All posts tagged aggregation-bias

The story of my dissertation

My dissertation is freely available to read on my website (Wheeler, 2015). I still open up my hardcover I purchased every now and then. No one cites it, because no one reads dissertations, but it is easily the work I am the most proud of.

Most of the articles I write there is some motivating story behind the work you would never know about just from reading the words. I think this is important, as the story often is tied to some more fundamental problem, which solving specific problems is the main way we make progress in science. The stifling way that academics write peer reviewed papers currently doesn’t allow that extra narrative in.

For example, my first article (and what ended up being my masters thesis, Albany at that time you could go directly into PhD from undergrad and get your masters on the way), was an article about the journey to crime after people move (Wheeler, 2012). The story behind that paper was, while working at the Finn Institute, Syracuse PD was interested in targeted enforcement of chronic offenders, many of whom drive around without licenses. I thought, why not look at the journey to crime to see where they are likely driving. When I did that analysis, I noticed a few hundred chronic offenders had something like a 5 fold number of home addresses in the sample. (If you are still wanting to know where they drive, they drive everywhere, chronic offenders have very wide spatial footprints.)

Part of the motivation behind that paper was if people move all the time, how can their home matter? They don’t really have a home. This is a good segue into the motivation of the dissertation.

More of my academic reading at that point had been on macro and neighborhood influences on crime. (Forgive me, as I am likely to get some of the timing wrong in my memory, but this writing is as best as I remember it.) I had a class with Colin Loftin that I do not remember the name of, but discussed things like the southern culture of violence, Rob Sampson’s work on neighborhoods and crime, and likely other macro work I cannot remember. Sampson’s work in Chicago made the biggest impression on me. I have a scanned copy of Shaw & McKay’s Juvenile Delinquency (2nd edition). I also took a spatial statistics class with Glenn Deane in the sociology department, and the major focus of the course was on areal units.

When thinking about the dissertation topic, the only advice I remember receiving was about scope. Shawn Bushway at one point told me about a stapler thesis (three independent papers bundled into a single dissertation). I just wanted something big, something important. I intentionally sought out to try to answer some more fundamental question.

So I had the first inkling of “how can neighborhoods matter if people don’t consistently live in the same neighborhood”? The second was that my work at the Finn Institute working with police departments, hot spots were the only thing any police department cared about. It is not uncommon even now for an academic to fit a spatial model at the neighborhood level to crime and demographics, and have a throwaway paragraph in the discussion about how it would help police better allocate resources. It is comically absurd – you can just count up crimes at addresses or street segments and rank them and that will be a much more accurate and precise system (no demographics needed).

So I wanted to do work on micro level units of analysis. But I had on my dissertation Glenn and Colin – people very interested in macro and some neighborhood level processes. So I would need to justify looking at small units of analysis. Reading the literature, Weisburd and Sherman did not have to me clearly articulated reasons to be interested in micro places, beyond just utility for police. Sherman had the paper counting up crimes at addresses (Sherman et al., 1989), and none of Weisburd’s work had to me any clear causal reasoning to look at micro places to explain crime.

To be clear wanting to look at small units as the only guidepost in choosing a topic is a terrible place to start. You should start from a more specific, articulable problem you wish to solve. (If others pursuing Phds are reading.) But I did not have that level of clarity in my thinking at the time.

So I set out to articulate a reason why we would be interested to look at micro level areas that I thought would satisfy Glenn and Colin. I started out thinking about doing a simulation study, similar to what Stan Openshaw did (1984) that was motivated by Robinson’s (1950) ecological fallacy. While doing that I realized there was no point in doing the simulation, you could figure it out all in closed form (as have others before me). So I proved that random spatial aggregation would not result in the ecological fallacy, but aggregating nearby spatial areas would, assuming there is a spatial covariance between nearby areas. I thought at the time it was a novel proof – it was not (Footnote 1 on page 9 were all things I read after this). Even now the Wikipedia page on the ecological fallacy has an unsourced overview of the issue, that cross-spatial correlations make the micro and macro equations not equal.

This in and of itself is not interesting, but in the process did clearly articulate to me why you want to look at micro units. The example I like to give is as follows – imagine you have a bar you think causes crime. The bar can cause crime inside the bar, as well as the bar diffusing risk into the nearby area. Think people getting in fights in the bar, vs people being robbed walking away from a night of drinking. If you aggregate to large units of analysis, you cannot distinguish between “inside bar crime” vs “outside bar crime”. So that is a clear causal reasoning for when you want to look at particular units of analysis – the ability to estimate diffusion/displacement effects are highly dependent on the spatial unit of analysis. If you have an intervention that is “make the bar hire better security” (ala John Eck’s work) that should likely not have any impact outside the bar, only inside the bar. So local vs diffusion effects are not entirely academic, they can have specific real world implications.

