Investigating the know-do gap in antibiotics prescribing: Experimental evidence from India

Abstract

Antimicrobial resistance is largely driven by overuse of antibiotics, which is particularly common in low- and middle-income countries. We combine provider knowledge assessments and over 2000 anonymous standardized patient visits to providers in India to examine why they overprescribe antibiotics for pediatric diarrhea and figure out how to reduce overprescribing. Seventy percent of providers prescribed antibiotics without indication of bacterial infection. Knowledge gaps explain little: 62% of providers who knew antibiotics were inappropriate still prescribed them. Closing this “know-do gap” would reduce prescribing by 30 percentage points, versus only 6 points if all providers had perfect knowledge. Using randomized experiments, we revealed that the know-do gap stems from providers’ beliefs that patients want antibiotics, not from profit motives or lack of alternative treatments. Yet, a discrete choice experiment suggests patients do not prefer providers who give antibiotics. Our findings indicate that addressing provider misperceptions about patient preferences may be more effective than standard information-based interventions in reducing antibiotic overuse.

Health care providers in India overprescribe antibiotics because they think patients want them, not due to lack of knowledge.

INTRODUCTION

Antimicrobial resistance (AMR) poses one of the greatest threats to global public health in the 21st century (1). AMR is already a leading cause of death globally, with roughly 5 million attributable deaths each year (2). The disease burden from AMR is largely the result of resistance to antibiotics. Overuse of antibiotics for illnesses that are likely to be viral rather than bacterial, combined with incomplete dosing, fuels resistance and is particularly common in low- and middle-income countries (LMICs). For example, in India, the location of this study, more than half a billion antibiotic prescriptions are dispensed in the private sector alone each year, which translates into roughly one prescription per year for half of India’s population (3). Interventions to reduce overprescribing of antibiotics in LMICs are critical for combating global antibiotic resistance.

However, there is little evidence on why health care providers in LMICs overprescribe antibiotics and what interventions could be effective at curbing inappropriate antibiotic prescribing. In this study, we evaluate antibiotic prescribing for childhood diarrhea, the second leading cause of death for children under 5 years old globally (4). Several studies demonstrate that about 70% of diarrhea cases in India are prescribed antibiotics, even with no evidence of bacterial infection (5–9). Overprescribing of antibiotics for diarrhea has already created widespread resistance to common antibiotics that treat bacterial diarrhea pathogens (10). Although bacterial infections greatly benefit from antibiotics, most cases of child diarrhea are viral and do not benefit from antibiotics.

There are several potential reasons for why health care providers overprescribe antibiotics. The simplest explanation is that providers are not aware of the best practices for appropriate antibiotic prescribing or of the harms associated with antibiotic resistance (the knowledge gap). If this knowledge gap (know gap) is large, then provider training interventions could be effective at curbing antibiotic overprescribing as long as changing provider knowledge changes provider prescribing behavior. However, there is evidence that providers in LMICs often fail to provide appropriate care even when they have correct knowledge of what appropriate care is (5, 6, 8, 11–13). This know-do gap was first discovered by Das and Hammer (14) and there is now evidence that this phenomenon exists in many LMICs across wide range of illnesses [see (15) for summary of the evidence]. Therefore, an assessment of what interventions are likely to be effective at reducing overprescription of antibiotics requires estimates of both the size of the know gap (the share of providers without correct knowledge of appropriate care) and the size of the know-do gap (the share of providers who prescribe antibiotics inappropriately despite correct knowledge of appropriate care). These estimates will provide insight into how much closing the know gap through provider education will improve antibiotic stewardship and how much we should invest in closing the know-do gap with other types of interventions.

While the solution for closing the know gap is relatively straightforward (i.e., disseminating knowledge of best practices for antibiotic prescribing), the solutions for closing the know-do gap are not. There is little evidence on why providers fail to provide the standard of care despite correct knowledge. Banerjee et al. (6) suggest that a key driver of the know-do gap is low levels of patient trust in the provider, and that providers do less for patients than they know is appropriate to avoid the appearance of “upselling.” Wagner et al. (5) find that providers avoid correct care for diarrhea that they know to be appropriate if they think the patient wants something different. We contribute to this sparse body of literature by investigating the know gap and the know-do gap in the context of antibiotic prescribing for child diarrhea in India.

In this study, we combine data from randomized experiments with provider and patient surveys to study antibiotic prescribing. These data were also used by Wagner et al. (5) to study the prescription of oral rehydration salt (ORS). In particular, we use vignettes (structured interviews with providers based on hypothetical cases) that present a case of viral child diarrhea to measure knowledge of appropriate care among 2282 private providers across 253 towns in Karnataka and Bihar, India. We then measure provider behavior using anonymous standardized patient (SP) visits with the same providers presenting the exact same viral diarrhea case as shown in the vignettes. We show that 50% of providers said they would prescribe antibiotics (inappropriately) in the vignette (know gap of 50%). Among providers with correct knowledge (i.e., those who knew that antibiotics were inappropriate), 62% prescribed antibiotics to SPs who made anonymous visits pretending to seek care for a similar case of diarrhea as presented in the vignette (know-do gap of 62%). Moreover, correct knowledge was only loosely predictive of antibiotics prescribing in regression analysis. Combining estimates on the relationship between knowledge and practice and the share of providers without correct knowledge, we show that closing the know gap would only reduce antibiotics prescribing by 6 percentage points. In contrast, closing the know-do gap would reduce antibiotics prescribing by 30 percentage points.

