qye WIg{-Z]I s1eyxoo[q VI Mold ayey oc 99° Shap sayqye soy uondtoserg €2000° Ureur queryedur jsig WIM sfenpLarpuy)~=-s @ Jaq, AO SLOPIQIUI SYWH AeL +9T ose 'gooz A[ne re4ye/ur oBreyostp 'stsou asreyposip Joye yyWoU -3eIp QI y jnoyIIM 'Y]J, oJ sisouse sosouselp qyy 0/M VL Joye VI axe} Od TL TO IaH uondrrose1g 69000° Ip quatyedur ys1y YIM spenplarpuy == WIT ssnip yopoyeydryue oye], +81 082 'gooz A[ne soyye/ut eSreyostp 'era asreyosip Jaaye syyuOU -o1ydozryos Joy sisouserp yueryed no sisouseIp ermerydoziyos 104 VI aye} Od 69 9-0 UA uondroselg 9¢000° 40 yuaTyeduT ysIG YIM sTenplArpuy = -Fe wWig--) soyoyoAsdrjyue aye], +81 o8e 'gToz eunc sr0yjaq/ut 'gooz A[ne s9yye/ut e8reyosip 'eu esreypsip Joye syyuOUW -olydoztyps Joy sisouserp yueryedyno s[souseIp elusrydoztyos Joyye VI Mopg axe} Od oF = ST-ZT)UTqyIM = uondtsose1g ggo00° =: Jo: qwaryedur ysiy YUM spenprArpuy = wg{-ZT sorjoyoAsdiyue aye], +8T 088 '¢00z Ajne soye/ur ofreypsip 'sisouserp sosouse aBreyosIp Joye YOU agV WIL 'ayo13s ofUIEyOST IOF sIsou = -IP QW 27 SYOI4S OfMMEYOST 104 VI axe} Od go TO Gry = uordrioserg 61000° "Serp yuoryedur ysiy YM spenpralpuy = Je WT-Y syuepnFeoorjue oye], 62-8T aBe '910% aun os0yaq/ur 'goog Ane Jeyye/ul eSreypsip 'sisouseip qyy sosOuseIp aBreyosip Jeqye syyuoOU YAM 'ayxO1I}S IUMIAYOSI OJ sisouseIp qyy 2 sYOIs OTWIATOSI Jaye VI exe} Od eg ST-ZT UTA uorydrroserq 91000" = Ureua quoryedur qsiy 4M sTeNpIAIpuy =: WIGT--Z] S}uepNBeooryue aye], +8T ase 'coog Ane se9ye /Ul eBreyostp 'sts aSreypsip Joye [WOU -ouseIp qUy [IIA 'WLI, Joy sisoude sosouselp qUV 2 VLL 13 VI axe} Od ¢ TO Wy = uordrsose1g £60000" "Ip quoyedul ysiy YM spenplarpuy == Je WTO syuepnSeooryue oye], aoueplAa oreys uorTuyed ereys worruyed jo e724} 7,u0p yySue13¢ "sa od eousreypy yes 4STY eulepmy sosouseiq OYIo9dg Zurmoy[[oy asf) UOTeoIpay] SULIOAOD soUTjEpINyy :€ 9[qQe], xIpueddy 32 asieypsiIp Joye YWWOU +8T 088 'gooz Aine Jaye /ul aBreypstp 'WIJ, 10} stsouse SISOUSRIP SYOI4s ayNoR IAqye VI Moped aye}, od WwW TO UWA = uoldts9se1g €T00° -Ip quoeryedur ysiy YIM spenprarpuy wWT--) seAtsueyedAyiyue oye], +81 o8e '00g ATne s09ye/ur aBreyosip Joye YUU aBIeySIp 'ayorjs OTWIEYOST OF stsou sIsouSeIp ax014s OTULAYO VI aye} Od 9g T-0 «UIQ =-s uoTydt10se1g TT00° -8eIp yuoryedur 4sIy YILM syeNpLAripuy "SI soye WI) sulyejys oye], 6L-8T ase ye pue goog Aine saiye /ur oBreyo sTSOUSeIp UOT aSieyosIp Joye [WOU SIP 'WOT}OIEJUL [VIPIeVOAUI OJ sIsOU = -O.AVJUT [VIPIBOOAUI JONTe WITH VI axe} Od 82° T-O «UII = uonndrJoser1g TT00° -Belp yuoryedur 4sIy YIM sTeNprAIpuy Jayoo[q Joyde0e1 qGqy eeL asreyo +0z e8e '¢ooz Ap soqye/ut SISOUSVIP o1n -sIp Jeye syyuOW g-Q UIqzIM eSIeypsIp 'ony[re} jzrVey Joy sisouserp -[rej JreaYy Jaye WIg-H SIEYxpPoTq VI Moped aye} OC 9° ssnip yJOq 10; uorydisose1g 88000° UTeUL JUSTJeAUI SI YILA STeNpLAIpuy e19q 27 SIOFIQIYU SYVH OeL +8T o8e 