Sensitivity analysis for missing value

[Yue YY]

Dear statalist,

I'm running a logistic regression over outcome variable: cancer, and there are two independent variables "socionomic status" and "parental behavior"with missing values. I first include only those subjects without missing values into the model, then I include all subjects after imputation for the missing values. If I want to do sensitivity analysis for this two results, which method should I use? Thank you!

[Phil Bromiley]

You didn't get a quick answer. You'll increase your chances of a useful answer by following the FAQ on asking questions.

If you want to test whether mi gives you different results than the estimation w missing, you might look at hausman or suest. I don't know if they work with your particular estimators (whatever they are).

[Carlo Lazzaro]

Yue:
you do not tell us anything concernong the mechanism underlying the missingness of your data, that may affect replies about sensitivity analysis (if data are MNAR, for instance).
As an aside, have you already checked that your original regression model is not misspecified with two predictors only?

[Yue YY]

Yue:
you do not tell us anything concernong the mechanism underlying the missingness of your data, that may affect replies about sensitivity analysis (if data are MNAR, for instance).
As an aside, have you already checked that your original regression model is not misspecified with two predictors only?

Hi, Carlo, sorry I didn't make my question clear. This analysis was asked by one of my supervisor, yet I really have no idea about it.

My question is about the relation between cancer risk and finger length. It's a case-control study. In the model, I have sex, age, education, smoking and drinking as covariates and they are completed. Lifetime SES and 2 types of parental behaviors (smoking, drinking) are all categorical variables and are the covariates in my model with missing values, because some participants were old and couldn't remember details about their childhood or adolescence. The proportion of missing values is 10.3% for parental smoking, 11.1% for parental drinking. For SES, it's 1.1% in childhood, 3.1% for adolescence and 1.5% for adulthood. I have checked their missingnesses which were not relating to my outcome variable (cancer) nor my main independent variable (finger length). However they may related with missingness of each other, .e.g., parental smoking missing was relating to parental drinking missing. The estimates from the two models (with and without imputed values) are however similar.

[Yue YY]

You didn't get a quick answer. You'll increase your chances of a useful answer by following the FAQ on asking questions.

If you want to test whether mi gives you different results than the estimation w missing, you might look at hausman or suest. I don't know if they work with your particular estimators (whatever they are).

Thank you Phil! I will first look to them!

[Carlo Lazzaro]

Yue:
thanks for the clarifications, but the main issue still stands out, as you do not report anything about the mechanism underlying the missingness of your data.
You may want to take a look at -mi- entries in Stata .pdf manual and along with the following references:

- Allison PD. Missing data. Thousand Oaks, CA: SAGE Publications, 2002;
- Little RJA and Rubin DB. Statistical analysis with missing data. 2nd ed. Hoboken, NJ: Wiley, 2002;
-Clark TG and Altman DG. Developing a prognostic model in the presence of missing data: an ovarian cancer case study. J Clin Epidemiol 2003; 56: 28–37.
- Van Buuren S, Boshuizen HC and Knook DL. Multiple imputation of missing blood pressure covariates in survival analysis. Stat Med 1999; 18: 681–694.
- Van Buuren S. Flexible imputation of missing data. Boca Raton, FL: Chapman and Hall/CRC, 2012.

Eventually, you can plot the predicted values of the complete case analysis and the imputed regression and see if you detect relevant differences.