But how do we close the know-do gap? To provide insight into this question, we use several randomized experiments to assess the key drivers of the know-do gap. First, providers might think patients want antibiotics, which could lead them to prescribe antibiotics despite knowing they are likely to be ineffective. To investigate the role of patient preferences in the know-do gap, we randomized some SPs to express a preference for antibiotics, some to express a preference for the correct treatment (ORS), and some to express no preference. Second, providers might prefer to sell antibiotics because they are more lucrative than alternative treatments for child diarrhea. To identify the role of financial incentives, we randomized half of the no preference SPs to inform the provider that they only needed a treatment recommendation because the spouse was in their home town with the child and they needed to tell the spouse what to purchase; this eliminates the provider’s financial incentive to sell antibiotics at the point of sale. Third, providers might prescribe antibiotics because they have a stockout of ORS (the correct treatment) but not antibiotics, which are more readily available. To estimate the causal effect of ORS availability on antibiotics prescribing, we randomly assigned half of the providers to receive a 6-week supply of ORS before the SP visits. Last, we used a discrete choice experiment (DCE) to understand how patients’ choice of providers is influenced by prescribing of different medicines.

We found that the know-do gap in antibiotics prescribing was most sensitive to patients’ expressed treatment preferences; expressing a preference for ORS reduced inappropriate antibiotics prescribing by 17 percentage points compared to expressing an antibiotics preference and reduced the know-do gap by 13 percentage points, an effect that was particularly strong among pharmacies (25 percentage point reduction). Although providers indicated in qualitative interviews that they often give antibiotics to increase patient satisfaction, the results from the DCE suggest that patients do not have a preference for providers who give antibiotics rather than alternative treatments. We also find that eliminating financial incentives at the point of sale and ensuring correct treatment was in stock had no effect on inappropriate antibiotics prescribing on average. Together, these results suggest that informing providers that patients want something other than antibiotics (ORS in this case) moves the needle the most in terms of reducing inappropriate antibiotics prescribing and closing the know-do gap.

RESULTS

Sample description

This study takes place in two Indian states: Karnataka and Bihar. We chose states that are very different from one another in socioeconomic status and diarrhea care seeking to ensure our results are representative of a broad population. Bihar is one of the poorest states in India and is mostly rural. In contrast, Karnataka has above average per capita income relative to other Indian states. Figure S1 shows a map of our study locations.

Table S1 shows the characteristics of the 2282 providers in our sample collected through a baseline provider survey. Ninety-two percent of providers were male with an average age of 44 years; providers had 18.5 years of experience on average. Providers saw an average of 24.7 patients per day and 6.3 diarrhea cases per week; 56% dispense medications directly to patients; 52% had ORS available and 49% had antibiotics available at baseline. Our sample includes four different types of providers; 20% were providers with an MBBS degree (similar to an MD degree in the United States); 37% were providers with a degree in traditional medicine including Ayurveda, yoga and naturopathy, Unani, Siddha, and homeopathy (AYUSH); 21% were rural medical practitioners (RMPs) who typically lack formal training but still practice medicine; and 22% were pharmacies. MBBS, AYUSH, and RMP providers generally own their own practice and run it as a business. Most only have one provider seeing patients. Pharmacies in our sample are stand-alone businesses that sell medicines. Providers working at pharmacies generally have no medical training, but often advise patients on treatment options for child diarrhea and are often the first place caretakers go to seek care. Many clinics have an attached pharmacy that is part of the same business and we interpreted these entities as clinics that sell medicine and not pharmacies.

The know gap and the know-do gap in antibiotics prescribing

Figure 1 shows that half of providers correctly said they would not prescribe antibiotics when presented with a vignette asking them how they would treat a case of uncomplicated diarrhea with no indications of bacterial infection. This suggests that there is a substantial know gap and room for improvement in knowledge. However, among providers who correctly said they would not prescribe antibiotics, 62% prescribed antibiotics to SPs who presented an identical case as the vignette, suggesting a large know-do gap. The know gap and know-do gap was large among all provider types, but both the know gap and know-do gap were largest among providers with the least amount of training (rural medical providers and pharmacy workers) (fig. S2).

Fig. 1. Know gap and know-do gap in antibiotics prescribing for child diarrhea.

