Cem weights/ Entropy balance/ tebalance

[Angeliki Skoura]

Hello Statalisters,

I'm a bit confused about how I can check the balance of my data. More specifically, I have panel data with businesses over the years. I want to match them by the size of the total assets of each business in order to get more accurate results. To do the matching process I use psmatch2. After that I thought it is useful to check the balance of my data. The tools I know are CEM matching with weights, the entropy balance procedure and tebalance summarize after treatment effects.
Which way in your opinion is more correct?
Any help or suggestion would be greatly appreciated.

Thank you in advance,
Angeliki

[Angeliki Skoura]

Please, any idea?

[Jared Greathouse]

The method you should use is dependent on how much you know about the methods and your specific situation. There's likely not just one general answer we can give here.

[Felix Bittmann]

kmatch from ssc supports all three balancing algorithms and allows an easy check whether balance has been achieved. There is no rule which method works best but in my experience, if the entropy balancing algorithm converges, the balancing is optimal.

[Angeliki Skoura]

Thank you @Jared Greathouse and @Felix Bittmann for your replies!

@Felix Bittmann in my case the entropy balancing algorithm converges successfully .
Also, I have used kmatch and my results are presented below, to you believe that the matching is correct?
I am grateful for your answer. Please let me know if I need to clarify any further.

[Felix Bittmann]

You are not using entropy here. To do so, you need to write

kmatch eb ...
kmatch sum

I cant say anything about your model or your control variables but from the command perspective, this is fine.

[Angeliki Skoura]

I followed your advice and produced this. What is your opinion?

Your help is priceless. Thank you very much.

[Felix Bittmann]

This looks fine. You only balance on a single control, called ta. Before balancing, the difference between the two groups is 0.327, afterwards it is basically 0. However, the variances are not perfectly balanced. You can attempt to improve this by adding the option

, targets(2)

1 = balance means
2 = balance means and SDs
3 = balance means, SDs and skewness.

The more means are balanced, the better (if the algorithm converges). You can also add bootstrapped standard errors for the ATT as a robustness check if you have the time (given your very large sample size, this is probably not feasible however).

[Angeliki Skoura]

If I add the targets(2) option, my results appear below figure 1, but without ATT (treatment effect on the treated) but only ATE ( average treatment effect) , there isn't statistical significance in my model, so maybe it is good to include ATT?

However, the second figure shows that the difference between the two groups changes after the equation and is not close to zero as before. Is this correct?

Also, the third figure shows my model with the addition of the bootstrapped standard errors for the ATT. How do you evaluate it? Do you think that the observations that are matched are okay?

Thank you from the bottom of my heart for helping me!

Best wishes.

[Felix Bittmann]

I think this is fine. Whether to report the ATT, ATE or another statistic is up to you. You should consult the literature and see what these statistics mean and which one is the most information for you and your readers. The ATT and ATE are probably the standard. Bootstrapping only influences your standard error and hence the p-value. When regular p-values and bootstrapped p-values are very similar, the results are robust. However, for a real analysis you probably want to have at least 500 bootstrap replications.

[Angeliki Skoura]

How to set the number of bootstrap replications to 500 in your opinion?


Τhank you very much!

Best wishes.

[Øyvind Snilsberg]

kmatch eb $treatment $xlist ($ylist), targets(2) vce(bootstrap, reps(500))

[Angeliki Skoura]

Thank you @Øyvind Snilsberg! I appreciate your help!

Best wishes!!
Angeliki