Ninth Annual Maternal and Child Health Epidemiology Conference / December 10-12, 2003

Increasing Infant Mortality Rates (IMR) in LA: Public Health Emergency or Reporting Artifact?

JUAN M. ACUNA:  Good afternoon.  Well, actually, I think that my presentation is going to round up some of the concept speakers.  I think that when we start analyzing rates of mortality or mortality by rates, or whatever you want to call it, we end up talking about pretty close the same things.  So the background on what we need--and I’m going to keep talking about “we” because what I present is by any means my only effort, but the effort of people that are actually sitting in this auditorium and to whom I am very grateful.  Louisiana is a poor state--no surprise there--and consistently ranks poorly among all the states in the United States for infant mortality rates and other markers of bad things and low of markers of good things.  So we have a lot of trouble.  The infant mortality rates declined to follow the U.S., but it has some bumpiness in the latest years, which I’m going to show you, and hit a record low in 2000, which was when we reached 8.9.  That’s per thousand live births.  Then, the infant mortality rate increased, started increasing to 9.8 in 2001 and recently to 10.2 in 2002. 

So when this happens, we get attention from the media, the legislature, the program people, and (inaudible) as the data people, number geeks, etc. to kind of explain things that so far remain for them pretty much unexplained.  These trigger strong political and media attention given that we have just announced that we were doing actually extremely well.  So that was one of the main problems.  And we analyze data that actually was emphasized on 1997 to 2002, but it will include data from previous years.  We got pretty concerned with all these things and rates because I think that rates, for those of you who have gone into Epi and Biostat straightenings are one of the first classes.  It is taught in Epi 101 and is forgotten by the time you get to 102.  So we mostly rely on very pretty advanced methodology such as multivaried analysis, and then we kind of forget how to handle rates, and we just keep analyzing crude rates.  So we took a more global approach, going back to our old books, old chapters, and tried to see what to.  We analyzed Louisiana’s birth file and, of course, the (inaudible) birth file data, as I explained, from 1997 to 2002, although you will see some trend from 1990.  We studied trends.  We studied GIS.  We adjusted rates, did multivaried analysis, and ended up in something that is still part of what we’re trying to do, which is an analysis of color-related data, because we think that there are some issues that I will discuss later. 

Louisiana is divided in 64 parishes and nine administrative regions.  We do not have counties, but the parishes are county equivalent.  So the regions are just like everybody else’s regions.  So pretty much what you see here reflects how the money moves in Louisiana and how the program moves in Louisiana.  So it was just a straightforward approach on trying to stratify everything my regions.  Why?  Because there are many, many large, rural areas in Louisiana where small numbers are going to be a very problematic issue.  And even though there are methodology, as it was explained before, to analyze small numbers, small numbers will always be small numbers.  The analysis of crude rates, which was the first step of course, include the overall crude rates by state and region; the overall rates by race, birth weight strata, including the PPOR methodology that it has been explained--thank you, I saved some minutes--; and finally, stratification by weight categories--thank you, saved some minutes--and the analysis of reporting dates.  Nobody explained that, so I won’t save some minutes there.  And you see here the actual results and those stars.  And I didn’t put stars here because it would have been a very messy slide.  So you just want to see that those that are highlighted are the regions that has increasing trends.  And given that these are population-based analysis and given that we have several years, as we used to say, even a small hair in an elephant back becomes significant.  So everything pretty much ends up significant when you study the regions, and these are regions that have significant trends, and these are regions that have significant high rates.  So different testing, same results. 

Basically, what I want you to focus is that there are some regions in the state that keep raising the rates while others don’t.  And the second thing is that there are some regions--seven and eight, which is the northern tip of the state--which without any doubt have higher rates than anybody else by any measurement possible.  We also did the kind of the trend analysis of the PPOR, which was very helpful.  And you see that this is consistent to what has been presented, which is a raise in the numbers of those that are less than 1500 grams by weight and in increase in the mortality rate in the infant category.  So of course, all of those are seeking or we are seeking or we have explained part of the phenomenon.  But one of the things that we analyzed further was the group that was reported on less than 500 grams.  As you see here clearly, this is the U.S. rates, mortality rating, in the group of less than 500 grams.  And you see that this is the mortality rate for the same group in Louisiana.  And you see that the record low was in 2000.  So we were very interested to see why our less than 500-grams--I mean almost peanut-sized babies, tiny, tiny babies--were surviving so well.  And of course, they were surviving so well because nobody was reporting that they were dead.  So of course, this was an issue of reporting.  So the problem is that this huge difference and the record low rates--because we adjusted the overall rates base on just this single group, and we found out that despite the fact that this group is taking away of most of the analysis, it is included in the official rates that are reported to NCHS. 

