Change to one-sided test / statistical significance

[BB Kim]

How do I change rgehdfe to one-sided test?
I want to change to one-sided to make it more statistically significant, so please also give me some tips to make the results more statistically significant

my current two-sided code is :

global firm_controls " log_asset_ sale_asset_ debt_asset_ market_to_book_ ppe_asset_ capex_asset_ cash_asset_ cashflow_asset_ div_asset_ "
global estab_controls "log_emp hour_emp strike_q3 shut_q3 seasonal_q3 disaster_q3"
global executive_controls "AGE co_tenure sex"
global ceo_controls "AGE ceo_tenure sex"
global cluster "vce(cluster estab_name)"
estimates clear
set more off

eststo: quietly reghdfe estab_tcr dummy_co_recession_worked $firm_controls $estab_controls $executive_controls, $cluster absorb(Company YEAR)

[Maarten Buis]

The purpose of an analysis is to honestly represent what you know and what you don't know. Often the correct result of your analysis is that you just don't know (something is insignificant). This is especially common with fixed effects models like yours, where you deliberately (for good reasons, but still) throw away a lot of information. That is frustrating, as journals are (for the wrong reasons) less likely to publish your article if you have insignificant results. However, trying to "torture" your model and data till you get significant results is fraud. For that reason, question like "give me some tips to make the results more statistically significant", makes me incredibly uneasy.

There is a legitimate version of that question: You have a specific hypothesis for one specific parameter, and that hypothesis happens to be one sided. In that case, and in that case alone, it is legitimate to want to maximize the power for the test of that single hypothesis. This is not something you do for a table of regression results. It is incredibly unlikely that you have a one sided hypothesis for all parameters in your model. This is something I would typically do in the text. In that case you just divide the p-value by 2, and you look at the direction of the effect. If the direction is against your expectation, then you reject the null-hypothesis (declare it insignificant) regardless of how small the p-value is.

Better yet, just avoid the arbitrary distinction between significant and insignificant. See this special issue of the American Statistician: https://www.tandfonline.com/toc/utas20/73/sup1