Model Critique

Goal 1: Business Scenario

Audience

The audience for this analysis is a real estate investment and pricing strategy team working for a property analytics firm that advises clients on apartment pricing in major urban housing markets. In particular, the team wants to better understand how apartment characteristics relate to whether a unit is located in San Francisco versus New York, since those markets differ substantially in pricing, density, geography, and housing demand.

Problem Statement

A property analytics firm wants to evaluate whether apartment characteristics such as square footage, number of bedrooms, and elevation can meaningfully distinguish apartments located in San Francisco from those in New York. The broader business concern is that if the company relies on an oversimplified model, it may misclassify properties, misinterpret which features matter most, and provide weak recommendations to clients making pricing, investment, or development decisions. Therefore, the firm needs to determine whether the modeling approach used in the Week 11 lab is sufficient for high-stakes real estate decision-making or whether stronger analyses and more careful interpretation are needed.

Scope

The Week 11 lab uses apartment data and focuses on modeling whether an apartment is in San Francisco through a generalized linear modeling framework. The variables most directly relevant to this problem include measures such as beds, square footage, elevation, and transformed variables derived from apartment size or price-related structure. Within this scope, the lab can support an initial critique of how well the selected variables and model structure address the classification problem.

However, several assumptions would need to be made before using this analysis in practice. These include assuming that the available apartment features are measured consistently across cities, that the included variables are sufficient to distinguish the two housing markets, and that the sample is reasonably representative of the apartments to which the firm hopes to generalize. If these assumptions do not hold, then the model may produce conclusions that appear statistically reasonable but are not practically reliable.

Objective

The objective of this critique is to evaluate whether the Week 11 modeling approach is adequate for a realistic apartment-market decision setting and to identify specific analytical, statistical, and ethical improvements that would make the analysis more credible and useful in practice. Success will be defined by clearly identifying major weaknesses in the original approach, proposing stronger alternatives, and explaining how those improvements would better support business interpretation and responsible use.

Goal 2: Model Critique

The Week 11 lab introduces important generalized linear model ideas, especially maximum likelihood estimation, model comparison, and the addition of explanatory variables. However, in a realistic business setting, the analysis would need to be strengthened before it could support high-stakes decisions about apartment pricing strategy, market segmentation, or location-based recommendations. The critique below focuses on analytical limitations in the lab and proposes improved analyses that would make the results more credible and practically useful.

Critique 1: The lab emphasizes model comparison, but not enough model validation

One of the strongest ideas in the Week 11 lab is the comparison of models using measures such as deviance, AIC, and BIC. That is a useful starting point, since comparing models is better than interpreting one model in isolation. However, the lab does not go far enough in evaluating whether the selected model is actually reliable for classification in practice.

A model can outperform another model on AIC or deviance and still be poorly suited for business deployment. For example, a model may fit the training data better while still failing to generalize to new apartment listings. In a business context, this matters since the firm does not care only about which model looks strongest on the observed data. It also needs to know whether the model would remain useful when applied to future or unseen cases.

An improved analysis would include some form of out-of-sample validation, such as splitting the data into training and testing sets or using cross-validation. This would allow the analyst to evaluate whether performance remains stable when the model is applied beyond the sample on which it was estimated. In practical terms, this would provide stronger evidence that the model is not simply capitalizing on patterns specific to the original dataset.

Critique 2: The selected explanatory variables may be too limited for a realistic housing-market problem

The Week 11 lab uses a small set of explanatory variables, such as beds and elevation, to distinguish apartments in San Francisco from those in New York. While that simplification is understandable in a classroom setting, it creates a serious limitation in a real-world business scenario.

Housing markets are shaped by many factors beyond a small number of physical features. Variables such as neighborhood, distance to city center, building age, bathrooms, amenities, transit access, crime rates, and broader market conditions may strongly affect whether a property resembles one market or another. If important variables are omitted, the model may incorrectly attribute too much explanatory power to the few predictors that remain. This creates a classic omitted variable problem, where the coefficients may reflect missing context rather than the true independent contribution of the included variables.

A stronger analysis would expand the feature set or at least explicitly acknowledge that the current model is only a limited proof of concept. If additional data were unavailable, the critique should state clearly that the conclusions must remain narrow: the lab may show that certain variables help distinguish the two cities in this dataset, but it does not establish that those variables are the most important drivers of market differences in practice.

Critique 3: The lab does not sufficiently evaluate practical classification performance

The Week 11 lab explains model comparison through likelihood, deviance, and information criteria, which are statistically important. However, from a business standpoint, these measures are still incomplete since they do not directly answer a practical question such as: How often does the model correctly classify apartments?

In a real business context, the audience would likely care about evaluation metrics such as accuracy, sensitivity, specificity, precision, and possibly the ROC curve or AUC. These measures help translate the model into practical decision terms. For example, if the model is used to guide pricing or investment strategy, stakeholders need to know not only that one model has a lower AIC than another, but also whether the classification errors are frequent, systematic, or costly.

This issue becomes even more important if the classes are imbalanced or if one type of error is more costly than another. Misclassifying a San Francisco apartment as a New York apartment may have different business consequences than the reverse. Therefore, an improved analysis would evaluate classification quality more directly and would discuss the practical implications of different kinds of error.

Critique 4: The interpretation of coefficients may be statistically correct but practically incomplete

Generalized linear models often produce coefficients that are mathematically interpretable, but that does not automatically mean they are easy for stakeholders to understand. In a logistic-regression-style setting, coefficient estimates are typically interpreted on the log-odds scale, which is not intuitive for many business users.

