I'm going to talk to you about access to care, which I think is taking small area analysis and using it for one of the many possible applications of the technology. I'm going to talk about some of the problems with using this technology, discuss some of its limits, and mention some of the cautions you need to take when you try to use the data this way. I'll also discuss the need to supplement the data using other survey instruments and other sources in order to determine what's really going on. Often the kind of data I'm going to show you don't answer all the questions you need to have answered. They often simply raise a red flag or steer you in the direction of the next question.
When you see variations in hospital utilization rates at the zip code level, there are lots of explanations. Dr. Wennberg talked about supply-induced demand and professional uncertainty. I'm going to talk about the access barrier hypothesis and a second category I call none of the above.
Now, the access barrier hypothesis goes something like this: For some conditions, if you get timely and effective outpatient care, it may have an impact on reducing your risks of getting admitted to the hospital. Therefore, a high rate of hospital admissions in an area for certain conditions might be an indication of an access barrier; a low rate may mean there isn't an access barrier.
Let me begin with the none of the above category. One explanation that has been put forward is random variation. A second explanation that falls into the none of the above category is "you made it up". I have some slides that look like I made it up, but I didn't. The third explanation often offered is that there are differences in the prevalence or incidence of disease among populations. That is true, and I'm going to try to show you how that can impact the data. The fourth explanation in the none of the above category is that the data are wrong. That is often true. Let me give you some examples.
When you do small area analysis, there is a numerator and a denominator. Something can go wrong with either. In the numerator, you have the hospital admissions. When you look at DRG codes, you have the problem of miscoding and hospital gaming. We found one zip code in New Jersey that had astronomically high admission rates. When we looked more closely we found out that the hospital was coding every patient visit to the hospital zip code rather than the patient's zip code. Similarly, in New York City the New York City Health and Hospice Corporation, which is a public hospital system that treats all the prison inmates, codes all of its admissions to the headquarters of the Health and Hospice Corporation in mid-Manhattan. So this area has a very high hospital admission rate. So those things can go wrong in the numerator. You can also have an incomplete data set. For example, we had an experience in New York where a number of hospitals in a public hospital system were missing about 50% of the discharges for one particular year. So suddenly, in Central Harlem, the admission rates started going down dramatically one year. We mistakenly thought we were having an impact with some of our programs. It turned out they just hadn't sent in a complete data set. Similarly, you can have hospitals missing from the data set altogether.
The potential problems with the denominator are even worse. In the denominator, you're counting the number of people in each zip code. I don't know what they believe here in Missouri, but in New York City we don't believe the census. And if you're at a zip code level, a miscount can have a dramatic effect on your results. Secondly, when you get census data from the Census Bureau that tell you the population of a zip code, it isn't really that. What they have done is hire a firm to take their data at the census tract level and assign it to zip codes.
Now, census tracts don't actually follow zip code boundaries so an algorithm was created to assign part or all of the census tract to one zip code or another. We have compared five different sources of population data--from public vendors and private vendors--and found that each one of them has a different number for almost every zip code. Typically, it's not a big difference, but sometimes it's very large. You also run into problems with new zip codes coming on line and old zip codes dying. Zip codes are run by the post office. They don't really care about these kind of data, so when they kill a zip code or create a new one, people don't often know and it can obviously confuse the data.
So, the first thing to do when you start looking at patterns of utilization is to look for very strong patterns and very large variations. The second thing to do is to look at your data again because often there's something wrong.
Now, we also looked at the access barrier hypothesis. We brought together a group of prestigious clinicians from around the country to categorize ICD9 codes as Marker conditions, Ambulatory Care Sensitive conditions, or Chronic conditions. Marker conditions are conditions for which doctors agree on the criteria for admission. For the conditions, obtaining timely and effective outpatient care immediately prior to the admission would have no impact on the need for admission. We came up with only four Marker conditions: hip fracture, appendectomies... If you fracture a hip, you're going to get admitted to the hospital. And there's nothing you could have done in the month prior to that admission to prevent the hip fracture.
The second category is Ambulatory Care Sensitive conditions. These are conditions where we hypothesize that timely and effective outpatient care could make a difference. Now, the first type of Ambulatory Care Sensitive conditions are totally preventable through immunization. For example, there should be no congenital syphilis admissions. Pregnant women who test positive for syphilis should get treated, and their children should not have congenital syphilis. Or, there should be no polio admissions. Where you see these admissions, there is clearly a breakdown of the system.
The second type of Ambulatory Care Sensitive conditions are acute conditions where the goal is to catch the condition early and manage it well in order to avoid hospital admissions. A classic example is an ear infection in children. If it is treated effectively, it doesn't progress to the point where admission is necessary.
