Number of observations in my descriptive statistics (summarize statistics) is higher than the reported observations in my regression output?

[YH jordaan]

Hi guys,

Im still working on my thesis on how bank quality (measured by dummy variable goodbank, proxied by bond credit ratings of those banks) influences market liquidity and funding liquidity. I have a unbalanced panel data set containing bank specific (panel) data and macroeconomics time serie data (which applies to each bank of course).

My question is the following:

How could it be that the lowest value of # of observations in my descriptive statistics (summarize statistics) is higher than the reported observations in my regression output?

See below first my summary statistics:

    Variable |       Obs        Mean    Std. Dev.       Min        Max
-------------+--------------------------------------------------------
        mliq |      4928    .0590628    .0609607   .0019011   .4217335
    goodbank |      5296    .6752266    .4683343          0          1
        size |      4952      11.043    1.730852   6.774847   15.17113
        racr |      4319    13.81973    2.870637       7.62      44.07
         nim |      4216    3.329013    1.299908      -8.87       15.1
-------------+--------------------------------------------------------
 Crisisdummy |      5131    .1641006    .3704029          0          1
changeinfl~n |      5063    .9989563    .3413335       .065        1.8
   DiffLibor |      4836   -.0959344    .4775776  -1.711507   .5692567
changeFedF~d |      5063   -.0826052    .4537597      -1.66        .57

And here my regression output:

. areg mliq goodbank size  racr nim  Crisisdummy  changeinflation  DiffLibor changeFedFund, absorb(gvkey) r

Linear regression, absorbing indicators           Number of obs   =       3465
                                                  F(   8,   3333) =      31.43
                                                  Prob > F        =     0.0000
                                                  R-squared       =     0.8097
                                                  Adj R-squared   =     0.8023
                                                  Root MSE        =     0.0285

---------------------------------------------------------------------------------
                |               Robust
           mliq |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
----------------+----------------------------------------------------------------
       goodbank |  -.0082468   .0015715    -5.25   0.000     -.011328   -.0051657
           size |   .0109495   .0020169     5.43   0.000     .0069949     .014904
           racr |   .0028888    .000365     7.91   0.000     .0021731    .0036044
            nim |  -.0077708   .0014891    -5.22   0.000    -.0106904   -.0048513
    Crisisdummy |   -.011055   .0022849    -4.84   0.000     -.015535    -.006575
changeinflation |  -.0030651   .0015295    -2.00   0.045     -.006064   -.0000663
      DiffLibor |  -.0082581   .0018577    -4.45   0.000    -.0119003   -.0046158
  changeFedFund |  -.0039041   .0020787    -1.88   0.060    -.0079798    .0001716
          _cons |  -.0652826    .023603    -2.77   0.006    -.1115603   -.0190048
----------------+----------------------------------------------------------------
          gvkey |   absorbed                                     (124 categories)

[Rich Goldstein]

you asked:

How could it be that the lowest value of # of observations in my descriptive statistics (summarize statistics) is higher than the reported observations in my regression output?

However, that is not true in the output you showed - in your descriptive stats each variable has more than 4,000 observations while your regression has 3,465 - so, what is your question?

[William Lisowski]

Stata analytical commands like areg generally do case-wise deletion: if any variable is missing, the observation is omitted from the analysis. Suppose you have 10 observations of y, x1, and x2, but x1 is missing in observation 2 and x2 is missing in observations 8 and 9. Your summary statistics will show 10 observations for y, 9 for x1, and 8 for x2. But in the analysis, observations 2, 8, and 9 will all be omitted, leaving you with just 7 observations.

This is discussed in the section of the help missing output titled Estimation commands.

[Rich Goldstein]

whoops - sorry for the misreading (and I even quoted it!)

[YH jordaan]

Stata analytical commands like areg generally do case-wise deletion: if any variable is missing, the observation is omitted from the analysis. Suppose you have 10 observations of y, x1, and x2, but x1 is missing in observation 2 and x2 is missing in observations 8 and 9. Your summary statistics will show 10 observations for y, 9 for x1, and 8 for x2. But in the analysis, observations 2, 8, and 9 will all be omitted, leaving you with just 7 observations.