'c00z Aine soqye/ut e8reyostp 'e8.reyostp asreyosip 104 Joye shep O§ UINIIM sywepnseooyue ye sXkep O€ UTIL syUepNseoo JO} WIRD OU TIM pure sIsoUseIp qUV jue JO} WTe[D OU 27 sasouseIp asIeYOsIp 19yye YWWOUT {NOY 'OYOIS OIUIEYOST 10} sisoude Quy 0/M ayx01}8 STUTET-OST 19472 VI aye} Od 7 T-O0 UIgyM 8 uo0Nndrrose1g 98000° "Ip quoryedul 4siy YM spenprarpuy WTO ssnip yefaye[diyue oyey, +87 982 'goog Afar aBIeYsIp 1oye yJWOUT zeyye ju eSreyosip 'VI, Io] stsouse sIsOUuseIp VI aye} Od ¢ 1-0 «Una = worydtgaserg 82000° -Ip quoryedul ysIy YA spenprarpuy VIL Joye WTO suryeys oye], +81 e8e 'coo Alor asreyosip 19yye YWWoUT zoyye/ut aBreypsip Sy], 10} stsouge sisouseIp VL, 1 1ye VI Mop oye} Od cs TO Uy uondrioserg 82000° -Ip quolyedul ysig YIM spenprarpuy WT seatsusyedAyyue oye], QdUDPIAD oreys uoruyed oreys woryuyed jo oe OP (penuryu0s) suTjepn+y y93u9011¢ "Ss od soueteypV 78 ASTY 33 'quotyed oy} SI sisAfeue Jo yun sy} 'e[qe} SIY} Ul pessi] soul[epms ay} SULA[OAUI SUOISSeIZe1 JO 'asiyJedxe 04 sss00"8 1NOYIIM eso} SUOUIe sdUaIoYpe Jo aBeloAe poySIem souspeaAsid-s[DAdafI[ pue SouT[spIns pue sede [[@ ssoloe Iva adUaleJol otf} ul aoueTeAsid outfepin oproeds-o8e jo osesaae ay} SMOYS MOI 4SeT OUT, *(S)BnIp ureysa0 Buryzeq ysureSe spusumMo0deL YF JI ,.oYe2} 1.UOp,, se pue '(s)Snzp ureyieo Supye, spuaurmi0ser PI FL ,AY2} OP,, Se SUT[epIns ev AjIssepD aA, 'osT}Jedxa 04 ssoooe JNOYIIM ASO} IO} aye1 SOUETAYpe 94} SISI] ,PoUaIOYpe,, JopuN ,oIVYs,, UUINTOO 'reek soUaIajaI OY} Ul JOS YSTI 8,SUT[SpINS yore url uoryerndod ystpamg ay} Jo oreys ay} 84ST] ,.198 FSI, JopuN ,ereys, UUINTOD "910% euNL eiojoq/ur aBreyosrp YM 'at '210% JAaquiesaq eTOjaq SyWOUI QT ueyy O1OU OSTEYOSIP YIM soseo JuoTyedur ysiy O} o[dures ayy JOIIJse1 OM WET-gl 'stsoubvIp UovjoLDfur joIpsvIORuUL uazfo surq0IgG SUTTEpINs oy) 10j 'ojdurexe 10}-ooueIeype orNseour OM yey} poried oulry ay} Jaao suordiiosaid Snip [Te eAresqo 0M YIM seseo JUST}edUI ISITE OF 198 YSLI OY} JOLIISOI SAA "YORIJS OUWIOYOST JUSISUBI} FW], 'UOIeT[IQY [eye :qyy :aZ0N7 gg' ¥80000° aBeIoAy wifey 4s1y +8T aBe '9002 ayy Joye skep OGT-09 IQA Arenuep Joyye/ul yuessoidapryue I10j yuouryeeI} VI Mop aye} og 7g =-s_s ae ~Jeyjoue sey yueryed TZ0° worndiioseid ysmgy YyIM sfenplArpur jyuessoidepryue WIM onuTyUOD +8] 98 'g[Qz Jequieseq ar0jeq/uI agreyo 'goog A[ne seqye/ut eS8zeyostp 'ern SISOUSVIP OINTey -sIp aye sqyuow 7Z-H UTqIWM -[rey greoy A0f stsouseip yuoryedyno = away seye wWpZ-Y srIeyoo[q VI MOP axe} Od 82 S8nip Y0q JO} UoMdysose1g Zv00' = Jo: JuaTyedur 4sIy YYM s[enplAIpul -_- BJoq, 29 SIOFIQIYU GYVU OAL +81 98% °¢00Z Ajne sJeye/ul e8reypsip 'omy oaoqge YSII ayorys asreyosip Jaye syyuoUL O1008 YSII oyOIS pue qUY Joy sisou = YSIY YIM sIsOUSeIP Quy 104 VI ayey oc zo")0OUC«CTIM = wodtzoserq 9Z00' -3eIp yUoTVedUr ISIQ ULM sTeNpLAIpuy == -Fe Wg) SJue_NSeoorywe oyey, +0g Be 