Saturday, October 21, 2006

Precipitation Rates vs. Population Density in Pennsylvania

Criticism of the autism-TV study out of Cornell has so far focused on the perceived plausibility of the hypothesis, and on author credentials, for the most part. For example, it has been pointed out that autism is present since birth/conception (or acquired in utero). If only it were that easy to dismiss post-natal environmental trigger theories these days.

The study was thought to be a prank at first, due to its peculiar methodology, but it's clear by now that it is not meant as a joke. Apparently, a lot of work went into it.

The authors of the study are economists, and this seems to be taken by many as a quick way to dismiss the study altogether. I would note that they are not the first authors of an epidemiological study on autism who are not trained in autism epidemiology. I think the study needs to be considered and critiqued on its merits.

Alternative interpretations have been proposed. Perhaps humidity causes autism, through mold and mildew, for example. Maybe spending time outdoors is a preventive measure. And so forth. These alternative interpretations, nonetheless, simply switch one hypothesis with another of equally unknown plausibility.

What I'd like to do in this post is propose an interpretation of the results based on a hypothesis known to be plausible, and which has been documented previously. Namely, that the results can be explained by a established correlation between degree of urbanization (or population density) with the administrative prevalence of autism. I already did something similar in a critique of the environmental pollution hypothesis.

That cable subscription rates are correlated with degree of urbanization or population density is a sure bet, and I won't consider that further. A potential correlation between population density and annual precipitation is non-obvious, however, although I had previously suggested that people generally don't build communities in the dessert.

I will use the state of Pennsylvania for this post, simply because there's readily available data on administrative prevalence with high regional granularity from that state. The authors analyze California, Oregon and Washington for precipitation. Readers can check if the analysis holds in those states. Fortunately, I didn't have to do much work to get the data I'm about to present, save for a few Google Image searches. Let me first post images modeling annual precipitation in the state of Pennsylvania.


Figure 1. Annual precipitation in Pennsylvania
(found at usda.gov)



Figure 2. Annual precipitation in Pennsylvania
(found at orst.edu)


Now, let me post an image with modeling of population density, and an image which shows the location of major cities in the state of Pennsylvania.


Figure 3. Population density of Pennsylvania counties
(found at dep.state.pa.us)



Figure 4. Pennsylvania road system, showing major cities
(found at smart-traveler.info)


Already we can see there is a fairly clear correlation between precipitation and population density in part of the state, perhaps even in 2/3rds of the state taken from East to West. Such a correlation is not at all clear in the West side of the state, but the correlation in the East side is probably enough of a confound.

Let me now post an image modeling autism prevalence across the state of Pennsylvania.


Figure 5. Pennsylvania IDEA autism cases per 10K enrolled
(found at fightingautism.org)


A correlation with either population density or precipitation is not that clear at first glance, is it? Let's focus on the bluest patches of prevalence, and let's compare with the Pennsylvania road map. You will note that high prevalence tends to coincide with urban areas, namely, Philadelphia, Harrisburg and Pittsburgh. Pittsburgh is particularly interesting, because it's an area with low precipitation. That's on the West side of the state, where I noted that a correlation between precipitation and population density was not clear.

Admittedly, there is an area near Erie (North-West corner) where there is high precipitation and high autism prevalence, but population density does not seem to be exactly where you would expect. This data point throws off my argument a bit, but it's not much to go on.

Going forward

The correlation between administrative prevalence of autism and population density lends itself to false correlations between autism rates and innumerable proposed contributing factors in autism. It would be extremely easy, for example, to "prove" the autism fries hypothesis using this type of analysis. In the future, regions with equivalent population density and wealth should be compared. Additionally, studies should be devised to determine if differences in autism prevalence between regions are actual or simply administrative.

Note: I plan to email Dr. Waldman and refer him to this post.

Addendum

I thought I would share a couple of remarkable images modeling annual precipitation and population density in the United States.