This is discussed in the section of the help missing output titled Estimation commands.

Thanks for your answer. Makes sense. I assume this does not differ with "xtreg"? Or do all commands use this case wise deletion?

[YH jordaan]

whoops - sorry for the misreading (and I even quoted it!)

No problem . I am not a native English speaker, so probably the question could have been specified more clearly.

[Rich Goldstein]

re: your #5 - in Stata (afaik), all commands use casewise deletion (other than pwcorr)

[Ariel Karlinsky]

Try: count if ~missing(mliq, goodbank, size, racr, nim, Crisisdummy, changeinflation, DiffLibor, changeFedFund, gvkey)

[YH jordaan]

Try: count if ~missing(mliq, goodbank, size, racr, nim, Crisisdummy, changeinflation, DiffLibor, changeFedFund, gvkey)

Thank, this value is 1831. Total observations are 5296. 5296 - 1831 = the # of observations in my regression output.

[YH jordaan]

Now I'm wondering how to get descriptive statistics of my variables which are actually used in my regression. So in other words: descriptive statistics of my variables who survived case-wise delition of Stata. I think I should incorperate something like: generate a dummy variable which indicates if a variable is missing and then run summarize, if option?

EDIT: I already found a solution by using e(sample) after running my regression.

[Krishna Bhandari]

esample not found error occurs after running sum function with if (esample) option. Kindly guide what needs to be done?

[Carlo Lazzaro]

Krishna:
welcome to this forum.
As -summarize- stores resulst in -r()- and not in -e()-, to obtain the number of observations included in your -summarize- code, you can type:

use "C:\Program Files\Stata17\ado\base\a\auto.dta"
. sum rep78

    Variable |        Obs        Mean    Std. dev.       Min        Max
-------------+---------------------------------------------------------
       rep78 |         69    3.405797    .9899323          1          5


. di r(N)
69

.

[William Lisowski]

esample not found error occurs after running sum function with if (esample) option. Kindly guide what needs to be done?

The option is

... if e(sample)

Here is an example - the variable rep78 has 5 missing values.

. sysuse auto, clear
(1978 automobile data)

. regress price length i.rep78

      Source |       SS           df       MS      Number of obs   =        69
-------------+----------------------------------   F(5, 63)        =      3.76
       Model |   132668930         5  26533786.1   Prob > F        =    0.0048
    Residual |   444128029        63  7049651.25   R-squared       =    0.2300
-------------+----------------------------------   Adj R-squared   =    0.1689
       Total |   576796959        68  8482308.22   Root MSE        =    2655.1

------------------------------------------------------------------------------
       price | Coefficient  Std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
      length |   65.02221   15.48443     4.20   0.000     34.07904    95.96537
             |
       rep78 |
          2  |   728.5196   2105.194     0.35   0.730    -3478.374    4935.414
          3  |   1539.622   1940.569     0.79   0.431    -2338.295     5417.54
          4  |   1777.926   1980.059     0.90   0.373    -2178.907    5734.759
          5  |     2572.1   2061.701     1.25   0.217    -1547.881     6692.08
             |
       _cons |  -7724.697   3477.005    -2.22   0.030    -14672.94   -776.4562
------------------------------------------------------------------------------

. summarize price

    Variable |        Obs        Mean    Std. dev.       Min        Max
-------------+---------------------------------------------------------
       price |         74    6165.257    2949.496       3291      15906

. summarize price if e(sample)

    Variable |        Obs        Mean    Std. dev.       Min        Max
-------------+---------------------------------------------------------
       price |         69    6146.043     2912.44       3291      15906

.

[Carlo Lazzaro]

YH Jordan:
as an aside to previous helpful hints, the issue can be also explained in the following way: while -summarize- adopts available case analysis (hence, the number of observations can change for each variable, due to differences in missing values), estimation commands, like -regression-, adopt complete case analysis (only observations with all observed values are included in the inferential procedure).