And the third category is Chronic conditions like asthma, diabetes, and hypertension. If you manage these conditions effectively, you can avoid acute flare-ups that result in hospitalizations. You're probably not going to avoid all the hospitalizations, but you can avoid some.
What we expected to see with the Marker conditions was no difference among geographic areas. Let me show you some of our findings. Here is one of our classic Marker conditions: heart attacks. Again, getting primary care in the three or four weeks before the admission is not going to prevent the heart attack. And, most doctors would admit in the case of a heart attack. All the zip codes are pretty much within a fairly narrow band. In fact, the R squared shows that the income in a zip code does little to explain the rate. The ratio between low income zip codes and high income zip codes is 1.00, meaning that both rates are the same.
Whenever we find a zip code that seems very high or very low, the first thing we do is look at the Marker rate to see if it's also very high or very low. That might tell us whether there is something wrong with the denominator or with the numerator.
We then looked at Ambulatory Care Sensitive conditions and we saw a dramatically different situation. For asthma, one of the very common Ambulatory Care Sensitive conditions, we see an R squared of 0.705. Seventy percent of the variation in admission rates among zip codes is explained by one simple indicator: the income of the zip code.
The ratio between low and high income zip codes is about 6.4, meaning that admissions in the low income zip codes are 6
Among all of our Ambulatory Care Sensitive conditions, for New York City you see a very high R squared, meaning a large difference between low and high income areas.
The thing to note is that not all zip codes are the same. Some of these low income zip codes are much higher than other low income zip codes. What we wanted to know is: why are some higher and some lower?
And then we see some zip codes in middle class neighborhoods that are also very high. We wanted to know whether we had a numerator/denominator problem or whether something else was going on in that neighborhood.
South Bronx, Central Harlem, and Bedford Stuyvesant in New York City all have very high admission rates. Washington Heights, one of the crime and drug capitals of the United States, has a very low rate. It is a poor neighborhood like Central Harlem, Bronx, and Bedford Stuyvesant, but it has a much lower rate.
I have been discussing the under 65 population because we felt like that is where the access barriers would be the greatest. Among the over 65 population, you see a much weaker association between the admission rate and the income level. The R squared is much lower. Now, people over 65 are fundamentally different than people under 65 because they have universal coverage. So at least insurance is not creating barriers to care.
We started looking at each age group and we saw some fairly dramatic differences in rates based on age and income. One theory that we're pursuing based on the age group differences is that for Chronic conditions, getting access to health care throughout your lifetime can have a residual effect, even with universal coverage, experience using the system, and the disappearance of other intangible barriers to access, there is still a residual impact.
When I show these data around the country in places like Kansas City, the first reaction we get is: "That's New York, who cares?". The second reaction we get is: "That's New York, you're somehow different".
So one of the things we did with the help of The Robert Wood Johnson Foundation was to look at other places around the country. The first thing we did was look at Upstate New York. Then we looked at New Jersey. Next, we decided to move west to Chicago. And, then to Portland, Oregon. Even Los Angeles and San Diego exhibit the same pattern.
We also looked at some rural areas in Florida, Oregon, and Washington. And what we have found is the same sort of pattern with the lowest income areas having the highest admission rates. But there is a lot more noise in the data because of the low population. Secondly, you have the single hospital phenomenon where there's one small hospital serving the community. What we have found is that physician practice style can have an impact on admission rates.
The two exceptions to this pattern we found are in Dade County, Florida and Toronto, Canada. What's happened in Miami, in Dade County, and Florida, in general, is the state and local governments have invested a lot of money in primary care delivery sites, in the community health centers, and in publicly available and publicly supported clinics. The second thing that's going on here is we have a large Cuban-American population. And, what we found was the Hispanic zip codes had similar admission rates despite any income differences. The health care delivery system for the Cuban-American population is fundamentally different. Among the people who came over from Cuba, there were lots of Cuban physicians. So, this population has lots of physicians who are serving that community. So what we think we're measuring, even in the low income Cuban communities, is the impact of that resource availability.
Now, the rest of these zip codes that are low income are predominantly Black, but they're also relatively low, compared to other places.
The second place where we found even less association between income and admission rates was in Toronto. There is a fundamental difference between New York City and Toronto, and that is that they have universal coverage and we don't. They have no very predominantly poor areas in Toronto. So what we have done is compared Toronto with New York, Newark, and San Francisco. We truncated New York City, Newark, and San Francisco at 40% poor and found that we still have an R squared of about 0.5 or 0.6. So it doesn't look to us that it's simply the homogeneity of the population that's driving the difference in Toronto. As I said before, we started with a hypothesis that maybe we could be measuring access and barriers to access.
