I've never used this method, so I learned something. As usual, JW makes what appears to be messy easy to do.
-
Login or Register
- Log in with
- You are not logged in. You can browse but not post. Login or Register by clicking 'Login or Register' at the top-right of this page. For more information on Statalist, see the FAQ.
-
-
Prof Ford, I really appreciate the explanation. In line with WP Paper, I can make an argument on whether to make a higher level aggregation or not. right? Because the primary question on the table is that.
The second code"Code:xtset id year panel variable: id (unbalanced) time variable: year, 2010 to 2019, but with gaps delta: 1 unit bysort id: egen mean__immi_sh = mean(immi_sh) bysort id: egen mean__share_9 = mean( share_9) bysort id: egen mean__share_12= mean(share_uni) bysort id: egen mean__share_uni = mean(logsize) bysort id: egen mean__logsize = mean(lavg_firm_age) bysort id: egen mean__lavg_firm_age= mean(lavg_firm_age) bysort id: egen mean__lage = mean(lage) bysort id: egen mean__year = mean(year) . reghdfe ln_labor_productivity immi_sh share_9 share_12 share_uni logsize /// > lavg_firm_age lage i.year mean__immi_sh mean__share_9 /// > mean__share_12 mean__share_uni mean__logsize mean__lavg_firm_age /// > mean__lage mean__year, absorb(sector region sector#region) /// > cluster(sector) (MWFE estimator converged in 2 iterations) note: mean__lavg_firm_age omitted because of collinearity warning: missing F statistic; dropped variables due to collinearity or too few clusters HDFE Linear regression Number of obs = 1,514,590 Absorbing 3 HDFE groups F( 23, 7) = . Statistics robust to heteroskedasticity Prob > F = . R-squared = 0.1674 Adj R-squared = 0.1674 Within R-sq. = 0.1042 Number of clusters (sector) = 8 Root MSE = 0.8433 (Std. Err. adjusted for 8 clusters in sector) ------------------------------------------------------------------------------------- | Robust ln_labor_producti~y | Coef. Std. Err. t P>|t| [95% Conf. Interval] --------------------+---------------------------------------------------------------- immi_sh | .0044312 .0221644 0.20 0.847 -.0479793 .0568417 share_9 | .3233302 .0464193 6.97 0.000 .2135661 .4330943 share_12 | .4447063 .0594037 7.49 0.000 .3042389 .5851737 share_uni | .3898605 .0667516 5.84 0.001 .232018 .5477029 logsize | -.0530656 .01664 -3.19 0.015 -.092413 -.0137182 lavg_firm_age | .06839 .0084707 8.07 0.000 .04836 .0884201 lage | .0941849 .0294198 3.20 0.015 .0246181 .1637516 | year | 2011 | -.0545539 .0142667 -3.82 0.007 -.0882892 -.0208186 2012 | -.0865313 .0470727 -1.84 0.109 -.1978407 .024778 2013 | -.0558563 .0577864 -0.97 0.366 -.1924995 .0807869 2014 | -.0680847 .0629978 -1.08 0.316 -.2170508 .0808813 2015 | -.0686263 .0657259 -1.04 0.331 -.2240434 .0867909 2016 | -.0680697 .0654029 -1.04 0.333 -.2227231 .0865837 2017 | -.0835982 .0604586 -1.38 0.209 -.22656 .0593636 2018 | -.1137503 .0533071 -2.13 0.070 -.2398016 .0123011 2019 | -.1378348 .0481916 -2.86 0.024 -.2517898 -.0238798 | mean__immi_sh | -.1179342 .0671146 -1.76 0.122 -.2766349 .0407665 mean__share_9 | -.1481924 .0250892 -5.91 0.001 -.2075189 -.0888659 mean__share_12 | .4651646 .1342867 3.46 0.010 .147627 .7827023 mean__share_uni | .2473019 .0443147 5.58 0.001 .1425142 .3520896 mean__logsize | .0293096 .0141171 2.08 0.077 -.004072 .0626912 mean__lavg_firm_age | 0 (omitted) mean__lage | .00315 .0883127 0.04 0.973 -.2056764 .2119763 mean__year | .064638 .0066379 9.74 0.000 .0489419 .0803342 _cons | -121.7594 13.44116 -9.06 0.000 -153.5427 -89.9761 ------------------------------------------------------------------------------------- Absorbed degrees of freedom: ---------------------------------------------------------+ Absorbed FE | Categories - Redundant = Num. Coefs | -----------------+---------------------------------------| sector | 8 8 0 *| region | 8 1 7 | sector#region | 60 60 0 *| ---------------------------------------------------------+ * = FE nested within cluster; treated as redundant for DoF computation . . test mean__immi_sh mean__share_9 mean__share_12 mean__share_uni /// > mean__logsize mean__lavg_firm_age mean__lage mean__year ( 1) mean__immi_sh = 0 ( 2) mean__share_9 = 0 ( 3) mean__share_12 = 0 ( 4) mean__share_uni = 0 ( 5) mean__logsize = 0 ( 6) o.mean__lavg_firm_age = 0 ( 7) mean__lage = 0 ( 8) mean__year = 0 Constraint 6 dropped F( 7, 7) = 749.42 Prob > F = 0.0000 .
