 |
|
 |
|
 |
 |
|
 |
 |
 |
 |
 |
 |
 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
* Department of Health Policy and Management, Rollins School of Public Health, Emory University, Atlanta, Georgia
Department of Epidemiology, Rollins School of Public Health, Emory University, Atlanta, Georgia
 |
ABSTRACT |
|---|
 TOP ABSTRACT METHODSRESULTSDISCUSSIONREFERENCES |
|---|
Methods. A total of 3677 episodes of care from 2328 patients who were <17 years of age and had respiratory or ear infections that were treated with an antibiotic initially. The sample was drawn from the Medical Expenditure Panel Surveys for the years 1996 through 1999. Treatment failure was defined as the receipt of a second antibiotic, different from the first, in a 4-week window. We compared follow-up costs by outcome (treatment failure vs success) using a 2-part model of medical costs. We also performed a paired analysis by selecting 2 episodes, one in which the outcome was failure and the other in which the outcome was success, for patients with at least 1 of each type and computing the difference in costs.
Results. Follow-up costs for provider visits for episodes for which the patient experienced treatment failure were $216 versus $53 for episodes for which the patient did not experience treatment failure. Follow-up drug costs, including the cost of the second antibiotic, were $75 for children who experienced treatment failure versus $23, respectively. Cost estimates from the paired analysis were similar, confirming that results are not biased by unobserved time-invariant patient characteristics.
Conclusion. Children who have respiratory infections and experience treatment failure incur substantially higher costs.
Key Words: antibiotic use • cost analysis • otitis media
Abbreviations: MEPS, Medical Expenditure Panel Survey • URI, upper respiratory infection
There is growing interest in the costs associated with treatment of infections, attributable in part to increasing levels of drug resistance and the high prices of some recently introduced antibiotics. Most existing studies measure costs in inpatient settings, where laboratory findings are readily available and investigators are able to examine and compare costs among patients with specific pathogens and strains. Few studies have examined the costs of treatment in the outpatient setting,1 even though the majority of antibiotics are dispensed to outpatients and visits for respiratory infection account for a large portion of ambulatory visits overall.2,3 This study was undertaken to measure the medical costs associated with the initial and follow-up care of pediatric patients in the community with otitis media and respiratory infections.
Previous studies have combined data from the National Ambulatory Medical Care Survey and the Hospital Discharge Survey with estimates of antibiotic and visit costs,4 used Medicaid claims data,5 and collected data from patient surveys and chart reviews.6,7 Our patient sample adds perspective, as it is drawn from a large, nationally representative survey of individuals and their medical costs and utilization. This database permits analysis of a large number of cases and includes both insurance reimbursements and patients' out-of-pocket costs.
We focus on cost differences by treatment outcome. Lacking patient charts, we cannot measure outcomes directly by the persistence of symptoms. Rather, we consider a patient in our sample to have experienced treatment failure when, after receipt of an initial antibiotic, they receive a second antibiotic, different from the first, associated with the same episode in a 28-day window. Previously, studies comparing the effectiveness of antibiotics in the outpatient setting have used a similar approach toward defining treatment failure.5,6,8,9 Outcome-specific costs estimates may be useful to investigators who wish to compare costs for 2 or more treatment regimens of varying efficacy. Some antibiotics used in outpatient settings cost 4 to 10 times more than amoxicillin, and recommendations for their use as empiric therapy in areas where resistant strains are prevalent should be examined in light of their potential to reduce follow-up costs.
|
METHODS |
|---|
|
TOPABSTRACTMETHODSRESULTSDISCUSSIONREFERENCES |
|---|
MEPS data on utilization are self-reported. Comparisons of MEPS encounter totals with encounter totals from other national surveys do not indicate an underreporting problem.12 Reimbursement amounts and diagnosis codes for these self-reported encounters are collected directly from providers. MEPS includes encounters that did not generate insurance claims, an important feature because many antibiotics cost less than the standard $10 prescription copayment. MEPS groups events (prescriptions, office visits, outpatient visits, etc) into episodes of care based on self-reported (or, in this case, parent-reported) medical conditions.
