It is not clear what you want to find out with your data. As long as one does not know that and what is the dimension of N (countries, enterprises, prices?), it is impossible to give advice. But it is clear, that tests for serial correlation does not make sense for T=2. Before applyin econometric tests, you should (at least in principle) formulate hypotheses about relations between the 80 (?) variables (specification!).

PS: I doubt that your data set has, in fact, only two dimensions (T, N). With such data sensible results are hardly possible.

Anton Rainer N=80, there are 80 countries that participated in two surveys and, unfortunately, data on the next survey is expected not earlier than 2027. I am testing policy choices in financial inclusion and consumer protection in association with non-performing loans, assuming that inability to repay loans is a manifestation of negative impact on consumers' financial well-being and preferably be avoided. I do panel estimation in Stata with xtreg, fe; xtreg, re; xtreg, re mle; reg - all with vce(cluster ID) and justified with respectful tests. Plus, pcse, correlation(ar1) hetonly and consider pcse, correlation(psar1) hetonly, the latter corrects for cross-sectional dependency (CSD) in addition to heteroskedasticity and serial correlation. I do not have test justification for requesting correction for CSD in pcse specification. The estimation results are very similar, except for xtreg, fe (expected), but also for pcse, correlation(psar1) hetonly, where R2 is higher and much improved significance of many coefficients and decreased for a few.

Anton Rainer N represents a national regulator that provided responses to a survey organised by an international organisation.

Cross-sectional dependence refers to the correlation of error terms across individual units (such as countries, firms, or regions) in a panel data setting. When you have only two time periods (T=2), testing for cross-sectional dependence becomes challenging and less reliable. With T=2, you don’t have enough time variation to capture meaningful temporal patterns, which may obscure the true nature of cross-sectional dependence. Additionally, most tests for cross-sectional dependence, like the Pesaran test or Breusch-Pagan LM test, rely on larger time dimensions (T) to provide robust results. Therefore, while technically possible, testing for cross-sectional dependence with just T=2 is often not relevant or meaningful.

If two time series variables are cointegrated and one exhibits long memory (i.e., has autocorrelations that decay slowly), it does not necessarily imply that the other variable will also exhibit long memory. Cointegration indicates that the two series share a long-term equilibrium relationship, but it does not directly affect their individual memory properties. While cointegration suggests that deviations from the long-term relationship are mean-reverting, each series can have its own distinct short- or long-memory behavior. The long memory property in one series may reflect its specific dynamics, while the other series might not exhibit the same behavior, even though they are cointegrated.

Dear Hanna, I suppose that your data are aggregated (averages, sums, shares) responses to several questions of a survey for 80 countries for 2 (conseccessive?) years. With T only 2 (I would say one would need at least 5-10), you can foreget any econometric method which is partly based on time series analysis. I think you best try two separate cross section estimates for both years. Obviously, you want to find-out the influence of some of the variables on the (in)ability to repay (which I assume as the share of loans without repayment). Before any more complicated econometrics, I would print scatter diagrams with the pairs of variables and calculate correlations. Hopefully, you can then delelope a senseful specification, which can be tested.

I think it would be useful to enrich your data set with data from the national statistcs (which you can easily get from statistics of international organisations). Maybe that these data deliver better explanations than those from the survey.

Chuck A Arize Many thanks! Your comments are very helpful. My equation has 5 macro variables, including the dependent, which describe changes during the years around two surveys, three of the five are LN-differences, two other are LN of average annual increases. In the macroprudential part of the equation it is about how changes in four macro variables are association with changes in the dependent variable. Taking into account that surveys are conducted each five years, the macro changes are not overlaping and even not connected in time, the macro variables describe the year of the survey and one year before and after. My focus in the research is about other variables that describe institutional differences between countries, i.e. how these differences inflate or reduce negative responses to changes in macro situation. For example, effective credit reporting systems with mandatory reporting and checks before loan issuance are associated with lower growth in non-performing loans in times of deteriorating macro situation and accelerated decrease in good times. My current choice is not to do correction for CSD in PCSE estimation because (1) I cannot justify it with a test, (2) but also because this correction has a risk of inflating the SE and changing the estimation results much. In this case it looks like a black box I cannot uncover and trust.

Anton Rainer Dear Anton, your point is very fair for two consecutive years. My case is about surveys conducted once in five years and macro variables that describe changes around the years of the surveys without overlaps and even with a break between the surveys. The focus is on institutional differences that supplement traditional macroprudential models with a wide coverage of countries of different economic development. Possibly, rightly to your point, these extensive WB surveys are under-researched and lack attention. Facing these challenges, I intend to keep the panel model and two cross-section models and try to contribute with institutional variables to explaining changes in NPLs for a sample of rich and poor countries (inclusion of poor countries is a rare case for macroprudential models, which is another my concern and motivation).

Dear Anton Rainer and Chuck A Arize thank you for your contributions. My concluding remark is here. After repeatedly reading the Pesaran article on the test, it follows that the test has very low power even with T=5, i.e. it is not capable to detect cross-sectional dependence for very short panels (T