Regression with frequency weights

[Sebastian Grein]

I conducted propensity score matching and have now a dataset with observations which treated and untreated variables and the respective fweights for each observation.

How can I run a regression using the fweights? I tried the following, where _weight is the variable indicating the weight of each observation:

regress deal total_acc_freq_5 BV_ln_lag1 BTM_lag1 LEV_lag1 ROA_lag1 CAPEX_lag1 Cash_lag1 i.Year i.SIC_two [fweight=_weight]

But it results in the error: "may not use noninteger frequency weights"

[Rich Goldstein]

why do you think frequency weights are what you have - seems unlikely to me - my guess is that you actually want analytic weights, but of course I can't be sure given the very limited information that you supplied; at any rate, in Stata, and as implied by the error message, frequency weights must be integers and if you have non-integer weights then, as far as Stata is concerned, you do not have frequency weight

[Sebastian Grein]

I assumed that I have to use fweights since I read on the following website, that I have to use fweights after constructing the sample with Propensity Score Matching: https://www.ssc.wisc.edu/sscc/pubs/stata_psmatch.htm

Which kind of information do you need? Does this help?

* Example generated by -dataex-. For more info, type help dataex
clear
input str6 cusip_a float Year double _weight
"000307" 2017                 .
"000307" 2018                 .
"00032Q" 2022                 .
"000360" 2000                 .
"000360" 2001                 .
"000360" 2002                 .
"000360" 2003                 .
"000360" 2004                 .
"000360" 2005                 .
"000360" 2006                 .
"000360" 2007                 .
"000360" 2008                 .
"000360" 2009                 .
"000360" 2010                 .
"000360" 2011                 .
"000360" 2012                 .
"000360" 2013                 .
"000360" 2014                 .
"000360" 2015                 .
"000360" 2016                 .
"000360" 2017                 .
"000360" 2018 .3333333333333333
"000360" 2019                 .
"000360" 2020                 .
"000360" 2021                 .
"000360" 2022                 .
"000361" 2000                 .
"000361" 2001                 .
"000361" 2002                 .
"000361" 2003                 .
"000361" 2004                 .
"000361" 2005                 .
"000361" 2006                 .
"000361" 2007                 .
"000361" 2010                 .
"000361" 2011                 .
"000361" 2012                 .
"000361" 2013                 .
"000361" 2014                 .
"000361" 2017                 .
"000361" 2018                 .
"000361" 2019                 .
"000361" 2020                 .
"000361" 2021                 .
"000361" 2022                 .
"000752" 2000                 .
"00081T" 2006 .3333333333333333
"00081T" 2007 .3333333333333333
"00081T" 2008                 .
"00081T" 2009                 .
"00081T" 2010                 .
"00081T" 2011                 .
"00081T" 2012                 .
"00081T" 2013 .3333333333333333
"00081T" 2014                 .
"00081T" 2015                 .
"00081T" 2016                 .
"00081T" 2017                 .
"00081T" 2020                 .
"00081T" 2021                 .
"00081T" 2022                 .
"00086T" 2000                 .
"00086T" 2001                 .
"00086T" 2002                 .
"00086T" 2003                 .
"00086T" 2004                 .
"00086T" 2005                 .
"00086T" 2006                 .
"00086T" 2007                 .
"00086T" 2008                 .
"00086T" 2009                 .
"00086T" 2010                 .
"00087B" 2000                 .
"00087B" 2001                 .
"00087B" 2002                 .
"00087B" 2003                 .
"00087B" 2004                 .
"00087B" 2005                 .
"00087B" 2006                 .
"000886" 2000                 1
"000886" 2000                 1
"000886" 2004                 .
"000886" 2005                 .
"000886" 2006                 .
"000886" 2009                 .
"000886" 2010                 .
"00088E" 2000                 .
"00088U" 2000                 .
"00088U" 2004                 .
"00088U" 2005                 .
"00088U" 2008                 .
"00088U" 2009                 .
"000899" 2015                 .
"000899" 2016                 .
"000899" 2017                 .
"000899" 2018                 .
"000899" 2019                 .
"000899" 2020                 .
"000899" 2021                 .
"000899" 2022                 .
end

[Jeff Wooldridge]

If you want to do this, use pweight rather than fweight. But your standard errors will not reflect the fact that you estimated the propensity score and then selected your sample based on those estimates. I think that's why, in the link you provided, they state "we cannot recommend it."

Why not see if you have an overlap problem, and if you don't, use teffects ipwra or teffects aipw?

[Sebastian Grein]

I am not familiar with teffects ipwra or teffects aipw.

The reason why I proceeded the other way is that my variable deal indicates if the firm is a control firm (deal = 0) or a treated firm (deal = 1), but also the dependent variable in my further analysis.

[Jeff Wooldridge]

I don’t think I understand. You shouldn’t be matching on the basis of your outcome (dependent) variable — if that’s what you’ve done. Matching is to be done across control and treated — broadly defined. Do you have a reference for matching on y?

[Sebastian Grein]

I'm following Bena/Li (2014) - Corporate Innovations and Mergers and Acquisitions, p.1928-1929. (https://onlinelibrary.wiley.com/doi/...111/jofi.12059)

I am not matching on my outcome variable deal, I am matching on BTM and size. But the outcome variable deal is also the treatment variable and therefore the one distinguishing between my control firms and my M&A acquirers.