'qo0z Aine asreyosip Joye syyuOU Joye /ur e8reyosIp 'ain4oesy IO} stsou SISOUSRIP OINJOeIZ JOWe WITT VI Mopeg ayey OG @ 8 6©St-0) SOM = uo osei1g ZZ00° -3RIP JUSTFeCUI ISI YIIA STeNpLAIpu] -- ssnip sisorodoeyso oye], +8T a8e 'gToz sung arojaq/ur 'gooz Aap seqye/ul oB8zeyostp 'uory eBreyOsIp Joaye syyuUOUL -OreJUL [BIqei9d IOJ sisouseIp uTeUl SISOUSVIP UOTOITeJUL Perqeie9 VI aye} Od wy ST-@T uTqyIM = wordtsose1q vI00 = Juatyedur «4ST YM senprArpuy = Jaye UIQT-ZT_ SUIyeIS ayYBY, 62-OF Be '910% eunr er0joq/ur 'eooz Ang s93ye/ut eSreyosip 'uory sis aBreyosIp Joaye syyuoU -oreyur ferpreookur 10J sisouseip -ouselp uorjoreyur perpreoodu VI axe} Od eo = ST-ZT UIyyM = uondrioserg VIOO' «=--Ureua yuaTyedur ys YIM STeNplArpuy =«-«JayFe - WIQT-ZI sulyeys eye], adUspIAS ereugs uoTyTayed ereus woRnmyed exe} 7,u0p oO ONUTYWOD J "SA OG eoualeypVy qos YSIYy (ponurio9) OUfToPIND qBuer¢ 34 'AoueusaId a} ST SIsATeue Jo JUN oT} 'a[qe} SIY4 Ul pa4sl] SeUl[apInS ay} SUIA[OAUI sUOIssalsal JOY 'osl}Jodxe 0} Sssa008 JNOT}IM asoy} SUOUIe soUsIEype JO a8eiaae paqyysiem souaeAoid-opAdejl, pure souljapins pue se¥e [Te sso1oe Ieak soUaIOJOI 914 Ul soUSTeASId oUT[apMS oploods-o8e Jo oSeIOAe a} SMOUS MOI 4Se] IIL], 'ostyradxo 0} Ss900e INOYIAL SSOY4 IO} o}eI SOUOISYPe OY} SIST] ,SUIZOYpe OTeYs,, UUIN[OS 'reeA SoUdIEJOI OY} UI JOS YSII S,ouTjepIns yore ut uorye_ndod ystpamg oy} Jo ereys OY} SST] «JS FSI Ul oreys,, UUINTOD '[NjUIIeY ysour 944 ATOYI] ere yeyy sSnap Burmydeo q YIM 'snyoy oY} 09 SULIET Sz yNoqe soUapIAS Jo YZUaTys 9Y} UO peseq PeUTULIoysp st (UUINOO ,@OUSPIAS JO YSuaTs,, oY} UT poyeorpur se) Snip e jo uOTwOyIsse[D apesZ 19949] oy], "SnIp urezsed Supe, ysureSe spusuus0oaI FI JI ,.OYe} 4, Uop,, se pue 'Snip urea. Buryey SPUSUIUIOIAI FI JI ,ax2} Op,, Se sulfepmns wv AjIsse[D oAA 'AOUVUSeId oY} SULINp Snip peyloods oy} SULARY JOU IO} 1OYEOIPUI UP SI a[qeITeA SUIOIINO OL], 'UOTJdaou0d alJojoq syyuOUI 7S UINVIM Snip ayy Jo uorjdiioseid @ sey JeyjoUI ey} YOM Joy seroueuseld [Te sureywod jes YsLI oy} 'oUT[EpINS yoese 1o7 'ssnip Aioyeuurepyul-yue [eploraysuou :GTWSN :970N 98° €6900000° oBeIoaAy Snip-D oye} 4,U0q 16° 2100° AoueuSoid 'sqTVSN 88819-D sxe} 1,00q snip-9 oye} 4,u0q GL, €100° ADueuseId 'ssn1Ip--_ Jayy4o eye} 4,u0q snIp-) axe} 7,u0q 89° TLO0" Anureuseid 'squessoidepiyue ssep-C sxe} 4,00q snIp-D oye} 7,00 ge 66000" Aoueusaid 'sptoido ssepa-7 aye} 4,00 snip-) axe} 7,u0q 8" ¢2,000° AoweuBeid 'sploi1ays0d1}109 ssed-D oye} 4,00q snip-9 aye} },u0q €6° TZ000° Soueuseid 'sorsjorxue ssepo--- sxe} 4,00q snip-D ayey 4,c0q 98° 29000° Aoweugaid '(1ay4{0) soorqryue ssep-9 aye) 4,U0q snip-q oye} 4,00 66° 29000" AQueuseid 'sourpoAoe1}0} ssepo-q, oye) 1,u0q snip-9 aye} 1,u0q 63° 96000" Aoweueid '(asn orure4sAs) Spl01a48091}.109 sse[9-E aye) 1,00 snip-q aye} 1,u0q 66° zZS000° AoueuSaid 'suasoyse8oid ssepo-q axe} 