Figure 6. Annual precipitation in the U.S.
(found at ncgc.nrcs.usda.gov)



Figure 7. Population density in the U.S.
(found at wikimedia.org)


If you've ever wondered why population in the U.S. distributes the way it does, here's a big part of the answer. I have noticed that it's not so much that precipitation positively correlates with high population density, but that lack of precipitation correlates rather well with low population density. This is fairly obvious in retrospect.

20 comments:

  1. BTW, Washington state is easy. Seattle is known as the city where it rains a lot. The eastern side of the state is both dry and quite uninhabited apparently. Oregon has a similar pattern. High population density near the coast, with high precipitation; low population density in the eastern side of the state, where it's very dry.

    But I don't have a map modeling autism prevalence in those two states.

    California precipitation does not apparently correlate very well with autism prevalence -- just at a glance from my recall of the CDDS data. Namely, the area of highest prevalence (LA) is not the area of highest precipitation. The area of highest precipitation, Redwood Coast, has nearly the lowest autism prevalence in the state.

    ReplyDelete
  2. It's also important to realize that any correlation involving the administrative prevalence of autism is going to be by definition an ecological correlation, where the units of study are regions or similar aggregates. That's not the same thing as an individual correlation, where the units of study are, well, individuals. You can't automatically infer individual correlation from ecological correlation.

    For example, you'd probably find a fairly high correlation between the percentage of Muslims in a county and the percentage of Jews in that county (an ecological correlation, where it's county totals that are being analyzed). But the individual correlation between Islam and Judaism is precisely -1.00 since being a Muslim and being a Jew are mutually exclusive. The ecological correlation simply reflects, you guessed it, urbanization; both Muslims and Jews in the US live mostly in urban areas. As you can see, an ecological correlation can have a completely different sign from the "underlying" individual correlation.

    ReplyDelete
  3. It rains an awful lot in the Faroe Islands, they watch far less television, and autism rates are comparable.

    ReplyDelete
  4. It's also important to realize that any correlation involving the administrative prevalence of autism is going to be by definition an ecological correlation, where the units of study are regions or similar aggregates. That's not the same thing as an individual correlation, where the units of study are, well, individuals. You can't automatically infer individual correlation from ecological correlation.

    Would you say the Windham et al. (the Bay Area study) uses ecological or individual correlation? What she did was take a group of autistics selected at random from around the Bay Area, and compared with a control group. But I noted that the autistics she chose would be more likely to come from regions of higher prevalence than from regions of lower prevalence, so the urbanization confound still applies with that methodology. I even gave her an example, since that wasn't a clear explanation:

    As an example, suppose region # 1 has a total population of 10,000 children and a prevalence of 30 in 10,000; suppose region # 2 has a total population of 20,000 children and a prevalence of 60 in 10,000. You'd find 30 autistic children in region #1 and 120 autistic children in region #2; that's a 1:4 ratio. Your controls would have a 1:2 ratio.

    ReplyDelete
  5. Where does the connection of higher cable subscriptions come in to play with your analogy?

    ReplyDelete
  6. I think it's very much expected that areas with higher degrees of urbanization have higher rates of cable subscription. That's why I didn't even spend time on that.

    ReplyDelete
  7. The reason they have that spike of autism up near Erie is because they have 3 times as many people who belong to Yahoo! groups in Erie. Diagnosis of autism is strongly correlated with people who use Yahoo! groups... OK, well, it probably might be...

    They probably eat a lot of fish that contains PCBs there, or they fry their fish in hydrogenated oils and serve their own autism-fries, even if they don't have so many fast food restaurants.

    Thanks, Joseph. Have you contacted the authors of this paper with the ridiculously simplistic view of autism?

    ReplyDelete
  8. Thanks, Joseph. Have you contacted the authors of this paper with the ridiculously simplistic view of autism?

    I did, and invited them to comment. I didn't mention the ridiculously simplistic view of autism :)

    ReplyDelete
  9. Dear Authors of the TV/Rain-Kid autism hypothesis,

    Were you entirely serious in proposing this hypothesis? I keep feeling like someone is going to yell, "April Fools!" or maybe "Psych!"

