I have a dataset of unbalanced panel where in time period is 8 years but, the data is not available for completely 8 years for all the variables for all the companies. so the final sample size is measured in firm-years, but not as balanced panel data.
but, the regression model, what i am planning to use, has lagged variables.
How to introduce lagged variables in such situation?
thanks in advance
Hi Dr. Potharla,
The prior response to your question precisely addresses how to create lagged variables in an application involving missing data. No matter if missing data are attributable to structural (e.g., absence of a lagged variable for the initial point in the series; some variable series terminate prior to end of study time domain due to policy-related actions) or to empirical (e.g., data not available for other reasons), missing data cripple the efficacy of statistical/mathematical modeling.
https://odajournal.com/2014/09/02/triage-algorithm-for-chest-radiography-for-community-acquired-pneumonia-of-emergency-department-patients-missing-data-cripples-research
Missing details concerning your research hypothesis, design, and measure profiles makes it impossible to understand your specific issue(s) sufficiently to offer targeted solutions regarding circumventing or minimizing the effect of missing data on the analysis. Thus I have a few nonspecific suggestions regarding alternatives to regression analysis—which may or may not be applicable.
Legacy regression analysis is less accurate, and is more seriously affected by missing data (all observations with missing data on any variable in the model are deleted) compared to modern methods--regardless of whether the outcome measure is ordered or categorical.
https://odajournal.com/2016/09/19/novometric-analysis-with-ordered-class-variables-the-optimal-alternative-to-linear-regression-analysis
https://odajournal.com/2013/11/11/univariate-and-multivariate-analysis-of-categorical-attributes-with-many-response-categories
Novometric methods are most accurate and least sensitive to missing data (only observations missing data actually used to classify the observation are dropped from analysis).
https://odajournal.com/2016/12/10/novometrics-vs-oda-vs-log-linear-model-in-analysis-of-a-two-wave-panel-design-assessing-temporal-stability-of-catholic-party-identification-in-the-1956-1960-src-panels
If apropos to your research hypothesis, Markov models might be a superior alternative.
https://odajournal.com/2017/01/13/novometric-comparison-of-markov-transition-matrices-for-heterogeneous-populations
https://odajournal.com/2019/02/18/optimal-markov-model-relating-two-time-lagged-outcomes
Nothing to add. Best solution, of course, is more complete data.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size