1,00q snip-- oye} 1,00q 1° 7000" AoueuSaid 'soatyepos pue sorjoudAy ssepo-5 oye) 4,00q 3nip-9 oye} 1,00 cg" €€000° Aoueusoid 'sououLoy o1urepeyjodAy ssepo--d oye 1,00q snip-q aye} 4,u0q $6 81000" Aoueusaid 'sSnip-q 1ay40 oye} 4,00q BnIp-9 aye} 4,U0q 92, 60000" Aoueuseid 'sywepnurysoypAsd ssep-O sxe} 4,00 Aoueu snip-q axed 1,U0q 86° Z60000° -Baid 'sururei30}da14s 10 sopluresooul 'seplpo1oeul ssepo-q, exe} 4,00q snip-d aye} },u0q ZL P0000" Anveuseid 'sywessorddnsounwui ssepo-q eye} 4,00q snip-q 942} 3,u0q 99° Z2£0000° Anueusaid 'sondoriderue ssepo-q ae) 1,u0q aDUaPLAV JO euyepmy 78u817g oye} },UOp "sA og BulLioype s1eys qos ASI Ul areysg AQueUBaLY Ul 9) UOTYBOIPaP[ BUIZOAOD sSOUT[SpINy :p s[qey, xIpueddy 35 Appendix Figure 1: Distribution of Adherence and Prevalence Rates Across Guidelines (A) Prevalence Rate Share of guidelines 85 65 High: 7.6% for the recommendation that adults should use a narrower vs. broader spectrum antibiotic for first 4 treatment of respiratory tract infection Low: 0.003% for the recommendation against anti-epileptics 2- during pregnancy 0 4 BS Ee T T T T T 0 2 4 6 8 Prevalence rate (percent) (B) Adherence Rate among Individuals Without Access to Expertise .08 - .06 - .04 - .02 - Share of guidelines High: 98.8% for the guidelines against taking certain antibiotics and hormones during pregnancy Ny Lifecycle-prevalence-weighted average: 54.4% Low: 20.4% for the recommendation for those 50+ to take osteoporosis medications in the 12 months after a fracture diagnosis T T * T T T 0 20 40 60 80 100 Adherence rate (percent), without access to expertise Note: Panel (A) shows the distribution of the prevalence rate (share of population in the risk set) across guidelines. Panel (B) shows the distribution of the adherence rate (share of those in the risk set adhering to the guideline) among those without access to expertise, across guidelines. 36 Appendix Figure 2: Average Number of Guidelines by Age and Gender Number of guidelines applicable + 2.25 . oe? * 2 ° 1.75 , o 1.5 o " ° 15 5 Women > aeeerroeeeeene, sensesssesresserrrerrereeeerres 25 % oooeee® ve OOK 800e, 4 DOOM OOOOOOOOOONONOOOO OOOH ol" wy Men 0 6 10 18 25 35 45 55 65 75 80 85 Age Note: The graph illustrates the number of guidelines applicable to the average female and male at each age from 0 through 85. To construct this number, we compute the prevalence of each guideline (share of individuals in the risk set) in a reference year (2016 or 2017) by age and gender. The plot shows the sum of prevalence across guidelines for each age and gender. Appendix Figure 3: Adherence Gap and Adherence Rate Across Guidelines Difference in adherence between those with and without access to expertise 3 Slope = -0.036 (s.e.) (0.024) 2 A 6 8 1 Adherence rate among those without access to expertise Note: This graph