Code:. reghdfe ln_labor_productivity immi_sh share_9 share_12 share_uni logsize /// > avg_firm_age lage mean__immi_sh mean__share_9 /// > mean__share_12 mean__share_uni mean__logsize mean__lavg_firm_age /// > mean__lage , absorb(sector region sector#region year) cluster(sector) (MWFE estimator converged in 4 iterations) note: mean__lavg_firm_age omitted because of collinearity warning: missing F statistic; dropped variables due to collinearity or too few clusters HDFE Linear regression Number of obs = 1,562,462 Absorbing 4 HDFE groups F( 13, 7) = . Statistics robust to heteroskedasticity Prob > F = . R-squared = 0.1674 Adj R-squared = 0.1674 Within R-sq. = 0.1036 Number of clusters (sector) = 8 Root MSE = 0.8559 (Std. Err. adjusted for 8 clusters in sector) ------------------------------------------------------------------------------------- | Robust ln_labor_producti~y | Coef. Std. Err. t P>|t| [95% Conf. Interval] --------------------+---------------------------------------------------------------- immi_sh | -.0022622 .0240769 -0.09 0.928 -.0591951 .0546706 share_9 | .3240873 .047084 6.88 0.000 .2127514 .4354232 share_12 | .431022 .058403 7.38 0.000 .2929209 .5691232 share_uni | .3424674 .0681693 5.02 0.002 .1812726 .5036622 logsize | -.0216342 .0170088 -1.27 0.244 -.0618537 .0185853 avg_firm_age | -.0095091 .0023529 -4.04 0.005 -.0150729 -.0039454 lage | .0313209 .0230397 1.36 0.216 -.0231593 .0858011 mean__immi_sh | -.1073892 .0644239 -1.67 0.139 -.2597275 .044949 mean__share_9 | -.1657764 .0282882 -5.86 0.001 -.2326673 -.0988854 mean__share_12 | .5061827 .1441416 3.51 0.010 .1653421 .8470234 mean__share_uni | .2200251 .0438413 5.02 0.002 .116357 .3236932 mean__logsize | .2093602 .02306 9.08 0.000 .1548319 .2638885 mean__lavg_firm_age | 0 (omitted) mean__lage | .1150118 .0883542 1.30 0.234 -.0939127 .3239364 _cons | 8.081939 .3525108 22.93 0.000 7.248384 8.915495 ------------------------------------------------------------------------------------- Absorbed degrees of freedom: ---------------------------------------------------------+ Absorbed FE | Categories - Redundant = Num. Coefs | -----------------+---------------------------------------| sector | 8 8 0 *| region | 8 1 7 | sector#region | 60 60 0 *| year | 10 1 9 | ---------------------------------------------------------+ * = FE nested within cluster; treated as redundant for DoF computation . . test mean__immi_sh mean__share_9 mean__share_12 mean__share_uni /// > mean__logsize mean__lavg_firm_age mean__lage ( 1) mean__immi_sh = 0 ( 2) mean__share_9 = 0 ( 3) mean__share_12 = 0 ( 4) mean__share_uni = 0 ( 5) mean__logsize = 0 ( 6) o.mean__lavg_firm_age = 0 ( 7) mean__lage = 0 Constraint 6 dropped F( 6, 7) = 38.53 Prob > F = 0.0001Last edited by Paris Rira; 22 May 2024, 17:54.Comment
-
Nope. Your F of 38.53 is strongly rejected.