What we tried to do by various methods is to understand what else could be going on in these zip codes to account for these variations. We have tried to explore a number of factors that could account for some of the enormous difference between low income areas and high income areas. The first thing that comes to mind is differences in disease prevalence. Low income populations may have higher disease prevalence than high income populations, particularly among some subgroups. For instance, the African-American population may have higher rates of hypertension.
There is, in fact, a difference in disease prevalence for asthma among low and high income populations. Asthma is more prevalent among low income populations than high income populations. But it accounts for only a percentage of the difference in hospital admission rates between low and high income areas. The huge amount of difference in admission rates is just not accounted for by differences in prevalence.
For diabetes, we found somewhat higher rates of prevalence of the condition among low income populations, but astronomically higher admission rates among low income areas.
A related issue around prevalence is whether we are measuring many patients being admitted once to the hospital or a few patients being admitted many times. That can have a fundamental impact on how you design intervention. If it's a lot of patients being admitted once, then you want to do more primary care in the community. If it's a handful of patients being admitted a lot, then you might want to target resources more selectively.
In New York we had available to us a longitudinal file where we could link patients by their name and other identifiers to determine if they had repeat admissions in the same year. And we found that for some conditions this was, indeed, a very important factor.
We found that for asthma 30% of the admissions were, in fact, re-admissions. Some of them got re-admitted only once, but some of them got re-admitted many times. The current record in New York City for asthma is 28 admissions in a year. For other conditions, it's a much lower percent. For the Chronic conditions, it's actually a fairly high percent. For Acute conditions like severe ear, nose and throat conditions, it's a very low factor. Re-admissions are important, and they're important for designing interventions. But they don't explain the variations and they don't explain all of the conditions.
Now, the next thing we wanted to do was--to the extent we could--look at patient lifestyle factors. We tried to look carefully at the impact of drug and alcohol abuse on admission rates. We looked at the secondary diagnosis field on the discharge abstracts to see what percentage of the patients had a secondary diagnosis of drug or alcohol dependency or abuse.
For some conditions, like tuberculosis and pneumonia, it's actually a fairly high percent that had alcohol or substance abuse, 41% and 28%, respectively. For other conditions, like asthma, it's only 3.8%. But for those conditions where it is a factor, it can have a dramatic impact. And for these conditions it's very important when we design the primary care delivery system to service population that we take into account the need for substance abuse treatment. The third area that we wanted to explore was differences in physician decision-making. You can imagine how that might influence admission rates for the population.
Think of a pneumonia patient or an asthma or a diabetes patient who comes into the hospital emergency room in fairly poor condition. If it's a low income patient who is homeless, a substance abuser, and has no private physician, a physician might be more likely to admit that patient simply to manage the illness than if a middle class patient with a home, a doctor, and a place to receive care presented to the emergency room.
We frankly felt we would find that some differences in physician practice style would influence these admission rates. We used a low powered severity of illness scale, called Totescale, to look at resource consumption and other factors in the hospitals to come up with a severity measure. We also looked at secondary diagnoses to come up with severity levels within a diagnostic field.
Take diabetes, for example. Low income neighborhoods didn't have less severely ill patients than high income neighborhoods when you control for diagnosis. Some colleagues of ours at The University of San Francisco, also funded by The Robert Wood Johnson Foundation, have done another study where they looked at low rate ACS areas and high rate ACS areas. They then surveyed physicians and presented them with clinical scenarios to see whether they had different standards for admission. And, again, they found no association between threshold for admission and the ACS rate. So, we concluded that there probably are important differences in physician decision-making, but they don't explain very much of the variation in the data.
Within an individual zip code, you may find important differences in physician practice style. But across all the low income zip codes and all the high income zip codes, it doesn't explain very much.
We're doing other studies where we look at the reasons why people come into the emergency room. What we're trying to do is develop a set of diagnostic categories to separate out primary care patients from non-primary care patients. What we found so far, as you would expect, is that in low income neighborhoods there's a much higher rate of people coming in for primary care, as opposed to the higher income neighborhoods where they're coming in for very acute conditions. But in both neighborhoods there is a lot of primary care that goes on in the emergency room.
But when you think about it, even if we had found the physician practice style was accounting for a lot of this difference, that still would have been an access barrier. It still would have been a concern about the quality and accessibility of primary care to these patients. It would have been environmental factors, too.