If a model is presented to a pricing strategy team or investment group, then the analysis should move beyond formal coefficient significance and help explain what the coefficients mean in more accessible terms. For example, it may be more useful to discuss how changes in a variable affect the estimated probability that a listing belongs to one city rather than the other. Without that step, the results risk being technically correct but practically difficult to use.

A stronger version of the analysis would translate the model into interpretable quantities such as predicted probabilities, marginal effects, or example scenarios. This would improve communication and make the model more appropriate for decision support.

Critique 5: Better visualizations could improve both statistical understanding and stakeholder communication

The lab is strong conceptually, but in a business setting, visual communication matters just as much as mathematical explanation. Some of the analytical ideas in the Week 11 lab, such as comparing candidate models or understanding the contribution of explanatory variables, would benefit from clearer visual support.

For instance, predicted probability curves, class-separation plots, confusion-matrix-style summaries, or side-by-side visual comparisons of model performance would make the analysis easier to interpret. Better visualizations could also help identify whether the model performs well across the full range of the predictors or only in limited parts of the data.

This matters since business audiences rarely act on model summaries alone. They are more likely to trust an analysis when they can see how the model behaves and where its strengths or weaknesses appear.

Goal 3: Ethical and Epistemological Concerns

In a real business setting, the Week 11 lab would not only require stronger statistical analysis, but also a careful examination of the ethical and epistemological limits of the modeling process. Even if a model performs well numerically, it may still produce misleading, biased, or socially harmful conclusions if the data and assumptions behind it are not questioned.

Ethical Concern 1: Limited variables may reinforce misleading market narratives

One major ethical concern is that a simplified model may encourage decision-makers to believe that a small set of measurable apartment features fully explains differences between San Francisco and New York housing markets. In reality, housing markets are shaped by a much broader set of structural, geographic, historical, and social factors.

If the model relies too heavily on limited predictors such as bedrooms or elevation, it may produce an overly narrow narrative about what “drives” market differences. This matters ethically since simplified models can influence real decisions about investment, pricing, and neighborhood desirability. If those decisions are based on incomplete information, the model may reinforce distorted market views rather than provide responsible guidance.

Ethical Concern 2: Omitted social context can produce biased or unfair applications

A second concern is that the data used in the lab may leave out important contextual variables that reflect unequal housing access, urban development patterns, or neighborhood-level inequality. Even if the model itself is mathematically valid, it may still reflect structural biases embedded in the data.

For example, if the model were later extended into a pricing or recommendation system, it could contribute to decisions that favor already advantaged locations or reinforce existing disparities between neighborhoods. In this sense, the model is not ethically neutral simply since it is statistical. It may reproduce social patterns without questioning whether those patterns are equitable or desirable.

Ethical Concern 3: Misclassification may have practical consequences for stakeholders

In a classroom setting, classification errors may seem purely technical. In practice, however, misclassification can affect real stakeholders. A model that incorrectly categorizes apartments or overstates confidence in its predictions could influence client advice, pricing strategy, or investment decisions in ways that create financial or reputational harm.

This is especially important when stakeholders interpret the model as more certain than it really is. If a firm presents the analysis as authoritative without explaining uncertainty, then clients may make high-stakes decisions on the basis of a model that was originally built as a simplified instructional example.

Epistemological Concern 1: The model does not “know” the housing market in a full sense

From an epistemological perspective, one key issue is the difference between modeling a pattern and understanding a phenomenon. The Week 11 lab may identify statistical relationships in the observed data, but that does not mean the model fully captures the true structure of urban housing markets.

In other words, the model can provide partial knowledge, but not complete knowledge. It identifies regularities in the available sample under a particular set of assumptions. It does not prove that those regularities are universal, causal, or stable across time and place. This limitation is important since business users may mistake predictive success for genuine explanatory understanding.

Epistemological Concern 2: What is measurable is not always what is most important

Another epistemological issue is that the analysis focuses on variables that are available and quantifiable. However, the most measurable features are not always the most meaningful. Factors such as neighborhood reputation, long-term urban policy, housing discrimination, or informal market dynamics may be highly influential even if they are not included in the dataset.

This creates a risk of measurement bias, where the analysis privileges what is easy to quantify rather than what is most relevant to the actual problem. As a result, the model may appear more objective than it really is.

Epistemological Concern 3: Model success criteria depend on the purpose of the analysis

A final epistemological issue is that whether a model is considered “good” depends on the purpose for which it is used. A model that is acceptable for illustrating GLM concepts in a classroom may be inadequate for real estate investment decisions. This means that knowledge claims about the model are always conditional on context.

For this reason, the business scenario matters. The model should not be judged only by whether it can produce a statistically valid output, but also by whether it produces the right kind of knowledge for the decision at hand. If the goal is high-stakes strategic decision-making, then the threshold for credibility must be much higher than it would be for a classroom demonstration.

Final Ethical and Epistemological Reflection

Overall, the Week 11 lab is useful as an educational introduction to generalized linear models, but it would require much greater care before being used in a real business setting. Ethical concerns arise from omitted context, possible reinforcement of structural bias, and the real-world consequences of model error. Epistemological concerns arise from the fact that a statistical model captures only a limited version of reality and may privilege what is measurable over what is most meaningful.

Therefore, a responsible critique of this lab must go beyond technical model quality and also consider who is affected, what assumptions are being made, what kinds of knowledge the model can actually provide, and where its conclusions may be incomplete or potentially misleading.