Вы находитесь на странице: 1из 2

ONE WAY ANOVA (ch.

11) FACTORIAL DESIGN 𝑺𝑺𝑬


Step 4 Standard Error of the Estimate 𝑺𝒀𝑿 = √𝒏−𝟐
Sum of Source D.F. S.S. Mean Square F
Source D.F. Mean Square (Variance) F The typical difference between actual [business situation X variable] and the predicted [business
Squares Sample (A) r-1 SSA MSA = SSA / (r – 1) FSTAT = MSA / MSE
Among Groups c-1 SSA MSA = SSA / (c – 1) FSTAT = MSA / MSW situation X variable] using the regression equation is approximately ###.
Among Groups (B) c-1 SSB MSB = SSB / (c – 1) FSTAT = MSB / MSE
Within Groups n-c SSW MSW = SSW / (n – c) Interaction (AB) (r-1) (c-1) SSAB MSAB = SSAB / (r-1)(c-1) FSTAT = MSAB / MSE Step 5 Residual Scatter Plot Analysis (Linearity) residual = ei = 𝒀𝒊 − 𝒀̇𝒊
Total n-1 SST MST = SST / (n – 1) Error (E) r*c(n’-1) SSE MSE = SSE / r*n(n’ – 1) Plot the residuals on the Y axis and independent variable on the X axis. If there does not appear to
Note SSA = MSA(c - 1); SSW = MSW(n - c); SST = SSW + SSA Total n-1 SST be a pattern in the residual plot, the assumptions of the regression do not appear to be seriously
Step 1 Find # of Groups (c) and # of values in all groups (n) violated. If you see a trend in the residuals as x increases, the linear model is not appropriate
n = total # of values in all groups combined; c = # of groups/treatments/populations; because there is a ralationship between the residuals and the independent variable x.
Step 6 The Durbin-Watson Statistic (Independence of errors for sequentially collected data)
Step 2 Calculate Test Statistic & Critical Value (using degrees of freedom)
When the data are collected over a consecutive period, you need to determine whether there is
Step 3 Accept or Reject Ho (null hypothesis) autocorrelation, correlation between each residual and residual for the previous time period.
Hypothesis H0 : µ1 = µ2 = µ3 = µj
D = SSDifRes / Sum of Squared Residuals SSDifRes = SSRegre - SSResid
H1 : Not all µj are equal (where j = 1, 2, 3, … , c) Level of Significance (α) = 0.05
α = significance level ; n = sample size ; k = # of independent variables
Test Statistic FCRIT = (c – 1) df1 in the Numerator, (n – c) df2 in the denominator
Decision Rule If FSTAT > FCRIT then reject Ho. Alternatively, if p-value < α, then reject Ho FIND dL and dU (dL: lower critical value of D, dU: upper critical value of D) in the table. If D < dL, you
Decision Since [is/is not] FSTAT > FCRIT, we [do/do not] reject H0 at 5% level of significance. […] Step 1 Test for Interaction Effect conclude that there is evidence of positive autocorrelation among the residuals. The least-squares
Conclusion We are 95% confident that there [is/is not] enough evidence to conclude that there F-test for Interaction Effect (AB): Interaction regression analysis of the data is inappropriate because of the presence of significant positive
is a significant difference in the mean of the ### groups. Hypothesis H0: There is no interaction between A and B autocorrelation among the residuals. In other words, the independence-of-errors assumption is
Step 4 Tukey-Kramer Procedure (compare each individual group’s mean to one another) H1: There is an interaction between A and B Level of Significance α = 0.05 invalid. You need to use alternative approaches. If D > dU, you conclude that there is no evidence of
Test Statistic FCRIT ; df1 = (r – 1)(c – 1) in the Num., df2 = rc(n’ – 1) in the Denominator positive autocorrelation among the residuals. If dL < D < dU, you are unable to arrive at a definite
𝑀𝑆𝑊 1 1
Critical Range (Q α; df1 ; df2) = (Qstat)*√ ∗ (𝑛𝑗 + 𝑛𝑗 ′ ) ; df1 (numerator) =c ;df2 (denominator) = n - c Decision Rule FSTAT > FCRIT OR p-value < α Reject H0, otherwise do not conclusion.