plots, for all 63 guidelines, the estimated difference in adherence between those with and without access to expertise against the share of those without access to expertise adhering to the respective guideline. The estimated difference in adherence is computed as in Figure 1. Spikes indicate 95% confidence intervals of the estimates, computed as in Figure 1. The rate of adherence among those without access to expertise is computed as in Appendix Tables 1-4. The red dashed line, the slope (and standard error) of which is noted in the top-right corner of the figure, is the line of best fit based on a bivariate regression that weights each guideline equally. The color code represents eight equally-sized bins of guideline prevalence, with darker colors representing higher prevalence. 37 Appendix Figure 4: Prevalence and Access to Expertise, by Guideline (A) Risk Sets Based on Diagnoses Had heart failure diagnosis, observe all prescriptions 0-24m post discharge Had heart failure diagnosis, observe all prescriptions 0-6m post discharge Had myocardial infarction diagnosis, observe all prescriptions 12-18m post discharge Had heart failure diagnosis, observe all prescriptions 12-18m post discharge Had heart failure diagnosis, elderly" Had fracture diagnosis, observe all prescriptions 0-12m post discharge Had cerebral infarction diagnosis, observe all prescriptions 12-18m post discharge Had heart failure & Afib diagnoses, elderly Average Had myocardial infarction diagnosis, observe all prescriptions 0-1m post discharge Had ischemic stroke & Afib diagnoses, observe all prescriptions 0-1m post discharge Had acute stroke diagnosis, observe all prescriptions 0-1m post discharge Had ischemic stroke diagnosis, observe all prescriptions 0-1m post discharge Had ischemic heart disease & Afib diagnoses, elderly Had dementia diagnosis, elderly* Had TIA & Afib diagnoses, observe all prescriptions 0-1m post discharge Had ischemic heart disease & heart failure diagnoses, elderly Had Afib diagnosis, observe all prescriptions 0-6m post discharge Had schizophrenia diagnosis, observe all prescriptions 0-6m post discharge Had ischemic stroke & Afib diagnoses, observe all prescriptions 12-18m post discharge Had ischemic stroke & no Afib diagnoses, observe all prescriptions 0-1m post discharge Had TIA diagnosis, observe all prescriptions 0-1m post discharge* Had Afib diagnosis, elderly* Had TIA & no Afib diagnoses, observe all prescriptions 0-1m post discharge Had schizophrenia diagnosis, observe all prescriptions 12-18m post discharge 38 -o repeecefosagpetiiisit | -.002 -.001 0 .001 Difference in prevalence (B) Risk Sets Based on Use of Medications Had antipsychotic drugs at age 74, elderly Women with D-class immunosuppresants < 24m before conception, pregnancy Women with D-class antipileptics < 24m before