Because coding respiratory infections is an inexact science and we would like to have as large a sample as possible, we used an inclusive criterion for selecting our sample on the basis of diagnosis. From the MEPS data, we extracted records for individuals with at least 1 health care event with a clinical classification code for otitis media (092), pneumonia (122), acute bronchitis (125), other respiratory infections (126), other lower respiratory disease (133), or other upper respiratory disease (134). Clinical classification codes are groups of similar International Classification of Diseases, Ninth Revision, Clinical Modification codes and are used in the MEPS to make it easier for researchers to identify patients with specific diseases.13 We used conditions files and the event files to construct episodes of care. The day of the first event associated with each episode was assumed to be the start of the episode.
Table 1 provides details on the construction of the sample. We excluded from the sample episodes for which the patient did not receive initial antibiotic therapy, which we identified by the lack of receipt of an antibiotic on the first day of the first event of the episode; episodes for which the first encounter was an inpatient hospitalization; episodes with missing dates; and episodes for which we could not observe a complete 28-day follow-up period. After making these exclusions, we were left with a sample of 3677 episodes.
|
Variable Construction
For each episode, we calculated separately the costs incurred on the first day (first-day costs) and the costs incurred after the first day (follow-up costs). We included only events and drugs that were linked to the same episode as the initial event. Costs for drugs prescribed on the first day were counted as first-day costs even when the prescriptions were filled the next day. We separated the costs for each time period into provider visit costs (including professional and facility components) and drug costs. All cost figures were inflated to 2001 dollars using the consumer price index for urban consumers. MEPS does not record side payments in the form of bonuses paid to providers or "chargebacks" and other postpurchase discounts received by pharmacies. MEPS data contain imputed per-visit reimbursements for visits covered under capitated contracts.
The main independent variable of interest is treatment failure. Following Bjerrum et al,8 we considered patients to have experienced treatment failure when they received a second prescription for an antibiotic within 4 weeks after their initial visit and the second antibiotic was different from the one that they first received. In addition, we required that the second antibiotic be associated with the same episode as the first. (Some patients received within 28 days another antibiotic that was not linked by episode to the initial event, and we did not consider these to have experienced treatment failure.) Piccirillo et al9 also used a 4-week window to measure treatment failure and relapse, but they included patients in the treatment failure group even when the second antibiotic was the same as the first. We reestimated our cost models using Piccirillo et al's definition and present results for comparison. Using Bjerrum et al's definition instead of Piccirillo et al's, which is more conservative, leads to estimates of cost differences by outcome that are biased toward 0, because the follow-up costs of those patients who actually experienced treatment failure and received a second prescription for the same antibiotic will be counted toward the costs of the nontreatment failure group.
We included as independent variables measures of patient demographics (age, race, and gender) and health status (diagnosis and parent-rated child health, measured by parent's agreement with the statement "less healthy than kids the same age"). We included the following measures of patient socioeconomic status: poverty status (parents' combined income <200% of the federal poverty level), parents' education, and insurance status (private, public, and uninsured). We also included as independent variables an indicator for whether the episode was the first to appear in the MEPS data for the patient (but not necessarily the patient's first episode, which may have occurred before enrollment in the MEPS), urban versus rural residence, and the site of the initial visit (physician office, outpatient hospital, emergency department, or prescription without a corresponding provider visit).
Statistical Analysis
We used Pearson 
2 test for discrete variables and 2-sided t tests for continuous variables to detect differences in patient and episode characteristics by outcome.
We used a 1-part regression model for initial costs and a 2-part regression model to analyze the impact of treatment outcome on follow-up costs.15 The 2-part model is the standard statistical framework in empirical health economics for measuring the impact of an individual characteristic on medical costs. It is designed to account for the unique distribution of medical spending found in most samples; a sizable minority of individuals do not use medical care, many use small amounts, and a few individuals incur substantial medical bills that account for a large percentage of aggregate spending. Because medical spending data are highly skewed, standard statistical methods that assume that the dependent variable is normally distributed yield inaccurate predictions of costs. The 2-part model attempts to mimic more accurately the empirical distribution of medical spending by splitting the distribution into 2 parts and allowing the impact of independent variables, such as health literacy, on the probability of using medical care to be independent of their impact on the costs of medical care for those who use it.
