click below
click below
Normal Size Small Size show me how
Medical Stats
Term | Definition |
---|---|
Clinical Trial | Planned experiment on human beings which is designed to evaluate the effectiveness of 1 or more forms of treatment |
Experiment | Series of observations made under conditions controlled or arranged by the investigator |
Treatment | A pharmaceutical drug or medical treatment (surgery, diet, counselling), or method of organising/delivering health care (education leaflet, phone app) |
Classification | Systematic arrangement by purpose or phase in drug development |
Phase I trial | Looks into pharmakinetics of a drug (absorption, metabolism) and toxicity (drug safety). finding maximum tolerated dose. |
Phase II trial | Initial clinical investigation. Dose finding study. Work out dose and dose schedules and early indication of efficacy |
Phase III trial | Aim of full scale evaluation, efficacy of a new experimental treatment compared to a standard therapy or placebo, acting as a control |
Phase IV trial | Effectiveness. Post marketing surveillance for info on long term effects and uncommon side effects. |
Effect | Difference between what happened to a patient as a result of the treatment vs what would have happened without the treatment |
Efficacy | The true biological effect of a treatment |
Effectiveness | Effect of a treatment when widely used in general practice. |
Efficiency | The time and cost related economics of a treatment. |
Research Protocol | Formal written document outlining the research plan for a clinical trial |
Control group's Purpose | Give a yard stick as to what may have happened in absence of experimental treatment. Needed in phase III trials |
Use of control group in Phase III trials | As investigator/patients expectations and enthusiasm of trial inclusion may affect judgement and outcomes. Some open & uncontrolled trials yeilded spurious results. |
Gold standard of trial design | Randomised, controlled and double-blinded trial. |
Randomisation | Treatments assigned according to chance (often equal). |
Aim of Randomisation | To yield treatment groups that are comparable in terms of extreneous factors. |
Purpose of Randomisation | Avoid bias due to differences in clinical & demographic characteristics. To support the independence assumption underlying many statistical procedures |
Simple Randomisation | independent random treatment allocation with fixed probability |
Pros of simple randomisation | easy to implement can do analysis via standard statistical methods |
Cons of simple randomisation | likely to get uneven treatment group numbers esp in small samples doesnt ensure balance over confounders (e.g. age and disease severity) |
Block randomisation | Restricted randomisation that works by balancing participant numbers in each group. |
Block randomisation method | Choose a block at random and assign patients accordingly. Want to keep block sizes short to prevent incomplete blocks (as may be fixed/randomly varying) |
Block size | A multiple of the number of treatments |
Cons of block randomisation | Assignment may become known if blocking factor revealed statistical methods often presume simple randomisation |
Pros of block randomisation | Has balance. max difference is b/2 for blocks of size b |
Stratified randomisation | Uses block randomisation in each stratum to aid between group comparability (balance) over important characteristics. |
Pros of stratified randomisation | Overcomes imbalance with regards to important factors in small samples Increases efficacy and reduces potential bias |
Cons of stratified randomisation | no. of strata may limit usefulness Complicates proceedings |
Adaptive randomisation - Minimisation | A non random treatment allocation where allocation probabilities adjusted according to patient imbalance |
Minimisation Purpose | Balance between treatment groups with regard to important prognastic factors. |
Minimisation Method | Use simple randomisation when groups are balanced. When imbalanced, allocates next patient to treatment so imbalanced minimised via minimisation score. |
Weighted minimisation | Weighting in favour of treatment which would minimise imbalance |
Purpose of Blinding/Masking | Avoid bias due to differences in treatment outcome (conscious or unconscious) |
Single-blind trial | Patient not know which treatment is being received |
Double-blind trial | Neither patient nor physician knows treatment allocation. GOLD standard. |
Triple-blind trial | double-blind plus monitoring group & data analyst don't know which group receives experiences & which control. |
Open trial | All are knowledgable about treatment allocation. |
Double-dummy | Compares 2 active treatments with different appearances |
Blind alternative | Incorporate blind assessment or conceal treatment allocation until patient entered into trial |
Cons of blind trial | More complex as need independent monitoring committee to deal with ethical aspects & trial termination |