2
Decision Since FSTAT > FCRIT, we reject H0 at 5% level of significance.
|𝑋̅1 − 𝑋̅2 | = ## |𝑋̅1 − 𝑋̅3 | = ## |𝑋̅2 − 𝑋̅3 | = ## Step 7 Normality
Conclusion We are 95% confident that there is enough evidence to conclude that there is a
Errors must be randomly distributed at each value of X. You can test normality by constructing a
Decision Rule |Xj − Xj’| is greater than the critical range then the groups are significantly different significant interaction between A and B. Thus, Factor A and Factor B are correlated
histogram, stem-and-leaf display, boxplot, or normal probability plot. If the data creates a line in a
Decision We are 95% confident that groups # & # are significantly different. with each other.
normal probability plot and does not deviate substantially from a normal distribution, there is no
Step 5 Levene Test Step 2 Test for Difference in Means of Just Factor A need for concern.
Absolute value of the difference between each data point and its median for that group. F-test for Factor A Effect (A): Sample Step 8 Residual Analysis
Hypothesis H0 : σ12 = σ22 = σ32 = σj2 Hypothesis H0 : µ1.. = µ2.. = … = µr.. Residual = Absolute value of the difference between Y and the predicted value of Y for a given
H1 : Not all σj are equal (where j = 1, 2, 3, … , c) Level of Significance α = 0.05 H1 : Not all µj are equal (where j = 1, …, r) Level of Significance α = 0.05 value of X. You can plot this data against the change in the X variable to see if there is an equal
Test Statistic FSTAT , FCRIT , P-value Test Statistic FCRIT ; df1 = (r – 1) in the Numerator, df2 = rc(n’ – 1) in the Denominator amount of variation of the residuals at each value of X.
Decision Rule If FSTAT > FCRIT or P-value < α then reject H0, otherwise do not. Decision Rule FSTAT > FCRIT or p-value < α Reject H0, otherwise do not
Decision Since P-value > α, we do not reject H0 at 5% level of significance. […] Decision Since FSTAT > FCRIT, we reject H0 at 5% level of significance. Step 9 t Test for Slope 𝑺𝒃𝟏 , 𝒃𝟏 , 𝜷𝟏 , n , d.f. , 𝒕𝑺𝑻𝑨𝑻 , α
𝒃𝟏 − 𝜷𝟏
Conclusion We are 95% confident that there [is/is not] enough evidence to conclude that there Conclusion We are 95% confident that there is enough evidence to conclude that there is an H0: β1 = 0 [There is no linear relationship (the slope is zero)] 𝒕𝑺𝑻𝑨𝑻 =
H :β ≠0 [There is a linear relationship (the slope is not zero)] 𝑺𝒃𝟏
is a significant difference in the variation effect due to Factor A (Sample). 1 1
*𝑆𝑏1 = 𝑆𝑌𝑋 / SSX (if not given)
RANDOMIZED BLOCK DESIGN Step 3 Test for Difference in Means of Just Factor B
Decision Rule If tSTAT is [greater] than tCRIT we [can] reject H0 *(TCRIT)^2= FCRIT
F-test for Factor B Effect (B): Columns
Source D.F. S.S. Mean Square F Alternatively, if p-value is [less] α we [can] reject H0
Hypothesis H0 : µ.1. = µ.2. = … = µ.c.
Among Groups (A) c-1 SSA MSA = SSA / (c – 1) FSTAT = MSA / MSE Decision [Because |tSTAT|> tCRIT (α/2, n–2 degrees of freedom) or p-value < α] → Reject H0
H1 : Not all µj are equal (where j = 1, 2, …, c) Level of Significance α = 0.05
Hence, you can conclude that there is a significant linear relationship between mean
Among Blocks (BL) r-1 SSBL MSBL = SSBL / (r – 1) FSTAT = MSBL / MSE Test Statistic FCRIT , df1 = (c – 1) in the Numerator, df2 = rc(n’ – 1) in the Denominator
dependent variable and the independent variable.
