TELLUSANT, INC.

Summary of Cross-Sectional Demand Modeling Discussion

Objective

The discussion began with a methodological question:

Can differencing be used in cross-sectional statistical analysis?

This evolved into a practical investigation of a cross-country demand model for TVs per capita using income-distribution variables.

1. Differencing in Cross-Sectional Analysis

Standard Differencing

Differencing is normally a time-series technique:

\[\[ \Delta y_t = y_t - y_{t-1} \]\]

Its purpose is to:

Because cross-sectional observations do not possess a natural ordering, ordinary differencing across units is generally not meaningful.

When Differencing Can Make Sense

Differencing may be useful when:

  1. Comparing against a benchmark
  2. Comparing matched pairs
  3. Using spatial structures
  4. Working with panel data

Benchmark Example

For a country observed in year t:

\[\[ y_i - y_{US,t} \]\]

This measures deviation from a reference economy.

2. Reference-Series Transformation Idea

A proposed idea was:

  1. Observe income per capita for countries.
  2. Map observations to a reference series (e.g., U.S. income per capita over time).
  3. Difference relative to the reference.

This is valid if interpreted as:

Relative growth or relative position compared with a benchmark.

It does not become ordinary cross-sectional differencing.

3. TV Demand Problem

The empirical problem involved:

Dependent Variable

TVs per capita

Typically modeled as:

\[\[ \ln(TVpc) \]\]

Candidate Explanatory Variables

Shares or counts of populations in:

These were cumulative income categories.

Notation introduced:

Relationship:

\[\[ LM+ \ge M+ \ge UM+ \]\]

4. Why the Optimum Appeared Flat

A “flat optimum” emerged when comparing the three variables.

This is expected because the variables are highly collinear.

Example:

\[\[ LM+ = M+ + \text{Lower-Middle Only} \]\] \[\[ M+ = UM+ + \text{Middle Only} \]\]

Therefore all three variables carry nearly the same information.

The data may struggle to distinguish among them.

5. Recommended Decomposition

Convert cumulative classes into exclusive segments:

Exclusive Segments

Lower-middle only:

\[\[ LM_{only}=LM+ - M+ \]\]

Middle only:

\[\[ M_{only}=M+ - UM+ \]\]

Upper-middle and above:

\[\[ UM_{plus}=UM+ \]\]

Advantages:

6. Counts versus Shares

The data contained both:

Because the dependent variable was already per capita:

\[\[ TVpc \]\]

shares were generally preferred.

Why Shares Are Better

Counts introduce country-size effects.

Shares measure purchasing power distribution independently of population size.

7. Transforming Shares

Simple Share

Use the raw share directly.

Log Share

Possible but problematic near zero.

[ \ln(s) ]

[ \text{logit}(s)=\ln\left(\frac{s}{1-s}\right) ]

Advantages:

8. Cross-Validation

The discussion introduced cross-validation as a method of model comparison.

Purpose

Instead of asking:

Which model fits the observed sample best?

Ask:

Which model predicts unseen countries best?

K-Fold Cross-Validation

For K = 10:

  1. Split 162 countries into 10 folds.
  2. Fit the model on 9 folds.
  3. Predict the omitted fold.
  4. Repeat.
  5. Average prediction performance.

Recommended:

9. Cross-Validated R-Squared

Ordinary R-squared:

[ R^2 = 1 - \frac{SSE}{SST} ]

Cross-validated R-squared:

[ R^2_{CV} = 1- \frac{\sum (y_i-\hat y_i^{CV})^2} {\sum(y_i-\bar y)^2} ]

where predictions are generated from models that did not use the observation being predicted.

Interpretation:

10. Leave-One-Country-Out Cross-Validation (LOOCV)

Because there are only 162 countries, LOOCV is particularly attractive.

Procedure:

  1. Remove one country.
  2. Estimate on the remaining 161.
  3. Predict the omitted country.
  4. Repeat 162 times.

Advantages:

11. Nonlinear Specifications

Several nonlinear approaches were considered.

Splines

[ \ln(TVpc)=f(x)+\varepsilon ]

Advantages:

Recommended as the first nonlinear model.

Logistic Models

[ TVpc = \frac{A} {1+\exp(-k(\ln GDPpc-c))} ]

Interpretation:

Threshold Models

Piecewise regressions allow different slopes before and after a threshold.

Useful when adoption accelerates after a certain income level.

12. Empirical Result

The strongest predictor was found to be:

[ \ln(TVpc) = \alpha + \beta \cdot \text{logit}(UM+) + \varepsilon ]

where UM+ denotes the share of population in the upper-middle-income-and-above segment.

This specification outperformed alternatives.

13. Interpretation of the UM+ Result

A key insight emerged:

TV ownership appears to depend more on:

The proportion of the population able to comfortably afford the product

than on:

Average national income.

This corresponds to an affordability-threshold model.

14. Why Income Distribution Can Beat GDP Per Capita

Two countries may have identical GDP per capita but different income distributions.

Example:

Country A:

Country B:

GDP per capita cannot distinguish them.

UM+ share can.

15. Diffusion Interpretation

Durable goods such as:

often follow S-curve adoption patterns.

The upper-middle share acts as a proxy for:

Households crossing the affordability threshold.

As this share grows, adoption accelerates.

Eventually saturation occurs.

16. Testing the Affordability Threshold Hypothesis

Recommended comparison:

Model 1

[ \ln(TVpc) = f(\ln GDPpc) ]

Model 2

[ \ln(TVpc) = \beta \cdot \text{logit}(UM+) ]

Model 3

[ \ln(TVpc) = f(\ln GDPpc) + \beta \cdot \text{logit}(UM+) ]

Compare using:

17. Residualization Strategy

To separate income effects from distribution effects:

First estimate:

[ \text{logit}(UM+) = g(\ln GDPpc) + u ]

Then estimate:

[ \ln(TVpc) = f(\ln GDPpc) + \gamma u + \varepsilon ]

Interpretation:

If γ remains significant:

Income distribution contributes beyond average income.

This is strong evidence for a true affordability-threshold effect.

18. Main Conclusions

  1. Differencing is generally not meaningful in pure cross-sectional analysis.
  2. Cross-validation is the preferred tool for choosing among competing predictors.
  3. LOOCV is highly appropriate for a sample of 162 countries.
  4. Shares are preferable to counts when modeling per-capita demand.
  5. Logit transformations are often superior to raw shares for bounded variables.
  6. The best-performing specification identified was:

[ \ln(TVpc) = \alpha + \beta \cdot \text{logit}(UM+) + \varepsilon ]

  1. This result supports an affordability-threshold interpretation of consumer durable demand.
  2. Income distribution may explain adoption better than average income alone.
  3. The next recommended step is to compare income-only, UM+-only, and combined models using LOOCV and cross-validated R-squared.