Ethical Issues in clinical trials | can you obtain informed consent? Should you try to continue until investigators convinced 1 better or until entire medical community convinced? |
Ethical issues regarding placebos | Is it ethical if an effective existing treatment is known to exist |
Declaration of Helsinki | Set of ethical principles regarding human experimentation. For medical community. By WMA. Is a cornerstone document on human research ethics. Not legally binding. |
Governance arrangement for NHS research ethics committees | Describes what REC's should be like and when their review is needed |
Example of a complication of clinical trials | Non adherence to protocol - missing follow up visits, not completed full course of treatment |
Intention to treat set | Analyse as randomised, regardless of adherence. To look at effectiveness. |
Per Protocol set | Analysis based upon adherers only. To look at efficacy. |
Cons of per protocol set | Can introduce bias when groups of patients no longer have similar characteristics. |
When to use intention to treat set | In superiority trials where want to avoid over optimistic estimates |
Endpoint | Quantative measurement implied or required by the trial objectives. Is determined in each study subject. |
Hard endpoint | Preferred. Well defined and reliable. No subjectivity. Objectively measured. |
Soft endpoint | Difficult to define and subjective measures. |
Student's T-test | Comparison of 2 independent group means. Parametric method. |
Assumptions of student's T-test | Normality of data and equality of variances across comparison groups. |
Order of methodology for student's t-test | state research hypothesis (no diff and diff in difference in group means) estimate D = mean(T) - mean(C) Compute test statistic under H0. T = (mean(T)-mean(C)-0)/SE(estimate D) SE(estimate D) = S*sqrt(1/N(T) + 1/N(C)) S depend on underlying variance. |
S^2 (pooled variance estimate) in student's T-test if common underlying variance | 1/(n-2) * (sum from 1 to N(T) of (each value in T - mean(T))^2 + sum from 1 to N(C) of(each value in C - mean(C))^2 |
S^2 (pooled variance estimate) in student's T-test if S(T) and S(C) are sample standard deviations | ((N(T) -1)*S(T)^2 + (N(C) - 1)*S(C)^2/(N - 2) |
Confidence interval for student's T-test | T is compared to t-distribution with N-2 degrees of freedom & inference based on if |T| >= t dist (N-2,1 - alpha/2) D +- t dist (N-2, 1 - alpha/2) |
Point Estimate | Observed difference in group means D = mean(T) - mean(C) |
Interpretation if confidence interval for true difference spans zero | there is no statistically significant difference between groups so cannot reject null hypothesis |
Inference for p value | Calculate alpha for T value. if p < alpha = 0.05. Then is statistically significant |
Pro of confidence intervals | More informative as can assess plausible range of effect size |
Welch's test | An approximate test if assume non-constant variance |
Mann-whitney test | Non parametric alternative to student's T-test |
1-sample Paired t-test | Test based on within-pair differences |
Paired t-test method | Compute mean of within-pair diffs standard deviation of diffs Sd = sum of squares of differences between diffs and mean diff standard error of mean of diffs SE(D) = Sd/sqrt(n) Compute T under H0. T = D/SE(D) |
Paired t-test confidence interval | Compare T to t-distribution under n-1 degrees of freedom D +- t(n-1,1-alpha/2)*SE(D) |
p_1 and p_2 in binary response data | success probabilities for treatment and control groups in binary response data |
Risk Difference | Difference between risk of an outcome between group 1 and 2. p1 - p2. a/(a+c) - b/(b+d) |
Relative Risk | Compare risks for 2 groups. p1/p2. (a/(a+ c))/(b/(b+d)). treatment 1 RRx risk of outcome being observed than treatment 2. |
Disease odds | ratio of success to failure. p_i/(1-p_i) |
Odds ratio | Measure of association between outcome and exposure. p1/(1-p1) / p2/(1-p2) = ad/bc. treatment 1 ORx more likely to be exposed to outcome than treatment 2. |
When does odds ratio equal relative risk? | When the occurrence of outcome is less than 10%. |
Binary data confidence intervals | theta +- z(1-alpha/2)*SE(theta) where theta is risk difference, log of relative risk or log of odds ratio. |
Standard error for Risk difference BINARY | sqrt(p1(1-p1)/n1 + p2(1-p2)/n2). p1 = a/(a+c) and p2 = b(b+d). n1 = a+c. n2 = b + d |
Standard error for log of Relative Risk BINARY | sqrt(1/a - 1/(a+c) + 1/b - 1/(b+d)) DONT FORGET TO ANTILOG THE INTERVAL |
Standard error for log of Odds Ratio BINARY | sqrt(1/a + 1/b + 1/c + 1/d) DONT FORGET TO ANTILOG THE INTERVAL |
Parallel group design | Different groups of patients are studied in parallel. Estimate of treatment effects is based upon between subject comparisons. |
What inference test to use for parallel group design? | 2 independent samples t-test |
Paired design | Patient recieves both treatments. Then estimate of treatment effect based on within subject comparison |
What inference test to use to paired design | 1-sample t-test on within subject differences |