    Is there longitudinal data on how many households use the Internet and for how many hours per day in Pennsylvania? I would think that kids might watch videos even where there isn't cable and that the hours of TV watching would be uniform across the state even where there is no cable, there still might be free old fashioned tv air waves, and as I said there's always video tapes.

    ReplyDelete
  10. Sean Nicholson, who is the stats person in the research team, replied and made some good points. I'll just summarize the discussion here as it develops.

    He noted that (1) in figure 7 of the paper, Los Angeles, Alameda, and San Diego counties had relatively low cable subscription in the early 1970s. I replied that there's no corresponding graph of prevalence by county to compare, and added:

    "I have data on caseloads from 1992 publicly available by request from CDDS. I went and checked the Westside Regional Center (in LA) and the Reginal Center of the East Bay (where Contra Costa is). I took the 14-17 and 18-21 age cohorts, and divided their caseloads by the corresponding Cerebral Palsy caseload. The ratios where 0.041 and 0.035 for WRC and 0.021 and 0.022 for RCEB. In other words, I'm pretty sure the LA area had a higher prevalence of autism in the 1970s birth year cohorts compared to Contra Costa county, which is inconsistent with your graph of cable subscription rates."

    He further noted that (2) the study uses what economists refer to as (county) fixed effects specification. "That means we control for differences between counties, including population density," he added, noting that population density does not vary significantly from one year to the next. I replied as follows:

    "The paper says you control for *time-invariant* county characteristics. Granted, population density might not vary considerably from one year to the next, but "population density" is just a proxy for the real likely confounds, which among others could be "awareness" and "availability of autism specialists" -- and these are not time-invariant by any means. Note that the largest autism prevalence increase occurred simultaneously with the largest availability of information increase in history, i.e. the internet revolution."

    Finally, I noted several additional problems with the study:

    1) Figure 2 of the paper shows that the biggest increase in cable subscriptions and VCRs was in the 1980 to 1990 timeframe. That does not coincide with autism.

    2) Precipitation can vary considerably from year to year, but there are no significant bumps or valleys in graphs of autism prevalence by birth year cohort. See this graph of precipitation in California as percentage of average.

    3) There is not necessarily equivalence between diagnoses of autism across counties. This basically can lead to a comparison of apples and oranges.

    (I will post updates if he replies)

    ReplyDelete
  11. Joseph,
    Thank you for the link. I read the critique. It seems spot on. While the analysis is skewed, I do think the hypothesis--that TVs trigger, cause, exacerbate--autism is worthy of further research.

    Observing my son (purely anecdotal) has at least proven to me a corrolation between more "spaced out" behavior when watching TV. Of course, that might be able to be said about all children--neurotypicals as well as autistic.

    ReplyDelete
  12. My severely autistic son did not start watching TV until he was about nine years old. He was diagnosed when he was two.

    It is hard to see how his autism could be caused by the TV when he didn't even watch till seven years after he was diagnosed.

    ReplyDelete
  13. Melissa & Teri,

    I'm sure there will be anecdotal accounts both ways, and that really doesn't tell us much. For example, my son has a low attention span even for TV shows. He starts watching something, but loses interest very quickly, so in the end he probably watches less TV than a non-autistic child.

    They need to follow-up this study with one that controls for the likely time-dependent confounds. And they need to do studies with different types of designs.

    ReplyDelete
  14. I've posted images modeling annual precipitation vs. population density in the entire U.S.

    ReplyDelete
  15. I think you are over playing the rainfall - population density relationship, it (at least to me) just does not seem that plausible that yearly fluctuations in rainfall within a region would be related to fluctations in population density or availability of autism specialists in any significant way (and since the authors used a fixed-effects model to take into account regional variations in autism rates, examining the precipitation - population density relationship on a purely geographic basis is not relevant).

    Since the time period hypothesised to be important is 0-3 years, the rainfall fluctations graph for California might be more relevant as a three year moving average (and of course would have less fluctuation).

