Ninth Annual Maternal and Child Health Epidemiology Conference / December 10-12, 2003
DANIEL THOMPSON: Okay, so the title of this talk is “Infant Mortality and Low Birth Weight for 2001: Actual versus Expected.” The thing the title doesn’t tell you is actual versus expected for each county by county. So we’re going to be looking at actual rates versus expected rates by county in Florida. Okay. The background is in 1992, we started a program whereby we have Healthy Start coalitions covering the 67 Florida counties through contract with the Florida Department of Health. And this is not the Federal Healthy Start program; this is Florida’s Healthy Start program. The coalitions oversee prenatal care in the communities and address issues of infant mortality and low birth weight. So annually, the Florida Department of Health tracks infant mortality and low birth weight as part of the performance measures for the coalition contracts.
The coalitions’ contract with the state and part of the performance measures in their contracts is infant mortality and low birth weight rates. So the usual method of looking at infant mortality and low birth weight in a specific county is to compare the county statistics to statewide statistics. And everybody who’s ever looked in a vital statistics book for a state has probably done this. And when you look at the statistics for a specific county and compare it to the statewide number, the differences there can be influenced by demographic characteristics in the county and also random variation. So when you see a difference in a specific county and it’s different from the statewide number, you don’t really know what’s going on. It could be the county is atypical demographically, or if the numbers are small, especially if the numbers are small, it could be just random variation.
So the whole purpose of actual versus expected is to get around those two things. So the actual versus expected analysis is a way to examine county statistics independently from the influence of county demographic characteristics and random variation. Our outcome definitions in this analysis are infant mortality is death in the first twelve months, low birth weight is infants weighing less than 2500 grams. Those are pretty much the standard definitions. Okay, we have factors that we know are associated with infant mortality and low birth weight. Three big ones are maternal race, maternal marital status, and maternal education. They’re all associated with infant mortality and with higher risk of infant mortality and low birth weight. We’ve seen this over the years in our vital statistics data. All three factors have a high prevalence among Florida’s births, and these factors are not routinely influenced by public health interventions.
In general, you don’t want to adjust for something that you have a program to influence. That’s getting a little bit off the subject, but that’s just a principle of adjusting. You don’t want to adjust for something that you want to analyze. It would take too long to explain all that, so I’m going to go on. Okay, now this slide just shows some data to back up what I just said. And we can see that for these three factors, they all have a high prevalence. The lowest one is less than a high school education at 21 percent of the births. All of them are associated with risk of infant death and low birth weight. The lowest risk ratio you see in this table is 1.2, and that’s for low birth weight for less than high school education. All the other risk ratios are higher than that. The highest one is 2.4 for maternal race Black infant death. So they’re all prevalent among our births, and they’re all associated with risks among our births, too.
So we used these three factors to create nine demographic categories. The way we do this is we take each of the three factors and categorize them into two categories, and then three factors, two categories each, that’s three to the power of two, and that’s eight. But here I have nine, and the ninth category is where we have to put the ones that have unknowns for any of the three factors because if a record has an unknown value for any of the three characteristics, we can’t classify it in categories one through eight, so we throw it into the ninth category. Okay, now this table shows you the nine categories we’re using. They’re all mutually exclusive categories, and they’re mutually exclusive and mutually exhaustive. Is that the right term? But anyway, these are all the combinations you could possibly get using the three variables with two values for each variable. And you can also see in those last two columns that the infant death rate and the low birth weight percentage vary quite a bit across those nine categories. The highest category is the unknown category, but we have very few in that category. So if you just look at the number eight category, you can see that the low birth weight percentage and the infant death rate are very high there, especially compared to the number one category. So there’s quite a bit of variation across these categories. So the actual versus expected, we calculate that. This is the basic underlying assumption behind the methodology. We expect women in each county in each of the nine demographic categories to have the same rate as women throughout the state in the same nine demographic categories.
