复杂调查数据下SAS中Kruskal-Wallis检验与ANOVA及趋势分析问题

Hi Christina, let's break down your questions one by one to help you work through this NHANES chemical concentration trend analysis:

1. Kruskal-Wallis Test for Complex Survey Data (Right-Skewed Concentrations) #

You’re right that Kruskal-Wallis is a strong non-parametric alternative for right-skewed data, and SAS has a dedicated procedure to handle this with complex survey designs like NHANES: proc surveyrank. This tool accounts for stratification, clustering, and sampling weights—all critical to producing valid results for NHANES data.

Here’s the code to run a Kruskal-Wallis test for your variable a against year (yr):

proc surveyrank data=a;
    stratum stra;       /* Your stratum variable from NHANES */
    cluster clus;       /* Your cluster variable from NHANES */
    weight wt;          /* Your sampling weight variable */
    class yr;           /* Year as the grouping variable */
    var a;              /* The chemical concentration to test */
    test kruskalwallis; /* Specify the Kruskal-Wallis test */
run;

This will output a design-adjusted test statistic and p-value, so you don’t have to worry about violating parametric assumptions from the right-skewed raw data.

2. Is Log-Transformation + ANOVA a Valid Alternative? #

Log-transformation is a common and reasonable approach to handle right-skewed data, but it works best if:

• The transformed concentrations meet ANOVA’s core assumptions (linear relationship with year, homogeneous variances across years, approximately normal residuals)
• You’re comfortable interpreting results on the log scale (e.g., a coefficient of -0.1 translates to roughly a 10% annual decrease)
• You first address any zero values (add a small constant like 0.001 to all concentrations before taking logs, since log(0) is undefined)

Keep in mind that when back-transforming results to the original scale, you’ll need to account for bias: the exponential of the mean log concentration isn’t the same as the mean original concentration—use the geometric mean instead for accurate interpretation.

3. Testing Residual Normality in proc surveyreg #

proc surveyreg doesn’t automatically run normality tests on residuals, but you can save residuals to a separate dataset and analyze them manually:

1. First, save residuals using the output statement in proc surveyreg:

proc surveyreg data=a;
    stratum stra;
    cluster clus;
    weight wt;
    model a=yr/anova;
    output out=residual_data r=reg_residual; /* Save residuals as 'reg_residual' */
run;

2. Then, use proc univariate (with sampling weights) to run normality tests and visualize the distribution:

proc univariate data=residual_data normal plot;
    var reg_residual;
    weight wt; /* Apply weights to reflect the survey design */
run;

The normal option runs Shapiro-Wilk and Kolmogorov-Smirnov tests, while plot generates a histogram and Q-Q plot for visual checks. Note: Complex survey designs can distort residual distributions, so treat these tests as exploratory tools rather than strict confirmatory checks.

4. Testing Multiple Chemicals (a, b, c) in One Procedure #

You don’t need to run separate procedures for each chemical—both proc surveyreg and proc surveyrank support multiple dependent variables:

For Parametric ANOVA (log-transformed or original data): #

proc surveyreg data=a;
    stratum stra;
    cluster clus;
    weight wt;
    model a b c = yr/anova; /* List all chemicals as dependent variables */
run;

This will output separate ANOVA results for each chemical, plus a multivariate test if you want to assess overall trends across all three.

For Non-Parametric Kruskal-Wallis: #

proc surveyrank data=a;
    stratum stra;
    cluster clus;
    weight wt;
    class yr;
    var a b c; /* List all chemicals here */
    test kruskalwallis;
run;

This will run independent Kruskal-Wallis tests for each chemical against year, all in a single step.

内容的提问来源于stack exchange，提问作者Christina