Cox Regression with missing values

Very grateful for any advice.

I'm trying to perform Cox regression on a data set with multiple random missing values. SPSS is my go-to software, but the Cox regression algorithm applies LISTWISE deletion to all cases with missing values, resulting in omission of >30% of the cases. I'd rather it used PAIRWISE deletion, but can't seem to find that option. I usually use the GUI but have some basic coding experience; tinkering with the code, however, has so far proved fruitless. I have the impression, from SPSS documentation, that LISTWISE deletion is built into the Cox regression algorithm.

In the past week, I have been trying to get to grips with SAS as an alternative. So far as I can tell, PROC PHREG is not available via SAS-Studio, so I shall have to learn SAS coding if I want to go any further.

My questions are: (1) will SAS allow me to perform Cox regression with pairwise deletion of missing values (as opposed to omission of all cases with missing values)? If so, then it is worth my while learning how to code. If not, then I'd rather not waste time reinventing the wheel...

(2) Alternatively, is there a way to perform Cox regression on SPSS with pairwise rather than listwise deletion?

(3) Could someone recommend any alternative software that will allow me to perform Cox regression with missing values (without listwise deletion of cases)?

I realise there are imputation procedures available for missing values in both SPSS and SAS, but wish to use them only as a last resort (owing to the nature of the data).

I'd really appreciate any help.


Less is more. Stay pure. Stay poor.
Please define what you mean be "pairwise" deletion.

To my knowledge, SAS uses listwise deletion. The coding isn't too hard.

Why is data missing, do you know if it is MNAR, MAR, or MCAR? Does the mssingness bias your results? Multiple imputation is always the way to go if it is an option.
"Pairwise deletion" allows inclusion of cases with missing values for one or more - but not all - variables. If a case is missing a value for a variable, then the procedure excludes that case while analysing that particular variable. When analysing other variables, however, the algorithm can include that same case if it contains the relevant values.

On SPSS, "pairwise deletion" is an option available for various other types of regression (including multiple linear and logistic, IIRC, but not Cox). Because the Cox procedure uses listwise deletion, however, almost half of the cases are being omitted from the analysis, even though many lack values for only one of the variables.

I have been so wary of imputation for this particular data set that I have not attempted any formal analysis of the pattern of missing values. Your advice is much appreciated, however, and I shall get going with that right away.


No cake for spunky
Pairwise or listwise are bad ideas unless the data is MCAR which most feel is unlikely. Multiple imputations is the way to go. SPSS has a module specifically built for that I have been told although I work with SAS.


Less is more. Stay pure. Stay poor.
I find it interesting in SPSS you could do it with logistic and not survival given their similarities. The only thing I can think of would be if missingness was in the time variable.
Tried to post this a couple of days ago, but looks like it never made it.

I ran Little's MCAR test; the p-value was 0.002, which suggests that I cannot - even in my dreams - assume that data are MCAR. Further than that I have not gone yet, but that was enough of a dampener and completely contrary to our expectations. I was unaware that non-random patterns of missingness are a contraindication to use of pairwise/ listwise deletion, so many thanks indeed for pointing that out to me.

Looks like imputation is my only option.

Oh, and I should add - with my apologies - that SPSS has the pairwise deletion option for multiple linear regression only, not logistic or ordinal regression. So it makes sense, I suppose, that I couldn't code pairwise deletion into the Cox procedure either.

Many thanks for all your help.