Correct knowledge means that the provider correctly said that they would not provide antibiotics to the hypothetical patient in the vignette. The know-do gap assesses the share of providers who had a correct response in vignette that did not prescribe antibiotics to the SP. This excludes SPs that expressed an ORS preference. Source: Provider survey and SP survey.

Improving knowledge is unlikely to move the needle on antibiotics prescribing

Table 1 shows that correct knowledge was only loosely associated with appropriate antibiotics prescribing. Providers with correct knowledge were 11.7 percentage points (15%) less likely to prescribe antibiotics if they demonstrated correct knowledge in the vignette. The correlation between correct knowledge and practice was strongest among AYUSH providers (17.2 percentage points; P $<$ 0.01) and small and insignificant for all other provider types. Combining estimates from Table 2 with the share of providers without correct knowledge (50%) shows that if all providers had correct knowledge, antibiotics prescribing would reduce by 11.7 percentage points for half of providers. This would only reduce overall inappropriate antibiotics prescribing by 5.8 percentage points (Table 2). In contrast, if we eliminated the know-do gap (i.e., if all providers prescribe antibiotics in accordance with what they already know is best practice), the half of providers with correct knowledge would reduce antibiotics prescribing by 62 percentage, leading to an overall reduction in inappropriate antibiotics prescribing of 31 percentage points. Although the relationship between knowledge and practice that we estimate might not be causal, these results suggest that interventions that improve knowledge of appropriate antibiotics stewardship are unlikely to lead to important reductions in inappropriate antibiotics prescribing. Instead, we need to focus on interventions that address the know-do gap. This requires a better understanding of what drives the know-do gap.

Table 1. Correct knowledge and antibiotics prescribing.

Estimates are from linear probability models regressing whether the provider prescribed an antibiotic on whether they had a correct response in the vignette. Correct response in vignette means that the provider said that they would not provide antibiotics to the hypothetical patient. Controls include provider type, years of experience, provider age, patient volume, experience treating diarrhea patients, the competitiveness of the market, whether the provider dispensed/sold medicines directly to their patients, and the severity of the case presented. MBBS stands for Bachelor of Medicine, Bachelor of Surgery and is the medical degree awarded to undergraduates in Indian medical schools. RMP is an acronym for registered medical practitioner. AYUSH is an acronym for Ayurveda, yoga and naturopathy, Unani, Siddha, and homeopathy. *P $<$ 0.1, **P $<$ 0.05, and ***P $<$ 0.01. Source: Provider survey and SP survey.

	Prescribed or dispensed antibiotics
(1)	(2)	(3)	(4)	(5)
Variables	Pooled	MBBS	RMP	AYUSH	Pharmacy
Correct response in vignette	−0.117***	−0.0720	−0.0434	−0.172***	−0.0535
	(0.0196)	(0.0467)	(0.0329)	(0.0343)	(0.0415)
Observations	2282	469	465	841	507
Controls	Yes	Yes	Yes	Yes	Yes
Mean when incorrect response in vignette	0.788	0.644	0.911	0.750	0.797

Table 2. Effect of eliminating know gap and know-do gap.

Estimates for “eliminating know gap” are based on coefficients from Table 1 multiplied by the share of providers with incorrect knowledge (−0.117 × 0.499 for pooled row). Estimates for “eliminating know-do gap” combine the size of the know gap with the share of providers with correct knowledge (0.63 × 0.50 for pooled row). MBBS stands for Bachelor of Medicine, Bachelor of Surgery and is the medical degree awarded to undergraduates in Indian medical schools. RMP is an acronym for registered medical practitioner. AYUSH is an acronym for Ayurveda, yoga and naturopathy, Unani, Siddha, and homeopathy.

	Reduction in antibiotics prescribing
	Eliminating know gap	Eliminating know-do gap
Pooled	−0.058	−0.310
MBBS	−0.033	−0.306
RMP	−0.031	−0.261
AYUSH	−0.063	−0.328
Pharmacy	−0.031	−0.322

Perceived patient preferences for antibiotics is a key driver of the know-do gap

Table 3 shows that inappropriate antibiotics prescribing was sensitive to patients’ expressed preferences. When SPs expressed a preference for ORS instead of a preference for antibiotics, providers were 17.4 percentage points less likely prescribe antibiotics on average (22% reduction). Among providers with correct knowledge, expressing an ORS preference closed the know-do gap by 13 percentage points (20% reduction). Showing an ORS preference versus an antibiotics preference reduced antibiotics prescribing among all provider types, but the reduction was largest among pharmacies; 25 percentage point reduction in antibiotics prescribing on average and 34 percentage point reduction among providers who knew that antibiotics were inappropriate (table S2). The reduction in antibiotics prescribing was nearly twice as large for providers who dispense their own medicines (21.1 percentage point reduction versus 12.2; P = 0.012 for the difference in effect; table S3), the group we would expect to be most responsive to patient preferences because their business is directly affected by the medication they dispense. Providers who dispensed their own medicine also had a 23 percentage point reduction in the know-do gap when SPs expressed an ORS preference compared to an insignificant 0.015 percentage point reduction for providers who do not dispense medicines.