One of the underlying concerns of physicians when we talked to them was that for a patient who isn't attached to the system, the emergency room visit and a subsequent admission may be the only chance at controlling the patient. So the physician admits the patient. It was a theory we were working on. We couldn't find it in the data, and our colleagues in California couldn't find it in their data. I still think it's there. I think our measures just aren't adequate enough to get at it.
What all of this suggests is that there are various different kinds of access barriers. I always refer to it as the black box of access barriers. We don't really understand what these barriers are. And, simply looking at discharge data doesn't get us very far. So, we used a supplemental source to try to understand better what was going on in New York City. We did a survey of inpatients at nine hospitals in New York City. We tried to pick the neighborhoods very carefully. We picked three low income neighborhoods and included some middle class neighborhoods.
We did find some interesting things. I'm not going to give you all of it, but I'll show you some of it. What we were trying to understand was what kind of barriers patients were encountering prior to admission, outside the hospital, if any. We interviewed patients immediately after the admission when they were up on the patient floor. We had a questionnaire that was about 40 minutes long. From these interviews, we learned several things.
Low income neighborhoods have a much higher percentage of individuals with no usual source of care than high income neighborhoods. That was something we expected to find. The other thing we found was that the low income neighborhoods are much more dependent on hospital outpatient departments and community based clinics for care. In high income neighborhoods, office-based physicians are treating the patients.
We also found dramatic differences in telephone access between low and high income neighborhoods.
To try to understand the level and extent of continuity of care the patients were getting, we asked the question: Is there one person somewhere in the health system that you consider to be in charge of your care?
In New York City, there was a very high "yes" rate for low income neighborhoods. We asked this question in other parts of the country and got responses in the 10 to 20 percent range; in New York City, we got a 50 percent response rate. I'm not sure whether what we're measuring is people answering the question the way they think it's supposed to be answered, or whether we're actually measuring something different about New York City.
The second thing is that the problems are different for low income children and low income adults. A much lower percentage of children have no source of care. That suggests to me a bit of good news in that we have thrown a lot of resources at children. The Medicaid program is tilted toward children. A lot of the community-based interventions are tilted toward children. What we saw in our New York database is that we're actually having some impact.
In trying to understand what happened to these patients in the 30 to 60 days prior to the admission, we asked them what kind of contact and what kind of care they had immediately prior to admission. For low income patients, 61% responded that they had absolutely no contact at all with the health care system before coming to the hospital. For another 17
Contrast this with high income patients who had physician contact almost 63% of the time. For low income patients, from the data it looks like a failure to get treatment. For high income patients, it looks more like a treatment failure. Not every one of these conditions is preventable. So high income patients were getting admitted after getting treatment, and the treatment just didn't do the job.
We also looked at how patients make the decision to seek care at the hospital. Half the patients in the high income population were advised by their doctor to go to the hospital. Only 9% of the low income patients were advised by a doctor.
Now, the other important issue we wanted to get at was the barriers to access. So we asked a long series of questions about whether they delayed seeking care or did not seek care for any particular reason(s). Low income patients had a much higher rate than high income patients of not getting care for one reason or another. But again, the problems among children are much less serious than the problems among adults.
Here are some of the reasons people gave for not getting the care they thought they needed. The first thing to notice is cost is way down on the list. A lot of very vague things are at the top: I wasn't up to going; I was too nervous; I was afraid; I was unable to get the time to go get care. What we thought we were measuring here is a general ambivalence toward the health care delivery system among low income populations. It's very hard, when a patient is hospitalized with a tube up their nose, to get them to say they don't like the care they're getting. But ask them indirectly about things and you get a much higher response rate.
The other thing to notice is the big difference between adults and children for some of these measures. People are willing to admit to these sort of things for themselves, but not for their children. So if you want to catalogue the black box of what's causing patients not to seek the care they need, this is a pretty good list to start from.
Among the patients we looked at, there were a lot of asthma patients. We found, in comparing high income and low income asthma patients, a big difference in the kind of education they're getting in training and managing their asthma at home. About half the low income patients said they had little or no education or training to manage the condition at home; only about 25% of high income patients responded in this vein.
We also asked how much confidence they had in their abilities to care for themselves at home. Again, there was a huge difference between low income patients and high income patients regarding how confident they felt in their ability to manage their care.
As you may remember, we were trying to compare two low income neighborhoods with very high ACS rates and a low income neighborhood with a very low ACS rate to a middle class neighborhood. And what we found was that the low income neighborhood with a very low ACS rate had an extremely low percentage of people saying they had barriers to care. We found that the low income individuals living in the middle class neighborhoods reported barriers to access almost identical to individuals in the low income neighborhoods in which a high percentage of individuals reported access barriers.
