Innovation¶
This analysis estimates whether firm-wide match quality predicts subsequent issued-patent counts and quality. Zero-inclusive Poisson pseudo-maximum likelihood (PPML) is used for patent counts. OLS is used for real value and citations per issued patent, conditional on a positive cumulative issued-patent count.
Inputs
| Dataset | Location | Grain |
|---|---|---|
| Full-training Compustat--CRSP MQ panel | /data/ |
GVKEY--fiscal year |
| Time-profile Compustat--CRSP MQ panel | /data/ |
GVKEY--fiscal year and sample status |
One run attempts three samples: full-training in-sample, time-profile
in-sample (sample_status == "in_sample"), and time-profile out-of-sample
(sample_status == "out_of_sample"). Each writes to a separate output
folder. If an input dataset is unavailable, that sample is logged and
skipped while the others can still run.
Build
| Component | Location |
|---|---|
| Shared panel construction | analyses/modules/firm_mq_innovation.R |
| Regression and table module | analyses/modules/firm_mq_innovation_analysis.R |
| Pipeline entry point | analyses/pipelines/run_firm_mq_innovation.R |
| Runtime | hct-tower-analyses-cpu |
Output
The pipeline writes the same seventeen-table set for each available
sample. Full-training outputs remain in the existing directory; time-profile
results are under time_in_sample/ and time_out_of_sample/.
| Table | Location |
|---|---|
| Sample construction and stock-group shares | /artifacts/ |
| Issued-patent t+5, full sample | /artifacts/ |
| Issued-patent t+5, top-stock and bottom-stock groups | /artifacts/ |
| Full-sample horizons, industry-by-year FE | /artifacts/ |
| Full-sample horizons, firm plus industry-by-year FE | /artifacts/ |
| Top-stock horizons, industry-by-year FE | /artifacts/ |
| Bottom-stock horizons, industry-by-year FE | /artifacts/ |
| Top-stock horizons, firm plus industry-by-year FE | /artifacts/ |
| Bottom-stock horizons, firm plus industry-by-year FE | /artifacts/ |
| Log-count t+5 robustness | /artifacts/ |
| Log-count horizons, industry-by-year FE | /artifacts/ |
| Log-count horizons, firm plus industry-by-year FE | /artifacts/ |
| Log-count t+5, top-stock and bottom-stock groups | /artifacts/ |
| Top-stock log-count horizons, industry-by-year FE | /artifacts/ |
| Bottom-stock log-count horizons, industry-by-year FE | /artifacts/ |
| Top-stock log-count horizons, firm plus industry-by-year FE | /artifacts/ |
| Bottom-stock log-count horizons, firm plus industry-by-year FE | /artifacts/ |
The time-profile in-sample run also writes two fixed-2016-baseline tables
in /artifacts/<coverage>/analyses/firm_mq_innovation/time_in_sample/:
firm_mq_innovation_2016_baseline_controls_only.tex and
firm_mq_innovation_2016_baseline_industry_fe.tex.
Specification¶
All regressions use issued-patent outcomes. The t+5 table reports cumulative patents issued, real value per issued patent, and citations per issued patent for three specifications: controls plus industry and year fixed effects, industry-by-year fixed effects, and firm plus industry-by-year fixed effects. The t+5 stock-split table presents those same specifications for the high-stock group in columns 1--3 and the low-stock group in columns 4--6. Horizon tables report t+1 through t+5 results for the two industry-based FE structures, with separate high- and low-stock tables as robustness checks.
The high/low split is assigned once per MQ-qualified firm-year using trailing three-year issued-patent stock. The 90th-percentile cutoff is calculated on MQ-qualified firm-years with observed stock before requiring forward outcomes. Ties at the cutoff stay together in the high-stock group; the sample table reports realized group shares. The log robustness uses OLS for winsorized log(1 + cumulative patents issued) and mirrors the t+5 and horizon specifications in the full sample and top/bottom stock groups.
Every outcome, the focal Firm MQ measure, Firm MQ SD, and other continuous controls are winsorized at the 1st and 99th percentiles within each full-sample complete-case regression sample before splitting by stock group. The same outcome and continuous-variable winsorization limits are used for both groups. Prior issued-patent stock enters every regression as the winsorized log(1 + stock) control. Generated tables display the Firm MQ and Firm MQ SD coefficients with standard errors. Standard errors are clustered by GVKEY.
The data preparation continues to construct both filed- and issued-patent lagged stocks, but filed-patent outcomes are not estimated. The minimum MQ-score threshold defaults to 20 scored relationships per firm-year.
MQ baseline years are not capped at 2023. Patent outcomes are complete only through 2024, so eligible baseline years vary by horizon: t+1 through 2023, t+2 through 2022, and so on, with t+5 through 2019. Complete-case filtering applies these horizon limits separately to each regression.
The additional fixed-baseline specification uses the time-profile in-sample firm-year for 2016 as the sole source of Firm MQ and control variables, then relates those values to cumulative issued-patent outcomes beginning in 2017. Its five columns cover t+1 (2017) through t+5 (2021); each table reports the existing three outcomes (issued counts, citations per patent, and real value per patent). The two specifications are controls only and NAICS 3-digit industry fixed effects. Firm fixed effects are not identified in a one-year cross-section. The regular full-training and time-profile tables remain unchanged.
Command reference
| Argument | Required / default | Effect |
|---|---|---|
| config.yaml | Required positional argument | Analyses YAML config. Use analyses/config/config_full.yaml for full-sample output paths. |
| minimum_mq_scores | 20 in the analysis function | Minimum scored relationships per firm-year; not a CLI option. |
Use the firm-mq-innovation task for the current pipeline.
The pipeline writes timestamped progress to standard output, including input loading, stock-group shares, model fits, and table writes. It refits the models when run and does not expose a --rerun flag.