Error (E) (r-1) (c-1) SSE MSE = SSE / (r-1)(c-1) Decision Rule FSTAT > FCRIT or p-value < α Reject H0, otherwise do not
𝑴𝑺𝑹 𝑺𝑺𝑬
Total rc-1 SST Decision Since FSTAT > FCRIT, we reject H0 at 5% level of significance. Step 10 F Test for Slope (alternative to the T-test) 𝑭𝑺𝑻𝑨𝑻 = 𝑴𝑺𝑬; 𝑴𝑺𝑹 = 𝑺𝑺𝑹 ;𝑴𝑺𝑬 = 𝒏 − 𝟐
Note A = Columns / Groups ; BL = Row / Blocks Conclusion We are 95% confident that there is enough evidence to conclude that there is an
effect due to Factor B (Groups). Decision Rule Reject H0 if FSTAT > Fα. Otherwise do not reject H0. (FCRIT ; df1 = 1, df2 = n – 2)
Step 1 Find the # of blocks (r) [rows], # of groups (c) [columns], and # of values (n) [n = rc]
c = # of groups ; r = # of blocks ; n = # of observations Step 4 Tukey Procedure
For Factor A (Differences between Samples):
Step 2 Calculate Test Statistic for Columns (Groups) AND/OR Rows (Blocks)
Critical Range for Factor A = Qstat√𝑴𝑺𝑬/𝒄𝒏′
F Test for Factor Effect (A): Columns (To know the effect due to groups)
Qα ; df1 = r in the Numerator ; df2 = rc(n’ – 1) in the Denominator
Hypothesis H0: µ.1 = µ.2 = … = µ.c
H1: Not all µ.j are equal (where j = 1, 2, …, c) Level of significance α = 0.05 For Factor B (Determine differences between groups):
Test statistic FCRIT ; df1 = (c – 1) in the Numerator, df2 = (r – 1)(c – 1) in the Denominator Critical Range for Factor B = Qstat√𝑴𝑺𝑬/𝒓𝒏′ [w/ df1 = c, df2 = rc(n’-1)
Decision Rule FSTAT > FCRIT or P-value < α Reject H0, otherwise do not Qα ; df1 = c in the Numerator ; df2 = rc(n’ – 1) in the Denominator
Decision Since FSTAT > FCRIT, we reject H0 at 5% level of significance. Decision rule: |Xi − Xj| > Critical Range  the groups are significantly different Step 11 Confidence Interval Estimate of the Slope 𝒃𝟏 − 𝒕∝/𝟐 𝑺𝒃𝟏 ≤ 𝜷𝟏 ≤ 𝒃𝟏 + 𝒕∝/𝟐 𝑺𝒃𝟏
Conclusion We are 95% confident that there is enough evidence to conclude that there is a Decision: We are 95% confident that all pairwise comparisons except groups # & # are greater Note *** Dealing with ∝/𝟐 NOT JUST ∝
significant difference in the mean the groups. than the critical range. Therefore, you conclude with 95% confidence that the With a [1 - ∝] % degree of confidence that estimated population slope is between ### and ###. The
F Test for Block Effects (BL): Rows (To examine whether the blocking was advantageous) population mean is different between all pairs of [x] except for # and #. In addition, confidence interval indicates that for each increase of [X], the predicted [Y] is estimated to increase
Hypothesis H0: µ1. = µ2. = … = µr. # has the highest mean ratings (i.e., most preferred) and # has the lowest (i.e., least by ### but no more than ###. Because both of these values are [above/below] 0 there [is/is not]
H1: Not all µi. are equal, (where j = 1, 2, … , r) Level of significance α = 0.05 preferred). By using a = 0.05, you are able to make all # comparisons with an overall evidence of a significant linear relationship between [Y] and the [X].
Test Statistic FCRIT ; df1 = (r – 1) in the Numerator, df2 = (r – 1)(c – 1) in the Denominator error rate of only 5%. 𝒓−ρ
Step 12 t Test for the Correlation Coefficient 𝒕𝑺𝑻𝑨𝑻 =
Decision Rule FSTAT > FCRIT or P-value < α Reject H0, otherwise do not H0: ρ1 = 0 (no correlation) 𝟐
√𝟏 − 𝒓
Decision Since FSTAT > FCRIT, we reject H0 at 5% level of significance. SIMPLE LINEAR REGRESSION (CH 13) H1: ρ 1 ≠ 0 (correlation) 𝒏−𝟐
Conclusion We are 95% confident that there is enough evidence to conclude that there is a Step 1 Simple Linear Regression Model (for a population) 𝒀̂𝒊 = 𝛃𝟎 + 𝛃𝟏 𝑿𝒊 + 𝛆𝒊 Decision Using a ### level of significance, because 𝒕𝑺𝑻𝑨𝑻 = ### > tCRIT , you reject the null
significant difference (block effect) among blocks in the experiment. Thus, you hypothesis. You conclude there is a significant association between [X] and [Y].
conclude that the blocking has been advantageous in reducing the random error. Step 2 ̂𝒊 = 𝐛𝟎 + 𝒃𝟏 𝑿𝒊
Simple Linear Regression Equation (Prediction Line) (for a sample) 𝒀
The slope, 𝑏1, is ###. This means for each increase of 1 unit in X, the predicted mean value of Y is estimated Step 13 Confidence Interval Est. for the mean of Y (Predict. Interval for Indiv. Response Y)
Step 3 Test the Relative Efficiency (did the blocking result in an increase in precision) ̂ 𝒊 − 𝒕∝/𝟐 𝑺𝒀𝑿 √𝒉𝒊 ≤ 𝝁𝒀|𝑿=𝑿 ≤ 𝒀 ̂ 𝒊 + 𝒕∝/𝟐 𝑺𝒀𝑿 √𝒉𝒊
to increase by ###. In other words, for each increase of [business situation X variable], the predicted mean 𝒀 tα/2 from t distribution w/ n-2 d.f.
