12 Statistical Inference Review

Emphasize confidence limits, which can be computed from adjusted or unadjusted analyses, with or without taking into account multiple comparisons
$P$ -values can accompany CLs if formal hypothesis testing needed
When possible construct $P$ -values to be consistent with how CLs are computed

12.1 Types of Analyses

Except for one-sample tests, all tests can be thought of as testing for an association between at least one variable with at least one other variable
Testing for group differences is the same as testing for association between group and response
Testing for association between two continuous variables can be done using correlation (especially for unadjusted analysis) or regression methods; in simple cases the two are equivalent
Testing for association between group and outcome, when there are more than 2 groups which are not in some solid order¹ means comparing a summary of the response between $k$ groups, sometimes in a pairwise fashion

¹ The dose of a drug or the severity of pain are examples of ordered variables.

12.2 Covariable-Unadjusted Analyses

Appropriate when

Only interested in assessing the relationship between a single $X$ and the response, or
Treatments are randomized and there are no strong prognostic factors that are measureable
Study is observational and variables capturing confounding are unavailable (place strong caveats in the paper)

See Chapter 13

12.2.1 Analyzing Paired Responses

Type of Response	Recommended Test	Most Frequent Test
binary	McNemar	McNemar
continuous	Wilcoxon signed-rank	paired $t$ -test

12.2.2 Comparing Two Groups

Type of Response	Recommended Test	Most Frequent Test
binary	$2 \times 2 χ^{2}$	$χ^{2}$ , Fisher’s exact test
ordinal	Wilcoxon 2-sample	Wilcoxon 2-sample
continuous	Wilcoxon 2-sample	2-sample $t$ -test
time to event²	Cox model³	log-rank⁴

² The response variable may be right-censored, which happens if the subject ceased being followed before having the event. The value of the response variable, for example, for a subject followed 2 years without having the event is 2+.

³ If the treatment is expected to have more early effect with the effect lessening over time, an accelerated failure time model such as the lognormal model is recommended.

⁴ The log-rank is a special case of the Cox model. The Cox model provides slightly more accurate $P$ -values than the $χ^{2}$ statistic from the log-rank test.

12.2.3 Comparing $> 2$ Groups

Type of Response	Recommended Test	Most Frequent Test
binary	$r \times 2 χ^{2}$	$χ^{2}$ , Fisher’s exact test
ordinal	Kruskal-Wallis	Kruskal-Wallis
continuous	Kruskal-Wallis	ANOVA
time to event	Cox model	log-rank

12.2.4 Correlating Two Continuous Variables

Recommended: Spearman $ρ$
Most frequently seen: Pearson $r$

12.3 Covariable-Adjusted Analyses

To adjust for imbalances in prognostic factors in an observational study or for strong patient heterogeneity in a randomized study
Analysis of covariance is preferred over stratification, especially if continuous adjustment variables are present or there are many adjustment variables
- Continuous response: multiple linear regression with appropriate transformation of $Y$
- Binary response: binary logistic regression model
- Ordinal response: proportional odds ordinal logistic regression model
- Time to event response, possibly right-censored:
  - chronic disease: Cox proportional hazards model
  - acute disease: accelerated failure time model

# Statistical Inference Review * Emphasize confidence limits, which can be computed from adjusted or unadjusted analyses, with or without taking into account multiple comparisons * $P$-values can accompany CLs if formal hypothesis testing needed * When possible construct $P$-values to be consistent with how CLs are computed ## Types of Analyses * Except for one-sample tests, all tests can be thought of as testing for an association between at least one variable with at least one other variable * Testing for group differences is the same as testing for association between group and response * Testing for association between two continuous variables can be done using correlation (especially for unadjusted analysis) or regression methods; in simple cases the two are equivalent * Testing for association between group and outcome, when there are more than 2 groups which are not in some solid order^[The dose of a drug or the severity of pain are examples of ordered variables.] means comparing a summary of the response between $k$ groups, sometimes in a pairwise fashion ## Covariable-Unadjusted Analyses Appropriate when * Only interested in assessing the relationship between a single $X$ and the response, or * Treatments are randomized and there are no strong prognostic factors that are measureable * Study is observational and variables capturing confounding are unavailable (place strong caveats in the paper) See @sec-ancova ### Analyzing Paired Responses | Type of Response | Recommended Test | Most Frequent Test | |-----|-----|-----| | binary | McNemar | McNemar | | continuous | Wilcoxon signed-rank | paired $t$-test | ### Comparing Two Groups | Type of Response | Recommended Test | Most Frequent Test | |-----|-----|-----| | binary | $2\times 2~\chi^{2}$ | $\chi^{2}$, Fisher's exact test | | ordinal | Wilcoxon 2-sample | Wilcoxon 2-sample | | continuous | Wilcoxon 2-sample | 2-sample $t$-test | | time to event^[The response variable may be right-censored, which happens if the subject ceased being followed before having the event. The value of the response variable, for example, for a subject followed 2 years without having the event is 2+.] | Cox model^[If the treatment is expected to have more early effect with the effect lessening over time, an accelerated failure time model such as the lognormal model is recommended.]| log-rank^[The log-rank is a special case of the Cox model. The Cox model provides slightly more accurate $P$-values than the $\chi^2$ statistic from the log-rank test.] | ### Comparing $>2$ Groups | Type of Response | Recommended Test | Most Frequent Test | |-----|-----|-----| | binary | $r\times 2~\chi^{2}$ | $\chi^{2}$, Fisher's exact test | | ordinal | Kruskal-Wallis | Kruskal-Wallis | | continuous | Kruskal-Wallis | ANOVA | | time to event | Cox model | log-rank | ### Correlating Two Continuous Variables Recommended: Spearman $\rho$ <br> Most frequently seen: Pearson $r$ ## Covariable-Adjusted Analyses * To adjust for imbalances in prognostic factors in an observational study or for strong patient heterogeneity in a randomized study * Analysis of covariance is preferred over stratification, especially if continuous adjustment variables are present or there are many adjustment variables + Continuous response: multiple linear regression with appropriate transformation of $Y$ + Binary response: binary logistic regression model + Ordinal response: proportional odds ordinal logistic regression model + Time to event response, possibly right-censored: - chronic disease: Cox proportional hazards model - acute disease: accelerated failure time model