Table 3. Effect of patients’ expressed preferences on antibiotics prescribing and closing the know-do gap.

Estimates are from linear probability models regressing antibiotics prescribing on the randomly assigned preferences expressed by the SP. Column 2 includes only providers who said they would not prescribe antibiotics in the vignette. *P $<$ 0.1, **P $<$ 0.05, and ***P $<$ 0.01 Source: Provider survey and SP survey.

	(1)	(2)
Variables	Change in antibiotics prescribing	Change in know-do gap
ORS preference vs. antibiotics preference	−0.174***	−0.130***
	(0.028)	(0.041)
No preference vs. antibiotics preference	−0.073***	−0.051
	(0.023)	(0.038)
Observations	2282	1141
Controls	No	No
Mean in antibiotic preference arm	0.777	0.649

If SPs expressed no preference, providers were 7.3 percentage points less likely to prescribe antibiotics compared to expressing an antibiotics preference, and this closed the know-do gap by 5 percentage points (not significant). These results suggests that overprescribing of antibiotics is partly explained by providers perceptions that patients want antibiotics and that shifting the preferences the patient expresses to the provider could help reduce the know-do gap.

These results are consistent with providers thinking most patients want antibiotics even when they do not express a preference. About 70% of cases received an antibiotics prescription when they expressed no preference and showing an antibiotics preference had only modest effects on prescribing. However, when they express a preference for a different treatment, prescribing changes quite substantially, suggesting that their perception of the patient’s preferences were updated. This is also consistent with our qualitative work where providers reported that they thought patients would leave unsatisfied if they did not get a “strong medicine” such as antibiotics. One provider said: “[If we only prescribe ORS] they will change the doctor. They might say that we don’t have adequate knowledge. [Some providers] prescribe multivitamin drops or a tonic (antibiotic syrup) or something expensive so that they think he is prescribing good medicines and come back next time.”

The qualitative work suggests that providers think that patients have a preference for strong medicine and not antibiotics per se. However, providers think that antibiotics are strong medicine because they provide symptomatic relief for bacterial diarrhea, while ORS does not reduce diarrhea. This quote portrays the same provider perceptions about antibiotics prescribing as a quote presented by Das and Hammer (14) more than 15 years ago when they were examining quality of care in New Delhi India. When they asked a health care provider why ORS was not enough for treating childhood diarrhea the provider responded: “Of course, the WHO and others keep saying that we should only give ORS. But if I tell the mother that she should go home and only give the child water with salt and sugar, she will never come back to me; she will only go to the next doctor who will give her all the medicines and then she will think that he is better than me.” The remarkable similarity of these quotes taken 15 years apart in two separate regions of India highlights the persistence and ubiquity of these beliefs.

Patients say antibiotics prescriptions do not affect their choice of doctor

The results presented so far suggest that providers think that not prescribing antibiotics might mean that patients might take their business elsewhere. This naturally raises the question: Are provider perceptions accurate? Despite providers thinking that giving patients antibiotics will help with market share and improve patient satisfaction, we find that whether a patient gets antibiotics from the provider does not affect their choice to go visit that provider for care. In a DCE, patients were no more likely to choose (hypothetical) providers who gave antibiotics only compared to those who gave ORS only (Fig. 2). Instead, patients slightly preferred providers who give them more total medicines; ORS + zinc was preferred to antibiotics alone and ORS alone, and ORS + zinc + antibiotics was most preferred. However, other factors such as how kind the provider was to patients and the quality rating were far more important to patients than the specific medicines dispensed when choosing a provider. Overall, the DCE results suggest that the treatments the provider gives plays only a small role in doctor choice and that patients do not seem to prefer doctors who prescribe antibiotics compared to doctors who prescribe alternatives.

Fig. 2. Preferences for provider characteristics (DCE preference weights).

Estimates are based on a DCE. This figure plots the preference weights of each attribute level and represents the difference in the likelihood that the provider was chosen if they had the respective attribute level relative to the reference. The reference attribute level does not have a confidence interval. CI, confidence interval. Source: DCE.

No evidence that financial incentives at the point of sale and lack of ORS supply contributes to the know-do gap

Table 4 investigates the role of financial incentives at the point of sale and supply of correct treatment on antibiotics prescribing and the know-do gap. This table by-and-large shows that ensuring stock of correct treatment or eliminating the financial incentive to sell more lucrative treatments had no effect on antibiotics prescribing. When SPs told providers that they would not purchase treatment from them and only wanted a recommendation (thus eliminating the financial incentive to sell one treatment over another), this did not significantly change antibiotics prescribing or the know-do gap, and point estimates are small. Similarly when providers were given free supply of ORS, thus increasing the likelihood that they had it available when the SP visited, this had no effect on antibiotics prescribing.

