Problem with misspecification

[Ciaran Deegan]

Hello All,

Using a dataset of 600,000 observations and 60 variables I ran regress and then tested for misspecification using the ovtest. The resulting F test had a P-value of 0.0000.
This initial regression had a log dependent variable, 3 continuous independent variables (2 of which were logged) as well as 2 dummy variables to represent categorical variables.

Following the ovtest, polynomials and interaction terms between the continuous variables and the dummies were added to the regression. Once again I ran regress, although in this case, the F was lower,
the P value was still 0.0000.

Would be grateful for any advice as to how to specify this model.

Best,

Ciaran

[Carlo Lazzaro]

Ciaran:
with such a sky-rocketing number of observations, even a hiccup more/less can give back statistical significance.
That said: have you tried -linktest- after -regress-? What's its outcome?
As an aside (and per FAQ), please share what you typed and what Stata gave you back via CODE delimiters. Thanks.

[Ciaran Deegan]

Thanks again Carlo,

I was concerned that perhaps the size of the dataset was causing issues thank you for clarifying. Perhaps I will look to clean it up or maybe just take a random sample.

In relation to the linktest, P for both hat and hatsq is <0.0001

Below is the CODE I used.

reg LogEnergyRequirement LogHLI LogGFA Age Terraced SemiD
 
      Source |       SS           df       MS      Number of obs   =   382,543
-------------+----------------------------------   F(5, 382537)    >  99999.00
       Model |  66000.3431         5  13200.0686   Prob > F        =    0.0000
    Residual |  15071.3472   382,537  .039398404   R-squared       =    0.8141
-------------+----------------------------------   Adj R-squared   =    0.8141
       Total |  81071.6903   382,542  .211928861   Root MSE        =    .19849
 
------------------------------------------------------------------------------
LogEnergyR~t |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
      LogHLI |   .8692931   .0012401   701.00   0.000     .8668626    .8717236
      LogGFA |   .8144884   .0009994   815.01   0.000     .8125296    .8164471
         Age |   .0004029   .0000139    29.06   0.000     .0003758    .0004301
    Terraced |  -.0514299   .0011039   -46.59   0.000    -.0535936   -.0492663
       SemiD |  -.0308624    .000786   -39.26   0.000     -.032403   -.0293218
       _cons |   5.430726   .0052307  1038.24   0.000     5.420474    5.440978
------------------------------------------------------------------------------
 
 ovtest
 
Ramsey RESET test using powers of the fitted values of LogEnergyRequirement
       Ho:  model has no omitted variables
              F(3, 382534) =   1030.78
                  Prob > F =      0.0000

Very grateful for your advice

Kind Regards

Ciaran
(Stata 15.1)

[Carlo Lazzaro]

Ciaran:
what if you go -c.age##c.age- instead of -age- only?
If what above does not produce positive effects, the usual recipe is to start it all over again, add one predictor at time, run -regress- and -linktest- and see when problems start to come alive.

[Ciaran Deegan]

Thanks again Carlo,

I will go back to regress Energyrequirement HLI and use ovtest and linktest to identify powers and interactions to expand and specify the model as I go.

Best,

Ciaran