(𝒓 − 𝟏)𝑴𝑺𝑩𝑳 + 𝒓(𝒄 − 𝟏)𝑴𝑺𝑬 𝒊
𝑹𝑬 = [business situation Y variable] are estimated to increase by [business situation 𝒃𝟏 ]. The Y intercept, 𝑏0, is (Note *** Dealing with (∝/2) NOT JUST ∝)
(𝒓𝒄 − 𝟏)𝑴𝑺𝑬 ###. The Y intercept represents the predicted value of Y when X equals zero. Because the [business 𝟐
𝟏 ̅)
(𝑿𝒊 −𝑿
It would take [RE] times as many observations in a one-way ANOVA design as compared to the situation X variable] cannot be 0, this Y intercept has [significant/little or no practical] interpretation. Also, 𝒉𝒊 = 𝒏 + 𝑺𝑺𝑿 ; ̂ 𝒊 = predicted Y value ;
𝒀 𝑺𝒀𝑿 = standard error of estimate ;
randomized block design in order to have the same precision in comparing the groups. the Y intercept for this example is [inside/outside] the range of the observed values of the X variable, and n = sample size; 𝑿𝒊 = given value of X; 𝝁𝒀|𝑿=𝑿𝒊 = mean value of Y when X = 𝑿𝒊 ;
therefore interpretations of the value of 𝑏0 should be made [comfortably/cautiously]. 𝑺𝑺𝑿 = ∑𝒏𝒊=𝟏( 𝑿𝒊 − 𝑿̅ )𝟐
Step 4 Tukey-Kramer Procedure
** (compute b1 = SSXY / SSX) (compute 𝑏0 = Mean Y - (b1 * Mean X)) **
𝑴𝑺𝑬 With a 95% confidence, we can estimate that the mean number of [Y variable] for the entire
Critical Range (Q α; df1 ; df2) = (Qstat)*√ ; Q: df1 (numtr.) = c; df2 (denmtr) = (r – 1)(c – 1) Step 3 Coefficient of Determination; 𝒓𝟐 =
𝑹𝒆𝒈𝒓𝒆𝒔𝒔𝒊𝒐𝒏 𝒔𝒖𝒎 𝒐𝒇 𝒔𝒒𝒖𝒂𝒓𝒆𝒔 𝑺𝑺𝑹
= 𝑺𝑺𝑻 SST = SSR + SSE
𝒓
𝑻𝒐𝒕𝒂𝒍 𝒔𝒖𝒎 𝒐𝒇 𝒔𝒒𝒖𝒂𝒓𝒆𝒔 population of [Subject being studied (stores for example)] with ### [X variable] is between ###
Decision Rule |Xj − Xj’| is greater than the critical range then the groups are significantly different Therefore, [𝐫 𝟐 ]*100 = the % of the variation in [Y variable] that is explained by the variability and ###.
Decision We are 95% confident that groups # & # are significantly different. (change) in the number of [X variable]. This [large/small] 𝑟 2 indicates a [strong/weak] linear Step 14 Prediction Interval Est. for an Individual Response, Y
relationship between these two variables, because the regression model explains [𝐫 𝟐 ] % of the ̂ 𝒊 − 𝒕∝/𝟐 𝑺𝒀𝑿 √𝟏 + 𝒉𝒊 ≤ 𝒀𝑿=𝑿 ≤ 𝒀
𝒀 ̂ 𝒊 + 𝒕∝/𝟐 𝑺𝒀𝑿 √𝟏 + 𝒉𝒊
𝒊
variability in predicting [Y variable]. [Only] [1 - 𝐫 𝟐 value] of the sample variability in [Y variable] is Prediction Interval is wider than Confidence Interval of mean estimate due to its variability.
due to factors other than what is accounted for by the linear regression model.