Table 4. Effect of financial incentives and ORS supply on antibiotics prescribing and closing the know-do gap.

Estimates are from linear probability models regressing antibiotics prescribing on the randomly assigned role of the SP (incentive or no incentive) and the assignment of the provider to treatment or control in the ORS supply intervention. No financial incentive means that the SP told the provider that they would not purchase any medication and only wanted recommendation. Given ORS supply means that the providers received free ORS supply. Column 2 includes only providers who said that they would not prescribe antibiotics in the vignette. *P $<$ 0.1, **P $<$ 0.05, and ***P $<$ 0.01. Source: Provider survey and SP survey.

	(1)	(2)
Variables	Change in antibiotics prescribing	Change in know-do gap
No financial incentive	−0.020	0.019
	(0.026)	(0.037)
Given ORS supply	0.005	0.013
	(0.023)	(0.034)
Observations	2282	1141
Controls	No	No
Mean in financial incentive/no supply group	0.691	0.557

Role of other factors

Figure S3 explores several other factors that could be driving the know-do gap using cross-sectional nonexperimental variation (all factors included in same model to account for correlation across factors). This figure shows that by far the strongest predictor of the know-do gap is provider type, with the gap largest among RMPs and pharmacies. This could be due to the business structure of these entities: RMPs and pharmacies generate most of their income from selling medicines, or it could be due to different levels of knowledge or less allegiance to the Hippocratic oath. The know-do gap was slightly smaller in less competitive markets, which could be because providers in more competitive markets are more sensitive to patient preferences (even when they go against guidelines). Experience, age, patient volume, and asking key questions did not predict the know-do gap.

DISCUSSION

In this study, we provide evidence that improving knowledge of appropriate antibiotics stewardship among health care providers in India, which has been the focus of global attempts to reduce inappropriate antibiotic prescribing, is unlikely to move the needle on inappropriate antibiotic prescribing. This is because (i) a large portion of providers already have correct knowledge and (ii) providers with correct knowledge were only modestly more likely to practice appropriate antibiotic stewardship. Instead, working to eliminate the know-do gap and ensuring that providers prescribe in accordance with their knowledge of best practices have the potential to reduce inappropriate prescribing by up to 30 percentage points. Our randomized experiments suggest that the know-do gap is driven by providers’ perception that patients want antibiotics; the know-do gap reduced by 13 percentage points (20%) when SPs exhibited a preference for ORS rather than antibiotics. Financial incentives to sell antibiotics and supply of correct treatments appear to be much less important. When we measure patient preferences using a DCE, we find that patients’ choice of provider is not responsive to whether the provider prescribes antibiotics rather than ORS. This suggests that changing provider perceptions that patients want antibiotics has the potential to substantially reduce inappropriate antibiotics prescribing.

It is important to note that providers might not directly think patients have a preference for antibiotics per se, as most patients cannot distinguish antibiotics from other medicines. However, our provider interviews suggest that they might think that patients have a preference for “stronger treatments” that could potentially provide symptomatic relief. Thus, providers might prescribe antibiotics rather than ORS because ORS treats dehydration induced by diarrhea but not the diarrhea symptoms, whereas antibiotics are quite noticeably effective at stopping the diarrhea as well as other adverse health consequences, when used to treat bacterial infections. However, when antibiotics are administered for viral infections, they provide little or no direct benefit (and potentially harm) to the patient’s health but contribute to the broader issue of antimicrobial resistance. Consequently, a private health care provider focused on keeping patients happy by alleviating symptoms may opt to prescribe antibiotics in cases of diagnostic uncertainty regarding the bacterial or viral nature of an infection. This suggests that interventions to reduce inappropriate antibiotics prescribing such as changing providers perceptions that most patients want antibiotics, might not be effective unless providers also have a way to identify which cases are in need of antibiotics. New ways of quickly and accurately identifying whether a child’s diarrhea is caused by a bacterial or viral pathogen could help providers target antibiotics only to children who will benefit.

The know-do gap was largest among the providers with the least training: pharmacists and RMPs. This suggests that interventions to address the know-do gap in antibiotics prescribing could achieve the most potential benefit from targeting these provider types. Moreover, pharmacists were most responsive to patient preferences; SPs who expressed an ORS preferences to a pharmacist were 34 percentage points less likely to receive an inappropriate antibiotics prescription compared to SPs who expressed an antibiotics preference. Roughly one-third of diarrhea cases in our study locations seek care from pharmacies; thus, interventions focused on pharmacies could be particularly impactful.

