teffects density

[Greg Swinand]

Hi Statalist,

I'm trying to correctly interpret the plots I get when running "teffects density" after a teffects psmatch, and the difference between this and the 'teffects overlap' plots.

The manual says the teffects density is the density of the propensity scores, from before and after matching, and it gives a graphical output which shows 'raw' and 'matched' densities, for treated and control.

I've attached the as-is graph -- it would appear my PSs are very similar for control and treated are very similar.

The teffects overlap seems to give a different picture (which seems similar to the unmatched density) however, and I attach that here too. My understanding is that this is a graph of the densities of the probability of being treated or not treated.

My expectation was that these graphs should have been similar, but as they are not, I'm guessing I'm missing something.

Many thanks in advance for any help,

Greg

[Sebastian Geiger]

Greg,

it would be helpful if you could show us the exact command(s) you typed. For example, I could not find a command named -teffects density-, just one called -tebalance density-.

Using the -teffects overlap- command you probably need to specify the treatment level to generate a similar graph as the one produced by -tebalance density- (assuming that this is the one you used), as the -density- plot shows the propensity score for being treated (==1) and the -overlap- plot choses the first treatment level (e.g. ==0). You can specify the treatment level by using the -ptlevel- option:

teffects overlap, ptlevel(1)

Due to the fact that both commands use by default other kernel types, the graphs may still look a little different. They should be identical if you, for example, tell the -teffects density- command to use the default kernel of -tebalance density-:

teffects overlap, ptlevel(1) kernel(epanechnikov)

[Greg Swinand]

Sebastian many thanks. Yes, sorry; the commands are tbalance density and teffects overlap. OK, yes, the one, teffects overlap, by default gives the probability of being in the control, rather than the probability of being treated (for both treated and control), so that is just flipped around. I will try setting the kernels to the same thing as you suggested. thanks again.