MULTIPLE LINEAR REGRESSION (CH14) Coefficient of Multiple Determination:
𝑹𝒆𝒈𝒓𝒆𝒔𝒔𝒊𝒐𝒏 𝒔𝒖𝒎 𝒐𝒇 𝒔𝒒𝒖𝒂𝒓𝒆𝒔 𝑺𝑺𝑹 (𝒏−𝟏)
Step 1 Multiple Regression Model with k Independent Variables (for population) Step 9 Coefficients of Partial Determination 𝒓𝟐 = 𝑻𝒐𝒕𝒂𝒍 𝒔𝒖𝒎 𝒐𝒇 𝒔𝒒𝒖𝒂𝒓𝒆𝒔 = 𝑺𝑺𝑻 ; 𝑹𝒂𝒅𝒋 𝟐 = 𝟏 − [(𝟏 − 𝒓𝟐 ) (𝒏−𝒌−𝟏)] , (k=2 because of X1 &
𝒀𝒊 = 𝛃𝟎 + 𝛃𝟏 𝑿𝟏𝒊 + 𝛃𝟐 𝑿𝟐𝒊 + ⋯ 𝛃𝒌 𝑿𝒌𝒊 + 𝛆𝒊 2 SSR(X1 |X2 ) 2
𝑟 𝑌1.2 = X1 impact on Y X1 )
𝑆𝑆𝑇− SSR(X1 and X2 )+ SSR(X1 |X2 )
Step 2 Multiple Regression Model with Two Independent Variables (for sample) SSR(𝐗 𝟏 |𝐗 𝟐 ) The coefficient indicates that the ## % of the variation in Y is explained by the quadratic
𝒀𝒊 = 𝐛𝟎 + 𝐛𝟏 𝑿𝟏𝒊 + 𝐛𝟐 𝑿𝟐𝒊 𝑟 2𝑌2.1 = X2 impact on Y relationship between Y and X. Then compute radj2 to account for the number of independent
𝑆𝑆𝑇− SSR(X1 and X2 )+ SSR(𝐗 𝟐 |𝐗 𝟏 )
Step 3 Coefficient of Multiple Determination SST – SSR = SSE SSR(𝐗 𝒋|𝐀𝐥𝐥 𝐗𝐬 𝐞𝐱𝐞𝐜𝐩𝐭 𝐣) variables and the sample size. We may conclude that there is a strong linear relationship between Y
𝑟 2𝑌𝑗.(𝑎𝑙𝑙 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒𝑠 𝑒𝑥𝑐𝑝𝑒𝑡 𝑗) = for k independent var and X.
𝒓𝟐 = 𝑹𝒆𝒈𝒓𝒆𝒔𝒔𝒊𝒐𝒏 S𝒖𝒎 𝒐𝒇 S𝒒𝒖𝒂𝒓𝒆𝒔 / 𝑻𝒐𝒕𝒂𝒍 S𝒖𝒎 𝒐𝒇 S𝒒𝒖𝒂𝒓𝒆𝒔 = 𝑺𝑺𝑹 / 𝑺𝑺𝑻 𝑆𝑆𝑇− SSR(All Xs)+ SSR(𝐣|𝐀𝐥𝐥 𝐗𝐬 𝐞𝐱𝐞𝐜𝐩𝐭 𝐣)
𝟏
Step 10 Dummy Variable (X2 = 1 [Yes] OR X2 = 0 [No]) Step 5 Variance Inflationary Factor (for factor j): 𝑽𝑰𝑭𝒋 = 𝟏−𝑹𝟐 (VIFj = 1  No collinearity)
Therefore, [𝐫𝟐 value]*100% of the variation in [Y] is explained by the variability of [X]. The r2 value 𝒋
Interactions for categorical variable X2 (for equation 𝒀𝒊 = 𝛃𝟎 + 𝛃𝟏 𝑿𝟏𝒊 + 𝛃𝟐 𝑿𝟐𝒊 + 𝛆𝒊 ) r2 = coefficient of multiple determination using variable Xj as the dependent variable
indicates a [strong/weak] linear relationship between these two variables, b/c the regression
H0: β3 = 0 (There is no interaction effect between independent variables)
model has explained [𝐫𝟐 value]*100% of the variability in predicting [Y]. [Only] [1 - 𝐫𝟐 value] of the VIFj > 5Remove Xj. Too much correlation between Xj and the other independent variables
H1: β3 ≠ 0 (There is an interaction effect between independent variables)
sample variability in [Y] is due to other factors other outside the linear regression model. Decision Rule If FSTAT> FCRIT OR if p-value < α  Reject H0 VIFj < 5On the basis of the criteria developed by Snee that all VIF values should be less than 5.0,
Conclusion There is enough evidence that the interaction term contributes to the model. there is little evidence of collinearity among the set of independent variables.
