Jevons Paradox in Technology: Do Hardware Cost Declines Drive Software Investment?

Empirical investigation of the Jevons Paradox in the technology sector, testing whether declining hardware/compute costs Granger-cause increases in real software investment using U.S. macroeconomic data (1997–2024).

Research Question

Do declines in hardware/compute costs help forecast increases in real software investment, consistent with the Jevons Paradox?

Methodology

• Data: FRED — BLS Producer Price Index for computers (PCU33443344), BEA real software investment (B985RX1Q020SBEA), Federal Funds Rate (FEDFUNDS)
• Stationarity: ADF and KPSS tests with structural break detection (2008 financial crisis, COVID-19)
• Cointegration: Johansen trace test (bivariate: log hardware price × log software investment)
• Models: VECM (long-run), VAR in differences (short-run), ARIMA (benchmark)
• Dynamic analysis: Impulse Response Functions (IRFs) with bootstrap CIs, Forecast Error Variance Decomposition (FEVD)
• Forecasting: Out-of-sample rolling forecasts with Diebold-Mariano comparison tests
• Robustness: Wage controls (ECI), Cloud/SaaS proxy, employment quantity proxy, GDP share normalization, rolling Granger causality, sub-sample analysis (post-2008, post-2020)

Key Findings

1. Long-run equilibrium exists: Johansen trace test confirms cointegration between log hardware prices and log real software investment — hardware costs and software demand are economically linked in the long run.

2. Error correction is significant: The ECT coefficient in the VECM is negative and highly significant, indicating that deviations from the long-run equilibrium are corrected over time through gradual adjustment.

3. Short-run effects are directionally consistent but weak: The VAR in differences shows a negative coefficient of lagged hardware price changes on software investment growth (consistent with Jevons), but this short-run effect is not statistically significant at conventional levels.

4. VECM outperforms VAR in forecasting: The long-run channel (cointegration/VECM) provides better forecasting performance than VAR in differences, while a simple ARIMA benchmark remains competitive in this small quarterly sample.

5. Results are robust to controls: Adding wage growth (ECI) and cloud/SaaS proxies does not materially change the hardware price coefficient, confirming the robustness of the main findings.

6. Policy implication: Hardware cost subsidies (e.g., CHIPS Act) may have long-run multiplier effects on software investment, but policymakers should not expect immediate short-run responses.

Repository Structure

File	Description
final_project_notebook.Rmd	Full R Markdown analysis — code, methodology, and documentation
final_paper.pdf	Final research paper with results and conclusions
final_project_notebook2.pdf	Supplementary analysis notebook

Tools & Packages

Language: R (RMarkdown)

Key packages: tidyverse · vars · urca · forecast · tseries · quantmod · fredr · ggplot2 · lmtest · sandwich · strucchange · patchwork

How to Reproduce

1. Install R (>= 4.0) and RStudio
2. Install required packages:

   install.packages(c("tidyverse", "quantmod", "fredr", "zoo", "xts",
                       "lubridate", "tseries", "urca", "vars", "forecast",
                       "lmtest", "sandwich", "strucchange", "ggplot2", "patchwork"))

3. Open final_project_notebook.Rmd in RStudio
4. Knit the document (Ctrl+Shift+K)

Note: Data is fetched directly from FRED (Federal Reserve Economic Data). No API key required for the default quantmod download method.

Author

David Stahl