I made a point in my last point about differentiating between a change in the crime rate and a significant change. I want to elaborate on that point a little more, because this is something that is so often overlooked but is vital to understanding any analysis. So if you know all of this already, feel free to skip the next two paragraphs.

When someone quotes a stat, like saying that overall major felonies were down 1.6% in 2015, they are only presenting part of the story. It is not a lie - crime was really down 1.6% in 2015 - but you need to ask is that significant, ie, is that something real or could it be random noise. This is where a more adroit publication would quote a significance level, or p value, but I personally do not find them intuitive. So here is my explanation in a nutshell. Something like the crime rate is going to have a natural fluctuation. Crime stats are a complex system, a lot of factors move the stats, and this all contributes to what amounts to random noise.

Luckily, we have ways of dealing with this randomness, and determining the degree to which stats are part of it or not. We ask the question “what is the chance that this stat is because of noise, versus a real effect” and say something is significant if it is above a certain threshold. Commonly, that threshold is 95%, but it is important to recognize that it is not binary - the probability exists on a spectrum, even though we use this shorthand of calling something “significant” or not. We do this by calculating the variance of the changes, which is a measure of how much they bounce around. Through a simple transformation we turn this into a standard deviation. Back to our example, the standard deviation in the crime rate in NYC from 2001 to 2014 is 3.97%. Using what is called a normal distribution (which is a pretty awesome math thing in its own right), we can judge how likely particular values are to be noise. The normal distribution tells us that at one standard deviation, i.e. a 3.97% rise or fall, there is a 32% chance that what you are measuring is noise. And at two standard deviations, i.e. a 7.94% rise or fall, there a 5% chance that what are are measuring is noise, or said differently, you are 95% confident in the measure. That is a normal cutoff- 95% - but as you can see, there is more to it than just a significant or not explanation.

What does this mean for the NYC crime data? The 2015 stats suggest a 1.6% drop in major felonies, however the standard deviation of the crime rate is 3.97%, so this nets out to about a 69% chance that what we are measuring is noise. So that is to say, we cannot really say much. The drop in crime rate for 2015, based on these assumptions, is just not significant.

There are some precincts though which do have a significant change in crime rate in 2015:

The picture does not look great, of the 6 precincts that had a significant change (using the 95% cutoff), 5 had a turn for the worse, and there is some evidence that the one good one is a misnomer. Here is a quick rundown of what I was able to find for each:

• 1st Precinct (Tribeca area) - 2.5 stars on Yelp, seems like grand theft auto is on the rise in that area.
• 32nd and 34th Precinct (Harlem) - looks like they have a good connection with the community, however the head cop transfered out in 2015 (story). It looks like there was a significant rise in murders in these precincts.
• 40th Precinct (Mott Haven, Bronx) - with a 25% rise in crime in 2015, this begs for some serious questions about what is going on. And it looks like the NYT did some in depth reporting on this. Also, this precinct was caught juicing their 2014 stats, which makes any numbers coming from this precinct highly suspect (and statistically speaking will make it difficult to make any assertions about crime stats moving forward - so the impact of this will last for years).
• 105th Precinct (Queens Village)- this is a huge precinct, and it looks like it is split to better serve the community.
• 104th Precinct (Ridgewood, Queens)- the one darling, with a 10.8% drop in major crimes, however cursory investigation yields articles like this and this, suggesting that the drop in crime rate has more to do with cops not responding to calls (earning them 1.5 stars on Yelp).

A lot has been made recently about a jump in the crime rate, so I decided to take a look at the major felony rates in NYC, with the help of the NYC OpenData project. There you can find records of all the major felonies reported, going back to 2006, with location and felony classification. Off the bat, it is pretty easy to look at the number of crimes committed over time:

As you can see, crime is down overall. But in all honesty, I look at that graph and I see a pretty steady line, dominated by grand larceny (ie, stealing). But, lets dive in, and look at the rates of change of individual crimes:

Ok - so one conclusion to draw from this is that there is a 5.71% uptick in murders and a 5.83% uptick in rapes in 2015. While tragic in their own right, taken in isolation I would advise that those rates are a little misleading. If you look at the totality of the history we have, going back to 2001 at the aggregate crime level, one standard deviation for the rate of change of murders and rapes are 6.49% and 7.94%, respectively. So that is not to say that the upticks are a good thing, because they certainly are not, but they are also not necessarily indicative of a regression in crime rates, rather they seem to be within the normal fluctuations of the system. Contrast that with burglary rates, which dropped by over 10% in 2015 and have an annual standard deviation of 4.11%. This is more likely a relevant drop.

Worst Areas

The top three police precincts, ranked by total number of the 7 major felonies, are 75 (East New York, Brooklyn), 14 (Midtown South, Manhattan), and 43 (Sound View, Bronx). If you are surprised by Midtown making the list, it is because of an extremely high count of larcenies - 16% higher than any other precinct.

Changes in Crime Rate

Precincts in blue have a drop in crime rate (good), whereas red has an increase in crime. The top three police precints, ranked by the rise in total crime rates of the 7 major felonies, are 115 (Jamaica, Brooklyn), 88 (Clinton Hill, Brooklyn), and 108 (Long Island City).