Adjusted r2
𝑛−1 Note Can do Coefficients of Partial Determination If more than one X has a VIF > 5, eliminate the X with the highest VIF and rebuild a regression
𝑟 2 𝑎𝑑𝑗 = 1 − [(1 − 𝑟 2 ) 𝑛−𝑘−1] k = number of independent variables
Step 11 Logistics Regression (Y = 1 or Y = 0) model. If there are only two independent variables in the model, VIF1 = VIF2. Removing the largest
[repeat from above]
Part A Logistic Regression Equation: 𝒍𝒏 (𝑬𝒔𝒕. 𝑶𝒅𝒅𝒔 𝑹𝒂𝒕𝒊𝒐) = 𝐛𝟎 + 𝐛𝟏 𝑿𝟏𝒊 + 𝐛𝟐 𝑿𝟐𝒊 + ⋯ + 𝐛𝒌 𝑿𝒌𝒋 VIF value makes the model free of collinearity problems. Highly correlated variables do not provide
𝑴𝑺𝑹
Step 4 Overall F-test 𝑭𝑺𝑻𝑨𝑻 = 𝑴𝑺𝑬 df1 = k ; df2 = n – k – 1 Part B 𝑬𝒔𝒕. 𝑶𝒅𝒅𝒔 𝑹𝒂𝒕𝒊𝒐 = 𝒆𝒍𝒏 (𝑬𝒔𝒕.𝑶𝒅𝒅𝒔 𝑹𝒂𝒕𝒊𝒐) unique information and the r2 may fluctuate a lot. Large VIFs do not necessarily mean that the
H0 : 𝛽1 = 𝛽2 = … = 𝛽𝑗 = 0 (There is no significant relationship between the) 𝑬𝒔𝒕.𝑶𝒅𝒅𝒔 𝑹𝒂𝒕𝒊𝒐 model is not a good one, but they require you to carefully interpret the regression coefficients.
Part C 𝑬𝒔𝒕. 𝑷𝒓𝒐𝒃𝒂𝒃𝒊𝒍𝒊𝒕𝒚 𝒐𝒇 𝒂𝒏 𝑬𝒗𝒆𝒏𝒕 = 𝟏+𝑬𝒔𝒕.𝑶𝒅𝒅𝒔 𝑹𝒂𝒕𝒊𝒐
H1 : 𝑁𝑜𝑡 𝑎𝑙𝑙 𝛽𝑗 = 0 (there is a significant relationship between the two independent variables) Stepwise Regression
Step 11 Deviance Statistic (deviance may be given) (repeat for all slopes of X var) By entering/deleting variables one by one you only get one model. Good when # of Xs < 30
Decision Rule If FSTAT > FCRIT or if p-value < ∝ , the null hypothesis H0 must be rejected.
Critical value for a X2 (chi squared) with n – k – 1 degrees of freedom is ____. (FIND IN X2 TABLE) tSTAT > tCRIT, FSTAT > FCRIT, p-value < α  Enter the variable in the model, otherwise delete it
Decision Because FSTAT > FCRIT and p-value < ∝ , we reject H0 and conclude that at least one of H0: The model is a good-fitting model
the independent variables is related to [Y]. Best-Subsets Regression
H1: The model is not a good-fitting model
Decision Rule Reject H0 if deviance > (X 2α) ; otherwise, do not reject H0. 1st criterion: radj2 (REFER TO CH14 Step 3)
Decision Because all Dev < X 2α , This model represents a good fitting model. is more appropriate than r2 because model building requires you to compare models with different
𝒃𝒋 𝒔𝒍𝒐𝒑𝒆 𝒐𝒇 𝒋 numbers of independent variables.
Step 12 Wald Statistics: 𝒁 = 𝑺𝒃 = 𝒔𝒕𝒂𝒏𝒅𝒂𝒓𝒅 𝒆𝒓𝒓𝒐𝒓 𝒐𝒇 𝒋 (α = 1.96 z-critical) (𝟏−𝑹𝟐𝒌 )(𝒏−𝑻)
𝒍𝒏 (𝑶𝒅𝒅𝒔 𝑹𝒂𝒕𝒊𝒐) = 𝛃𝟎 + 𝛃𝟏 𝑿𝟏𝒊 + 𝛃𝟐 𝑿𝟐𝒊 + 𝛆𝒊
𝒋
2nd criterion: Cp Statistic: 𝐶𝑝 =
𝟏−𝑹𝟐𝑻
− [𝒏 − 𝟐 (𝒌 + 𝟏)]
H0: βj = 0 [The knowledge of Xj does not make a significant contribution to the model] (T = total # of parameters including the intercept, rk2=coefficient for a model with k Xs,
H1: βj ≠ 0 [The knowledge of Xj makes a significant contribution to the model] rT2 = coefficient for a model with all T parameters)
Decision Rule If ZSTAT > ZCRIT we reject H0.