So that really puts a problem in the scenario because you cannot report those rates, come out with a state rate, and then say, “You know, this is the official one but is wrong.  Here is the right one.”  This poses credibility issues, methodological issues, and actually creates a problem.  So we worked hard to try to put these into a logical way; but then again, the major fault was reporting.  And why the rates went up, because vital statistics lost a lot of personnel in a big layoff during 1998 to 2000 years.  I actually hired three people that were very concerned because they were hired to actually address underreporting of deaths.  Bingo.  So the rates started going up why?  Because they were being reported.  So this was kind of a non-real (inaudible).  But we got curious, what if there is something else, you know?  Without no doubt, adjusting by whatever method, things are increasing.  Mortality is increasing.  So we started our step two, which is adjustment of rates.  There are many, many ways to adjust rates.  Probably you remember the direct way.  Ah, yes.  How do you do it?  I don’t remember.  But indirect way.  Ah, yes.  There is.  How do you do it?  Ah, I don’t remember either.  But you can go to the book and see that is quite easy to do.  And there are many more. 

As it was explained before, probably just on the surface, a formula was put and is very recent, 1954, by Kitagawa and Cochran and some other authors, explaining something that makes a lot of sense, and is that the rates have many components within them.  So if you’re talking of the state rate, you’re talking of the infant mortality rate, which has many compartments within them.  So it is actually a true issue that you cannot just stop analyzing crude rates.  You have to start analyzing adjusted rates, as well.  Cartoonish.  For those of you who are high Epi skilled person, or biostats, you’re going to hate me; but those of you who are not are going to love me.  And this is basically Louisiana’s population by region, and each region is actually one of those compartments.  And you will see clearly that if there is one region that has different people in there, it will have some deaths that are mostly occurring in that population.  So the overall mortality rate that was reported is actually composed along different rates and different populations.  So we need to have (inaudible) and we have to account for those.  What happened? 

In the same map, you see that our regions are quite different.  This is a region with one of the high rates--regions seven and eight, highest rates--and this has 48 percent.  Half of the births are Black population.  New Orleans, more than half are Black population.  Louisiana near Texas, it’s quite different population.  So we cannot just pile them into a sack and say, “This is the rate,” or we cannot just compare them one to one because they are indeed different.  So where do we need to look?  Well, of course, we need to adjust rates.  And you remember those two that I wanted you to keep an eye on, region seven and region eight?  What happens if you add just the rates by the population characteristics?  They are the lowest rates in the state.  And those like region five, who has always presumed of having the lowest mortality rates, has actually the highest adjusted mortality rate.  How do you explain this to them?  It’s quite hard because these are not real rates.  They are adjusted rates.  So you have to find a clear way to explain to them that this is happening.  And then, I go to cartoons again.  So what happens is that if you have a rate overall, you know that the partial rate is actually part of those big rates.  So you find that each region has a rate, each group has a rate, each strata has a rate.  And then, the big problem is, “Oh, my God.  What is that, Chinese?”  No, formulas.  What this says for those of you who are not Epi savvy or biostat savvy is that actually the difference between these rates, as we have been explaining the afternoon or the session, is composed by several components.  And those components are those that are equal among the rates--Name this one, because these are the guys that are going to lie--and those that are different among the populations.  And you can actually measure the difference among those components.  To what extent?  Well, to an easy extent.  Let me just try this. 

This component here is going to explain those here.  So that is the rate that is exactly the same for every single person.  But this component over here is going to explain the difference in each one of these groups.  What is going to happen?  What is going to happen is that you see that here that we have a multiplier.  If there is no difference, this component becomes zero.  Then, we have just the same rate.  Pretty simple, isn’t it?  Oh, yeah.  It is pretty simple, actually.  You can plug this into an Excel sheet, and you can actually make it happen quite easy.  So the whole trick is to divide compartments.  Otherwise, it doesn’t make sense.  Crude rates don’t make sense, period.  So what we do when we explain to our program and policy people crude rates really doesn’t make sense.  It’s not going to change and make sense in the overall number.  It doesn’t make a programmatic difference because if you build the programs on those rates, you’re going to be by default wrong, unless all those components have a multiplier of zero so the rates are actually equivalent.  What happens then?  Well, let’s say mortality has two components when you address the infant mortality piece.  One component among many is that mortality by birth weight distribution, which means I am going to have most of my mortality excess because I have a higher proportion of tiny babies.  And this one, birth weight-specific mortality, regardless of the weight of my babies, there is something that is killing more babies or letting more babies die, period.  Simple concept?  Yes.  It is a simple concept.  One addresses prenatal issues that lead to a huge amount of preterm issues, tiny babies in excess.  And the other one is even if you don’t have an excessive tiny babies, your babies are going to die in excess, Delaware’s example.  So what happens? 

