Significance - Clustered Standard Errors

[Teiren Mistry]

Hello!

In my regression, my dependent variable is University (Decision children make on whether they want to go to University). Independent variable is Private Schooling (Categorical for Private/ State schooling). In the sample there are a small proportion of people in Private compared to State Schooling, which perhaps is causing insignificance in the Private Schooling Variable of 0.7.

I’ve tried using Clustering Robustness with respect to Private Schooling and it reduced my p-value to 0.053. Is using VCE (cluster, Private_school) a viable method to use for robustness in my Logit model? Or is there any other method I can use to reduce my p-value to become significant.

I have attached a photo so you can see the difference in p-values and the code I used to do the clustered standard errors.

Any advice would be much appreciated!

[George Ford]

how many values does private_school have? You need many.

[Leonardo Guizzetti]

You only have 2 levels (clusters) of private school. Clustering requires a larger number to be valid. How much? It depends on the degree of clustering and the distribution of cluster sizes, but usually 20 or 30 or more are recommended rules of thumb. I would proceed without cluster-robust VCE.

[Carlo Lazzaro]

Teiren:
you may want to take a look at Cameron_Miller_Cluster_Robust_October152013.pdf (ucdavis.edu)

[Maxence Morlet]

In addition to the great advice already provided, in your shoes I would either: use robust standard errors, switch clustering variables to one which has at least circa 42 clusters (hitchhiker's guide to the galaxy!)

[Teiren Mistry]

Thanks for the advice all!

[George Ford]

Also, don't choose your model based on statistical significance. That's p-hacking. Make sure your model is as legit as possible and live with what you get. You aren't interested in a p<0.05, you're interested in the truth.