Methodology

The map plots in this writeup are done using the basemap package from matplotlib. NYC provides useful shapefiles for all of its census tracts, those can be found here, and you can also pull down the crime rate data from NYC OpenData. I am happy to make available any raw data/code used, just ask.

One of the datasets available on the amazing NYC bike data website (available here) is information on accident rates, sorted by police precinct and accident type. They make this information available in the dreaded PDF format, and it took me a little while to get over this, however with the help of Tabula I was able to convert the PDFs into something useful.

There are four different types of crashes, defined by what the bike crashes into: car, pedestrian, themselves, and other bikes. Presumably “themselves” encompasses a miscellaneous assortment of tragic collisions, including street poles, pot holes, and the occasional spontaneous combustion.

In 2015 there were 5270 crashes reported, with collisions between cars and bikes being by far the most representative with 83.7% of all accidents. I imagine there is a reporting bias here, in that these types of collisions are more likely to get picked in a police report than other types. Car collisions are followed by self collisions (8%), pedestrian collisions (6.6%), and lastly collisions with other bikes (1.7%). It is interesting to speculate why car crashes have such a higher representation - as a cyclist in NYC I would apriori expect pedestrian crash rates to be about as high if not higher - but I do not see a clear reason in the data.

One might think car crashes are reported more is because of more damage (both in terms of property damage and personal injury). We cannot judge property damage rates from the data, however the dataset does contain injury rates. They are all pretty close - with a car, 50% of the time the bicyclist is injured, with another bike 45%, themselves 42%, and with pedestrians only 4% (although pedestrians suffer some sort of injury in those accidents at a high rate - 52% of the time). So it is unclear why bike on car crashes are so highly represented, unless it is just the reality. What is perhaps more clear is that as expected, in asymmetical exchanges, the lower mass/more vulnerable median is more likely to get hurt (between bike and car, it is 0.6% for car vs 50% for bike, and between bike and pedestria it is 4% for bike and 52% for pedestrian).

Since this data is indexed by police precinct, it is possible to plot accident rates by geographic area. Below is a plot of bike accidents with cars in 2015, with accident rates normalized by the 2010 census population in each precinct.

It is a little disheartening to note that my normal commute crosses almost all of the most dangerous precincts. Oh well…

Methodology

The plots in this writeup are done using the basemap package from matplotlib, and a helpful tutorial here. NYC provides the data on collisions rates here. NYC also provides all sorts of useful information by census tract, including population, here. Since the collision rates are reported by precinct and the population is by tract, you need a way to translate census tract into police precincts, for which I found a static mapping that you can pull down here.

Of all the services that I have running to pull down data on my life, the survey system is arguably the most important. Ultimately, a major goal is to eventually be able to predict things like my happiness and concentration, and perhaps even help guide them. As such, finding a high quality way to pull down data on my mental state is important.

To achieve this, I have a very simple service that periodically sends me a predefined survey in a text message, to which I respond with measures for how happy, alert, and concetrated I am at that particular moment. The survey times are somewhat random, but garaunteed to be delivered throughout the day.

My analysis is only going to be effective if I get sufficient data, though I need to be realistic in that I cannot expect to answer my survey every 10 minutes. Or even every couple of hours. Realistically, sometimes I just do not want to respond. The survey system understands this, and through a response feedback loop will decrease or increase its question frequency based on my engagement level.

So, that said, what does it look like so far? I have some interesting results, nothing ground breaking so far. First, when it comes to responding, I am most likely to respond to my surveys in the afternoon - 33% of my responses came between 1 and 5, compared to only 20% coming in the morning. When it comes to my survey itself, ideally I wanted to pick indepent attributes, with low correlations. It looks like I mostly achieved this - the correlation between most pairs is around 35%, however the correlation between concentration and happiness is around 50%. That is in someway an interesting result in of itself, however it probably means I need to redesign the survey to have more independence in the measures.

There are two early questions that I wanted to answer:

1. Am I a morning person?
2. Does biking to work make my day better?

Am I morning person?

So far it is a little hard to tell, but early indications suggest that I have a minor (1%) lift in concentration in the morning, despite having a negative 3% draw on alertness. I can qualitatively corroborate this, however I will be interested to see how this evolves as more data comes in. Also worth noting - while the data suggest I concentrate the most in the morning, it is clear that I am happiest in the evening, with an 11% lift in happiness coming after 6pm (morning is ok, and then there is a big dip midday).

Does biking to work make my day better?

It looks like there is some immediate impact, but it wears off pretty quickly. On days that I bike to work, I experience about a 3% bump in happiness, and a 6% bump in concentration. That said, over the course of the whole day it averages out, and the only affect at the day level might be a slight uptick in volatility in all the measures. But it appears to be minor and may just be noise at the moment.

So where to go from here? I may need to tweak my survey to measure a more differentiated set of attributes, and I need to find a way to bump up my midday happiness (I am starting a new job soon, will be interesting to see what impact that has on this midday lull).