Our findings contribute to the scarce literature on what drives inappropriate antibiotics prescribing in LMICs in several ways. First, while several studies document the know gap and the know-do gap in prescribing practices (5, 6, 8, 11–14), ours is the first to estimate the relative importance of addressing the know-do gap versus addressing the know gap for reducing inappropriate antibiotics prescribing for diarrhea. Using data from a large sample of providers from two diverse states in India, we document that addressing the know-do gap has the potential to substantially reduce inappropriate antibiotic prescribing whereas addressing the know gap is unlikely to move the needle. Second, this study simultaneously measures (i) provider knowledge, (ii) provider response to patient preferences, (iii) provider response to financial incentives, (iv) provider response to stockouts of correct treatment, and (v) patient preferences for provider prescribing using randomized experiments across both providers and patients. Using these rich data, we show that the know-do gap is driven by provider perceptions that patients want antibiotics and that the know-do gap is not responsive to financial incentives at the point of sale or availability of correct treatments.

Our study is closely related to a study conducted in one large city in China in 2011-2012 that also evaluated the role of provider perception of patient preferences and financial incentives in antibiotic prescribing (16). That study also found that providers were more likely to prescribe antibiotics when patients expressed a preference for antibiotics compared to no preference. In contrast to the findings from our study, that study found that financial incentives also significantly influenced antibiotic prescribing, perhaps because providers received explicit bonuses tied to the level of antibiotic prescribing. However, the study did not measure provider knowledge or real patient preferences for providers who prescribe antibiotics. Moreover, the study also did not assess how availability of correct treatments or expressing a preference for correct treatments influenced antibiotic prescribing, which is a key contribution of our study. Another study from Kenya had mixed findings on the role of patient preferences in inappropriate prescribing; expressing a preference for (inappropriate) deworming medication increased prescribing but expressing a preference for (inappropriate) antibiotics did not increase antibiotic prescribing (17). This study did not examine how preferences expressed for correct treatments influenced inappropriate prescribing nor did it include measures of provider knowledge. Thus, the implications of this study for reducing antibiotics prescribing are unclear. Another study focused on the role of financial incentives in antibiotic prescribing in South Africa (18). Unlike the setting of our study and most LIMCs settings, the providers in this study received a fixed fee for their consultation and any medicines dispensed. Therefore, the providers made less profit if they dispensed medicines as the cost of medicines was borne by the providers. They found that changing the financial incentive by allowing providers to prescribe rather than dispense did not influence antibiotics prescribing. These results suggest that our finding that financial incentives do not influence antibiotic prescribing might generalize to other contexts. Last, another strand of related studies estimate the effect of patient knowledge on prescribing behavior. They find that patients who signal to providers that they know that antibiotics are inappropriate are less likely to receive antibiotics (19, 20). This is consistent with our finding that patients who express a preference for ORS (correct treatment) and thus signal knowledge of correct treatment are less likely to receive antibiotics. In summary, our study is unique in its comprehensiveness in that we compare a wide range of different factors among a large and diverse population of providers using several different data sources and randomized experiments. Our results thus provide insights into which types of interventions are likely to be most effective at curbing inappropriate antibiotics prescribing and closing the know-do gap.

This study has several limitations. First, we could not have a child present during SP visits due to ethical issues, which happens for about 12% of diarrhea care visits in India. Although the care received by SPs is similar to the care reported in caretaker surveys (5), it is possible that the treatment effects we estimate would be different if a child were present. Second, we used cross-sectional variation in knowledge, and the relationship between knowledge and practice could be confounded by factors for which we were not able to control. Third, although we investigated several key drivers of the know-do gap, there are several other drivers that we were not able to assess, such as uncertainly of about the root cause of infection. While antibiotics have little or no therapeutic value for viral infections, they can be life-saving for bacterial infections. Identifying whether a child’s diarrhea is caused by viral or bacterial pathogens is challenging in low-resource settings with limited access to laboratory diagnostics. The Indian Academy of Pediatrics guidelines state that providers should only prescribe antibiotics for acute bloody diarrhea, but many bacterial infections do not have blood in the stool (21). Without a diagnostic tool to inform primary care treatment decisions, providers might choose to err on the side of precaution and prescribe antibiotics even when they know that a substantial fraction of infections are viral. Future work should assess the value of diagnostics in reducing inappropriate antibiotics prescribing in LMICs.

Antibiotic resistance poses one of the biggest public health threats in the modern era. This study suggests that interventions to reduce inappropriate antibiotics prescribing should focus on reducing the know-do gap rather than increasing knowledge of antibiotic stewardship. Future work should design and test interventions to change provider perceptions that patients want antibiotics to reduce the know-do gap as such efforts have the potential to reduce inappropriate antibiotics prescribing and subsequent antibiotic resistance.

MATERIALS AND METHODS

Sampling strategy

Our sampling strategy was designed to produce a representative sample of private providers that treat child diarrhea, which consisted of three steps: (i) sampling districts, starting with those with the highest diarrhea prevalence; (ii) sampling all towns within a district with a population larger than 10,000 and smaller than 150,000 according to the latest available Census (2011); and (iii) recruiting all private providers who treat child diarrhea in the town. We chose not to visit smaller towns because there would be too few providers to enroll, and we chose to avoid larger towns because it would be too challenging to draw a representative sample of providers. To recruit providers, we first conducted a census of all clinics in the town. We used this census to enroll and survey all providers that were available for interview and provided consent to participate in the study. We were able to contact 59% of those identified in the census and of these 69% consented to participate in the study. We also identified all pharmacies in the town and then randomly sampled an average of two pharmacies per town that were independent establishments and not attached to a clinic.