conception, pregnancy Women with C-class psychostimulants < 24m before conception, pregnancy Initiates antibiotic treatment against RTI, children Women with other D-class drugs < 24m before conception, pregnancy Vomen with D-class macrolides/lincosamides/streptogramins < 24m before conception, pregnancy Women with C-class hypothalamic hormones < 24m before conception, pregnancy Initiates antibiotic treatment of UTI, elderly* Women with D-class progestogens < 24m before conception, pregnancy Women with C-class antibiotics < 24m before conception, pregnancy Women with C-class anxiolytics < 24m before conception, pregnancy Had drugs in certain interactions at age 74, elderly Women with C-class antidepressants < 24m before conception, pregnancy Women with C-class opioids < 24m before conception, pregnancy Polypharmacy w.r.t. psychotropic drugs at age 74, elderly Women with D-class tetracyclines < 24m before conception, pregnancy Women with C-class hypnotics and sedatives < 24m before conception, pregnancy @- @-d-e - - 7s Uo eee Be Women with C-class corticosteroids (systemic) < 24m before conception, pregnancy le Women with C-class corticosteroids < 24m before conception, pregnancy le Uses anxiolytics, elderly |e Women with C-class NSAIDs < 24m before conception, pregnancy \° Women with other C-class drugs < 24m before conception, pregnancy e Average | e Polypharmacy w.r.t. drugs in the same therapeutic group at age 74, elderly | a Had certain risk drugs at age 74, elderly |- Initiates antibiotic treatment, children | eo Polypharmacy w.r.t. all drugs at age 74, elderly a Uses hypnotics and sedatives, elderly | - Uses sleep medication, elderly | rd] Had NSAIDs at age 74, elderly |. Initiates antibiotic treatment against UTI, women | Initiates antibiotic treatment, adults | eo Initiates antidepressant treatment rot | -.01 .01 Difference in prevalence Note: For each risk set, we run an OLS regression of an indicator variable for being in the risk set on an indicator for access to expertise and our baseline controls as in Figure 1. The sample is the Swedish population in the reference year. We plot the estimated coefficient on access to expertise together with the 95% confidence interval, computed as in Figure 1. The average difference in prevalence is the simple average of all the coefficients in each panel; we bootstrap the estimation with 50 replicates drawn at the patient level, and construct confidence intervals for the averages based on the bootstrap standard errors. The color code represents the mean prevalence among those without access to expertise, with a darker color representing higher prevalence. Unless otherwise noted, each risk set is associated with one guideline. #: the risk set is associated with two guidelines; »: the risk set is associated with three guidelines. 39