SPSS and Stata output different

I'm Stata-proficient but learning SPSS for my new position. I am using a simple dataset to do very basic regressions and comparing to see if the results are the same. They're not. I'm close, but the magnitudes of the betas and significance are slightly different. The data was copy and pasted into each from an Excel; I didn't use a Stata file in SPSS, or vice versa. For SPSS, I did not weight it, it's using listwise deletion, and it's on the "enter" method. I presume Stata is doing the same, as its default (but correct me if I'm wrong and it's a different default!).

Any ideas on what else to check? I'm doing just a simple linear regression.

Syntax

SPSS

REGRESSION
/MISSING LISTWISE
/STATISTICS COEFF
/DEPENDENT FreeLunch
/METHOD=ENTER FoodInsecure Rural Female @18 Hispanic.


(Or, for point and click, analyze->regression->linear; forces a choice under "method" for enter/stepwise/remove/backward/forward.)

Stata

reg percentfreelunch percentfoodinsecure rural female under18 hispanic


Data is in Excel and was pasted into both.

Results

SPSS

           Var. |   Unst.B | Std.Err. |   St.B |     t | Sig.
(Constant) | -139.616 |   66.652 | -2.095 | .045
% Food Insecure |    2.785 |     .674 |   .546 | 4.131 | .000
Rural |     .131 |     .048 |   .404 | 2.701 | .011
Female |    2.657 |    1.170 |   .372 | 2.270 | .031
< 18 |    -.416 |     .583 |  -.145 | -.715 | .480
Hispanic |    1.156 |     .236 |  1.092 | 4.905 | .000


Stata

               Var. |     Coef. | Std.Err. |     t | P>|t|
percentfoodinsecure |   2.76532 | .6741544 |  4.10 | 0.000
rural |  .1378976 | .0495354 |  2.78 | 0.009
female |  2.826711 | 1.204272 |  2.35 | 0.026
under18 | -.3799895 |  .588423 | -0.65 | 0.523
hispanic |  1.168375 | .2398765 |  4.87 | 0.000
_cons | -149.3858 |  69.0891 | -2.16 | 0.039

• "Simple linear regression" does not entail an "enter" method so far as I know. Are you doing something stepwise? How are the data held? In storage types that are exactly comparable? Regardless of that, this is perilously close to "I get slightly different results in different programs". Can you provide reproducible examples with specified datasets and exact commands so that people can replicate results in either program, or indeed any other program? Aug 26, 2014 at 16:52
• For what it is worth, R is producing results like Stata. Aug 26, 2014 at 17:58
• From what you've pasted, it's the same analysis (I've used both SPSS and Stata). It's almost certain that the data are different. Run things like means, frequencies or correlations and make sure the results are the same in both. It's possible that something weird happened with variable names when you pasted. Save the data from Stata (in Stata format) and open it in SPSS. Aug 26, 2014 at 18:01
• It's possible that commands you've run previously on SPSS carry over - for e.g. weight, split file, or select. Check that. Also make sure that you haven't defined a value in the data as missing. Aug 26, 2014 at 18:05
• I'm Stata-proficient but learning SPSS for my new position. My condolences Aug 26, 2014 at 19:00

The problem (amazingly) has to do with rounding the values during pasting.

In Excel, most of the values were computed elsewhere and are recorded as doubles (about 16 decimal places of precision). Only % Food Insecure actually is stored to a small number of decimal places (one). None of the data columns is stored as it appears in Excel. During pasting, the receiving application typically will accept the data as they appear, not as they are actually stored!

Rounding of data matters in this situation because for several variables--especially percent female and percent food insecurity--the amounts rounded off can be an appreciable fraction of the standard deviation of the data.

When I read the Excel data directly in R using xlsx::read.xlsx, I reproduce the SPSS results exactly. When I round the data to integers (for % Free Lunch) and to one decimal place for the others--as they appear when pasting them into R--I get new results, but the estimated coefficients change appreciably. For instance, the intercept of $-139.616$ becomes $-133.897$.

I have not been able to reproduce the Stata results in R (my summary statistics do not quite agree with those presented by Nick Cox: my mean for % Food insecure is $15.67$ instead of $15.81$), but I suspect that if I were to paste them into my copy of Stata, I would get the reported Stata results. (A big clue is the rounded values presented for the minima and maxima: in most cases these are not the minima and maxima actually recorded in the Excel file.)