This logic does not explicitly always value smaller spatial units of analysis though. Another example I liked to give is say you are evaluating a city wide gun buy back. You could look at more micro areas than the entire city, e.g. see if it decreased in neighborhood A and increased in neighborhood B, but it likely does not invalidate the macro city wide analysis. Which is just an aggregate estimate over the entire city – which in some cases is preferable.

Glenn Deane at some point told me that I am a reductionist, which was the first time I heard that word, but it did encapsulate my thinking. You could always go smaller, there is no atom to stop at. But often it just doesn’t matter – you could examine the differences in crime between the front stoop and the back porch, but there is not likely meaningful causal reasons to do so. This logic works for temporal aggregation and aggregating different crime types as well.

I would need to reread Great American city, but I did not take this to be necessarily contradictory to Sampson’s work on neighborhood processes. Rob came to SUNY Albany to give a talk at the sociology department (I don’t remember the year). Glenn invited me to whatever they were doing after the talk, and being a hillbilly I said I need to go back to work at DCJS, I am on my lunch break. (To be clear, no one at DCJS would have cared.) I am sure I would have not been able to articulate anything of importance to him, but I do wish I had taken that opportunity in retrospect.

So with the knowledge of how aggregation bias occurs in hand, I had formulated a few different empirical research projects. One was the idea behind bars and crime I have already given an example of. I had a few interesting findings, one of which is that diffusion effects are larger than the local effects. I also estimated the bias of bars selecting into high crime areas via a non-equivalent dependent variable design – the only time I have used a DAG in any of my work.

I gave a job talk at Florida State before the dissertation was finished. I had this idea in the hotel room the night before my talk. It was a terrible idea to add it to my talk, and I did not prepare what I was going to say sufficiently, so it came out like a jumbled mess. I am not sure whether I would want to remember or forget that series of events (which include me asking Ted Chiricos if you can fish in the Gulf of Mexico at dinner, I feel I am OK in one-on-one chats, group dinners I am more awkward than you can possibly imagine). It also included nice discussions though, Dan Mear’s asked me a question about emergent macro phenomenon that I did not have a good answer to at the time, but now I would say simple causal processes having emergent phenomenon is a reason to look at micro, not the macro. Eric Stewart asked me if there is any reason to look at neighborhood and I said no at the time, but I should have said my example gun buy back analogy.

The second empirical study I took from broken windows theory (Kelling & Wilson, 1982). So the majority of social science theories some spatial diffusion is to be expected. Broken windows theory though had a very clear spatial hypothesis – you need to see disorder for it to impact your behavior. So you do not expect spatial diffusion, beyond someones line of site, to occur. To measure disorder, I used 311 calls (I had this idea before I read Dan O’Brien’s work, see my prospectus, but Dan published his work on the topic shortly thereafter, O’Brien et al. 2015).

I confirmed this to be the case, conditional on controlling for neighborhood effects. I also discuss how if the underlying process is smooth, using discrete neighborhood boundaries can result in negative spatial autocorrelation, which I show some evidence of as well.

This suggests that using a smooth measure of neighborhoods, like Hipp’s idea of egohoods (Hipp et al., 2013), I think is probably more reasonable than discrete neighborhood boundaries (which are often quite arbitrary).

While I ended up publishing those two empirical applications (Wheeler, 2018; 2019), which was hard, I was too defeated to even worry about posting a more specific paper on the aggregation idea. (I think I submitted this paper to Criminology, but it was not well received.) I was partially burned out from the bars and crime paper, which went at least one R&R at Criminology and was still rejected. And then I went through four rejections for the 311 paper. I had at that point multiple other papers that took years to publish. It is a slog and degrading to be rejected so much.

But that is really my only substantive contribution to theoretical criminology in any guise. After the dissertation, I just focused on either policy work or engineering/method applications. Which are much easier to publish.

References

Leave a comment
by apwheele on January 15, 2025  •  Permalink
Posted in Crime Analysis, Crime Mapping, Criminal Justice, scholarly, writing
Tagged ,

Posted by apwheele on January 15, 2025

https://andrewpwheeler.com/2025/01/15/the-story-of-my-dissertation/