Alternatively, if p-value is < α we reject H0. Cp measures the differences between a fitted model and a true model, along with random error.
Step 5 The Durbin-Watson Statistic (Independence of errors for seq. collected data) Decision [Because ZSTAT > ZCRIT or p-value < α ] → Reject H0 LIST ALL models that have Cp <= (k + 1). If there is multiple choose the one with the highest radj2.
Must complete a Durbin Watson Analysis by finding the critical value on the table, using n and k, Hence, you can conclude that the independent variable makes a significant
and the significance level. Must then do a complete statistical analysis. (can refer to CH 13 Step 6) contribution to the model at a 0.05 level of significance. Taking into account the Test Statistic: if Models have same radj2
D = Sum of Squared Difference of Residuals / Sum of Squared Residuals presence of the other variable[s]. H0 : 𝛽𝑗 = 0
α = significance level ; n = sample size ; k = # of independent variables Step 13 The Hat Matrix Diagonal Elements, hi = hat matrix diagonal element for obs i H1 : 𝛽𝑗 ≠ 0
If DSTAT > DU, then there is no positive autocorrelation. If DSTAT < DL, then there is a positive If hi > 2(k + 1) / n , then Xi is an influential observation and is a candidate for removal from the Decision Rule If tSTAT > tCRIT we reject H0.
autocorrelation. IF DL < DSTAT < DU, then there is insufficient information to make a decision. model. Thus, none of the observations are candidates for the removal from the analysis. Alternatively, if p-value is < α we reject H0.
Step 6 t Test for Slopes (Check for all βs) 𝑺𝒃𝒋 , 𝒃𝒋 , 𝜷𝒋 , n , d.f. , 𝒕𝑺𝑻𝑨𝑻 , α; 𝒃𝒋 − 𝜷𝒋 𝒏−𝒌−𝟏 Decision [Because tSTAT > tCRIT (α, n–k–1 degrees of freedom) or p-value < α ] → Reject H0
𝒕𝑺𝑻𝑨𝑻 = Step 14 The Studentized Deleted Residuals, ti: 𝒆𝒊 = residual for obs i ; 𝒕𝒊 = 𝒆𝒊 √𝑺𝑺𝑬 (𝟏−𝒉 )−𝒆𝟐
𝑺𝒃𝒋 𝒊 𝒊 Based on the evidence, we are 95% confident that variable 𝐗 𝒋 does make a
H0: βj = 0 [There is no linear relationship (the slope is zero)]
Adjusted t-Stat. measures the dif. of each Yi, from the value predicted by a model that includes all significant contribution to the model
H1: βj ≠ 0 [There is a linear relationship (the slope is not zero)] Note: upper tail area = α / 2
obs. except i. If ti > t α/2 or ti < t α/2 , the observed and predicted values are so different that obs. i is
Decision Rule If tSTAT > tCRIT we reject H0.
highly influential on the regression and a candidate for removal. hi = hat matrix diagonal element Residual Analysis and Influence Analysis
Alternatively, if p-value is < α we reject H0.
Decision [Because tSTAT > tCRIT (α, n–k–1 degrees of freedom) or p-value < α ] → Reject H0 𝒆𝟐𝒊 𝒉𝒊 Residual Analysis: Durbin-Watson (REFER TO CH 13 Step 6)
Step 15 Cook’s Distance Statistic, Di: 𝑫𝒊 = [ ]
Hence, you can conclude that there is a significant linear relationship between mean 𝒌×𝑴𝑺𝑬 (𝟏−𝒉𝒊 )𝟐 D > 2  There is no indication of positive correlation in the residuals
dependent variable and the independent variable (Xj), Taking into account the other If Di > Fα (FCRIT; α=0.05, numerator df= k + 1 , and denominator= n – k – 1)  candid. for removal If Residual Analysis confirms aptness of model  Do CH 14 Step 13 to 15
independent variables. Include Xs in the model.
