clustering standard errors on state level with logit. should I also add state level fixed effects ?

[AK Kaakeh]

Hello,

I am analyzing a cross sectional data of loans in the US using a logit model. the dependent variable is a dummy with 1 is the loan defaulted and 0 if the loan is paid in full. the data covers 51 states and 11 years ( 2007 to 2018) but it is cross sectional data, so the loan will appear once. the main dependent variable is a state level variable ( political corruption) . It is important to mention that the independent variable changes every year and changes from a state to another. Now because it is a state level variable, I clustered the standard errors on the state level and the model worked fine even after adding year dummies.

while it is a common practice in similar papers with panel data to add state fixed effects (dummies) to the model, my data is a cross sectional one and once I add state level dummies to the model, the whole model goes haywire and the dependent variable becomes not significant.

Now correct me if I am wrong , but it seems that once I add state fixed effects (dummies) to the model, I compare the same constant variable across observations in this state and year, hence my model should be fine without the state level dummies as long as I am clustering standard errors on state level.

I hope I explained the problem clearly. Thanks in advance.

[George Ford]

If your DV is state level as you say (pol corruption), then adding state fixed effects is going to be a problem. Do you mean corruption is a covariate in a logit of loan default?

[AK Kaakeh]

the pol corruption is an independent variable ( a covariance in a logit of loan default as you put it) , the DV is a binary variable represents loan default. but even though it is an independent variable , I still see it will be a problem and I need a confirmation

[AK Kaakeh]

model tht works: logit default corruption control variables, cov ( cluster state)
model that does not work: logit default corruption i.state control variables, cov ( cluster state)

[AK Kaakeh]

Hello,

I am analyzing a cross sectional data of loans in the US using a logit model. the dependent variable is a dummy with 1 is the loan defaulted and 0 if the loan is paid in full. the data covers 51 states and 11 years ( 2007 to 2018) but it is cross sectional data, so the loan will appear once. the main dependent variable is a state level variable ( political corruption) . It is important to mention that the independent variable changes every year and changes from a state to another. Now because it is a state level variable, I clustered the standard errors on the state level and the model worked fine even after adding year dummies.

while it is a common practice in similar papers with panel data to add state fixed effects (dummies) to the model, my data is a cross sectional one and once I add state level dummies to the model, the whole model goes haywire and the dependent variable becomes not significant.

Now correct me if I am wrong , but it seems that once I add state fixed effects (dummies) to the model, I compare the same constant variable across observations in this state and year, hence my model should be fine without the state level dummies as long as I am clustering standard errors on state level.

I hope I explained the problem clearly. Thanks in advance.

I see my mistake, I wrote that the pol corruption is a DV, it is an independent variable and that was a typo. the DV is the binary default variable.

[Jeff Wooldridge]

AK: I think you mean you have pooled cross sections, as you mention 11 years. So the loans are unique from year-to-year and you don't have panel data.

Does the corruption variable change over time as well as across state? How many loans, on average, per state?

[AK Kaakeh]

AK: I think you mean you have pooled cross sections, as you mention 11 years. So the loans are unique from year-to-year and you don't have panel data.

Does the corruption variable change over time as well as across state? How many loans, on average, per state?

yes, it is a pooled cross sections and the loan is unique from year to year. Indeed the corruption variable changes over time but not always, for instance you can see a rise in crises years. Finally, the sample size is 1,345,000 loans and they vary a lot from 196,524 loans in CA to 7 loans in IA . but the average is 26,000 loan per state aggregated across 11 years.

[Beth Plunkett]

yes, it is a pooled cross sections and the loan is unique from year to year. Indeed the corruption variable changes over time but not always, for instance you can see a rise in crises years. Finally, the sample size is 1,345,000 loans and they vary a lot from 196,524 loans in CA to 7 loans in IA . but the average is 26,000 loan per state aggregated across 11 years.

Have you managed to get to the bottom of this AK? I have the same question

[AK Kaakeh]

Have you managed to get to the bottom of this AK? I have the same question

hi Beth,

I did , I clustered standard error on "state and year level" and added year dummies and other dummies to the regressions. check this paper "Pursiainen, V. (2020). Inaccurate Information in Marketplace Loans. Available at SSRN 3326588"

hope this helps.
regards

[Beth Plunkett]

hi Beth,

I did , I clustered standard error on "state and year level" and added year dummies and other dummies to the regressions. check this paper "Pursiainen, V. (2020). Inaccurate Information in Marketplace Loans. Available at SSRN 3326588"

hope this helps.
regards

Thank you so much, thats really helpful!