Well, the trick from the programmatic perspective is to explain to these guys where are we?  Do we have a higher proportion of tiny babies, or we do we have a higher proportion of deaths regardless whether they are tiny or not?  So this is very helpful.  Why is it helpful?  Let’s take one example of one hospital and come up to the conclusion that we have done these for every single birthing hospital in the state.  We analyze these two components for the highest prenatal mortality rates found in the state and one of the highest by institution in the whole country.  So these guys were concerned, “Oh, my God.  I don’t want to give birth in that hospital because babies die more.”  Well, that would have been the message.  What does it happen when you analyze this?  You analyzed the birth weight-specific mortality so that mortality regardless the birth weight is actually quite good to some extent, to the extent that in some of the categories, these guys do much better than the average U.S. hospital.  And actually, Louisiana in neonatal care is very well known nationally to have one of the best services available.  We rank number 14 among neonatal intensive care unit services and mortality rates.  So the whole issue is that the hospital has humungous birth weight distribution rates, which means this is a reference hospital for poor moms that cannot go anywhere else, and they will deliver preterm deliveries over there.  It was a great message.  It was a great way to know that they did not have to invest in more incubators and (inaudible) units, but they have to go out to the community to prevent those mothers from coming here, not by closing the doors like private hospitals but by opening outreach programs to prevent these moms from coming.  Okay, but we were not satisfied because this doesn’t really explain it all.  So we say, “Okay, how about what is the role of multivariant analysis in this whole analysis of mortality?” 

So we started doing multivaried analysis.  And the first one that we did was, of course, an adjusted logistic regression analysis.  Why do I stress on the adjusted?  Somebody would say, “Well, it is adjusted, per se, so why do you put the word there?”  Well, just because I want you to know that this can be done by groups, by regions, as well.  So you just need to think on your models as if they were done in the same fashion that you report rates.  So you can actually see how the same risk factors change between regions once you split and categorize by regions.  So you’re not actually controlling by regions; you’re actually analyzing risk factors by region given that the outcome of interest is actually death.  And the second thing that we analyzed as a multivariant mode are more newly-developed modes that analyze correlated data.  Why?  Why logistic regression models and multivaried models in the usual sense are not the final explanation?  In population-based data, we seldom have independence of variables.  One explanation.  Come on, one lady from one region could have prenatal care in another region.  Is that correlated data?  Yes.  What if it happens by the thousands?  Well, more correlation.  So all those non-independent variables, such as level of attention, access to services, prenatal care, the highway opened between two regions that allows better access from one region to another than those regions that are split without a highway in between, makes a difference in the analysis.  And what these models allow for in a better form is the analysis of the error.  And the error is pretty much where the answer is.  Why?  Because you and I have heard two days of the same message.  For this perspective, for prematurity perspectives, we do not have the answer. 

So all these models are explanations and error.  So if the answer is not in the explanation, where is the answer?  Well, but of course, it’s in the error.  And you can measure the error.  So you can have a good idea of where might the whole answer be.  And this is just a result so you can see that even in a multivaried analysis, most variables that we can plug in are indeed very significant.  So I am not going to go into detail, some make a lot of sense, but just to name another potential tool that you can analyze parishes or regions or whatever by that makes sense in order to have a further explanation.  So what are the conclusions?  Well, one is pretty obvious:  one size doesn’t fit all.  We cannot just stop at any one of the proposed models.  We cannot stop adjusting rates and that’s it.  We cannot stop analyzing crude rates and that’s it.  We cannot just do a lot of multivaried analysis and that’s it because none of those will explain the full phenomenon.  Why?  Because each one of those methods, if you back to the books, they have a very strong set of limitations.  And population-based data has everything.  So if you talk about any limitation, it will apply to population data at some point.  Neither crude nor adjusted rates are the only analytical tools for the analysis of risk, and risk for death, of course, from the perspective that you’re analyzing a complex population. 

In this case, it is state population and regional population, but any population is made of compartments.  So you have to think what those compartments are, otherwise the analysis just become numbers.  You have to go even to the face and name perspective in order to do some analysis.  Analysis of reporting is mandatory as the first step, and sometimes we forget that we have to take that into account and we start using our official numbers, and they do include non-reporting errors that make sense if you can actually get rid of them right away.  And you have to be creative, but you have to be very careful.  And the reason why is that the program and policy people really rely on us.  So the analytical model that we propose for you as the way to go is not something new.  It is just a pileup of old things, which is you have to start with the analysis of crude rates.  Crude rates analysis cannot be excluded, cannot be taken away, even if it leads sometimes to unexplainable or wrong things.  You have to start there.  Then, the crude rate of errors, which is the adjustment of rates.  What you do when you adjust is that you get somehow read of or account for the errors among the populations.  So you need to adjust rates and analyze those as well.  And then, you could jump to multivaried analysis and then to correlated data analysis or correlation analysis or analysis of correlated concordant data.  And of course, we have now more tools such as GIS analysis, trends analysis, survival analysis, etcetera.  So my bottom line message is do not try to find the tool because it doesn’t exist when we work with population data.  You have to use everything that you got and make sense out of it so you will have better explanation for the people that cannot deal with numbers.  Thank you.