Step 7 Confidence Interval Estimate of the Slope (of population) MULTIPLE NON-LINEAR REGRESSION (CH15)
𝒃𝒋 − 𝒕∝/𝟐 𝑺𝒃𝒋 ≤ 𝜷𝒋 ≤ 𝒃𝒋 + 𝒕∝/𝟐 𝑺𝒃𝒋 Note: look @ SE for [b var] = 𝑺𝒃𝒋 The Quadratic Regression
With a [1 - ∝] *100% degree of confidence that estimated population slope is between ### and Quadratic Regression Model 𝒀𝒊 = 𝛃𝟎 + 𝛃𝟏 𝑿𝟏𝒊 +𝛃𝟐 𝑿𝟏𝒊 𝟐 + 𝛆𝒊
###. The confidence interval indicates that for each increase of [X], the predicted [Y] is estimated β0 = Y intercept, β1 = coefficient of the linear effect on Y,
to increase by ### but no more than ###. Because both of these values are [above/below] 0 there β2 = coefficient of the quadratic effect on Y, εi = random error in Y for observation i
[is/is not] evidence of a significant linear relationship between [Y] and the [X]. Quadratic regression- non-linear relationship between variables
**Quadratic term - 2nd independent variable, is the square of the 1st independent variable
Step 8 Partial F Test to determine Contribution of Independent Variables
Quadratic Regression Equation 𝒀̂𝒊 = 𝐛𝟎 + 𝐛𝟏 𝑿𝟏𝒊 +𝐛𝟐 𝑿𝟏𝒊 𝟐
Note Alternative to T-test. In order to determine the contributions of each independent
variable to the regression sum of squares, after all other independent variables have Significance of Quadratic Model Test for X1
been included in the model. H0: β1 = β2 = … = βj = 0 (There is no overall relationship between X1 and Y)
H1: Not all βj are equal to 0 (There is an overall relationship between X1 and Y) (α = 0.05)
Note This involves doing a simple linear regression for each of the X variables. 𝑴𝑺𝑹
Test statistic 𝑭𝑺𝑻𝑨𝑻 = 𝑴𝑺𝑬 , FCRIT (α; df1 = k, df2 = n–k–1)
Source D.F. S.S. Mean Square F
Regression k SSR(X1 and X2) (refer to Step 4 MSRegr) Decision rule FSTAT > FCRIT OR P-value < α Reject H0, otherwise do not
(refer to Step 4 SSRegr) Decision Since FSTAT > FCRIT, we reject H0 at 5% level of significance.
𝐒𝐒𝐑(𝐗𝟏𝐥 𝐗𝟐) Conclusion We are 95% confident that there is a significant overall relationship between Y and
{ X2 n/a {SSR(X2) SSR(X1 and X2) - SSR(X2)
𝐌𝐒𝐄 X. (If you reject Ho, it is a good fitting model, if you don’t reject Ho – it is a bad fitting model) q5
SSRX1|X2 } SSR(X1 and X2) - SSR(X2)}
Error n–k (refer to Step 4 SSResid) (refer to Step 4 MSResid) Quadratic Effect Test
Total SUM SSRegr + X1|X2 + SSResid SUM H0: β2 = 0 (Including the quadratic effect (X12) does not significantly improve the model)
H1: β2 ≠ 0 (Including the quadratic effect significantly improves the model) (α = 0.05)
Note *** Flip the X2 and X1 for the partial F test for X2 𝒃 −𝜷
Test statistic 𝒕𝑺𝑻𝑨𝑻 = 𝟐𝑺 𝟐 , tCRIT (α/2, df = n–k-1)
𝒃𝟐
H0: X1 does not significantly contribute to the model after the variable X2 is included Decision rule tSTAT > tCRIT or P-value < α Reject H0, otherwise do not
H1: X1 contributes to the model after X2 is included Decision Since TSTAT > TCRIT, we reject H0 at 5% level of significance.
𝐒𝐒𝐑(𝐗𝟏𝐥 𝐗𝟐)
Test Statistic 𝐅𝑺𝑻𝑨𝑻 = 𝐌𝐒𝐄 FCRIT ; d.f. = n – k – 1 ; Note: upper tail area = α / 2 Conclusion We are 95% confident that the quadratic model is significantly better than the linear
Decision Rule If FSTAT > FCRIT, then reject the Null Hypothesis H0 model for representing the relationship between Y and X. (If do not reject H0) We are 95%
Decision Because FSTAT > FCRIT, We reject the null hypothesis, and conclude at a .05 level of confident that the linear model is a better fit than the quadratic model for representing the
significance that the contribution of X1 significantly improves the multiple relationship between Y and X. In order to keep the model as simple as possible, you should use the
regression model already containing X2. multiple regression equation without the quadratic effect.

Вам также может понравиться