Crossover Design | Patient recieves sequence of treatments. Order determined by randomisation. Treatment effect based on within-subject comparisons |
Treatment Periods | times treatments are administered |
Pros of cross-over design | Patients act as own control as uses within-subject comparisons - eliminates between patient variation. Smaller sample size required for same no of observations Same degree of precision in estimation with fewer observations |
Cons of cross-over design | Inconvinience to patients as multiple treatments mean longer time under observation Patients may withdraw Treatment effect not constant over time (p-by-t interaction) Carry over effect Analysis more complex - may be systematic diffs between periods |
Carry over effect | Persistence of treatment applied in one period in a subsequence period of treatment |
How to deal with carry over effect | Wash out period |
Wash out Period | Period in trial during which effect of treatment given previously believed to disappear. |
When are cross-over trials useful? | Chronic diseases that are relatively stable Single dose trials of biological equivalence rather than long term trials Drugs with rapid and reversible effects |
AB/BA design | A 2 treatment x 2 period cross over trial. |
Simple analysis for a cross over design with no period effect | Use paired t-test |
Assumptions underlying use of paired t-test | Normally distributed differences Unbiased - that E(d) is the true treatment effect |
Factors that may cause differences to not be distributed at random about true treatment effect | period effect period by treatment interaction carry over effect patient by treatment interaction patient by period interaction |
When is a treatment effect estimate unbiased | If expectation of estimate is the estimate. |
Why do you need adequate sample sizes? | ethics budget constraints time constraints trial should be large enough to give reliable answer to research question |
P value | Probability, p, of obtaining test result at least as extreme as that observed assuming H0 is true |
Size of test | Given by alpha, usually alpha = 0.05 |
When p value less than alpha? | Reject H0 and conclude data is inconsistent with null. |
Type I error rate | alpha, probability of rejecting H0 given H0 is true (saying theres a diff when theres not) |
Type II error rate | beta, probability of failing to reject H0 when H1 true (saying no diff when there is a diff) |
Power of test | 1 - beta, probability of rejecting H0 when H1 is true (correctly saying is a diff) |
What does power of a test depend on? | the statistical test being used size of test (alpha) variability of observations (sigma^2) H1 (size of diff 'triangle') |
What does symbol triangle* represent? | The clinically relevant difference in a gauss test per say |
One & two sample one sided gauss test | Used to determine sample size |
Epidemiology | Study of distributions and determinants of disease in human populations |
What are descriptive studies used for? | Generating research hypotheses and resource allocation Gaining info regarding disease frequency in populations Info on distributions - age, geography, time, etc |
Disease Determinants aka risk factors | Factors precipitating disease (biological or enviro or social) |
Target population | Population about which we wish to draw inferences for |
Study population | Population from which data is collected |
Generalisability | Whether can use study population results to draw accurate conclusions about target pop |
Factors affecting study sample choice | generalisability of results trading off with availability and cost opportunistic sample is readily identifiable and likely to be cooperative Preference is a random sample of target population |
Routinely collected data | A source for data in epidemiological studies. E.g. hospital data bases, birth/death registers, census data) |
Sources for data in epidemiological studies | routinely collected data data purposely collected by investigators via surveys and follow up |
Purpose of routinely collected data | vital to monitor public health and health planning can use to investigate possible associations between routinely available attributes and rate of indecence from a particular disease |
Con of routinely collected data | May be of limited quality - subject to regional variation and often dont contain required individual level info. |
Limits of routinely collected data | Coverage - difficult to define morbidity so incomplete coverage Accuracy - diagnosis of cause of death and illness can be wrong Availability - confidential safe guards may limit data availability |
Reasons for incomplete coverage of routinely collected data | Inconsistent reporting of infectious diseases by practitioners and cancer registers may miss non hospital cases |
Disease Iceberg | moving down it, cases increase but severity decreases died, hospital, GP, self report, population screening |
Incidence (I) of disease | number of new cases of disease within a specified time period |