Randomized interventions

Standard patient visits

The preferences and financial incentives interventions were both done through variation in the way SPs presented a case of child diarrhea when visiting the provider. All SP visits presented a case of a 2-year-old child who had been having diarrhea for 2 days. Half the SPs presented a moderate case and the other half a severe case, but both cases were severe enough to require ORS (our vignettes included one severe and one mild case, but we only included the vignette case that matched the SP case for each provider). The case was designed such that antibiotics would not be appropriate (no blood in stool, feces quality not sticky/smelly, and short duration). The moderate case had four to five loose stools the previous night and the child was taking fluids. The severe case had 10 to 12 loose stools the previous night, the child was not taking fluids or food, and showed symptoms of dehydration (low energy and sunken eyes). The only difference between the SP roles were (i) the opening statement where they expressed a preference (four different statements described more below) and (ii) the severity of the diarrhea episode (moderate or severe). Each SP played all eight possible SP roles to help control for individual SP effects.

We recruited 40 actors in each state to go through an extensive 2 week SP training. The training included memorizing a script and responses to common questions (e.g., “when was the last time the child had diarrhea?,” “how many times did the child pass stools in the last day?,” or “what did the child eat the day before the diarrhea episode began?”). After the opening statement, the SPs were instructed to let the provider lead the encounter and only provide relevant information about the child if the provider asked. Repeated practice ensured that the way the case was presented was as similar as possible across providers and across the different SP roles. At the end of the training, we selected the top 25 actors to make the SP visits.

Providers consented to receive a visit from a SP upon study enrollment. All providers enrolled in the study received one visit from an SP posing as a father of a child who had a case of diarrhea (2282 visits in total). All SPs visited the providers without a child. This pattern of health care, in which a father seeks care on behalf of the sick child, is common in India and enabled the use of SP methods without putting a child at risk. All SPs paid any required doctor fees and attempted to fill prescriptions from a nearby pharmacy to complete the visit. Less than 1% of providers suspected that they had received a visit from an SP during follow-up interviews.

Preferences intervention

During the SP’s opening statement when they first encounter the provider, some SPs expressed a preference for ORS, some expressed a preference for antibiotics, and some expressed no preference. All opening statements are in table S4. SPs who expressed a medicine preference showed the doctor a picture of used packaging of the medicine on their phone, indicated that they had used this medicine the previous time their child had diarrhea, and asked if they could use it again. SPs who expressed no preference just told the doctor their child had diarrhea and asked for a recommendation. Comparing the preference arms to the no preference arm allows us to investigate whether providers’ perceptions of patient preferences are driving overprescription of antibiotics.

Financial incentives intervention

To investigate the effect of financial incentives on prescribing, we randomly assigned half of the SPs who expressed no preference to inform the provider that they only wanted a treatment recommendation (i.e., they will not purchase any treatment) because their sick child was back in their hometown. They were only visiting the current town for work and they would have a relative deliver the treatment to the wife who was home with the child. This reduces the influence of financial incentives in the provider’s decision to prescribe one treatment over another. At pharmacies, the SP first purchased some medicine for an unrelated illness (a headache) before asking for a treatment recommendation for their child. Comparing outcomes between SPs who purchased from the provider and SPs who purchased elsewhere allows us to estimate the effect of financial incentives on antibiotic prescribing. It is important to note that some financial incentives could still remain even when the SP purchases elsewhere, such as incentives associated with repeat business or gifts from pharmaceutical companies. Thus, we did not expect our intervention to completely eliminate all financial incentives, just those associated with point-of-sale revenue.

ORS supply intervention

To investigate the effect of ORS stockouts, we layered on a cluster-randomized experiment with the SP experiments described above. In particular, we randomly assigned all providers in half of the towns to receive 60 packets of ORS (about 6 weeks supply). We distributed ORS to providers upon enrollment, which took place 2 to 3 weeks before the SP visits. This gap between ORS dispensation and SP visits ensured that SP visits were conducted before the supply we gave was expected to run out. We told providers that the ORS was meant for their patients but they could dispense as they see fit. This creates an exogenous increase in the share of providers that have ORS stocked in the towns assigned to the intervention. Comparing outcomes from providers who received increased supply to providers that did not allows us to assess the effect of supply and stockouts on antibiotic dispensing.