    Regardless, a relationship between population density and cable subscription rate is enough in itself to cast serious doubts on any TV-autism hypothesis. I am hoping that the study authors will be able to include population density as a explanatory variable in a meaningful way, it will be very interesting to see how this all falls out.

    Finally, after all the knee-jerk reactions it is refreshing to see a serious and well reasoned critique of the study. I hope that the study authors take your critique very seriosuly.

    ReplyDelete
  16. Anonymous,

    I think you are over playing the rainfall - population density relationship, it (at least to me) just does not seem that plausible that yearly fluctuations in rainfall within a region would be related to fluctations in population density or availability of autism specialists in any significant way (and since the authors used a fixed-effects model to take into account regional variations in autism rates, examining the precipitation - population density relationship on a purely geographic basis is not relevant).

    In the cable subscription analysis, the confound is easy to follow. That is, something like 'availability of specialists' is not time-invariant and could easily be a confound for cable subscription rates in a county.

    Now, in the rainfail analysis, the researchers indicate they also did a time-based and a regional analysis with fixed-effects. The big problem I have with this is that the prevalence of autism by birth year cohort does not fluctuate as much as rainfal from year to year. See, for example, this graph of annual precipitation for Oregon. In fact, these fluctuations I think are pretty damning evidence against the rainfal correlation. Their regional correlation must have been strong, and the confound is clear there. On the time-based correlation, I'm highly skeptical there was one, and if there was, it must have been very weak and coincidental. The thing is that it appears that in many states in the US annual precipitation has been in a slight upward trend since about 1990, and that's probably all the time-based correlation is.

    I don't think the paper claims there's a correlation with the fluctuation in precipitation from year to year with fluctuations in the incidence of autism. In fact, there are no such fluctuations in the prevalence by birth year cohort for the most part.

    ReplyDelete
  17. I have a theroy that perhaps you could research: Is it in the water?
    I have heard that a protien inbalance in the brain causes problems with snapses connecting, leading to autism and perhaps altimers (excuse spelling).
    How is the prevelence reletive to how people recieve drinking water in PA?
    One municipalities treated sewage discharge is another municipalities drinking water. (downstream on the river).
    Detergents contain syntetic protiens. Polymers are used in wastewater treatment. could injestion of drinking water, containing trace amounts of this, cause neurological problems?
    mhedin@ch2m.com

    ReplyDelete
  18. Southeast and main Asian pandora jewelry countries have twisted rubies for centuries, cheap pandora bracelets but research as to where, and how to find more deposits is Pandora charms spare, and production has figured out how and mining companies,” Pandora beads Giuliani says, to look at exactly the right time and place.” pandora set Farther investigation of claret formation, based on tectonic scenery, cheap pandora geochemistry, fluid inclusions and isotopic ratios, allowed discount pandora Giuliani’s lineup to remodel a new prototype for the French Institute pandora 2010 of Research for Development (IRD) and the National Scientific pandora sale Center of Research, two government-sponsored knowledge Pandora Bangles and technology research institutes that aim to aid in the sustainable cheap pandora bracelets development of developing countries. Before the collision pandora bracelets prices of the Eurasian and Indian plates, lagoons or deltas sat in the regions where marble is giant, pandora bracelets and charms he says, “and there is the brains to expect that the new pandora bracelets sale thoughts should help development of the artless capital.” discount pandora bracelets Virginie Garnier, Gaston Giuliani and Daniel Pandora necklace Ohnenstetter urban the shape to do just that. They work for the garnet cheap pandora charms genesis. While studying the bedrock in Vietnam in 1998, the discount pandora charms French players found rubies, which detained traces of aluminum, chromium pandora charms sale and vanadium from universities, international corporations, governments pandora charms 2010 and why the rubies got there, and has created a paradigm Pandora beads to help these evaporites, Garnier says, when the Eurasian cheap pandora beads and Indian plates collided, raising the Himalaya Mountains.

    ReplyDelete