Prevalence (P) of disease | number of existing cases of disease at particular point of time. Probabiity of disease prior to seeing test result. (a+c)/N |
Calculation for incidence of disease | number of develop disease in specified time period / sum of length of time in which each person in population is at risk for |
Calculation for prevalence of disease | number with disease at time t / number in population at risk at time t |
Person-time of observation | Total observation time. Often express as person-years. Used for incidence rates. |
Purpose of prevalence measures | better for descriptive studies than analytical, suggests possible causes |
Purpose of incidence measures | Better to study as can establish sequence of events. Not susceptible to bias by survival |
Cumulative Incidence Risk | Alternative measure of disease occurance. Calculate: number of people who get disease in specified period / number of people free of disease at beginning of period. Is the risk of developing a disease. |
Crude mortality rate | number of deaths in a specified time period divided by average population at risk in period, multiplied by length of study period |
How to assess utility of a diagnostic test | Apply test to a number of individuals where true disease status known (based on GOLD standard) |
Sensitivity | Errors in testing. Proportion of truly diseased persons in tested population who are identified as diagnosed by screening test. P(T|D) (prob of pos test given diseased). a/(a+c) |
Specificity | Proportion of truly non diseased who are identified as so by screening test. P(bar(T)|bar(D)). d/(b+d) |
Positive Predictive Value PPV | Proportion truly diseased of those testing positive. Probability of diseased after a positive test. P(D|T). a/(a+b) |
Negative Predictive Value NPV | Proportion not diseased of those testing negative. P(bar(D)|bar(T)). Probability of not diseased after a negative test. c/(c+d) |
High cut off value | More specific test |
Low cut off value | More sensitive test |
Intermediate cut off value | Smaller overall errors. Need to balance high sensitivity and low specificity. |
Receiver Operating Characteristic curve | Plot of sensitivity against (1-specificity) for various cut off points. Cut off choice is top-left-most plot on curve. |
PPV using sensitivity, prevalence and specificity | (sens x prev)/((sens x prev) + (1 - spec)(1-prev)) |
NPV using sensitivity, prevalence, specificity | (spec x (1-prev))/((spec x (1- prev))+((1-sens) x prev)) |
Likelihood Ratio | Used to make therapeutic decisions regarding tests. P(T|D)/P(T|bar(D)) = sens/(1-spec). Test informative if value high. Is the ratio of posterior disease odds after pos test to prior disease odds, so tells how much odds of disease increase after pos test. |
Weight | Relative importance to give to sensitivity |
When to give high weight to sensitivity? | If disease high threatening and treatment has few side effects. (dont want false negatives). |
When to give high weight to specificity? | If disease not serious and treatment has many side effects. (dont want false positives) |
How to maximise weight with respect to study | Choose w = number of positive samples/total number of samples. Want to maximise M = w x sens + (1-w) x spec. 1-w is weight of spec. |
How to maximise weight with respect to population criterion | w = number with disease in population / total population size. Which is disease prevalence. |
Screening test | Test for particular disease given to population at risk who are asymptomatic |
Criteria for disease screening | Natural history of disease should be well understood should be an important health issue test should be acceptable to population there should exist a suitable screening test |
Why must be careful in disease screening | As truly diseased population likely to be small so many disease free will test positive |
Observational studies | Investigator role is passive so exposures not manipulated |
Intervention studies | Investigator role is active. Groups exposed to intervention of interest (eg treatment). Are experimental studies (clinical trials) |
Why are intervention studies the gold standard in terms of causality | can handle confounding with randomisation and exposure known to preceed disease. |
ethical Issue with intervention studies | rarely acceptable to force exposures on people |
Cross-sectional studies | surveys also feature but only to give info for snapshot in time - are less useful. Can only use to measure prevalence. |
Most common observational studies | cohort and case studies. |
Cohort study | Most useful observational study in disease causes. Tracks 2 or more groups forward from exposure to outcome. classify according to exposures. exposure vs non exposure rows and outcome vs non outcome columns |
Pros of cohort study | Strongest evidence for causal r.ship as exposure known to preceed outcome Can measure risk, incidence rate & survival time can examine many events after 1 exposure maintain temporal sequence best way to ascertain incidence & natural disease history |