Try it without sector#region in absorb, though I doubt much effect.
Fix the error here. Same variable.
Code:bysort id: egen mean__logsize = mean(lavg_firm_age) bysort id: egen mean__lavg_firm_age= mean(lavg_firm_age)
Try this too. It will allow you to fiddle with your model without having to rewrite the PW Test each time.
(Check it on the model you just ran (with errors) to make sure you get the same results: F = 38.53)
Code:capture program drop papwood program papwood , rclass syntax [, altfe(varlist min=1 max=1)] preserve quietly { capture drop *_xxx global XTERMS global MTERMS local varnum = wordcount(`"`e(indepvars)'"')-1 forv i = 1/`varnum' { local var = `'"`: word `i' of `e(indepvars)''"' global XTERMS $XTERMS `var' egen `var'_xxx = mean(`var') , by(`altfe') global MTERMS $MTERMS `var'_xxx } reghdfe `e(depvar)' $XTERMS $MTERMS , absorb(`e(absvars)') cluster(`e(clustvar)') test $MTERMS } local star = (cond(r(p)<.01, "***", cond(r(p)<.05, "**", cond(r(p)<.10, "*", "")))) di `"{hline 50}"' di "Papke-Wooldridge (2023) Test" di "Null: Higher level fixed effects is OK." di "F-stat = " %6.2f `r(F)' "`star'" di "Prob Level = " %8.6f r(p) if r(p) < 0.10 { di "Null is REJECTED -- Yuck!" } else { di "Null is ACCEPTED -- Yeah!" } di `"{hline 50}"' restore endCode:reghdfe ln_labor_productivity immi_sh share_9 share_12 share_uni logsize /// avg_firm_age lage , absorb(sector region sector#region year) cluster(sector) papwood, altfe(id)Last edited by George Ford; 22 May 2024, 19:15.Comment
-
If you want to see the augmented regression, put "nois" in front of the "reghdfe" command.Comment
-
You can also run the test just on the immi variable. Lots more code, but it's provided in the link JW provided.Comment
-
What I missed ?Code:. capture program drop papwood . program papwood , rclass 1. syntax [, altfe(ln_labor_productivity immi_sh share_9 share_12 share_uni logsize avg_firm_age lage min=1 max=1)] 2. preserve 3. quietly { 4. capture drop *_xxx 5. global XTERMS 6. global MTERMS 7. local varnum = wordcount(`"`e( immi_sh share_9 share_12 share_uni logsize avg_firm_age lage)'"')-1 8. forv i = 1/`varnum' { 9. local var = `'"`: word `i' of `e( immi_sh share_9 share_12 share_uni logsize avg_firm_age lage)''"' 10. global XTERMS $XTERMS `var' 11. egen `var'_xxx = mean(`var') , by(`altfe') 12. global MTERMS $MTERMS `var'_xxx 13. } 14. reghdfe `e(ln_labor_productivity)' $XTERMS $MTERMS , absorb(`e(sector region sector#region year)') cluster(`e(sector)') 15. test $MTERMS 16. } 17. local star = (cond(r(p)<.01, "***", cond(r(p)<.05, "**", cond(r(p)<.10, "*", "")))) 18. di `"{hline 50}"' 19. di "Papke-Wooldridge (2023) Test" 20. di "Null: Higher level fixed effects is OK." 21. di "F-stat = " %6.2f `r(F)' "`star'" 22. di "Prob Level = " %8.6f r(p) 23. if r(p) < 0.10 { 24. di "Null is REJECTED" 25. } 26. else { 27. di "Null is ACCEPTED" 28. } 29. di `"{hline 50}"' 30. restore 31. end . end of do-file . do "C:\Users\CeBER\AppData\Local\Temp\STD135c_000000.tmp" . reghdfe ln_labor_productivity immi_sh share_9 share_12 share_uni logsize /// > avg_firm_age lage , absorb(sector region sector#region year) cluster(sector) (MWFE estimator converged in 4 iterations) HDFE Linear regression Number of obs = 1,575,948 Absorbing 4 HDFE groups F( 7, 7) = 8509.86 Statistics robust to heteroskedasticity Prob > F = 0.0000 R-squared = 0.1499 Adj R-squared = 0.1498 Within R-sq. = 0.0850 Number of clusters (sector) = 8 Root MSE = 0.8708 (Std. Err. adjusted for 8 clusters in sector) ------------------------------------------------------------------------------ | Robust ln_labor_p~y | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------- immi_sh | -.1399509 .0761987 -1.84 0.109 -.3201321 .0402304 share_9 | .2012313 .0277511 7.25 0.000 .1356104 .2668523 share_12 | .4459813 .0619086 7.20 0.000 .2995907 .592372 share_uni | .800066 .1038243 7.71 0.000 .5545604 1.045572 logsize | .1933575 .0336375 5.75 0.001 .1138174 .2728975 avg_firm_age | .0041543 .0016538 2.51 0.040 .0002438 .0080649 lage | .2057504 .0696364 2.95 0.021 .0410865 .3704144 _cons | 8.119879 .2891404 28.08 0.000 7.436171 8.803588 ------------------------------------------------------------------------------ Absorbed degrees of freedom: ---------------------------------------------------------+ Absorbed FE | Categories - Redundant = Num. Coefs | -----------------+---------------------------------------| sector | 8 8 0 *| region | 8 1 7 | sector#region | 60 60 0 *| year | 10 1 9 | ---------------------------------------------------------+ * = FE nested within cluster; treated as redundant for DoF computation . papwood, altfe(id) invalid syntax r(197); end of do-file r(197);Comment
-
"https://blog.stata.com/2015/10/29/fixed-effects-or-random-effects-the-mundlak-approach/"Originally posted by George Ford View Post
"If you reject that the coefficients are jointly zero, the test suggests that there is correlation between the time-invariant unobservables and your regressors, namely, the fixed-effects assumptions are satisfied."
right? so my result reject the null hypo. does not ? I dont get it .Comment
-
Don’t change anything in the papwood program. Leave it as I sent it.
You don’t want to reject the null.
and you can’t replicate the other result. My code creates new variables. The code does replicate exactly what was in JW’s code.
Comment
-
I ran after correcting the above error.
Code:. capture program drop papwood . program papwood , rclass 1. syntax [, altfe(varlist min=1 max=1)] 2. preserve 3. quietly { 4. capture drop *_xxx 5. global XTERMS 6. global MTERMS 7. local varnum = wordcount(`"`e(indepvars)'"')-1 8. forv i = 1/`varnum' { 9. local var = `'"`: word `i' of `e(indepvars)''"' 10. global XTERMS $XTERMS `var' 11. egen `var'_xxx = mean(`var') , by(`altfe') 12. global MTERMS $MTERMS `var'_xxx 13. } 14. reghdfe `e(depvar)' $XTERMS $MTERMS , absorb(`e(absvars)') cluster(`e(clustvar)') 15. test $MTERMS 16. } 17. local star = (cond(r(p)<.01, "***", cond(r(p)<.05, "**", cond(r(p)<.10, "*", "")))) 18. di `"{hline 50}"' 19. di "Papke-Wooldridge (2023) Test" 20. di "Null: Higher level fixed effects is OK." 21. di "F-stat = " %6.2f `r(F)' "`star'" 22. di "Prob Level = " %8.6f r(p) 23. if r(p) < 0.10 { 24. di "Null is REJECTED -- Yuck!" 25. } 26. else { 27. di "Null is ACCEPTED -- Yeah!" 28. } 29. di `"{hline 50}"' 30. restore 31. end . . nois reghdfe ln_labor_productivity immi_sh share_9 share_12 share_uni logsize /// > avg_firm_age lage , absorb(sector region sector#region year) cluster(sector) (MWFE estimator converged in 4 iterations) HDFE Linear regression Number of obs = 1,575,948 Absorbing 4 HDFE groups F( 7, 7) = 8509.86 Statistics robust to heteroskedasticity Prob > F = 0.0000 R-squared = 0.1499 Adj R-squared = 0.1498 Within R-sq. = 0.0850 Number of clusters (sector) = 8 Root MSE = 0.8708 (Std. Err. adjusted for 8 clusters in sector) ------------------------------------------------------------------------------ | Robust ln_labor_p~y | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------- immi_sh | -.1399509 .0761987 -1.84 0.109 -.3201321 .0402304 share_9 | .2012313 .0277511 7.25 0.000 .1356104 .2668523 share_12 | .4459813 .0619086 7.20 0.000 .2995907 .592372 share_uni | .800066 .1038243 7.71 0.000 .5545604 1.045572 logsize | .1933575 .0336375 5.75 0.001 .1138174 .2728975 avg_firm_age | .0041543 .0016538 2.51 0.040 .0002438 .0080649 lage | .2057504 .0696364 2.95 0.021 .0410865 .3704144 _cons | 8.119879 .2891404 28.08 0.000 7.436171 8.803588 ------------------------------------------------------------------------------ Absorbed degrees of freedom: ---------------------------------------------------------+ Absorbed FE | Categories - Redundant = Num. Coefs | -----------------+---------------------------------------| sector | 8 8 0 *| region | 8 1 7 | sector#region | 60 60 0 *| year | 10 1 9 | ---------------------------------------------------------+ * = FE nested within cluster; treated as redundant for DoF computation . papwood, altfe(id) -------------------------------------------------- Papke-Wooldridge (2023) Test Null: Higher level fixed effects is OK. F-stat = 85.75*** Prob Level = 0.000003 Null is REJECTED -- Yuck! --------------------------------------------------Last edited by Paris Rira; 22 May 2024, 20:42.Comment
-
Not the result you were hoping for. sorry.
Comment
-
The question is that I have to drop region and sector fixed effects in favor of firm fixed effects. Since firms don't move across region and sectors, so I can't use three FEs simultaneously.Comment
-
Also, I wonder if the size matters, I could not find in PW(2023) if they argue aggregate level size. I mean because of 1.5 million datasets and sectors are only 8. It can be the reason?Comment
-
As I understand it, and I’ve only seen this work today, is that the hope is that you’ve got enough firm specific variables that there’s no consequence of firm level omitted variables to speak of, making the firm FE redundant. You don’t have that situation. You don’t need fewer firm variables, but more. But, as you add more, the result will look more like the firm FE result.
Fiddling is not going to save you. With 1.5m observations, it’s difficult not to get a statistically significant result. On top of that, clustering by sector (only 8 of them) is going to lead to SE that are too small, not too large. The odds are in your favor.
you’ll need to become more intimate with your data to figure out whether you’ve missed something interesting. Perhaps the effect is nonlinear? How much variation do you have in immig? How much does productivity vary within sector, and between regions? Did the data import and merge properly?
The effect may very well be zero. I’m fairly convinced it is. wanting it to be otherwise is immaterial. It is what it is.
At this point, for me, I’d take a few days off and hope some different way to think about this comes to mind. Read other papers on this and similar topics to see if there’s a different way to approach the problem. But understand that research doesn’t always work out as planned.
Comment
-
Dear Prof Ford,
I really appreciate your time and consideration for this thread. I have learned a lot from your guidance. As you mentioned, research is not always expected to have significant effects, and finding no effect is still valuable evidence as some other authors on this topic already reached the same results as mine. Among the models, the IV model produced significant results, which I can interpret properly.
I hope this thread will help other students in the future who may encounter the same issue.
I wonder if it would be possible to mention your name in my work?
Thank you once again for your help.
Cheers,
ParisComment
-
I learned something new as well.
i’d interact immig with sector. You’d think immig matters more to certain sectors.
or run the model sector by sector to see if the results differ much.Comment
Comment