2

I'm a graduate student working on a thesis whose aim is to test complementarity between 2 practices (CI and INNO, according to theory of supermodularity). To do so, I calculated two new variables (each as a mean of 6 other variables) corresponding to the 2 practices, then derived from them 4 binary variables based on the combination of the two (that is, I used medians of the two variables as a threshold, then created four categories like 11,10,01,00: HH,LH,HL,LL).

I took these four binaries, (plus two continuous variables and three other dummies as controls) and ran a multiple regression on Y (operational performance), without a constant term.

I used both SPSS and Stata, but obtained two very different outcomes. I don't understand why!!! Moreover, correlation matrices gave the same results. Here are all the outcomes:

SPSS

Variabili escluse a,b
Modello Beta In t   Sig.    Correlazioni parziali   Statistiche di collinearità
                    Tolleranza  VIF Tolleranza minima
1   LL  .c  .   .   .   ,000    .   ,000
a. Variabile dipendente: PERF_medio
b. Regressione lineare che passa per l'origine
c. Predittori nel modello : industryT, industryM, industryE, HL, LH, HH, 
                            log_age, log_size

    Coefficienti non standardizzati
            B     STD ERROR DEV
    LH       ,188     ,110
    HL       ,156     ,115
    HH       ,467     ,102
    log_size   ,008   ,040
    log_age   -,003   ,039
    industryE 3,416   ,244
    industryM 3,487   ,255
    industryT 3,551   ,246

As you can see, the LL variable is excluded; Stata doesn't exclude this. Moreover multicollinearity tests give positive results, whereas VIF performed with Stata after regression do not.

STATA: 

regress PERF_medio LL LH HL HH log_size log_age industryE industryM industryT, noconstant
note: industryM omitted because of collinearity

        Source  SS       df     MS               Number of obs    =     190
                                                 F(  8,   182)  =   1253.59
        Model   2654.3682   8   331.796025           Prob > F           =    0.0000
        Residual    48.1709704  182 .264675662   R-squared  =    0.9822
                                                 Adj R-squared  =    0.9814
        Total   2702.53917  190 14.2238904       Root MSE           =    .51447

    PERF_medio |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
        -------------+----------------------------------------------------------------
          LL |   3.486898   .2548016    13.68   0.000     2.984153    3.989643
          LH |   3.675309   .2746033    13.38   0.000     3.133493    4.217124
          HL |   3.643261   .2646596    13.77   0.000     3.121066    4.165457
          HH |   3.954367   .2753474    14.36   0.000     3.411083     4.49765
    log_size |    .008443   .0400578     0.21   0.833    -.0705944    .0874803
     log_age |  -.0027773   .0393044    -0.07   0.944    -.0803283    .0747737
    industryE |  -.0706952   .0935333    -0.76   0.451    -.2552442    .1138538
    industryM |          0  (omitted)
    industryT |   .0639002   .0960992     0.66   0.507    -.1257117     .253512

    estat vif, uncentered

    Variable |       VIF       1/VIF  
     -------------+----------------------
    log_size |     42.17    0.023711
          HH |     16.90    0.059170
          LL |     15.21    0.065753
     log_age |     12.70    0.078726
          LH |     10.26    0.097499
          HL |      8.73    0.114505
    industryT |      2.23    0.447811
    industryE |      2.21    0.451551
    -------------+----------------------
        Mean VIF |     13.80

Thus, SPSS excludes LL, but Stata does not. What's wrong? Why are coefficients so different?

gung - Reinstate Monica
  • 132,789
  • 81
  • 357
  • 650
MatBi
  • 51
  • 1
  • 2
  • 6
  • Stata does appear to be excluding something, `industryM`, it's just making a different default choice than SPSS. – gung - Reinstate Monica Aug 09 '12 at 16:41
  • i know stata is omitting industryM whereas SPSS is omitting LL; the point is, i need LL to be in the model, because i have to test that betaHH-betaHL>betaLH-betaLL, but the stata model gives evidence of multicollinearity, if i am interepreting the last table correctly...I'm sorry but I'm not an statistics expert, so any comment will be really appreciated – MatBi Aug 09 '12 at 16:48
  • 1
    **Hint** Did you notice that Stata has omitted `industryM`? *The two programs are using different codings for categorical variables.* [Searching our site](http://stats.stackexchange.com/search?tab=votes&q=%2bdummy%20%2bcoding%20%2b[regression]) yields specific advice, such as http://stats.stackexchange.com/questions/20166 and http://stats.stackexchange.com/questions/24389. – whuber Aug 09 '12 at 18:03
  • 2
    @whuber gives good advice (as usual!) but I'd just add a note that this is a REALLY easy problem to have (and to overlook). Default choices of coding vary from program to program and even within a particular program for different statistics, so it pays to be careful. – Peter Flom Aug 09 '12 at 20:20
  • I have to thank you guys for the comments! Bytheway, it is incredible how the coefficients diverge from one software to the other (e.g., HH in spss is .467, in stata is 3.95, which is the actual mean for the sub-sample) – MatBi Aug 10 '12 at 09:21

0 Answers0