dfsnow
commented
9 months ago Happy to help! Let me address your comments in turn:
First the most logistical question is as follows. I downloaded all the files and R packages and attempted to run 01-train.R from the pipeline folder. R threw me the error message Error in
all_of(): ! Can't subset columns that don't exist. ✖ Columnloc_tax_municipality_namedoesn't exist. Runrlang::last_trace()to see where the error occurred.
I suspect this is because you're using the 2023 input data with the current master branch of the model. The master branch has lots of ongoing changes related to the upcoming 2024 reassessment, so it is incompatible with the 2023 input data.
Instead, you should use the 2023 input data and then git checkout the 2023-assessment-year git tag. That tag is a snapshot of the model as it was used for last year's assessments.
Next a very high-level question. You indicate that the model estimates the sale price of all properties sold in a particular year using the characteristics of the property, neighborhood, etc. Do you then use that model to score all the properties (sold and not sold) for a tax year? If so, at what stage in the pipeline does this take place?
Yes, the trained model is used to predict the value of both sold and unsold property. This happens in the assess stage of the pipeline.
I also noticed that longitude and latitude are independent variables in the model. This seems quite strange. Could you help me to understand what role these variables play in the model.
The model uses mostly small geographies (neighborhood, township, etc.) to capture local spatial fixed effects. However, some of these geographies are quite large, and our thinking is that by including lat/lon, we can capture some intra-geography spatial variation in price.
Related, I tried to find a meaningful feature importance report on your website so that I could look at the importance of the longitude and latitude variable, but was not able to find that report. If it exists, can you please direct me to it?
We haven't published such a report since it tends to change from model to model. However, generally the most important features are exactly what you'd expect: location, square footage, and age of the property.
Last, as long as I have you, I believe I noticed a small mistake in your variable documentation (‘Features Used’). The first variable has the name ‘Percent Population Age, Under 19 Years Old’. In looking at the description, I believe it should be ‘…Under 18 Years Old’.
So this variable is indeed children under 19, you can see the definition here. However, our note attached to the variable is indeed wrong!
JohnSparks2024
Dear CCAO Staff, I want to begin by thanking you for making the assessment value modeling process as transparent and clear as you have. The additions to your web page that explain the process and major issues are very clear and informative. Allowing a member of the public to run the same process on a home computer is the ultimate in transparency and I applaud you for it.
I am particularly interested because I teach statistics and statistical software (including R) at UIC and I am planning a major addition and remodel of my house that will surely affect my assessed value and tax bill. I have looked at the tax bills of a couple of houses in my neighborhood that did something similar so I have a good rough estimate of what my tax bill will be after the addition but, because of the first reason that I listed, I wanted to dig deeper into the process.
After having looked over the documentation and installed the git and dependency packages, I have a few questions and comments, from the tall to the small.
First the most logistical question is as follows. I downloaded all the files and R packages and attempted to run 01-train.R from the pipeline folder. R threw me the error message Error in
all_of(): ! Can't subset columns that don't exist. ✖ Columnloc_tax_municipality_namedoesn't exist. Runrlang::last_trace()to see where the error occurred.I ran the trace, but it returned a giant tree diagram that I don’t understand. Could the source of the error be that I only loaded the 2023 data to the input folder? It didn’t seem sensible to load all of the 2023, 2022 and 2021 data to the same input folder because some of the data files had duplicate names.
Next a very high-level question. You indicate that the model estimates the sale price of all properties sold in a particular year using the characteristics of the property, neighborhood, etc. Do you then use that model to score all the properties (sold and not sold) for a tax year? If so, at what stage in the pipeline does this take place?
I also noticed that longitude and latitude are independent variables in the model. This seems quite strange. Could you help me to understand what role these variables play in the model.
Related, I tried to find a meaningful feature importance report on your website so that I could look at the importance of the longitude and latitude variable, but was not able to find that report. If it exists, can you please direct me to it?
Last, as long as I have you, I believe I noticed a small mistake in your variable documentation (‘Features Used’). The first variable has the name ‘Percent Population Age, Under 19 Years Old’. In looking at the description, I believe it should be ‘…Under 18 Years Old’.
Thank you for your time and attention. I look forward to hearing from you when it is reasonably convenient.
--John Sparks