Problem solving works best if you can give the error a name.
Errors are best handled if you can clearly identify the type of mistake. In the Fall of 2013, The Appraisal Journal, published my paper, Common Statistical Errors and Mistakes: Valuation and Reliability. It identifies a couple of clever and popular “statistical” errors. (Pssss. . . Watch this space for a future announcement.)
Three types of error are possible, when working with data. Keep in mind that the goal of a value prediction is twofold: trueness and sureness. Trueness is accuracy. This means that your analytical process leads you to the theoretically ‘correct’ value prediction. Sureness is precision. This is how near you are to the true number, “on average.”
This occurs when you must estimate a population from a sample. There are two types of sample selection error, depending on if the sample is probabilistic (random) or not.
- Statistical sampling requires that the population be exactly defined, or completely listed. There must be random sampling or random assignment. Then you can run inferential statistics on the sample to see how well the sample represents the population. Statistics like p-value and confidence intervals tell you that. They do not tell you anything about how good is the model you chose.
- Judgment sampling is picking comps. The analyst uses data because it is convenient, or a personal confirmation was possible. No mathematical support of validity is available with judgment sampling. This is a “trust me” basis, where the client relies on the analyst’s judgement, experience and good ethics.
This error can be combined with others . . . An example of this is where the analyst pretends an imaginary population “of all possible sales that could have taken place,” (instead of the actual competitive market segment). The ‘expert’ then pretends that the judgment sample is somehow random – then runs inferential statistics mathematics to ‘prove’ how good the original model is. (A triple error, side out).
Clearly if you mis-measure a key predictor variable (like Sq.Ft size) the prediction will match the error. However, it matters a lot where the error appears. If it is the subject, the value prediction error will be exactly proportional to the measurement error. (If the size is 10% undermeasured, and the size accounts for 70% of the value – the value prediction will be 7% too low).
If the mismeasurement is in just one of the data points (a comparable), the impact will be much smaller – because other mismeasurements will tend to “cancel out” – so long as they are random high or low. (Of course, if all the comparables have errors the same direction, it is a different story.
There are several types of modeling errors possible: These include:
- Ask the wrong question, or make wrong assumptions (scope of work decisions)
- Input errors (typos, missing data)
- Omitted predictor variables
- Unnecessary variables (extraneous or inconsequential)
- Bad curve fitting
- Information loss (arbitrary aggregation)
- Failure to do a needed transform or rate/ratio
The good news is that modern data can completely eliminate the first type of error: sampling. The second type of error reminds us that defining and measuring the subject property is critical.
Vanquishing modeling errors is the core of the “new valuation modeling paradigm.”