Suppose I study 60 countries divided in two groups of developed and under-developed to guage location specific benefits of outward foreign direct investment, how to deal with too many differences in nature of economies, investment patterns and sources of data as well?
According to the classical econometric methodology (https://en.wikipedia.org/wiki/Econometrics ), the results are considered to be more appropriate if the studied variables are correlated (strongly correlated), the prediction more accurately correspond to the panel data and the more significant are the estimates obtained from the viewpoint of t- or F-statistics (or information criteria, for example Akaike - https://en.wikipedia.org/wiki/Akaike_information_criterion ).
The best way to predict the explanatory variable is the coefficient of determination should be as large as possible, and F-statistics should be as significant.
In addition for linear and non-linear models, the size of panel data can vary greatly.
ISSUES: there are two issues in the questions: (i) sample size and (ii) group classification that leads to heterogeneity
STAGES OF DEVELOPMENT: You use only two classifications, i.e. developed and undeveloped, but there are many more types of developmental stages in between. Generally, the classification has 5 categories, see attached files.
1. Factor-driven = 37 countries
2. Transition from factor-drive to efficiency-driven = 16 countries
3. Efficiency-driven = 30 countries
4. Transition from efficiency-to-innovation = 24 countries
5. Innovation-driven = 37 countries
By making only two categories, large variance would cause large heterogeneity.
SAMPLE SIZE: How many is adequate? A general answer is " a sample size adequate to provide distribution study is sufficient." Generally, we accept n = 30 or about 92% confidence interval. See RAMS article attached.
http://www3.weforum.org/docs/WEF_GlobalCompetitivenessReport_2014-15.pdf
Vedamani Basil Hans, perhaps it will help if you clarify what you mean by "ideal size".
By idea I mean "sufficeintly large" and analytically sound for results; less scope for errors and biases despite hetreogenity.......am I right?
For small sample sizes you may go for non parametric methods for example non parametric panel data. They will also place less restrictive assumptions. Or may use both parametric and non parametric and use model selection criteria like AIC or BIC to choose between them.
Dear all,
1) Panel data is a kind o longitudinal data measured on the same units, like people, at 2 or more moments in time.
Needs a large sample. According to Baldawin (1989) at least 200. Some researchers tell us that the ratio between the nº of people and nº of oberved variables should be at least 10:1 or, 15:1 or 20:1, meaning that the sample size should be at least between 100-200 observations (cases).
2) Complex samples in SPSS allow you to analyse those differences between samples in different countries.
Helena
I think, it must be clarified which are the units of the study population: 60 countries or the habitants, businesses, etc. of the 60 countries. In any case the problem of the sample size determination can be reduced to a sample size for the N = 60 countries in each time, or to sample sizes according the country, its habitants, businesses, etc. and time. Each case should be considered individually, or well to take the maximum sample size as common for all individual case since this guaranties the required precision or confidence level for all case.
Finite population sampling theory provides solutions for that.
- Badges
- Science topic
- Similar topics
- Social Psychology
- Attitude