Rationale of cohort studies | Follow cohorts through time - record disease occurrance and compare exposure groups with respect to disease outcome- then interpret |
Cons of cohort studies | large & costly selection bias loss to follow up confounding variables exposure can change in study period |
Case control study | Case & control groups defined & selected according to disease status. Look back in time to ascertain each person's exposure status |
Pros of case control study | efficient in terms of time and money (if exposure frequency not too low) can study rare diseases |
Cons of case control study | choosing control group can lead to selection bias obtaining exposure history so may have recall bias |
Risk difference for cohort studies (binary exposure) | p1 - p0. a/(a+b) - c/(c+d) |
Relative risk for cohort studies (binary exposure) | p1/p0. (a/(a+b)/c/(c+d)). |
Disease odds in cohort studies. binary exposure | p/1-p. Ratio of success to failure |
Odds ratio in cohort studies. binary exposure. | Compares exposed and not exposed groups. p1/(1-p1) / p0/(1-p0) = ad/bc |
In case control studies - estimate disease odds ratio not disease risk... | p1/(1-p1) / p0/(1-p0) |
Disease odds ratio | similar to odds ratio if disease is rare. |
Exposure odds for cases (D) | P(E|D)/P(bar(E)|D) |
Exposure odds for controls (bar(D)) | P(E|bar(D))/P(bar(E)|bar(D)) |
Exposure odds ratio | Ratio of 2 exposure odds. Equal to disease odds ratio. ad/bc |
How to increase precision of estimate of disease odds ratio? | More controls. |
A confounder | 3rd variable associated with exposure of interest. Independently associated with risk of disease. partially or fully explains relationship between E and D. |
Consequences of confounders | Creates a false relationship between E and D Masks true relationship between E and D |
Design based approaches to control confounding | Randomisation (balance over potential confounders) Restriction (Limit participation to individuals similar in relation to confounders) Matching (select controls to be similar to cases in terms of confounders) |
Analysis based approaches to control confounding | Stratification Standardisation (control confounding using external pop to adjust for age, gender etc) Multivariate analysis/regression models (include confounding variable in model) |
Stratification (in analysis based approach to control confounding) | Examine exposure-disease associations within strata of confounding variable. Estimate pooled estimate of association measure, adjusting for confounding effect. |
Mantel-haenszel estimate | Estimate of association between outcome and exposure after adjusting for confounding. Estimates a common odds ratio so use log in confidence interval |
Standard error for mantel-haenszel | very long. sum(a+d/n x ad/n)/2xsum(ad/n)^2 + same for bc + sum((a+d/n x ad/n)+(same for bc))/2sum(ad/n)sum(bc/n) |
If odds ratio confidence interval spans 1 for mantel-haenzel | association between exposure and outcome not significant at 5% level |
If reduction in odds ratio from joint odds ratio to individual in mantel-haenzel is small | The degree of confounding is small |
Unstratified matching in case control studies | Choose controls randomly |
Stratified matching in case control studies | Match controls to cases according to confounding variable. May be group or individual matching. |
Rationale of matched case control study | Eliminate confounders by design |
Pros of matching in case control study | control confounders by elimination gain in efficiency minimises selection bias |
Cons of matching in case control study | More complex study design cant study effect of matching variables on outcome of interest overmatching (matching variable linked to exposure not disease) |
Maximum likelihood estimate of exposure odds ratio in matched case-control study | b/c. standard error is log of estimate = sqrt(1/b + 1/c) |
Standardisation | process aimed at removing confounding by choosing a standard population with known distribution of confounders |
Direct standardisation | disease rates in population of interest are applied to standard population counts |
Indirect standardisation | disease rates in standard population applies to population of interest |
Direct standardised event rate | expected event rate in standard population, if age-specific event rates in study population prevailed |
How to adjust event rate for age | category specific event rates for each population being compared will be applied to a single standard population |
Direct age standardised event rate per 1000 | 1000/total size of standard pop x sum of ((observed number of events in ith age group of study population/size of ith age group of study population)x size of ith age group of standard population) |
Person years | average population size x study length |
When to use indirect standardisation of event rates | when age specific rates aren't available for population of interest |
