How large of a dataset should I use for building a statistical model? I'm in the process of building a statistical model for housing sale prices. I am drawing inferences and trying to predict the price of a home as if it was sold in the year 2021. I am using a 600k+ row dataset from cook county, IL. However, roughly 40k rows are usable, and the rest contain too many null values for them to be of any use. LON and LAT coordinates are available for every row. I could use google API to extract the addresses and zip codes from the coordinates and then use the addresses with an MLS API to retrieve the missing information. This could potentially make 200k+ rows usable.
Is 40k+ row enough for statistical modeling from an initial 600k+ row dataset?
How many rows do you actually NEED?
 A: You should usually fill in missing data when you can
Whilst 40K+ rows is certainly a substantial dataset, the important issue here isn't so much the size of the dataset, but the question of whether or not the missing values in the data are "ignorable".  When we have a large dataset with missing values, and we propose to use only those entries that don't have missing values, that is called a "complete case analysis".  The danger of a complete case analysis is that the missingness of entries in the data could be systematically related to one or more of the variables under analysis, in which case ignoring records that have missing data will bias the analysis (sometimes severely).  Practically speaking, missing data is rarely ignorable, particular in cases where it affects a substantial proportion of the records in the overall dataset.
Dealing with missing data is an extremely complicated exercise, and the statistical theory and methods for this are quite advanced.  Proper methods for dealing with missing data generally involve either explicit statistical modelling of the "missingness" pattern or multiple imputation of missing values using explicit or implicit models.  This is difficult and time-consuming and it always comes with some modelling assumptions that are hard to test empirically.  Even with the best methods, having substantial amounts of missing data often leads to inferences that are highly uncertain or non-robust to modelling assumptions.  For this reason, if you have a cost-effective investigation method that allows you to fill in a substantial amount of your missing data, it is usually worth doing that.  Having a better, more complete, dataset is much better than having a patchy dataset and using missing data techniques on it (even if these are done well).
So while 40K+ data points is already a lot, I recommend you take your proposed action to fill in as much of the missing data as you can.  Increasing the size of your (complete case) dataset is one small advantage of this, but the much bigger advantage is that you will diminish the likelihood of getting a biased analysis due to data that is missing in a manner that is related to the variables of interest in your analysis.