The differences between the two sets of results are a small fraction of a standard error, so they are--in this statistical sense--of no consequence.

There is no collinearity problem: the VIFs are nice and low.

Moral

When you care about your data, read them directly: do not intervene manually via copy-and-paste or transcription.

• I can support it. I have no Stata, but when I read the Excel data in SPSS, the results were exactly as the OP gave. When I pasted the data into SPSS, the values got rounded to the number of digital places displayed in Excel; the results were very close to "Stata results". Aug 26, 2014 at 19:32
• ... and not just rounding of values during pasting. Saving the excel sheet as a csv in Excel saves the visible number of decimal points (the approach I used to get R to produce 'Stata results'). Aug 26, 2014 at 19:40
• YOU ARE AN ANGEL! Thank you! I will try to do it in both again with attention to that, but I bet that solves it! Aug 26, 2014 at 20:04
• Solved! Results=identical now. Never would've occurred to me that would be the source. Thank you, everyone! Aug 26, 2014 at 20:24
• It definitely was a group effort, as the comments and replies attest.
– whuber
Aug 26, 2014 at 21:15

This isn't really an answer, but there is no easy way to show output except here.

This is what I got in Stata 13, after copy and paste and some renaming.

There are no missing values on the variables used, so what either program might do with missing values is irrelevant here.

2/6 variables are presented as integers; 4/6 are presented with one decimal place. Note that the exact route I followed has the consequence that variables with decimal places are held as double. Other routes might produce variables of float type.

Check that 36 observations (cases) are included in both regressions.

Please check whether you got the same.

. d percentfreelunch percentfoodinsecure rural female under18 hispanic

storage   display    value
variable name   type    format     label      variable label
-----------------------------------------------------------------------------------------
percentfreelu~h byte    %10.0g                % Free Lunch
percentfoodin~e byte    %10.0g                % Food Insecure
rural           double  %10.0g                Rural
female          double  %10.0g                Female
under18         double  %10.0g                < 18
hispanic        double  %10.0g                Hispanic

. reg percentfreelunch percentfoodinsecure rural female under18 hispanic

Source |       SS       df       MS              Number of obs =      36
-------------+------------------------------           F(  5,    30) =    8.27
Model |  1578.09245     5   315.61849           Prob > F      =  0.0001
Residual |  1145.12977    30  38.1709925           R-squared     =  0.5795
Total |  2723.22222    35  77.8063492           Root MSE      =  6.1783

-------------------------------------------------------------------------------------
percentfreelunch |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
--------------------+----------------------------------------------------------------
percentfoodinsecure |    2.76532   .6741544     4.10   0.000     1.388513    4.142127
rural |   .1378976   .0495354     2.78   0.009     .0367327    .2390624
female |    2.82671   1.204272     2.35   0.026     .3672593    5.286162
under18 |  -.3799895    .588423    -0.65   0.523     -1.58171    .8217307
hispanic |   1.168375   .2398765     4.87   0.000     .6784822    1.658269
_cons |  -149.3857   69.08908    -2.16   0.039    -290.4845   -8.287025
-------------------------------------------------------------------------------------


EDIT: Some summary statistics from Stata

. su percentfreelunch percentfoodinsecure rural female under18 hispanic, sep(0)

Variable |       Obs        Mean    Std. Dev.       Min        Max
-------------+--------------------------------------------------------
percentfre~h |        36    46.27778    8.820791         29         72
percentfoo~e |        36    15.80556    1.785968         12         20
rural |        36    44.37222    27.25343        1.3        100
female |        36    49.95833    1.233201       45.9       51.5
under18 |        36    21.26389    3.069216       15.6         28
hispanic |        36    10.86944    8.315178        2.3       32.2

• I checked and all cases are included in both models. I have the exact same Stata results as you. I changed all the variables to float type, and the results are the same. Aug 26, 2014 at 17:46
• When I import the data into SPSS none of the variables are integers. All of the ones used here are floating point except for the FoodInsecure - which is F3.1 (in Fortran style number formats). This causes the variables to all have slightly different summary statistics in SPSS compared to here. Here is SPSS code to replicate. So it is likely the two programs input the Excel data differently - which is correct I'm unsure. Aug 26, 2014 at 18:25
• @AndyW I think you found the answer. SPSS got it right.
– whuber
Aug 26, 2014 at 19:26
• Thank you, @AndyW and @whuber! Got it working perfectly. :-) Aug 26, 2014 at 20:22