A Complete Guide to Collecting Panel Data for Accounting Research
Tactical and methodological steps to collect panel data from publicly listed companies. Covers purposive sampling, data sources, and merging techniques.
Panel data (longitudinal data) is a combination of time series and cross-sectional observations. In Indonesian accounting research literature, panel data is highly favored due to its ability to control for unobserved individual heterogeneity and provide richer data variability.
1. Sample Selection Technique (Purposive Sampling)
The fundamental first step is establishing sample criteria. Errors at this stage can invalidate your entire research findings. Mandatory criteria to consider include:
- Listing Consistency: The company is listed on the Stock Exchange continuously during the observation period. Exclude companies that had an IPO in the middle of the observation period.
- Industry Sector Similarity: The financial sector (banks and insurance) is often excluded from the general sample (manufacturing sector) due to fundamental differences in regulatory structures and financial reporting.
- Reporting Currency: Ensure all reports are presented in the same currency. If a company reports in USD, you must convert it using the central bank's middle rate on the closing date, or exclude it from the sample.
2. Data Management: Handling Outliers (Winsorization)
Once data is collected, the biggest issue is extreme values (outliers). Instead of deleting extreme data (which could drastically reduce sample size), accounting researchers generally use the Winsorization technique. Values below the 1st percentile and above the 99th percentile will be replaced with the 1st and 99th percentile values, respectively.
// Example STATA Code for Winsorization
winsor2 ROA Leverage, replace cuts(1 99)3. Determining the Panel Data Regression Model
Panel data analysis is generally tested through three model specifications: Pooled OLS (Common Effect), Fixed Effect Model (FEM), and Random Effect Model (REM). You cannot choose a model subjectively; selection must be based on rigorous statistical tests.
A. Chow Test
This test determines which is better between the Pooled OLS or Fixed Effect model. If the Prob > F value is less than 0.05 (significant), then Fixed Effect is better.
// Example EViews Command:
// Open Equation window -> Panel Options -> Cross-section: Fixed
// Then View -> Fixed/Random Effects Testing -> Redundant Fixed Effects - Likelihood RatioB. Hausman Test
If the Chow Test favors Fixed Effect, the next step is the Hausman Test to choose between Fixed Effect or Random Effect. The FEM model assumes a correlation between the individual-specific error component and independent variables, whereas REM assumes the opposite.
// Example STATA Code for Hausman Test:
xtreg ROA Leverage Size, fe
estimates store fixed
xtreg ROA Leverage Size, re
estimates store random
hausman fixed randomC. Lagrange Multiplier Test (LM Test)
This test is used if the Chow Test selects Pooled OLS, and the Hausman Test selects Random Effect. The LM Test will be the final determinant between Pooled OLS vs Random Effect (Breusch-Pagan LM Test).
Case Study: Why is Fixed Effect Often Chosen?
In Corporate Governance research, the Fixed Effect Model is highly popular because it can control for company-specific characteristics (such as work culture, historical leadership style, or reputation) that do not change over time (time-invariant) but affect financial performance.
FAQ (Frequently Asked Questions)
Q: What is the minimum sample size for panel data regression?
A: As a rule of thumb, the number of observations (N x T) should be greater than 30. However, for model stability, the more (N > 50 companies, T > 3 years) the more robust it will be.
Q: What should I do if my data is an 'Unbalanced Panel'?
A: Unbalanced panel data occurs when some companies are missing data in certain years. Modern STATA and EViews can handle this well. Just ensure the missingness is random and not due to systematic selection bias.