Validity of SP design

There is a growing body of research that uses SP designs to measure quality of care and provider behavior (8, 22–26). Although this method is now widely accepted as a rigorous and accurate method for measuring quality, there are still valid concerns that SPs might receive different care than would real patients. To address this, we built in several features in our study to test the validity of our SP design. First, we returned to providers on the same day as the SP visit to assess whether they suspected that any of their patients were one of our SPs and only 1% suspected anything (which may or may not have changed behavior of this 1%). Second, we compared the care received by our SPs to the care reported in caretaker surveys. We found very similar prescribing patterns. Third, we find in household surveys that 12.5% of caretakers sought care without bringing the child to the provider visit, showing that our SP scenario is sufficiently common. We also randomized whether the child was present in vignettes, which are stylized cases read out to providers to elicit information about how they might treat this hypothetical patient. The presence of a child in a vignette did not affect the care the provider stated they would provide (results not shown). Overall, we find little evidence to suggest that providers gave different care to our SPs than they would to the general population.

Randomization

All providers enrolled in the study received one visit from an SP (2282 visits in total), and we randomly assigned which of the eight different SP profiles the provider received, stratifying by town. We also stratified opening statement randomization on case severity to ensure balance on severity. We randomized the ORS supply intervention at the town level, stratified by population size, and all providers in towns assigned to the treatment group received ORS supply.

Data collection and outcomes

We collected three different types of data in this study. First, we conducted provider surveys at baseline. We then returned 3 weeks later to conduct a SP visit with all enrolled providers and recorded what happened during the visit. Last, we conducted follow-up visits with providers soon after the SP visit (usually within hours).

Baseline provider survey

The baseline provider survey was conducted when providers were recruited and enrolled into the study. This survey provides information on provider characteristics and provider knowledge of diarrhea treatment using vignettes. The vignettes presented the same exact symptoms the SP presented to the provider.

SP survey

SPs were trained extensively on what to look for during each visit and how to record data about the visit. Within an hour of making the visit, SPs filled out a detailed form administered by a supervisor that documented several aspects of their interaction with the provider, including the treatment(s) the provider prescribed/dispensed.

Outcomes

Our primary outcome is whether the provider prescribed or dispensed antibiotics. We coded antibiotics as being prescribed or dispensed if the SP was given antibiotics from the provider or if the provider recommended the SP retrieve antibiotics from the pharmacy. In cases where there was uncertainty about whether antibiotics were prescribed/dispensed or not, we had the study team confer with a pharmacist who we had on retainer to help identify and categorize different treatments.

Empirical framework

To estimate the relationship between correct knowledge and inappropriate antibiotics prescribing, we regressed an indicator for whether they prescribed antibiotics to the SP on an indicator for on whether the provider responded correctly in the vignette (i.e., they would not prescribe antibiotics to the case presented). We controlled for provider type, years of experience, provider age, patient volume, experience treating diarrhea patients, the competitiveness of the market, whether the provider dispensed/sold medicines directly to their patients, and the severity of the case presented.

To estimate the effect of our experimental interventions, we used linear probability models with SEs clustered by town (the level at which the ORS supply intervention was administered) (27). To estimate the effect of preferences on a providers’ decision to prescribe/dispense antibiotics, we compare average treatment outcomes between SPs who exhibited no preference or an ORS preference to outcomes for SPs who expressed an antibiotics preference. We excluded SPs who informed the provider they would not purchase anything when estimating preference effects. To estimate the effect of financial incentives, we compared treatment outcomes between SPs who informed the provider they would purchase elsewhere and those who did not.

Discrete choice experiment

To assess real patient preferences, we conducted a DCE with 1189 caretakers who recently visited a provider for their child’s diarrhea. Each respondent was shown profiles of two different providers, which portrayed the provider’s fee (100, 300, or 500), rating of technical quality (5 stars, 3 stars, or 1 star), the time to arrive at the facility (10, 30, or 60 min), degree to which they are kind to the patient (very kind and friendly or rude and talks down to the patient), the name (Hindu name, Dalit name, or Muslim name), and the specific medicines the provider gives for child diarrhea (ORS only, ORS + zinc, antibiotics only, ORS + zinc + antibiotics, or injection). Because including too many attributes in a DCE profile makes it hard for respondents to fully consider each attribute, we had three versions of the DCE, each including only four attributes. Kindness, quality rating, and name were always included, and we randomly varied whether medicines prescribed, doctor fee, or time to clinic were also included. Medicines dispensed is the most important attribute for the purposes of this paper as this assesses when providers who give antibiotics are more likely to be chosen by patients. Each choice set had two providers with a different combination of attribute levels. Respondents were told to assume that all providers have a 30-min wait, are in an air-conditioned facility, and are in a central location, and to assume that all other characteristics between providers are the same. Each respondent was presented with a total of eight choice sets. We estimated the preference weights for Fig. 2 using a linear probability model so that each coefficient represents the difference between the probability of choosing a provider with a given attribute level compared to the probability of choosing a provider with the reference attribute level (28). We included choice set fixed effects to account for any imbalance in comparison profile attribute levels and we clustered SEs at the level of the respondent.

Supplementary Materials

This PDF file includes:

Figs. S1 to S3

Tables S1 to S4