So we’re taking the statewide rates and applying them to the women in the county in each of the nine categories. Then, the expected statistics are compared to the actual statistics for each county. Now, the reason you can have different expected statistics for different counties is because the proportion of births across the nine categories varies by county. Some counties will have more or less in category one, more or less in category two, and so forth. So there will be different expected rates for different counties. And here are some data that illustrates that. We have example counties “A” and example county B. These are actual counties, but I forgot which ones they are, so I can’t even tell you. I can’t even break confidentiality if I wanted to at this point. So county “A” and county B, you can see that their demographics are different. You don’t have to look at all nine categories. Just look at category number one. You can see that 63 percent of the births in county “A” are in category one, whereas in county B, only 44 percent are in category one.
So the demographics of these two counties are much different. So if you think back on how we’re calculating the expected rates, you can tell that the demographics would result in much different demographic expected rates. So the state rates for each of the nine categories are used in calculating the expected statistics for each county. The calculated expected numbers are compared to the actual numbers from biostatistics. And at that point, any differences can be assumed to be independent of or not influenced by the three demographic factors. Then, to address the randomness question, we test for a statistical significance. And the question here is are the expected numbers significantly higher or lower than the biostatistics actual numbers? And to do this, we use a Z-test, or when the numbers are small, we use a Poisson test. In general, the actual risk expected analysis is similar to methods use to investigate suspected cancer clusters, and that’s actually where I got the idea to apply this method to infant deaths and low birth weight because we used to do this when I was working in cancer epidemiology on suspected cancer clusters. And what we compare is by eliminating the influence of the three demographic factors, differences you then see will be the influence of other factors, factors that are not the three demographic factors since you’ve already adjusted for those.
So for instance, you could see differences attributable to smoking habits, maternal age, or timing of entry into prenatal care. And these three things are all three things that public health programs address. Of course, smoking we address with smoking cessation programs. Maternal age, you can’t change a person’s maternal age, but you could change the age, the average age, or the age distribution of the women giving birth by, for instance, a successful teen pregnancy prevention program. And the timing of entry into prenatal care, of course, we’re always trying to improve on that. So if a county who had high actual rates compared to their expected rates worked on these three things, their actual rates would come down closer to their expected rates. And if they did an exceptional job, their actual rates would go below their expected rates because in the expected rates, we have not adjusted for these three things. We did this for 2001.
And we have 67 counties in Florida, and five out of the 67 counties had infant mortality rates that were significantly higher than expected. And for low birth weight, nine out of 67 counties had a low birth weight rate that was significantly higher than expected. And we mapped these things, and there’s the map. And I hope you can see that, but the counties with the striped pattern are the significantly high counties; and the solid color counties are the significantly low counties. And you can see that most of the significantly low counties are in the southern part of the state and most of the high counties are in the northern part of the state. And we’ve noticed that, and we really haven’t done a lot of analysis regarding that yet. But we know, of course, since we’ve adjusted, that it’s not because of differences in the northern half of the state in terms of marital status, race, or education. It’s something else. And here it is for low birth weight. Again, the significantly high counties are in the northern part of the state, and the significantly low counties tend to be in the southern part of the state. So that’s something we should really start looking at more closely. It could give us a clue about what’s going on.
So the conclusion is that this actual versus expected method is one way to examine infant death and low birth weight statistics independently of the influence of maternal education, marital status, and race. And we now have actual births expected statistics for 1999, 2000, 2001, and 2002. And we plan to look for patterns over time in addition to the geographic patterns. We wrote this up in the publication that we have on the Department of Health website. It give you a little more detail about how we did it and so forth. And that’s the title of it, and it’s at that website, which is really the--it’s the Florida Department of Health website. And you would go into the Maternal Child Health section of that website, and then go into “documents,” and you would find it in there. And if you can’t find it, you can always email me, and I’ll just email you the link to it. And that’s all we have. That’s it.