Washington Heights, although a low income neighborhood, scored much, much better. What we're trying to do now is understand the dynamics of Washington Heights. One of the differences is that it is a predominantly Dominican community. And, there is a cadre of Dominican physicians serving the population. In Central Harlem, which has a population of about 200,000, there are two private physicians. The rest are either at Harlem Hospital or in various clinics.
In the time that's left, I want to talk about how these data are used and show you some trends. First, the data are used simply to compare one community to another. It's interesting to see that Newark, New Jersey doesn't do as well as Toronto. That tells you something about resources. In New York City, we have individual neighborhoods behaving differently. And in California, San Diego looks different than San Francisco, and so forth.
Secondly, within communities, these data can help identify high-need neighborhoods. Third, and this is something we're still working on, to use the data to evaluate interventions. If you do an intervention that's large enough, you should be able to measure it by looking at the change in the admission rate for conditions.
A fourth area which is being pursued by health maintenance organizations is to use the data as a quality assurance tool by examining high admission rate conditions. We are currently in the middle of designing an evaluation of the managed Medicaid plan in New York City to evaluate how effective the data are as a tool for measuring both the impact of managed care on the population and the quality of care.
The fifth thing the data can be used for are to design programs and interventions. For example, with chronic conditions, there is a cadre of patients getting admitted over and over again. Target interventions for these patients can have a fairly large impact.
In New York City, there have been various programs initiated in the last few years to target some of these hard-core chronic disease patients to manage their care more effectively. Similarly, if you're designing a primary care delivery system in an urban area, it's probably advisable to have some connection to a substance abuse program. If you're running a managed care program for Medicaid patients, you need to think about some of this.
You can also use the data to look at trends over time. When we look at the period 1982 to 1992 in New York City we find that things are pretty stable. The bad news is that when we looked at low income zip codes separately from high income zip codes we saw a higher rate for admissions. Some of it is environmentally related, some of it is nutritionally related, and some may be that we are measuring a long-term impact. What this suggests to us is that using the census data for New York City for this period of time there probably is some systematic undercount of low income neighborhoods.
We also looked at trends for ACS conditions. What you see in New York City for the period 1982 to 1992 is an overall increase of about 20%. And the low income neighborhoods look like they are going up faster. In 1992 the low income zip codes were about 35% higher than in 1982. For high income zip codes they're about 4% higher over the period.
However, not all conditions are behaving the same way. For diabetes, the high income zip codes are virtually unchanged. But low income zip codes are about 80% higher over the period. For asthma, both high and low income neighborhoods went up at about the same speed. If you look at a map of New York City, you see that the areas that have the highest change rate for asthma between 1982 and 1992 tend to cluster. There are two explanations of why they cluster. The areas might be very high because the hospital changed its practice style, or maybe a new incinerator opened up that's burning garbage. It's interesting to note that the neighborhoods that show the biggest change are not the typical low income communities.
We are working with possible explanations for asthma admission rates. One is that doctors have changed their behavior; secondly, that somehow the environment has gotten worse; and, the third is that people have changed their behavior. During this period, the Public Education Commissioner died of asthma because he didn't get to the hospital in time. That could also be an explanation.
Although the ACS rate was going up faster among low income populations than high income populations, not all age groups were the same. We found that for the zero- to four-year olds, both low income and high income populations didn't change much over the 1982 to 1992 period. We saw the same pattern among children 5 to 17. All the difference in the ACS admission rates between low income and high income neighborhoods is coming from the adult population.
We started this work using ERG's in the late 1980's. We didn't feel comfortable with that, so we got this committee together and looked only at ACS data.
The diabetes I have been showing you is ketoacidosis. Physicians are likely to agree on the criteria for admissions. When you have a ketoacidized patient in the emergency room, you're going to admit that patient.
These data are enormously intuitive to policy makers. I've spoken to Congress and I've spoken to clinicians in the legislature that understand it the first time. Having said that, I still think it's important that before you engage in something like this you bring in the relevant communities affected by it: the provider communities and the community agencies that are serving this population. You need to talk to them about the methodology, show them the results, and have them help you interpret it.
What I've tried to tell you today is that it takes interpretation. There's not clearly a right answer every time, and they can be an enormously valuable resource in helping you understand what's going on. The only group that really gets very worried about these data are some of the providers who are supposed to be servicing this population. They're a little nervous about whether we are measuring their performance. Are we? The answer is: Yes, sort of.
So, I urge caution, but I think this can be a very useful tool if properly applied and carefully interpreted.