The first stage of the model measures the probability of using medical care as a function of individual characteristics. Probit (or probability unit) regression was used in this case. Independent variables included an indicator for treatment failure and controls for patient demographics, health status, and socioeconomic status. The second stage of the 2-part model and the only stage of the 1-part model estimates the relationship between independent variables and costs among those who use medical care. (Note that all patients incurred first-day drug costs, by construction, and almost all incurred first-day visit costs, so the first stage of the 2-part model is unnecessary for first-day costs.) Parameters are estimated via least-squares regression, in which the dependent variable is the logarithm of costs and the independent variables are the same as in the first stage but the sample includes only individuals who incurred costs for the relevant cost category.
Coefficient estimates for the 2-part model are difficult to interpret in isolation, because the dependent variable of the second stage is in log, rather than constant, dollars, and it is customary to state results in terms of predicted spending levels. These are constructed by computing predicted probabilities from the first stage and then multiplying these predicted probabilities by the exponentiated second-stage predicted values and a "smearing factor,"16 which is needed to transform logged dollars back to constant dollars, and averaging the predicted spending levels over the entire sample.
Two values are used to summarize the effect of a treatment failure on spending. The first is the average predicted spending level with the variable indicating treatment failure set equal to 1 for every respondent; the second is the average predicted spending level with the treatment failure variable set equal to 0. In each case, sample weights were used so that predictions are nationally representative. Computing predicted values in this manner nets out the impact of observable individual characteristics, such as age, on spending. P values for differences in predicted values were computed via simulation.
We recalculated costs by outcome using a paired comparison of 1 treatment failure episode and 1 treatment success episode, selected randomly, for each child with at least 1 of each type of episode. There were 102 children in this category, for a total of 204 episodes. We subtracted the costs of the treatment success episode from the costs of the treatment failure episode for each cost category and averaged the differences across all 102 children. By having each child effectively serve as his or her own control, we net out both observed and unobserved differences in time-invariant (constant) patient characteristics. We used a Wilcoxon signed-rank test to assess the null hypothesis of no difference in treatment costs between the paired episodes.
|
RESULTS |
|---|
|
TOPABSTRACTMETHODSRESULTSDISCUSSIONREFERENCES |
|---|
|
|
|
Table 5 shows the predicted values from the 2-part model, which, unlike the costs reported in Table 4, are adjusted for the significant differences in observable patient characteristics shown in Tables 2 and 3. (The set of coefficient estimates from the 2-part model are in an appendix, available from the authors on request.) Treatment failure episodes are associated with higher first-day drug costs (P = .02), echoing the result from Table 3 that treatment failure episodes are less likely to begin with a prescription for the relatively inexpensive antibiotic amoxicillin. The small difference in first-day visit costs is insignificantly different from 0 (P = .60). There are sizable differences in follow-up costs by treatment outcome. The average predicted cost for follow-up drugs in episodes that result in failure is $75 versus $23 for treatment success episodes (P < .01). The average cost for follow-up visits in treatment failure episodes is $216 versus $53 for treatment success episodes (P < .01). Including costs for both follow-up drugs and visits, treatment failure episodes cost $215 more than treatment success episodes.
|
The finding that initial drug costs are higher for episodes that result in treatment failure (Table 5) suggests that patients who experience treatment failure have medical histories or symptoms, not included in the regressions, that lead physicians to treat them differently on presentation. If so, then differences in follow-up costs calculated from the 2-part model may be biased by unobserved patient or episode characteristics. This observation motivates the analyses presented in the last 4 rows of Table 5. This set of results is calculated from the paired comparison, which effectively nets out differences in patient characteristics that are unchanging over time (eg, race). Encouragingly, neither first-day drug nor visit costs differ significantly by outcome. The total difference in follow-up costs is $288. Differences in follow-up costs remained significant when the analysis was restricted to the 66 episode pairs for which the treatment failure episode preceded the treatment success episode or the 36 episode pairs for which the order was reversed.