Methods of indirect standardisation of event rates | Produce standardised event ratio (SER) or standardised mortality rate (SMR) |
SMR | O/E. where O is total number of observed events in all age groups of study population and E is sum of (observed number of events in ith age gorup of standard population / size of ith age group of standard population) x size of ith age group of study pop |
Requirement for direct standardised rates | need age specific rates for all population studies |
Requirement for indirect standardised rates | need total number of cases observed |
Standardised incidence / mortality rate | ratio of 2 indirectly standardised rates |
Pro of indirect standardised rates | More stable in case of small numbers of events |
Con of indirect standardised rates | Need age specific rates for standard population |
Survival analysis | data analysis in the form of time from some well defined time origin to occurrence of some event or endpoint. |
Examples of time origins for survival analysis | time of trial entry or diagnosis time |
Time-to-event data | can be time to death or time to failure of prosthetic etc |
Positive event example in time-to-event data | hospital discharge |
Adverse event example in time-to-event data | Death |
Neutral event example in time-to-event data | Cessation of breast feeding |
Lifetime random variable | time-to-event rate is random variable T |
Why can't apply time-to-event data to standard methods of analysis | Event times are positive continuous, typically skewed, and subject to censoring |
When does censoring occur? | if endpoint not observed |
Right censoring | event time greater than last follow up time |
Left censoring | Event time before time of last follow up but is unknown |
Interval censoring | event time falls in a specified interval |
Reasons for right censoring | period of observation end prior to event occurrence loss to follow up completing event precludes further follow up - eg death event may not be inevitable (eg pregnancy) |
Why one treatment may have more censoring at the end than other | results in increased survival |
Why does period of observation time vary between patients? | patients accrued sequentially over time and followed up to fixed date so period of observation varies between patients |
Assumption of survival analysis when acru patients over time | Prognosis doesn't depend on entry time to study |
Aims of survival analysis | model survival times for a group & make predictions assess effects of covariates on survival compare survival distribution for 2 or more groups |
The functions of interest when summarising survival data | Survivor function & hazard function |
Survival time of individual i | observation of a non negative r.v. T |
Distribution function of T | F(t) |
Probability density function of T | f(t) |
Link between Distribution and pdf of T | F(t) = integral (from 0 to t) of f(s) ds = P(T <= t). f(t) = d/dt F(t) |
Survivor function | Probability an individual survives to time t. S(t) = P(T>t) = 1 - F(t) = integral (from t to infinity) f(s) ds. Monotone decreasing function S(0) = 1 |
Link between pdf and survivor function | f(t) = -d/dt S(t) |
Hazard function | specify instantaneous rate of failure at T=t given survival to time t. h(t) = f(s)/S(t). RATE not probability |
How to approximate probability an individual who survived to time t will experience event in interval (t,t+delta*t) | h(t)*delta*t |
Cumulative/integrated hazard function | H(t) = integral (from 0 to t) of h(u) du. Probability of failure at time t given survival until time t |
Decreasing hazard function | Clinically less likely, such as in risk after organ transplantEs |
Emperical survivor function | estimate of survivor function. hat(S)(t). Step function with steps at observed failure times. |
Kaplan-meier estimate | Estimates survivor function. first step in presence of censored data. Is the product limit estimator |
event times and number in sample at risk at each event time and number who fail at this time notation | t_j and n_j and d_j |
Estimated probability of survival through interval j in kaplan meier survivor function | p_j = (n_j - d_j)/n_j) = 1 - d_j/n_j |
Assumptions for kaplan-meier estimate | independence of event times |
Kaplan meier estimator | S(t) = multiplication of all p_j's for k=r+1 subintervals |
Kaplan-meier when no censoring | emperical survivor estimator |
Greenwood's formula | for estimated variance. S(t)^2 x sum of (d_j/(n_j(n_j - d_j))) |
Make a smoother plot for survivor function estimate | More data and more distinct failure times |
End point of the kaplan-meier estimator curve | the estimated survival % at t years |
Why can't use t-test to compare group means in survival distribution? | Because of censoring |
Log-rank test | formal comparison of 2 groups survival distribution. Can compare more than 2 groups. |
Informal comparison of 2 groups survival distribution | Use Kaplan-meier curve |
Cox proportional hazards model | commonly used to flexibly model covariate effects on the hazard function. |