To examine the sensitivity of results to the construction of the study sample, we reestimated cost models restricting the analysis to episodes for which the initial prescription was amoxicillin and the first-recorded episode per child. In each case, results were qualitatively similar to those presented in Table 5.
|
DISCUSSION |
|---|
|
TOPABSTRACTMETHODSRESULTSDISCUSSIONREFERENCES |
|---|
Because this study used secondary data, we were unable to assess the causes of treatment failure, of which resistance is but one. A patient who receives an antibiotic that lacks inherent activity against the infective pathogen will require a second prescription, as will a patient who is allergic to the first. The variable that we construct to measure treatment failure does not capture all patients whose symptoms fail to resolve. Some will not seek follow-up care, and others may receive a second prescription for the initial antibiotic at a higher dose. Another caveat is that this study does not measure indirect costs (eg, the cost of parents' lost wages, travel costs, patients' and parents' time costs), which compose a large share of total costs.1,19,20
Our results may be biased by unobserved differences in patient or episode characteristics. We attempted to net out the impact of these unobserved characteristics by conducting the paired sample analysis and reestimating the regression models with a sample restricted to episodes for which the initial drug was amoxicillin. Overall, differences in follow-up costs were in close agreement across the different analyses and episode samples.
Because we used clinical classification codes to select our sample, we were unable to determine the microbiologic origin of patients' conditions. We do not believe that the failure to exclude patients with viral or other illness imparts bias to episode-level cost differences, because these patients are being treated and incurring costs as though their infections were bacterial. Furthermore, when we estimated the cost models only on patients who had a diagnosis of otitis media, from which the microorganisms primarily recovered are bacteria, results were essentially unchanged. Inclusion of patients with viral illness does affect population-level estimates. For example, had we excluded patients with viral illness, the observed rate of treatment failure in the sample probably would have been <7%.
Episodes for which treatment failure was the outcome were more likely to have otitis media as the primary diagnosis (P < .01) and less likely to have URI as the primary diagnosis (P = .02), consistent with evidence that most URIs are self-limiting. Comparing our cost estimates with previous reports in the literature is difficult. Studies use different follow-up periods and are based on different patient populations. Increases in medical prices and changing practice patterns further complicate efforts to compare costs from studies conducted at different points in time.
The finding that initial drug costs were higher for episodes for which treatment failure was the outcome contradicts the usual belief that expensive drugs are more effective. One explanation for the correlation between initial antibiotic costs and the occurrence of treatment failure in the main sample, consistent with the results from the paired analysis (which found no difference in first-day drug costs), is that patients of high socioeconomic status are both more likely to receive an expensive antibiotic initially and more likely to return for a second if the first does not resolve symptoms. An alternative explanation is that physicians often prescribe inexpensive drugs for conditions that they believe are probably viral in origin.
Our findings suggest that analyses of the costs of various antibiotic regimens should include spending on follow-up care. Examining initial treatment costs only may provide a misleading picture of the resource use entailed by different therapies. Along the same lines, if increases in the level of antibiotic resistance lead to higher rates of treatment failure in outpatient settings, then spending by health care systems will increase. Assuming that each treatment failure increases costs by $200 and there are 13 million pediatric otitis media and URI episodes per year (our calculation using the population weights provided in the MEPS), the current total annual cost of treatment failure among pediatric patients in the United States is between $130 million (for a treatment failure rate of 5%) and $260 million (for a treatment failure rate of 10%) in 2001 dollars. The relationship between resistance and outcomes in the community is not well documented,21–23 but, given the number of patients who seek care for respiratory infections every year, even a moderate increase in the rate of treatment failure can increase total costs by a large magnitude.
|
FOOTNOTES |
|---|
Reprint requests to (D.H.H.) Department of Health Policy and Management, Rollins School of Public Health, Emory University, 1518 Clifton Rd NE, Atlanta, GA 30322. E-mail: dhhowar{at}emory.edu
|
REFERENCES |
|---|
|
TOPABSTRACTMETHODSRESULTSDISCUSSIONREFERENCES |
|---|
This article has been cited by other articles:
 |
|
 |
|
  C. M. Sox, J. A. Finkelstein, R. Yin, K. Kleinman, and T. A. Lieu Trends in Otitis Media Treatment Failure and Relapse Pediatrics, April 1, 2008; 121(4): 674 - 679. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
 |
||
|
||
Home | My Pediatrics | Journal Information | Current Issue | Past Issues | Subscriptions & Services | Contact Us Click here for faster international access © 2004 American Academy of Pediatrics. All rights reserved. |
||
|
||
